ZaighamMahmood Editor

Data Science
and Big Data
Computing
Frameworks and Methodologies

Data Science and Big Data Computing

ThiS is a FM Blank Page

Zaigham Mahmood
Editor

Data Science and Big Data

Computing
Frameworks and Methodologies

Editor
Zaigham Mahmood
Department of Computing and Mathematics
University of Derby
Derby, UK
Business Management and Informatics Unit
North West University
Potchefstroom, South Africa

ISBN 978-3-319-31859-2
ISBN 978-3-319-31861-5
DOI 10.1007/978-3-319-31861-5

(eBook)

Library of Congress Control Number: 2016943181

Springer International Publishing Switzerland 2016
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or
dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt
from the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained
herein or for any errors or omissions that may have been made.
Printed on acid-free paper
This Springer imprint is published by Springer Nature
The registered company is Springer International Publishing AG Switzerland

To
Rehana Zaigham Mahmood:
For her Love and Support

ThiS is a FM Blank Page

Preface

Overview
Huge volumes of data are being generated by commercial enterprises, scientific
domains and general public. According to a recent report by IBM, we create 2.5
quintillion bytes of data every day. According to another recent research, data
production will be 44 times greater in 2020 than it was in 2009.
Data being a vital organisational resource, its management and analysis is
becoming increasingly important: not just for business organisations but also for
other domains including education, health, manufacturing and many other sectors
of our daily life. This data, due to its volume, variety and velocity, often referred to
as Big Data, is no longer restricted to sensory outputs and classical databases; it also
includes highly unstructured data in the form of textual documents, webpages,
photos, spatial and multimedia data, graphical information, social media comments
and public opinions. Since Big Data is characterised by massive sample sizes, highdimensionality and intrinsic heterogeneity, and since noise accumulation, spurious
correlation and incidental endogeneity are common features of such datasets,
traditional approaches to data management, visualisation and analytics are no
longer satisfactorily applicable. There is therefore an urgent need for newer tools,
better frameworks and workable methodologies for such data to be appropriately
categorised, logically segmented, efficiently analysed and securely managed. This
requirement has resulted in an emerging new discipline of Data Science that is now
gaining much attention with researchers and practitioners in the field of Data
Analytics.
Although the terms Big Data and Data Science are often used interchangeably,
the two concepts have fundamentally different roles to play. Whereas Big Data
refers to collection and management of large amounts of varied data from diverse
sources, Data Science looks to creating models and providing tools, techniques and
scientific approaches to capture the underlying patterns and trends embedded in
these datasets, mainly for the purposes of strategic decision making.
vii

viii

Preface

In this context, this book, Data Science and Big Data Computing: Frameworks
and Methodologies, aims to capture the state of the art and present discussions and
guidance on the current advances and trends in Data Science and Big Data Analytics. In this reference text, 36 researchers and practitioners from around the world
have presented latest research developments, frameworks and methodologies, current trends, state of the art reports, case studies and suggestions for further understanding and development of the Data Science paradigm and Big Data Computing.

Objectives
The aim of this volume is to present the current research and future trends in the
development and use of methodologies, frameworks and the latest technologies
relating to Data Science, Big Data and Data Analytics. The key objectives include:
Capturing the state of the art research and practice relating to Data Science and
Big Data
Analysing the relevant theoretical frameworks, practical approaches and methodologies currently in use
Discussing the latest advances, current trends and future directions in the subject
areas relating to Big Data
Providing guidance and best practices with respect to employing frameworks
and methodologies for efficient and effective Data Analytics
In general, advancing the understanding of the emerging new methodologies
relevant to Data Science, Big Data and Data Analytics

Organisation
There are 13 chapters in Data Science and Big Data Computing: Frameworks and
Methodologies. These are organised in three parts, as follows:
Part I: Data Science Applications and Scenarios. This section has a focus on Big
Data (BD) applications. There are four chapters. The first chapter presents a
framework for fast data applications, while the second contribution suggests a
technique for complex event processing for BD applications. The third chapter
focuses on agglomerative approaches for partitioning of networks in BD scenarios; and the fourth chapter presents a BD perspective for identifying
minimum-sized influential vertices from large-scale weighted graphs.
Part II: Big Data Modelling and Frameworks. This part of the book also
comprises four chapters. The first chapter presents a unified approach to data
modelling and management, whereas the second contribution presents a distributed computing perspective on interfacing physical and cyber worlds. The third
chapter discusses machine learning in the context of Big Data, and the final

Preface

ix

contribution in this section presents an analytics-driven approach to identifying

duplicate records in large data repositories.
Part III: Big Data Tools and Analytics: There are five chapters in this section that
focus on frameworks, strategies and data analytics technologies. The first two
chapters present Apache and other enabling technologies and tools for data
mining. The third contribution suggests a framework for data extraction and
knowledge discovery. The fourth contribution presents a case study for adaptive
decision making; and the final chapter focuses on social impact and social media
analysis relating to Big Data.

Target Audiences
The current volume is a reference text aimed at supporting a number of potential
audiences, including the following:
Data scientists, software architect and business analysts who wish to adopt the
newer approaches to Data Analytics to gain business intelligence to support
business managers strategic decision making
Students and lecturers who have an interest in further enhancing the knowledge
of Data Science; and technologies, mechanisms and frameworks relevant to Big
Data and Data Analytics
Researchers in this field who need to have up-to-date knowledge of the current
practices, mechanisms and frameworks relevant to Data Science and Big Data to
further develop the same
Derby, UK
Potchefstroom, South Africa

Zaigham Mahmood

ThiS is a FM Blank Page

Acknowledgements

The editor acknowledges the help and support of the following colleagues during
the review and editing phases of this text:

Dr. Ashiq Anjum, University of Derby, Derby, UK

Anupam Biswas, Indian Institute of Technology (BHU) Varanasi, India
Dr. Alfredo Cuzzocrea, CAR-CNR & Univ. of Calabria, Rende (CS), Italy
Dr. Emre Erturk, Eastern Institute of Technology, New Zealand
Prof. Jing He, Kennesaw State University, Kennesaw, GA, USA
Josip Lorincz, FESB-Split, University of Split, Croatia
Dr. N Maheswari, School CS & Eng, Chennai, Tamil Nadu, India
Aleksandar Milic, University of Belgrade, Serbia,
Prof. Sulata Mitra, Indian Institute of Eng Science and Tech, Shibpur, India
Prof. Saswati Mukherjee, Anna University, Chennai, India
Dr. S Parthasarathy, Thiagarajar College of Eng, Tamil Nadu, India
Daniel Pop, Institute e-Austria Timisoara, West Univ. of Timisoara, Romania
Dr. Pethuru Raj, IBM Cloud Center of Excellence, Bangalore, India
Dr. Muthu Ramachandran, Leeds Beckett University, Leeds, UK
Dr. Lucio Agostinho Rocha, State University of Campinas, Brazil
Dr. Saqib Saeed, Bahria University, Islamabad, Pakistan
Prof. Claudio Sartori, University of Bologna, Bologna, Italy
Dr. Mahmood Shah, University of Central Lancashire, Preston, UK
Amro Najjar, Ecole Nationale Superieure des Mines de Saint Etienne, France
Dr. Fareeha Zafar, GC University, Lahore, Pakistan

I would also like to thank the contributors to this book: 36 authors and coauthors,
from academia as well as industry from around the world, who collectively submitted 13 chapters. Without their efforts in developing quality contributions,
conforming to the guidelines and meeting often the strict deadlines, this text
would not have been possible.

xi

xii

Acknowledgements

Grateful thanks are also due to the members of my family Rehana, Zoya,
Imran, Hanya, Arif and Ozair for their continued support and encouragement.
Best wishes also to Eyaad Imran.
Department of Computing and Mathematics
University of Derby
Derby, UK
Business Management and Informatics Unit
North West University
Potchefstroom, South Africa
14 February 2016

Zaigham Mahmood

Other Springer Books by Zaigham Mahmood

Cloud Computing: Challenges, Limitations and R&D

Solutions
This reference text reviews the challenging issues that present barriers to greater
implementation of the Cloud Computing paradigm, together with the latest research
into developing potential solutions. This book presents case studies, and analysis of
the implications of the cloud paradigm, from a diverse selection of researchers and
practitioners of international repute (ISBN: 978-3-319-10529-1).

Continued Rise of the Cloud: Advances and Trends in Cloud

Computing
This reference volume presents latest research and trends in cloud-related technologies, infrastructure and architecture. Contributed by expert researchers and practitioners in the field, this book presents discussions on current advances and
practical approaches including guidance and case studies on the provision of
cloud-based services and frameworks (ISBN: 978-1-4471-6451-7).

Cloud Computing: Methods and Practical Approaches

The benefits associated with cloud computing are enormous; yet the dynamic,
virtualized and multi-tenant nature of the cloud environment presents many challenges. To help tackle these, this volume provides illuminating viewpoints and case
studies to present current research and best practices on approaches and technologies for the emerging cloud paradigm (ISBN: 978-1-4471-5106-7).
xiii

xiv

Other Springer Books by Zaigham Mahmood

Software Engineering Frameworks for the Cloud

Computing Paradigm
This is an authoritative reference that presents the latest research on software
development approaches suitable for distributed computing environments. Contributed by researchers and practitioners of international repute, the book offers
practical guidance on enterprise-wide software deployment in the cloud environment. Case studies are also presented (ISBN: 978-1-4471-5030-5).

Cloud Computing for Enterprise Architectures

This reference text, aimed at system architects and business managers, examines
the cloud paradigm from the perspective of enterprise architectures. It introduces
fundamental concepts, discusses principles and explores frameworks for the adoption of cloud computing. The book explores the inherent challenges and presents
future directions for further research (ISBN: 978-1-4471-2235-7).

Contents

Part I
1

An Interoperability Framework and Distributed Platform for

Fast Data Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Jose Carlos Martins Delgado

Complex Event Processing Framework for Big Data

Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Rentachintala Bhargavi

41

Agglomerative Approaches for Partitioning of Networks in Big

Data Scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Anupam Biswas, Gourav Arora, Gaurav Tiwari, Srijan Khare,
Vyankatesh Agrawal, and Bhaskar Biswas
Identifying Minimum-Sized Influential Vertices on Large-Scale
Weighted Graphs: A Big Data Perspective . . . . . . . . . . . . . . . . . . .
Ying Xie, Jing (Selena) He, and Vijay V. Raghavan

Part II
5

Data Science Applications and Scenarios

57

79

Big Data Modelling and Frameworks

A Unified Approach to Data Modeling and Management in Big

Data Era . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Catalin Negru, Florin Pop, Mariana Mocanu, and Valentin Cristea

95

Interfacing Physical and Cyber Worlds: A Big Data

Perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
Zartasha Baloch, Faisal Karim Shaikh, and Mukhtiar A. Unar

Distributed Platforms and Cloud Services: Enabling Machine

Learning for Big Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
Daniel Pop, Gabriel Iuhasz, and Dana Petcu

xv

xvi

Contents

An Analytics-Driven Approach to Identify Duplicate Bug

Records in Large Data Repositories . . . . . . . . . . . . . . . . . . . . . . . . 161
Anjaneyulu Pasala, Sarbendu Guha, Gopichand Agnihotram,
Satya Prateek B, and Srinivas Padmanabhuni

Part III

Big Data Tools and Analytics

Large-Scale Data Analytics Tools: Apache Hive, Pig, and

HBase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
N. Maheswari and M. Sivagami

10

Big Data Analytics: Enabling Technologies and Tools . . . . . . . . . . 221

Mohanavadivu Periasamy and Pethuru Raj

11

A Framework for Data Mining and Knowledge Discovery in

Cloud Computing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
Derya Birant and Pelin Yldrm

12

Feature Selection for Adaptive Decision Making in Big Data

Analytics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
Jaya Sil and Asit Kumar Das

13

Social Impact and Social Media Analysis Relating to Big

Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
Nirmala Dorasamy and Natasa Pomazalova

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315

Contributors

Gopichand Agnihotram Infosys Labs, Infosys Ltd., Bangalore, India

Vyankatesh Agrawal Department of Computer Science and Engineering, Indian
Institute of Technology (BHU), Varanasi, India
Gourav Arora Department of Computer Science and Engineering, Indian Institute of Technology (BHU), Varanasi, India
Zartasha Baloch IICT, Mehran University of Engineering and Technology,
Jamshoro, Pakistan
Rentachintala Bhargavi School of Computing Sciences and Engineering, VIT
University, Chennai, India
Derya Birant Department of Computer Engineering, Dokuz Eylul University,
Izmir, Turkey
Bhaskar Biswas Department of Computer Science and Engineering, Indian Institute of Technology (BHU), Varanasi, India
Anupam Biswas Department of Computer Science and Engineering, Indian Institute of Technology (BHU), Varanasi, India
Valentin Cristea Computer Science Department, Faculty of Automatic Control
and Computers, University Politehnica of Bucharest, Bucharest, Romania
Asit Kumar Das Department of Computer Science and Technology, Indian Institute of Engineering Science and Technology, Shibpur, Howrah, West Bengal, India
Jose Carlos Martins Delgado Department of Computer Science and Computer
Engineering, Instituto Superior Tecnico, Universidade de Lisboa, Lisbon, Portugal
Nirmala Dorasamy Department of Public management and Economics, Durban
University of Technology, Durban, South Africa
Sarbendu Guha Infosys Labs, Infosys Ltd., Bangalore, India
xvii

xviii

Contributors

Jing (Selena) He Department of Computer Science, Kennesaw State University,

Marietta, GA, USA
Gabriel Iuhasz Institute e-Austria Timisoara, West University of Timisoara,
Timioara, Romania
Srijan Khare Department of Computer Science and Engineering, Indian Institute
of Technology (BHU), Varanasi, India
N. Maheswari School of Computing Science and Engineering, VIT University,
Chennai, Tamil Nadu, India
Mariana Mocanu Computer Science Department, Faculty of Automatic Control
and Computers, University Politehnica of Bucharest, Bucharest, Romania
Catalin Negru Computer Science Department, Faculty of Automatic Control and
Computers, University Politehnica of Bucharest, Bucharest, Romania
Srinivas Padmanabhuni Infosys Labs, Infosys Ltd., Bangalore, India
Anjaneyulu Pasala Infosys Labs, Infosys Ltd., Bangalore, India
Mohanavadivu Periasamy TCL Canada, Montreal, QC, Canada
Dana Petcu Institute e-Austria Timisoara, West University of Timisoara,
Timioara, Romania
Natasa Pomazalova Department of Regional Development, Mendel University,
Brno, Czech Republic
Durban University of Technology, Durban, South Africa
Daniel Pop Institute e-Austria Timisoara, West University of Timisoara,
Timioara, Romania
Florin Pop Computer Science Department, Faculty of Automatic Control and
Computers, University Politehnica of Bucharest, Bucharest, Romania
Satya Prateek B Infosys Labs, Infosys Ltd., Bangalore, India
Vijay V. Raghavan The Center for Advanced Computer Studies, University of
Louisiana at Lafayette, Lafayette, LA, USA
Pethuru Raj IBM Global Cloud Center of Excellence, Bangalore, India
Faisal Karim Shaikh IICT, Mehran University of Engineering and Technology,
Jamshoro, Pakistan
TCMCORE, STU, University of Umm Al-Qura, Mecca, Saudi Arabia
Jaya Sil Department of Computer Science and Technology, Indian Institute of
Engineering Science and Technology, Shibpur, Howrah, West Bengal, India

Contributors

xix

M. Sivagami School of Computing Science and Engineering, VIT University,

Chennai, Tamil Nadu, India
Gaurav Tiwari Department of Computer Science and Engineering, Indian Institute of Technology (BHU), Varanasi, India
Mukhtiar A. Unar IICT, Mehran University of Engineering and Technology,
Jamshoro, Pakistan
Ying Xie Department of Computer Science, Kennesaw State University, Marietta,
GA, USA
Pelin Yldrm Department of Software Engineering, Celal Bayar University,
Manisa, Turkey

ThiS is a FM Blank Page

About the Editor

Professor Dr. Zaigham Mahmood is a published author of 16 books, five of which

are dedicated to Electronic Government and the other eleven focus on the subjects
of Cloud Computing and Data Science, including: Cloud Computing: Concepts,
Technology & Architecture which is also published in Korean and Chinese languages;
Cloud Computing: Methods and Practical Approaches; Software Engineering
Frameworks for the Cloud Computing Paradigm; Cloud Computing for Enterprise
Architectures; Cloud Computing Technologies for Connected Government; Continued
Rise of the Cloud: Advances and Trends in Cloud Computing; Connectivity
Frameworks for Smart Devices: The Internet of Things from a Distributed Computing
Perspective; and Cloud Computing: Challenges, Limitations and R&D Solutions.
Additionally, he is developing two new books to appear later in 2017. He has
also published more than 100 articles and book chapters and organised numerous
conference tracks and workshops.
Professor Mahmood is the Editor-in-Chief of Journal of E-Government Studies
and Best Practices as well as the Series Editor-in-Chief of the IGI book series on
E-Government and Digital Divide. He is a Senior Technology Consultant at
Debesis Education UK and Associate Lecturer (Research) at the University of
Derby, UK. He further holds positions as Foreign Professor at NUST and IIU in
Islamabad, Pakistan, and Professor Extraordinaire at the North West University,
Potchefstroom, South Africa. Professor Mahmood is also a certified cloud computing instructor and a regular speaker at international conferences devoted to Cloud
Computing and E-Government. His specialised areas of research include distributed computing, project management and e-government.
Professor Mahmood can be reached at z.mahmood@debesis.co.uk

xxi

Part I

Data Science Applications and Scenarios

Chapter 1

An Interoperability Framework
and Distributed Platform for Fast Data
Applications
Jose Carlos Martins Delgado

Abstract Big data developments have been centred mainly on the volume dimension of data, with frameworks such as Hadoop and Spark, capable of processing
very large data sets in parallel. This chapter focuses on the less researched dimensions of velocity and variety, which are characteristics of fast data applications. The
chapter proposes a general-purpose distributed platform to host and interconnect
fast data applications, namely, those involving interacting resources in a heterogeneous environment such as the Internet of Things. The solutions depart from
conventional technologies (such as XML, Web services or RESTful applications),
by using a resource-based meta model that is a partial interoperability mechanism
based on the compliance and conformance, service-based distributed programming
language, binary message serialization format and architecture for a distributed
platform. This platform is suitable for both complex (Web-level) and simple
(device-level) applications. On the variety dimension, the goal is to reduce
design-time requirements for interoperability by using structural data matching
instead of sharing schemas or media types. In this approach, independently developed applications can still interact. On the velocity dimension, a binary serialization
format and a simple message-level protocol, coupled with a cache to hold frequent
type mappings, enable efficient interaction without compromising the flexibility
required by unstructured data.
Keywords Internet of Things IoT Big data Web services XML Coupling
Structural compatibility Compliance Conformance Distributed programming
Variety Velocity

J.C.M. Delgado (*)

Department of Computer Science and Computer Engineering, Instituto Superior Tecnico,
Universidade de Lisboa, Lisbon, Portugal
e-mail: jose.delgado@tecnico.ulisboa.pt
Springer International Publishing Switzerland 2016
Z. Mahmood (ed.), Data Science and Big Data Computing,
DOI 10.1007/978-3-319-31861-5_1

1.1

J.C.M. Delgado

Introduction

One of the fundamental objectives of any distributed data system is the ability to
perform the required amount of data exchange and computation in the available
timeframe, which translates into a required minimum data flow and processing
rates. Big data scenarios turn this into a harder endeavour due to several reasons,
including the following characteristics of data [1]:
Volume: high volume of data (more data to process)
Velocity: high rate of incoming data (less time to process data)
Variety: data heterogeneity (more data formats or data sources to deal with)
Big data developments have been mainly centred on the volume dimension, with
dynamic frameworks such as Hadoop [2] and Spark [3], capable of processing very
large data sets in parallel. This chapter focuses on the less researched dimensions of
velocity and variety, which are characteristics of fast data applications [4]. Typically, these involve too many entities, interacting and exchanging too many data, at
too high rates in a too heterogeneous environment. An entity can be a complex
application in a server or a very simple functionality provided by a small sensor, in
the context of what is usually known as the Internet of Things, abbreviated as IoT
[5, 6].
The European Commission [7] estimates that by 2020, the number of globally
connected devices will be in the order of 50100 billion devices. These will
generate big data, which many applications will need to process very quickly and
with low latency.
Variety means supporting a diversity of data sources, formats and protocols. Not
all devices are adequate to support Transmission Control Protocol/Internet Protocol
(TCP/IP) and all the features required to be part of the Web. Velocity requires
efficient data exchange and processing mechanisms. Together, they demand for
new data-level distributed interoperability mechanisms.
Current interoperability technologies rely on text-based data description languages, such as Extensible Markup Language (XML) and JavaScript Object Notation (JSON) [57], and high-level and complex protocols such as Hypertext Transfer
Protocol (HTTP) and Simple Object Access Protocol (SOAP). However, these
languages have not been designed for the high throughput and low latency that
fast applications require. Similarly, the big data solutions such as Hadoop emphasize the volume dimension and are not adequate for fast data [4]. In terms of
interoperability, these languages and protocols constitute specific solutions,
designed for the Web class of applications (many clients for each server, best effort
rather than real time) and do not allow an arbitrary set of computer-based applications to interact as peers.
What is needed is a new set of solutions that support the generic interoperability
of fast data applications, in the same way as web technologies have provided
universal interoperability for web applications. These solutions include native
support for binary data, efficient and full-duplex protocols, machine-level data

1 An Interoperability Framework and Distributed Platform for Fast Data. . .

and service interoperability and context awareness for dynamic and mobile environments, such as those found in smart cities [8]. Today, these features are
simulated on top of Web services: applications based on representational state
transfer (REST), HTTP, XML, JSON and other related technologies, rather than
implemented by native solutions. The problem needs to be revisited to minimize the
limitations at the source, instead of just hiding them with abstraction layers that add
complexity and reduce performance.
As a contribution to satisfy these requirements, this chapter includes the following proposals:
A layered interoperability framework, to systematize the various aspects and
slants of distributed interoperability
A language to describe not only data structures (state) but also operations
(behaviour), with self-description capability to support platform-agnostic
interoperability
A data interoperability model, on which this language is based, which relies on
compliance and conformance [9] instead of schema sharing (as in XML) or
previously defined data types (as REST requires)
A message-level protocol at a lower level than that of SOAP and even HTTP,
with many of the features included in these protocols implemented on top of the
basic interoperability model
The architecture of a node of a distributed platform suitable for fast data
applications
These features are the building blocks of a distributed interoperability platform
conceived to tackle the velocity and variety dimensions of distributed applications,
modelled as services. This platform is suitable not only for complex, Web-level
applications but also for simple, device-level applications.
This chapter is structured as follows. Section 1.2 describes some of the existing
technologies relevant to the theme of this chapter, followed by a description in
Sect. 1.3 of several of the issues concerning fast data. Section 1.4 describes an
interoperability framework with emphasis on fast data problems, namely, those
affecting variety (interoperability and coupling) and velocity (message latency). It
also presents a resource-based model to support structural compatibility, based on
compliance and conformance, and a service interoperability language that implements these proposals. Section 1.5 describes the architecture of a distributed
platform to support the resource-based model and the service-based language.
The chapter ends by discussing the usefulness of this approach, outlining future
directions of research and drawing the main conclusions of this work.

1.2

Background

Big data [1] has become a relevant topic in recent years, spurred by the everincreasing growth rate of data produced in this globally interconnected world and
by the always-pressing need of obtaining meaningful information from large data

J.C.M. Delgado

sets, in domains such as business analytics [10], healthcare [11], bioinformatics

[12], scientific computing [13] and many others [14].
Big data refers to handling very large data sets, for storage, processing, analysis,
visualization and control. The National Institute of Standards and Technology
(NIST) have proposed a Big Data Interoperability Framework [15] to lay down a
foundation for this topic.
A big push for this area came from the MapReduce programming model [16],
initially used for indexing large data sets and business intelligence over data
warehouses. The first motivating factor for big data was, thus, volume (data size),
using immutable and previously stored data. However, agile enterprises [17]
require almost real-time analysis and reaction to a large number of events and
business data, stemming from many sources and involving many data formats. This
means that the velocity and variety dimensions are gaining momentum [4]. A
survey of systems for big data, with emphasis on real time, appears in [18].
Enterprise integration models and technologies, such as service-oriented architecture (SOA) [19], REST [20] and enterprise service bus (ESB) [21], have not been
conceived for fast data processing and therefore constitute only as a best-effort
approach.
Besides the dimensions already described (volume, velocity and variety), other
Vs have also been deemed relevant in this context [13, 15], these being veracity
(accuracy of data), validity (quality and applicability of data in a given context),
value (of data to stakeholders), volatility (changes in data over time) and variability
(of the data flow).
Gone are the days when the dominant distributed application scenario consisted
of a Web encompassing fixed computers, both at the user (browser) and server
(Web application) sides. Today, cloud computing [22] and the IoT [6] are revolutionizing the society, both at the enterprise and personal levels, in particular in
urban environments [8] with new services and applications. For example, mobile
cloud computing [23] is on the rise, given the pervasiveness of smartphones and
tablets that created a surge in the bring your own device (BYOD) trend [24]. The
increasing use of radio-frequency identification (RFID) tags [25] in supply chains
raises the need to integrate enterprise applications with the physical world, including sensor networks [26] and vehicular [27] networks.
Cisco have set up a counter [28], indicating the estimated number of devices
connected to the Internet. This counter started with 8.7 billion devices at the end of
2012, increased to roughly 10 and 14 billion at the end of 2013 and 2014,
respectively, and at the time of writing (April 2015), it shows a figure of 15.5 billion,
with a foreseen value in the order of 50 billion by 2020. The Internet World Stats
site (http://www.internetworldstats.com/stats.htm), on the other hand, estimates
that by mid-2014, the number of Internet human users was around three billion,
almost half the worldwide population of roughly 7.2 billion people. The number of
Internet-enabled devices is clearly growing faster than the number of Internet users,
since the world population is estimated to be in the order of 7.7 billion by 2020
[29]. This means that the Internet is no longer dominated by human users but rather

1 An Interoperability Framework and Distributed Platform for Fast Data. . .

by smart devices that are small computers and require technologies suitable to
them, rather than those mostly adequate to full-fledged servers.

1.3

Introducing Fast Data

Fast data has a number of inherent issues, in addition to those relating to big data.
This section describes motivating scenarios and one of the fundamental issues, that
of interoperability. Other issues, stemming from the variety and velocity dimensions are discussed in Sects. 1.4.2 and 1.4.3 respectively.

1.3.1

Motivating Scenarios

Figure 1.1 depicts several scenarios in which large quantities of data can be
produced from a heterogeneous set of data sources, eventually with different
formats and processing requirements. For simplicity, not all possible connections
are depicted, but the inherent complexity of integrating all these systems and
processing all the data they can produce is easy to grasp.

Web servers

Hadoop cluster
Private cloud

Datacenter

Mobile
users

Public cloud

Sensor network

Mobile cloud

Users
Vehicular network

Fig. 1.1 An example of a set of heterogeneous systems with big data requirements

J.C.M. Delgado

Most big data applications today use best-effort technologies such as Hadoop [2]
and Spark [3], in which immutable data is previously loaded into the processing
nodes. This is suitable for applications in areas such as business analytics [30],
which attempt to mine information that can be relevant in specific contexts and
essentially just deal with the volume dimension of big data. However, this is not the
case for applications where many data sets are produced or a large number of events
occur frequently, in a heterogeneous ecosystem of producers and consumers. In
these applications, processing needs to be performed as data are produced or events
occur, therefore emphasizing the variety and velocity dimensions (fast data).
No matter which dimension we consider, big essentially means too complex,
too much, too many and too fast to apply conventional techniques, technologies and
systems, since their capabilities are not enough to handle such extraordinary
requirements. This raises the problem of integrating heterogeneous interacting
parties to a completely new level, in which conventional integration technologies
(such as HTTP, XML, JSON, Web Services and RESTful applications) expose their
limitations. These are based on technologies conceived initially for human interaction, with text as the main format and sub-second time scales and not for heavyduty, machine-level binary data exchange that characterizes computer-to-computer
interactions, especially those involving big data.
New solutions are needed to deal with these integration problems, in particular in
what concerns fast data requirements. Unlike processing of large passive and
immutable data sets, for which frameworks such as Hadoop are a good match,
fast data scenarios consist of a set of active interacting peers, producing, processing
and consuming data and event notifications.

1.3.2

Issues Relating to Interoperability

A distributed software system has modules with independent life cycles, each able
to evolve to a new version without having to change, suspend or stop the behaviour
or interface of the others. These modules are built and executed in an independent
way. Frequently, they are programmed in different programming languages and
target different formats, platforms and processors. Distribution usually involves
geographical dispersion, a network and static node addresses. Nevertheless, nothing
prevents two different modules from sharing the same server, physical or virtual.
Modules are usually designed to interact, cooperating towards some common
goal. Since they are independent and make different assumptions, an interoperability problem arises. Interoperability, as old as networking, is a word formed by the
juxtaposition of a prefix (inter) and the agglutination of two other words (operate
and ability), meaning literally the ability of two or more system modules to operate
together. In this context, an application is a set of software modules with synchronized lifecycles, i.e. compiled and linked together. Applications are the units of
system distribution, and their interaction is usually limited to message exchange.

1 An Interoperability Framework and Distributed Platform for Fast Data. . .

Applications are independent, and each can evolve in ways that the others cannot
predict or control.
The interaction between modules belonging to the same application can rely on
names to designate concepts in the type system (types, inheritance, variables,
methods and so on). A name can have only one meaning in a given scope, which
means that using a name is equivalent to using its definition. A working application
usually assumes that all its modules are also working and use the same implementation language and formats, with any changes notified to all modules. The application is a coherent and cohesive whole.
The interaction of modules belonging to different applications, however, is a
completely different matter. Different applications may use the same name for
different meanings, be programmed in different languages, be deployed in different
platforms, use different formats and without notifying other applications, migrate
from one server to another, change their functionality or interface and even be down
for some reason, planned or not.
This raises relevant interoperability problems, not only in terms of correctly
interpreting and understanding exchanged data but also in keeping behaviours
synchronized in some choreography. The typical solutions involve a common
protocol (such as HTTP), self-describing data at the syntax and sometimes semantics levels and many assumptions previously agreed upon. For example, XMLbased interactions, including Web services, assume a common schema. REST
proponents claim decoupling between client and server (the client needs just the
initial URI, Universal Resource Identifier). However, RESTful applications do
require previously agreed upon media types (schemas) and implicit assumptions
by the client on the behaviour of the server when executing the protocol verbs.
It is virtually impossible for one application to know how to appropriately
interact with another application, if it knows nothing about that application. Not
even humans are able to achieve it. Some form of coupling (based on shared and
agreed knowledge, prior to interaction) needs to exist. The goal is to reduce
coupling as much as possible while ensuring the minimum level of interoperability
required by the problem that motivated the interaction between applications.
Figure 1.2 provides an example of the kind of problems that need to be tackled in
order to achieve this goal.
Figure 1.2 can be described in terms of the scenario of Fig. 1.2a which refers to
the first seven steps of the process; the entire process being as follows:
1. Application A resorts to a directory to find a suitable application, according to
some specification.
2. The directory has a reference (a link) to such an application, e.g. B.
3. The directory sends that reference to A.
4. Application A sends a message to B, which it must understand, react and
respond according to the expectations of A (note the bidirectional arrow).
5. If message is unreachable, A can have predefined alternative applications, such
as B1 or B2. Resending the message to them can be done automatically or as a
result from an exception.

10
Fig. 1.2 Illustration of
some interoperability
aspects. (a) Before
migration of application
B. (b) After migration

J.C.M. Delgado

(a)
Server 1

Directory

2
5

B1

B2

B3

Server 2

Server 3

(b)
Server 1

A
5

3
4

proxy

B1

Directory

10

B2

B
6

Server 2

B3
Server 3

6. If B is reachable but somehow not functional, B itself (or the cloud that
implements it) can forward the message to an alternative application, such
as B3.
7. Application B can be migrated dynamically to another cloud, yielding the
scenario of Fig. 1.2b.
8. B leaves a reverse proxy as a replacement, which means that if A sends another
message to B (step 4), it will be automatically forwarded to the new B.
9. The response, however, will be sent to the original sender, A, including
information on the new location of B, which A will use for subsequent messages
(the message protocol must support this).
10. The proxy could be garbage collected; but this is not easy to manage in a
distributed system that is unreliable by nature. Therefore, the proxy can be
maintained for some time under some policy and destroyed afterward. If some
application still holding a reference to the old B sends a message to it, the
protocol should respond with a suitable error stating that B does not exist. A can

1 An Interoperability Framework and Distributed Platform for Fast Data. . .

11

then repeat steps 1 through 3, obtaining the new location of B from the
directory.
Figure 1.2 raises some interesting interoperability issues, namely, the question of
compatibility between an application and its alternatives, which must be able to
behave as if they were the original application. Therefore, we need to detail what is
involved in an interaction between two applications.

1.4

An Interoperability Framework for Fast Data

Having discussed the interoperability problem, a framework is needed to dissect its

various slants. This will enable us to discuss what the variety and velocity dimensions really mean and to derive a metamodel and a language to deal with them. This
section presents the relevant discussion.

1.4.1

Understanding Interoperability

In general, successfully sending a message from one application to another entails

the following aspects (noting that requests and responses reverse the roles of the
sender and receiver applications):
Intent. Sending the message must have an underlying intent, inherent in the
interaction to which it belongs and related to the motivation to interact and the
goals to achieve. This should be aligned with the design strategy of both
applications.
Content. This concerns the generation and interpretation of the content of a
message by the sender, expressed by some representation, in such a way that the
receiver is also able to interpret it, in its own context.
Transfer. The message content needs to be successfully transferred from the
context of the sender to the context of the receiver.
Willingness. Usually, applications are designed to interact and therefore to
accept messages, but nonfunctional aspects such as security and performance
limitations can impose constraints.
Reaction. This concerns the reaction of the receiver upon reception of a message,
which should produce effects according to the expectations of the sender.
Interoperability between two applications can be seen at a higher level, involving intentions (why interact and what reactions can be expected from whom), or at a
lower level, concerning messages (what to exchange, when and how). Detailing the
various levels leads to a systematization of interoperability such as the one
described in Table 1.1.

12

J.C.M. Delgado

Table 1.1 Levels of interoperability

Category
Symbiotic (purpose and intent)

Pragmatic (reaction and effects)

Semantic (meaning of content)

Syntactic (notation of
representation)

Level
Coordination
Alignment
Collaboration
Cooperation
Contract
Workflow
Interface
Inference
Knowledge
Ontology
Structure
Predefined type
Serialization

Connective
(transfer
protocol)

Messaging
Routing
Communication
Physics

Environmental
(deployment and
migration)

Management
Library
Platform
Computer

Main artefact
Governance
Joint venture
Partnership
Outsourcing
Choreography
Process
Service
Rule base
Knowledge
base
Concept
Schema
Primitive
application
Message
format
Message
protocol
Gateway
Network
protocol
Media
protocol
API
Utility
service
Basic
software
Computing
hardware

Description
Motivations to have the interaction,
with varying levels of mutual knowledge of governance, strategy and goals
Management of the effects of the
interaction at the levels of choreography, process and service
Interpretation of a message in context,
at the levels of rule, known application
components and relations and definition of concepts
Representation of application components, in terms of composition, primitive components and their serialization
format in messages
Lower-level formats and network protocols involved in transferring a message from the context of the sender to
that of the receiver

Environment in which each application

is deployed and managed, including the
portability problems raised by
migrations

Table 1.1 can be described as follows, using the category column as the top
organizing feature:
Symbiotic: Expresses the purpose and intent of two interacting applications to
engage in a mutually beneficial agreement. Enterprise engineering is usually the
topmost level in application interaction complexity, since it goes up to the
human level, with governance and strategy heavily involved. Therefore, it
maps mainly onto the symbiotic category, although the same principles apply
(in a more rudimentary fashion) to simpler subsystems. This can entail a tight
coordination under a common governance (if the applications are controlled by
the same entity), a joint venture agreement (if the two applications are substantially aligned), a collaboration involving a partnership agreement (if some goals
are shared) or a mere value chain cooperation (an outsourcing contract).

1 An Interoperability Framework and Distributed Platform for Fast Data. . .

13

Pragmatic: The effect of an interaction between a consumer and a provider is the

outcome of a contract, which is implemented by a choreography that coordinates
processes, which in turn implement workflow behaviour by orchestrating service
invocations. Languages such as Business Process Execution Language (BPEL)
[31] support the implementation of processes and Web Services Choreography
Description Language (WS-CDL) is an example of a language that allows
choreographies to be specified.
Semantic: Both interacting applications must be able to understand the meaning
of the content of the messages exchanged: both requests and responses. This
implies interoperability in rules, knowledge and ontologies, so that meaning is
not lost when transferring a message from the context of the sender to that of the
receiver. Semantic languages and specifications such as Web Ontology Language (OWL) and Resource Description Framework (RDF), map onto this
category [32].
Syntactic: This deals mainly with form, rather than content. Each message has a
structure composed of data (primitive applications) according to some structural
definition (its schema). Data need to be serialized to be sent over the network as
messages using representations such as XML or JSON.
Connective: The main objective is to transfer a message from the context of one
application to the other regardless of its content. This usually involves enclosing
that content in another message with control information and implementing a
message protocol (such as SOAP or HTTP) over a communications network
according to its own protocol (such as TCP/IP) and possibly resorting to routing
gateways if different networks are involved.
Environmental: Each application also interacts with the environment (e.g. a
cloud or a server) in which it is deployed, anewed or by migration. The
environments management application programming interface (API) and the
infrastructure level that the application requires will most likely have impact on
the way applications interact, particularly if they are deployed in (or migrate
between) different environments, from different vendors. Interoperability
between an application and the environment in which it is deployed usually
known as portability.
In most cases, not all these levels are considered explicitly. Higher levels tend to
be treated tacitly (specified in documentation or simply assumed but not ensured),
whereas lower levels are commonly dealt with empirically (ensured by running
software but details hidden by lower level software layers).
Syntactic is the most used category, because it is the simplest and the most
familiar, with interfaces that mainly deal with syntax or primitive semantics. The
pragmatic category, fundamental to specify behaviour, is mainly implemented by
software but without any formal specification.
Another important aspect is nonfunctional interoperability. It is not just a
question of sending the right message. Adequate service levels, context awareness,
security and other nonfunctional issues must also be considered when applications
interact, otherwise interoperability will be less effective or not possible at all.

14

J.C.M. Delgado

Finally, it should be stressed that, as asserted above, all these interoperability

levels constitute an expression of application coupling. On the one hand, two
uncoupled applications (with no interactions between them) can evolve freely and
independently, which favours adaptability, changeability and even reliability (such
that if one fails, there is no impact on the other). On the other hand, applications
need to interact to cooperate towards common or complementary goals, which
imply that some degree of previously agreed mutual knowledge is indispensable.
The more they share with the other, the more integrated they are, and so the
interoperability becomes easier, but coupling gets more complicated.
The usefulness of Table 1.1 lies in providing a classification that allows coupling
details to be better understood, namely, at which interoperability levels they occur
and what is involved at each level, instead of having just a blurry notion of
dependency. In this respect, it constitutes a tool to analyse and to compare different
coupling models and technologies.

1.4.2

The Variety Dimension

The previous section suggests that coupling is unavoidable. Without it, no interaction is possible. Our goal is to minimize it as much as possible, down to the
minimum level that ensures the level of interaction required by the applications
exchanging fast data. In other words, the main goal is to ensure that each application knows just enough about the other to be able to interoperate but no more than
that, to avoid unnecessary dependencies and constraints. This is consistent with the
principle of least knowledge [33].
Minimizing coupling maximizes the likelihood of finding suitable alternatives or
replacements for applications, as well as the set of applications with which some
application is compatible, as a consumer or as a provider of some functionality.
This is precisely one of the slants of the variety problem in fast data.
Figure 1.3 depicts the scenario of an application immersed in its environment, in
which it acts as a provider for a set of applications (known as consumers), from
which it receives requests or event notifications, and as a consumer of a set of
applications (called providers), to which it sends requests or event notifications.
Coupling between this application and the other applications expresses not only
how much it depends on (or is affected by the variety in) its providers but also how
much its consumers depend on (or are affected by changes in) it. Dependency on an
application can be assessed by the level and amount of features that another
application is constrained by. Two coupling metrics can be defined, from the
point of view of a given application, as follows:
CF ( forward coupling), which expresses how much a consumer application is
dependent on its providers, is defined as

An application

Backward coupling

15

Providers

Consumers

1 An Interoperability Framework and Distributed Platform for Fast Data. . .

Forward coupling

Fig. 1.3 Coupling between an application and its consumers and providers

P
CF

Upi
Tpi N i

i2P

j Pj

1:1

where P is the set of providers that this application uses and |P| denotes the
cardinality of P, Upi is the number of features that this application uses in
provider i, Tpi is the total number of features that provider i exposes and Ni is
the number of providers with which this application is compatible as a consumer,
in all uses of features of provider i by this application.
CB (backward coupling), which expresses how much impact a provider application has on its consumers, is defined as
P
CB

i2C

Uci
Tc M

jCj

1:2

where C is the set of consumers that use this application as provider and |C|
denotes the cardinality of C, Uci is the number of features of this application that
consumer i uses, Tc is the total number of features that this application exposes
and M is the number of known applications that are compatible with this
application and can replace it, as a provider.
The conclusion from metric 1.1 above is that the existence of alternative providers to an application reduces its forward coupling CF, since more applications
(with which this application is compatible, as a consumer) dilute the dependency.
Similarly, the conclusion from metric 1.2 above is that the existence of alternatives
to an application as a provider reduces the system dependency on it, thereby
reducing the impact that application may have on its potential consumers and
therefore its backward coupling CB.
Current application integration technologies, such as Web services [34] and
RESTful applications [35], do not really comply with the principle of least

16

J.C.M. Delgado

knowledge and constitute poor solutions in terms of coupling. In fact, both require
interacting applications to share the type (schema) of the data exchanged, even if
only a fraction of the data values is actually used. A change in that schema, even if
the interacting applications do not actually use the part of the schema that has
changed, implies a change in these applications, because the application that
receives the data must be prepared to deal with all the values of that schema.
Web services rely on sharing a schema (a document expressed in WSDL or Web
Services Description Language) and RESTful applications require data types that
have already been previously agreed upon. These technologies solve the distributed
interoperability problem but not the coupling problem. This is a consequence of the
classical document-style interaction, heralded by XML and schema languages, as
illustrated by Fig. 1.4a. This is a symmetric arrangement in which a writer produces
a document according to some schema, and the reader uses the same schema to
validate and to read the contents of the document. There is no notion of services,
only the passive resource that the document constitutes. We are at the level of data
description languages, such as XML or JSON.
Figure 1.4b introduces the notion of service, in which a message is sent over a
channel and received by the receiver. It is now treated as a parameter to be passed
on to some behaviour that the receiver implements, instead of just data to be read.
However, the message is still a document, validated and read essentially in the same
way as in Fig. 1.4a. We are at the level of Web services or RESTful applications.
(a)
Schema
read

read

Reader

Writer
refer to
Data
structure

write

Data
structure

Document
validate
& read

(b)
Schema
read

read

Sender

Receiver
refer to

Data
structure

send

Message
sent

channel

refer to
Message
received

receive,
validate &
read

Fig. 1.4 Interaction styles: (a) documents; (b) document-based messages

Data
structure

1 An Interoperability Framework and Distributed Platform for Fast Data. . .

17

The behaviour invoked can thus be exposed and implemented in various ways, but
in the end, the goal is still similar.
The schemas in Fig. 1.4 refer to type specifications and need not be separate
documents, as it is usual in XML schema and WSDL. In the REST world, schemas
are known as media types but perform the same role. The difference is that instead
of being declared in a separate document referenced by messages, they are usually
previously known to the interacting applications, either by being standard or by
having been previously agreed upon. In any case, the schema or media type must be
the same at both sender and receiver endpoints, which imposes coupling between
the applications for all its possible values of a schema, even if only a few are
actually used.
Another problem concerns the variety in the networks. The protocols underlying
the Web (TCP/IP and HTTP) are demanding in terms of the smaller devices.
Hardware capabilities are increasingly better, and efforts exist to reduce the
requirements, such as IPv6 support on low-power wireless personal area networks
(6LoWPAN) [36] and the Constrained Application Protocol (CoAP) [37], to deal
with constrained RESTful environments. Building on the simplicity of REST,
CoAP is a specification of the Internet Engineering Task Force (IETF) working
group CoRE, which deals with constrained RESTful environments. CoAP includes
only a subset of the features of HTTP but adds asynchronous messages, binary
headers and User Datagram Protocol (UDP) binding. It is not easy to have the Web
everywhere, in a transparent manner.

1.4.3

The Velocity Dimension

Velocity in fast data is not just about receiving, or sending, many (and/or large)
messages or events in a given timeframe. The fundamental problem is the reaction
time (in the context of the timescale of the applications) between a request message,
or an event notification, and the availability of the processed results. Real-time
applications usually depend on a fast feedback loop. Messages frequently take the
form of a stream. Complex event processing [38], in which a set of related events is
analysed to aggregate data or detect event patterns, is a common technique to filter
unneeded events or messages and thus reduce the real velocity of processing and of
computing requirements.
Several factors affect velocity, e.g. processing time and message throughput. In
our context, emphasizing interoperability, the main problem lies in throughput of
messages exchanged. We cannot assume that data have been previously stored with
a view to be processed afterwards.
Throughput depends on message latency (the time taken to serialize data and
then to validate and to reconstruct it at the receiver) and on network throughput (its
own latency and bandwidth). The most relevant issue for velocity is the message
latency, since the characteristics of the network are usually fixed by the underlying
communications hardware. Latency is affected by several factors, including the

18

J.C.M. Delgado

service-level platform (e.g. Web services, RESTful applications), the message

protocol (e.g. HTTP, SOAP, CoAP) and the message data format (e.g. XML,
JSON).
Web services, SOAP and XML are the most powerful combination and are
essentially used in enterprise-class applications, which are the most demanding.
But they are also the most complex and the heaviest, in terms of processing
overheads. RESTful applications, HTTP and JSON are a simpler and lighter
alternative, although less powerful and exhibiting a higher semantic gap in modelling real-world entities [39]. Nevertheless, application designers will gladly trade
expressive power for simplicity and performance (which translates into velocity),
and the fact is that many applications do not require all the features of a full-blown
Web services stack. REST-based approaches are becoming increasingly
popular [40].
In any case, these technologies evolved from the original Web of documents
[41], made for stateless browsing and with text as the dominant media type. The
Web was developed for people, not for enterprise integration. Today, the world is
rather different. The Web is now a Web of services [42], in which the goal is to
provide functionality, not just documents. There are now more computers
connected than people, with binary data formats (computer data, images, video
and so on) as the norm rather than the exception.
Yet, the evolution has been to map the abstraction of Web of services onto the
Web of documents, with a major revolution brought by XML and its schemas. The
document abstraction has been retained, with everything built on top of it. In the
interoperability levels of Table 1.1, XML (or JSON, for that matter) covers only the
syntactic category. The lack of support of XML for higher interoperability levels
(viz. at the service interface level) is one of the main sources of complexity in
current technologies for integration of applications. In turn, this imposes a significant overhead in message latency and, by extension, velocity.
Table 1.2 summarizes the main limitations of existing technologies that are
particularly relevant for this context.

1.4.4

Modelling with Resources and Services

Any approach should start with a metamodel of the relevant entities. In this case,
the organization of applications and their interactions are primordial aspects. The
interoperability levels of Table 1.1 constitute a foundation for the metamodel,
although this chapter concentrates on the syntactic, semantic and pragmatic
categories.
The interaction between different applications cannot simply use names,
because the contexts are different and only by out-of-band agreements (such as
sharing a common ontology) will a given name have the same meaning in both sides
of an interaction.

1 An Interoperability Framework and Distributed Platform for Fast Data. . .

19

Table 1.2 Main limitations of relevant interoperability technologies

Technology
XML

JSON
HTTP

SOAP

Web Services
(SOA)

REST

Main limitations
Text based, verbose, complex, poor support for binary, data description only
(no behaviour), syntax-level only (higher levels require other languages),
high coupling (interoperability achieved by schema sharing)
Simpler but less powerful than XML and the same limitations. Coupling is
also high since data types have to be agreed prior to data interaction
Optimized for Web browsing (scalable retrieval of hypermedia information)
but not for generic and distributed service-based interactions. Inefficient and
specific text-based control information format. Synchronous, committed to
the client-server paradigm, lack of support for the push model (serverinitiated interactions) and for binary data
Complex, significant performance overheads, XML-based (thus inheriting
XMLs problems), too high level, with specific solutions for nonfunctional
data and routing
Complex (based on WSDL, SOAP and XML), many standards to cover
distributed interaction aspects, high coupling (interoperability achieved by
sharing the WSDL document), lack of support for structure (flat service
space)
Forces a create, read, update, delete (CRUD) approach to model real-world
entities, causing a significant semantic gap in generic service modelling.
Coupling is disguised under the structured data and the fixed syntactic
interface. Structure of data returned by the server may vary freely, but the
client needs to have prior knowledge of the data schemas and of the expected
behaviour of the server

XML-based systems, for example, solve this problem by sharing the same
schema, as illustrated by Fig. 1.4. However, this is a strong coupling constraint
and contemplates data only. Behaviour (operations) needs to be simulated by data
declarations, as in WSDL documents describing Web services.
We need to conceive a more dynamic and general model of applications and
their interactions, which supports interoperability without requiring to share the
specification of the application interface (schema). The strategy relies on structural
type matching, rather than nominal type matching. This approach entails:
A small set of primitive types, shared by all applications (universal upper
ontology)
Common structuring mechanisms, to build complex types from primitive ones
A mechanism for structurally comparing types from interacting applications
Applications are structured, and, in the metamodel as described below, their
modules are designated as resources. Since applications are distributed by definition, sending each other messages through an interconnecting channel is the only
form of interaction. To make intra-application interactions as similar as possible to
inter-application interactions, all resources interact by messages, even if they
belong to the same application. Resources are the foundation artefacts of the
metamodel of applications and of their interactions as depicted in Fig. 1.5.

20

J.C.M. Delgado

Fig. 1.5 A structural metamodel of resources and of their interactions

This metamodel can be briefly described in the following way:

A resource is an entity of any nature (material, virtual, conceptual, noun, action
and so on) that embodies a meaningful, complete and discrete concept, making
sense by itself while being distinguishable from, and able to interact with, other
entities.
Resources that interact do so by sending each other messages. A message is itself
a resource that migrates (through some channel, usually a network, using some
protocol) from the context of the resource that sends it to the context of the
resource that receives it. Messages can be either requests or responses.
Operations (the equivalent of methods in objects) are first-class resources and
can be sent as message components. All instantiations of an operation (one for
each message received, with lazy creation) are closures [43], but distributed
messages (between applications) must be self-contained (no free variables). The
advantages of using operations as full-fledged resources instead of methods or
functions are uniformity of treatment (everything is a resource), support for
higher-order programming (operations as arguments to other operations) and

1 An Interoperability Framework and Distributed Platform for Fast Data. . .

21

transparent event processing (the occurrence of an event sends a message to a

resource, which can be an operation or any other kind of resource).
A link can target more than one resource, constituting a list of remote references,
for consumer-side redundancy. If a resource targeted by the link is unavailable,
the support system can automatically try to resend the message to the next in
the list.
A service is a set of logically related reactions of a resource to the messages it
can receive and react to (depends on the operations that are components of that
resource). A service can be seen as pure behaviour, although the implementation
of concrete reactions may depend on state, which needs a resource as an
implementation platform, or have side effects, changing that state.
Sending a request message from a resource A to a resource B (which eventually
responds) constitutes a service transaction between these two resources, in
which the resources A and B perform the roles of service consumer and service
provider, respectively. A service transaction can entail other service transactions, as part of the chain reaction to the message by the service provider. A
service is defined in terms of reactions to messages (external view) and not in
terms of state transitions or activity flows (internal view).
A process is a graph of all service transactions that can occur, starting with a
service transaction initiated at some resource A and ending with a final service
transaction, which neither reacts back nor initiates new service transactions. A
process corresponds to a use case of resource A and usually involves other
resources as service transactions flow.

The metamodel depicted in Fig. 1.5 has been conceived to cover the main levels
of interoperability described in Table 1.1 and as a contribution to solve the
limitations of existing technologies, summarized in Table 1.2.
Interoperability, in terms of the current technological developments, is based on
data description languages such as XML and JSON, which cover just the syntactic
category of Table 1.1. Semantics has been an afterthought, with languages such as
RDF and OWL [44]. Pragmatics (behaviour) has been dealt with through generic
programming languages or, under the process paradigm, with BPEL [31]. Services
have been modelled essentially with WSDL documents, which are no more than an
interface specification simulated on top of a data description language (XML). The
REST architectural style uses a fixed service specification at the syntactic level and
restricts variability to resources and their structure. In essence, no solution today
covers Table 1.1 adequately.
Resources in Fig. 1.5 model real-world entities, with their interfaces modelled as
services. Dynamically, services interact by messages, requests and responses, over
some channel with some protocol. These constitute transactions that build up
processes. Structured resources refer to others by component names, or roles, in
the sense of Description Logic [45], and allow semantics to be incorporated in the
model from scratch (details not described here due to space limitations). Operations
are modelled as resources, in the same manner as entities typically modelled by a

22

J.C.M. Delgado

class. This makes the metamodel simpler and more canonical and allows a finer
grain of application modelling.

1.4.5

Handling the Variety Dimension

Section 1.4.2 has shown that reducing coupling is a good approach to deal with the
variety problem, by increasing the potential base of consumers and providers. Most
resources are made interoperable by design, i.e. conceived and implemented to
work together, in a consumer-provider relationship. Therefore, coupling is high,
and it is not easy to accommodate changes and deal with variants of these resources.
For instance, Web services are a typical interoperability solution but work by
sharing schemas (Fig. 1.4b) and establish coupling for all the possible documents
satisfying each schema, even if they are not actually used. Searching for an
interoperable service is done by schema matching with similarity algorithms [46]
and ontology matching and mapping [47]. This does not ensure interoperability and
manual adaptations are usually unavoidable.
This chapter introduces a different perspective on interoperability, stronger than
similarity but weaker than sharing schemas and ontologies. The trick is to allow
partial (instead of full) interoperability, by considering only the intersection
between what the consumer needs and what the provider offers. If the latter
subsumes the former, the degree of interoperability required by the consumer is
feasible, regardless of whether the provider supports additional features or not.
When this is true, the consumer is said to be compatible with the provider or, more
precisely, that a resource X is compatible with a resource Y regarding a consumerprovider relationship. In this respect, two resource relationships are of relevance:
Consumer-provider: Compatibility between a consumer and a provider is known
as compliance [48]. The consumer must satisfy (comply with) the requirements
established by the provider to accept requests sent to it, without which these
cannot be validated, understood and executed. It is important to note that any
consumer that complies with a given provider can use it, independently of
having been designed for interaction with it or not. The consumer and provider
need not share the same schema. The consumers schema needs only to be
compliant with the providers schema in the features that it actually uses (partial
compliance). Since distributed resources have independent lifecycles, they cannot freely share names and types, and schema compliance must be tested
structurally, feature by feature, between messages sent by the consumer and
the interface offered by the provider;
Provider-provider: The issue is to ascertain whether a provider Y, serving a
consumer X, can be replaced by another provider Z in a way that the consumerprovider relationship enjoyed by X is not impaired. In other words, the issue is
whether Z is replacement compatible with Y. Replacement compatibility
between two providers is known as conformance [49, 50]. The provider must

1 An Interoperability Framework and Distributed Platform for Fast Data. . .

New provider
specification

Provider
specification

Consumer

23

New provider

Compliance

View provider as

View consumer as

Conformance

Compliance

Conformance

Compliance

View provider as

View consumer as

Conformance

New
consumer

New consumer
specification

Consumer
specification

Provider

Fig. 1.6 Resource compatibility, by use and replacement

fulfil the expectations of the consumer regarding the effects of a request (including eventual responses), therefore being able to take the form of (to conform to)
whatever the consumer expects it to be. Note that a provider may be conformant
to another with respect to one consumer but not with respect to another consumer. It all depends on the set of features used by the consumer. The reasons for
replacing a provider with another may be varied, such as switching to an
alternative in case of failure or lack of capacity of Y, evolution of Y (in which
case Z would be the new version of Y), or simply a management decision. The
important aspect to note is that Z does not need to support all the features of Y,
but just those that X actually uses (partial conformance).
Compliance and conformance are not symmetric relationships (e.g. if
X complies with Y, Y does not necessarily comply with X). Figure 1.6 illustrates
compliance and conformance, as well as its effects on reducing coupling, either by
limiting dependency to the required features or by increasing the set of possible
alternative providers or consumers.
A resource X, in the role of consumer, has been designed to interact with a
resource, in the role of provider, with a specification Y, which represents the set of
features that X requires as its provider. Nothing prevents the actual provider from
having additional features, as long as it includes this set.
Now consider a resource W, in the role of provider, which has been designed to
expect a consumer Z, which represents the set of features that W expects that the
actual consumer will access. Nothing prevents the consumer from accessing less
features, as long as they are included in this set.
X and W were not designed to interoperate, since what one expects is not what
the other is. They have been developed independently, or at least do not constitute a
full match of features (in other words, do not share the same schema). However,
interoperability is still possible, as long as the following holds:
Y complies with Z, at least partially. This means that the set of features that
Y represents is a subset of the set of features that Z represents. In turn, X cannot
access a feature that W does not expect to be accessed.

24

J.C.M. Delgado

Z conforms to Y, at least partially. This means that the set of features that
Z represents is a superset of the set of features that Y represents.
Although it seems that these two items assert the same thing, it is not the case
when optional features exist, such as optional arguments to functions or optional
fields in an XML document. For compliance, the consumer needs to specify only
the mandatory features (minimum set), whereas for conformance, all features
(mandatory and optional) must be supported by the provider (maximum set).
Now consider adding a new consumer A to provider W, or simply replacing X. A
does not have to bear any relationship to X, the previous consumer. It may even use
a completely different set of features of W (represented as specification B) as long as
B complies with Z and Z conforms to B. This also applies if A is a changed version
of X.
A similar reasoning can be made if X now uses a new provider, D, with the set of
requirements on consumers represented by specification C. Again, interoperability
is still possible as long as Y complies with C and C conforms to Y. Apart from this,
D and W need not bear any relationship. This also applies if D is a changed version
of W.
Reducing dependencies to the set of actually used features, instead of having to
share the full set of features (the schema) is an improvement over current technologies, typically based on XML or JSON documents. It also ensures interoperability,
whereas similarity-based approaches [46] do not.
We can now extend Fig. 1.4 with another interaction style, compliance-based
messages, as illustrated by Fig. 1.7.
Messages do not obey some external schema. Each message has one specific
value (most likely structured, but it is not a type, only one specific value) and its
own exclusive schema, which is nothing more than a self-description, without the
value variability that a type exhibits. This value and its description can be validated
against an infinite number of schemas, those that have this particular value included
in the set of their instances.
The receiver in Fig. 1.7 exposes a schema that defines the values it is willing to
accept. When a message is received, its schema is checked against the receivers
own schema. If it complies (satisfies all the requirements of the receivers schema),
the message can be accepted and processed. The advantage of this is that a resource
can send a message to all the resources with schemas that the message complies
Receiver
schema
read

check
Sender

Data
structure

check &
validate
Schema
send

Message
sent

Schema
channel

Message
received

Receiver
Compliance
checker

receive,
read &
map

Fig. 1.7 Extending the interaction styles of Fig. 1.4 with compliance-based messages

Data
structure

1 An Interoperability Framework and Distributed Platform for Fast Data. . .

25

with and, conversely, a resource can receive messages from any resource that sends
messages compliant with its receiving schema.
In other words, coupling occurs only in the characteristics actually used by
messages and not in all the characteristics of the schemas used to generate the
message or to describe the service of the receiving resource. Since the schemas of
the message and of the receiver are not agreed upon beforehand, they need to be
checked structurally. Resources of primitive types have predefined compliance
rules. Structured resources are compared by the names of components (regardless
of order of declaration or appearance) and (recursively) by the compliance between
structured resources with matching names. Since the order of appearance of named
component resources may different in the message and in the receiving schema,
there is the need to map one onto the other (bottom of Fig. 1.7).
This is a form of polymorphism that increases the range of applicability of both
sender and receiver, constituting a means to reduce coupling to only what is
actually used. Sender and receiver no longer need to be designed for each other.
Compliance ensures usability, whereas conformance ensures replaceability. When
a resource is able to interact with another, although not entirely interoperable with
it, we say that they have partial interoperability.
We also need to handle the velocity dimension, but, since it involves performance, this is included in the description of the platform, in Sect. 1.5.1.

1.4.6

A Service Interoperability Language with Support

for Variety

Data description languages, such as XML and JSON [57], merely describe data and
their structure. If we want to describe resources and their services, we can use
WSDL, but the resulting verbosity and complexity has progressively turned away
developers in favour of something much simpler, REST [20]. If we want a programming language suitable for distributed environments, we can use BPEL [31]
but again with an unwieldy XML-based syntax that forces programmers to use
visual programming tools that generate BPEL and increase the complexity stack.
JSON is much simpler than XML, and its popularity has been constantly
increasing, but the evolution of the dynamic nature of the Web, as shown by
JavaScript and HyperText Markup Language, version 5 (HTML5) [51], hints that
data description is not enough anymore and distributed programming is a basic
requirement. In the IoT, with machine-to-machine interactions now much more
frequent than human-to-server interactions, this is even of greater importance.
Current Web-level interoperability technologies are greatly constrained by the
initial decision of basing Web interoperability on data (not services) and text
markup as the main description and representation format. Using them to implement the metamodel of Fig. 1.5 leads to a complex stack of specifications and to
syntactic and semantic adaptations and compromises, to bridge the gap between the
metamodel and the chosen language and/or platform. A language conceived from

26

J.C.M. Delgado

scratch to implement this metamodel, providing native support for its characteristics, will be a better match. We have designed Service Interoperability Language
(SIL) [56] to provide this support.
Each resource in SIL is represented by a SIL Public Interface Descriptor (SPID),
which corresponds to a Web services WSDL but much more compact and able to
describe both structure and operations. It is obtained automatically from the
resource description itself by including only the public parts (public component
resources, including operation headings). Unlike XML or even JSON, there is no
separate schema document to describe a resource.
Program 1.1 shows a simple example of resources described in SIL, involving
sensors in the context of the IoT [6]. It includes a temperature controller
(tempController), with a list of links to temperature sensors with history
(tempSensorStats). Each of these has a link to a remote temperature sensor
(tempSensor). The lines at the bottom illustrate how these resources can be used
and should be included in some resource that uses them.
1.1 Describing and Using Resources Using SIL
tempSensor: spid {// descriptor of a temperature sensor
getTemp: operation (-> [-50.0 .. 60.0]);
};
tempSensorStats: {// temperature sensor with statistics
sensor: @tempSensor; // link to sensor (can be remote)
temp: list float; // temperature history
startStats <||; // spawn temperature measurements
getTemp: operation (-> float) {
reply sensor@getTemp<--; // forward request to sensor
};
getAverageTemp: operation ([1 .. 24] -> float) {
for (j: [temp.size .. temp.size-(in-1)])
out temp[j];
reply out/in; // in number of hours
};
private startStats: operation () {// private operation
while (true) {
temp.add<-- (getTemp<--); // register temperature
wait 3600; // wait 1 hour and measure again
}
}
};
tempController: {// controller of several temperature
sensors
sensors: list @tempSensorStats; //list of links to sensors
addSensor: operation (@tempSensor) {
t: tempSensorStats; // creates a tempSensorStats resource

1 An Interoperability Framework and Distributed Platform for Fast Data. . .

27

t.sensor in; // register link to tempSensor

sensors.add<-- @t; // add sensor to list
};
getStats: operation (-> {min: float; max: float;
average: float})
{
out.min sensors[0]@getTemp<--;
out.max out.min;
total: float : out.min; // initial value
for (i: [1 .. sensors.length-1) { // sensor 0 is done
t: sensors[i]@getTemp<--; // dereference sensor i
if (t <out.min) out.min t;
if (t >out.max) out.max t;
total t;
};
out.average total/sensors.length;
reply; // nothing specified, returns out
}
};
// How to use the resources
// tc contains a link to tempController
tc@addSensor<-- ts1; // link to a tempSensor resource
tc@addSensor<-- ts2; // link to a tempSensor resource
x: tc@sensors[0]@getAverageTemp<-- 10; // average last
10 hours
This program can be described in the following way:
The temperature sensor (tempSensor) is remote, and all we have is its SPID, the
equivalent to a Web services WSDL. For example, the SPID of
tempSensorStats can be expressed by the following lines:
tempSensorStats: spid {// temperature sensor with
statistics
sensor: @tempSensor; // link to sensor (can be remote)
temp: list float; // temperature history
getTemp: operation (-> float);
getAverageTemp: operation ([1 .. 24] -> float);
}
Resources and their components are declared by a name, a colon and a resource
type, which can be primitive, such as integer, a range (e.g. [1 .. 24]), float or user
defined (enclosed in braces, i.e. {. . .}). There are some resemblances to JSON,
but component names are not strings, and operations are supported as first-class
resources.
The definition of a resource is similar to a constructor. It is executed only once,
when the resource is created, and can include statements. This is illustrated by

28

J.C.M. Delgado

the statement startStats<|| in tempSensorStats. Actually, this is an asynchronous invocation (<||) of private operation startStats, which is an infinite loop
registering temperature measurements every hour. Asynchronous invocations
return a future [52], which will be later replaced by the returned value (in this
example, the returned future is ignored). Synchronous invocation of operations
is done with <, followed by the argument, if any.
Operations have at most one argument, which can be structured (with {. . .}).
The same happens with the operations reply value, as illustrated by operation
getStats. Inside operations, the default names of the argument and the value to
return are in and out, respectively. The heading of operations specifies the type
of the argument and of the reply value (inside parentheses, separated by ->);
Links to resources (indicated by the symbol @) are not necessarily URIs, not
even strings. They can also be structured and include several addresses, so that a
resource in a network (e.g. the Internet) can reference another in a different
network, with a different protocol (e.g. a sensor network). It is up to the nodes
and gateways of the network, supporting these protocols, to interpret these
addresses so that transparent routing can be achieved, if needed
Resource paths, to access resource components, use dot notation, except if the
path traverses a link, in which case a @ is used. For example, the path used in
the last line of Program 1.1 computes the average temperature, along the last
10 h, in sensor 0 of the controller.
Program 1.2 illustrates the compliance and conformance concepts, by providing
two additional temperature sensors (weatherSensor and airSensor) in relation to
Program 1.1. Only the additional and relevant parts are included here, with the rest
as in Program 1.1.
1.2 Example of Partial Interoperability with Structural Compliance
and Conformance
tempSensor: spid {
getTemp: operation (->[-50.0 .. 60.0]);
};
weatherSensor: spid {
getTemp: operation (->[-40.0 .. 50.0]);
getPrecipitation: operation (-> integer);
};
airSensor: spid {
getTemp: operation (->[-40.0 .. 45.0]);
getHumidity: operation (-> [10 .. 90]);
};
// tc contains a link to tempController
tc@addSensor<-- ts1; // link to a tempSensor resource
tc@addSensor<-- ts2; // link to a tempSensor resource
tc@addSensor<-- as1; // link to an airSensor resource
tc@addSensor<-- ws1; // link to an weatherSensor resource

1 An Interoperability Framework and Distributed Platform for Fast Data. . .

29

tc@addSensor<-- ws2; // link to an weatherSensor resource

tc@addSensor<-- as2; // link to an airSensor resource
x: tc@sensors[0]@getAverageTemp<-- 10; // average last
10 hours
s: {max: float; average: float};
s tc@getStats<--; // only max and average are assigned to s
temp: [-50.0 .. 50.0];
temp ws1@getTemp <--; // complies. Variability ok
temp ts1@getTemp <--; // does not comply. Variability
mismatch
In Program 1.2:
weatherSensor and airSensor conform to tempSensor, since they offer the same
operation and the result is within the expected variability. This means that they
can be used wherever a tempSensor is expected, which is illustrated by adding
all these types of sensors to tempController (through tc, a link to it) as if they
were of type tempSensor. Nonrelevant operations are ignored.
The result of invoking the operation getStats on tempController is a resource
with three components (as indicated in Program 1.1), whereas s has only two.
However, the assignment is still possible. Structural interoperability ignores the
extra component.
The last statement triggers a compliance check by the compiler, which issues an
error. The variability of the value returned by operation getTemp in tempSensor
(referenced by ts1) is outside the variability range declared for component temp.
Other features of SIL, not illustrated here for simplicity, include:
Delegation: An operation can execute some statements and then delegate the rest
of the execution of the request to another resource, by forwarding either the
received message or another one. The delegate will respond directly to the first
sender, the one that originated the request.
Context-awareness: Messages can be sent in two parts, data and context (separated by a with keyword). The latter is typically used to pass nonfunctional
information. Both parts are normal resources and obey the same compliance and
conformance rules as other resources. This applies to both request and response
messages, which means that SIL supports forward and backward context passing. The latter is useful both in normal responses and exceptions. Context
awareness is another way in which variety can be supported.
Reliability: The links in Programs 1.1 and 1.2 can target a list of resources. If for
some reason the protocol fails to communicate with the first resource in the link,
it will try automatically the following one. An exception is generated if all
alternatives fail. There is opportunity for recovery, at the level of the exception
handler (covering a set of statements) or at the application level.
Interface to other languages: Resources in SIL can invoke methods in objectoriented languages and have its operations invoked. The solution is similar to
XML-based systems, including Web services, and involves annotations. The

30

J.C.M. Delgado

SPID of a resource can be used to generate at compile time an interface in Java,

for instance, or reflection can be used to support access to resources and objects
only at runtime.

1.5

A Distributed Interoperability Platform

This section proposes a platform for distributed interoperability that revisits the
problem and constitutes a native solution for computer-based applications in a fast
data context, instead of being based on the classical technologies conceived originally for human browsing. Backward compatibility is lost, but this exploratory
approach has the potential to compensate with flexibility and adequacy.

1.5.1

Handling the Velocity Dimension

We used several techniques to reduce message processing latency, including:

Unlike XML or JSON, SIL is compiled. This corresponds to having two resource
representations, text (source) and compiled (binary), with coherence between the
two maintained by the compiler.
A binary format based on the tag, length and value (TLV) scheme used by
ASN.1 [53], which allows a very efficient traversal of a serialized data structure.
Messages can be sent with an adjustable level of metadata, from source (including SPID) to almost no metadata at all (just structures composed of primitive
resources). Metadata means overhead and does not need to be sent in full in all
messages.
Compliance checks (mappings between a message and the data structure
expected by the receiver) can be cached, so that subsequent messages of the
same type can be optimized.
Using efficient full-duplex transport level protocols with support for binary data,
such as Web sockets [51], with automatic session management.
Using a generic message-level protocol as the foundation, instead of an
application-level protocol such as HTTP, even in its HTTP/2 incarnation. The
protocol just deals with basic messages and not resource-level operations.
We have developed a compiler for SIL based on ANother Tool for Language
Recognition (ANTLR) [54], which is able to produce the SPID automatically and to
convert source to instructions and data in a binary format (TLV), or silcodes
(similar to Javas bytecodes). An interpreter, similar to a Java virtual machine
(JVM), executes these silcodes. The current implementation, in pure Java, is not
optimized and has a performance roughly 50 times slower than a JVM. Much of that
time is spent just on virtual method dispatch, the mechanism used to execute the
various silcodes. A C-based interpreter, for example, would be much faster,

1 An Interoperability Framework and Distributed Platform for Fast Data. . .

31

although harder to develop. To maintain flexibility and control of implementation,

we did not use a JVM.
The format of a compiled resource is exactly the same as a serialized resource, to
be sent as a message or stored in a persistent medium. It supports three levels of
metadata (self-description) information, in addition to the data itself: source code,
ontology (names and relationships) and no metadata. The latter is used when a
previous data-type checking was cached by the message receiver and a token
returned to the sender, to be used in subsequent messages. The protocol returns
an error if metadata is absent and there is no valid token in the message. The sender
then repeats the message, this time with metadata. The ontology-level metadata is
the usual case without cache optimization. The source code can be included, if
desired, to provide user-level readability.
Table 1.3 gives an idea of the size of resource representations in various
situations. The example refers to a data-only resource, so that XML and JSON
can be used for comparison. The two lines refer to the same resource, but with
longer and shorter component names, which have an impact on metadata size. That
is why the sizes in the last column are the same, and the size reduction with less
metadata is greater when names are longer. The sizes presented vary with the
concrete example, but this gives an idea of the values involved.
An end-to-end application performance comparison with SOA- and REST-based
solutions has not been done yet. Trivial examples tend to assess essentially the
application servers and network latency, which are not the distinguishing aspects.
A meaningful comparison needs to separate concerns, such as protocols, data parsing,
platform-agnostic execution (SIL) versus local programming language (e.g. Java),
professional versus research implementation and so on. This will be done in the near
future. In the meantime, it has been shown that parsing binary data markup (TLV) can
be at least an order of magnitude faster than parsing XML text markup [55].
Unlike HTTP, which goes down from the transport level up to the service level,
with verbs such as GET or PUT, the SIL protocol deals only with messages, addressed
to resources. Its goal is to support message-level interaction and nothing else. What to
do with the message is the receivers responsibility. Unlike SOAP, which defines
headers for additional information such as security, the SIL protocol delegates to
higher interoperability levels the responsibility to add extra information. This is done
by wrapping simple messages within other messages, with the extra information, in
what constitutes an open extension to the basic protocol. Both interacting applications
need to support the protocol extension mechanism, which may be standardized or be
specific to a pair of applications. The basic protocol remains simple.
Table 1.3 Comparison of the sizes (in bytes) of several resource representations, from source to
binary only
Names
Long
Short

XML
2595
1972

JSON
1631
1193

Silcodes
With source
1432
1014

With ontology
953
621

Without metadata
382
382

32

J.C.M. Delgado

The SIL protocol requires only an underlying addressable message-level transport protocol, which does not have to be reliable. Message level reliability, if not
provided by the transport protocol, can be implemented at the resource or application level. This is intended as a means to support heterogeneity, namely, in
non-TCP/IP protocols, such as those used in sensor networks [26], or in the IoT,
at the UDP level, with protocols such as CoAP [37]. In the opposite direction,
security can be provided by the underlying protocol itself, if it is complex enough to
support secure transport.
Table 1.4 illustrates some of the message types in this protocol, which implements basic transactions in a universal way. Any resource can be both sender and
receiver of messages.
There is no notion of session. Messages are independent of each other, with
coherence maintained by higher interoperability levels, namely, the pragmatic and
symbiotic categories in Table 1.1. However, the underlying transport protocol can
open a channel, such as a socket in the Web socket protocol [51], and maintain it
open for some predefined time (TTL or time to live) or until some consecutive
inactivity period occurs. These timings can be defined by any of interacting
applications (either of them can close the channel). If subsequently a message
needs to be sent between the interacting applications, the channel is open again,
automatically. Timings can be adaptive, depending on message traffic.
The correlation between request and response messages is made by a token value
generated at the senders server (by a pseudo-random number generator) for each
message sent. It is meaningful in the context of that server only. The receiver of the
request must copy this token to the response and forward it in delegation messages,
so that the delegate can also include it in the response. The implementation of
futures also uses this mechanism.

Table 1.4 Examples of message types

Message category/type
Request
React
React & respond
Asynchronous react & respond
Amendment
Cancel
Response
Answer
Resource fault
Protocol fault
Notification
Denial
Done
Cancelled

Description
Initial request in each transaction
React to message, no answer expected
React to message and answer/notify
Asynchronous data message. Return future immediately
Further information on an already sent request
Cancel the execution of the request
Response to the request
Data returned as a response
Data returned as the result of an exception
Error data resulting from protocol or partner failure
Information of completion status
Confirm rejection of request
Request completed but has no value to reply
Confirm cancellation of request

1 An Interoperability Framework and Distributed Platform for Fast Data. . .

1.5.2

33

Architecture of the Platform

The platform is essentially a distributed collection of SIL nodes, each confined to

one computing server. Each SIL node can connect to more than one network, as
shown in Fig. 1.8, which details the architecture of one node. This requires the
provision of some server or handler to receive and send messages through each
networks protocol. The protocol at message level (Table 1.1) is identical in all
cases.
The architecture of a SIL node can be briefly described as follows:
The application server (Jetty, in our implementation, but any other would do) is
the interface to the Internet. It can support several protocols, although we have
only catered for two, HTTP (to handle Web requests) and Web sockets
(to support the message protocol of SIL). There is a list of protocol handlers
and each in turn checks whether it recognizes the message format. The first one
to do so gets the message for further processing. All SIL messages begin with
SIL, encoded in UTF-8.
Other types of network protocols, such as those of sensor networks, can be
supported as long as there are handlers to use those protocols at the transport
level.
Sensor

Sensor

SIL
node

SIL
node

SIL
node

Sensor

Sensor

Sensor
network

Server

Internet

SOA or REST
service

Sensor
network
handler

Directory

HTTP
handler
Jetty
server

Resource
manager

Resources

Message
manager
Compliance
checker

Interpreter

Web
Socket
handler
Resource
serializer

SIL node

Fig. 1.8 Basic architecture of a SIL node, expressing its multi-network capability

34

J.C.M. Delgado

A message received (by any of the protocols) is handled by the message

manager, which determines the recipient of the message, the type of the message
(according to the SIL message protocol, described in Table 1.4), and whether a
compliance token is present (described in Sect. 1.5.1);
The resource manager implements the access to the structured resources registered in the directory, obtaining internal links (indices to a resource table) to
resources targeted by messages or by distributed links (the equivalent of URIs).
The directory is the root resource in each SIL node and implements several
operations, such as searching for a resource with a service that conforms to a
given SPID. The resource tree depicted in Fig. 1.8 shows only containment
relationships. Any resource can have a distributed link to another but only if it is
registered as globally accessible in a directory. This means that resources can be
locally reached from others during execution of a SIL program, but only
registered resources, directly in the directory, can be addressed by a global,
distributed link.
The compliance checker performs type compliance between the message and the
addressed operation or resource. It can do so in text or binary formats, since each
has all the information needed, as long as the metadata is also present. Naturally,
this is faster when done in binary. Messages that include a compliance token,
obtained in a previous checking, can skip this step.
If the resource targeted by the message is not an operation, the compliance
checker must go through the various operations defined in that resource, to find
one, which the message complies with. A protocol fault (Table 1.4) is returned if
none is found.
Once the target operation has been identified, a thread is created in the silcode
interpreter to execute that operations code, produced previously by the SIL
compiler.
If the operation needs to send a message to a resource in another SIL node, as
illustrated by Program 1.1, it passes the message (which is a resource) to the
resource serializer, which produces the same format as the compiler.

1.6

Usefulness of the Approach

Long gone are the days when most of the relevant data were structured and static,
enabling schema-based optimizations that have supported conventional databases.
The success of the Web-made data become much more dynamic and unstructured,
and now we are in the era of big (unstructured) data. In particular, fast data, along
with velocity and variety dimensions, make design-time optimizations increasingly
difficult, if applicable at all.
This is why the solutions for application interoperability need to become more
dynamic, with decisions taken at runtime, in an adaptive manner. The approach
taken by this chapter entails two essential slants:

1 An Interoperability Framework and Distributed Platform for Fast Data. . .

35

To reduce the interoperability requirements at the schema level, through partial

compliance and conformance. Interacting applications need not be designed to
work together, either by sharing schemas or by agreeing on specific data types at
design time. Any application that happen to comply with another can use it as a
consumer, and any application that conforms to another can replace it as a
provider. Although not a magical or universal solution, it can definitely contribute to better support variants of some basic design. This is a contribution to the
variety dimension.
To optimize statistically while maintaining the flexibility. Checking for compliance and conformance is just too slow to be done in all messages. However,
messages arriving at a high rate tend to repeat a data pattern, which means that
the mapping between the data types of the messages and the receivers interface
can be cached and managed automatically. A simple token such as a 64-bit
integer, returned by the receiver in the first messages response and used
subsequently by the sender in future messages of the same type, can optimize
performance without losing the benefits of dynamic type checking. When a
mapping is cached, metadata need not be checked and only binary data is
relevant (Table 1.3). The more repetitive message patterns are, the more effective this technique is. This is a contribution to the velocity dimension.

1.7

Future Research

The proposals contained in this chapter constitute exploratory work. There is a

prototype of the language compiler, interpreter and node platform, but much work
is still needed before the approach can be validated in a sufficiently large scale. The
following activities are currently in development:
Assessing the fault-tolerant capabilities of the platform. SIL supports faulttolerant resource links (with alternative resources), but state checkpointing and
recovery are not implemented yet.
Assessing the scalability capabilities of the platform and assuming a large
number of small-grained applications, characteristic of the IoT. This will be
carried out by simulation.
Tackling the volume dimension with a general-purpose mindset. Programming
models such as MapReduce force a specific organization of data processing,
from input data until results are produced. Our approach is to use any type of
service (a task, in MapReduce parlance) or combination thereof and start with
the last service (the one producing the final result) and proceed backward,
invoking other services asynchronously. The final result, a collection of data
(resource) initially containing just futures, is progressively filled in with the
partial results returned by these services. Services are automatically replicated
by each request. The nuance is to allow automatic replication and spreading of
resources and their services, throughout a pool of SIL nodes. SIL already

36

J.C.M. Delgado

supports distributed resources, since components of a resource can be remote,

accessible by a link. The idea is to allow automatic replication of immutable
components (a feature of SIL not described in the example of this paper), as well
as code. This allows not only the code-to-data migration approach of Hadoop but
also data-to-code (in which new data, such as streaming messages, are send to
remote nodes to be processed there) or even resource-to-node (in which complete resources, code and data, are migrated to new nodes just to spread the
computing load);
Carrying out a comparative study, with qualitative and quantitative assessment,
between the distributed platform proposed in this chapter and other dynamic
platforms for real-time data processing [18].

1.8

Conclusions

Currently available platforms for big data are mainly centred on the volume
dimension and usually assume that data already exist, are immutable and have
been loaded into the system. This may be adequate for batch applications but not for
processing dynamic data, namely, data streams. Although there are platforms trying
to process large sets of data or events in real time or near it [18], these are still very
focused on data and events and do not yet constitute a general-purpose platform to
interconnect a large number of heterogeneous services. This is the scenario when
we consider the Web and the IoT, in which the number of interacting entities and
messages is very large but do not constitute a coherent, or even related, large set of
data or events. This chapter intends to constitute a step towards such a generalpurpose platform.
Current integration technologies evolved from basic browsing to web services
by privileging backward compatibility over adequacy to computer-to-computer
interactions. As a result, they exhibit many limitations for systems in which the
number of entities is large and data granularity (messages) is small. New integration
approaches, with native solutions to the problems of today, need to be devised.
This chapter contends that new solutions to fast data should adopt the following
approaches:
Languages should describe not only passive data (XML, JSON) but also operations, describing structured resources that offer services, so that most of the
interoperability levels of Table 1.1 are covered natively, instead of having to
build languages over languages, which increases the complexity and reduces the
performance.
A resource-based metamodel should be used as the foundation, to minimize the
semantic gap in modelling real-world entities, in structure, state and behaviour.
Interoperability should be allowed to be partial, based on compliance and
conformance. This reduces service coupling with regard to sharing of schemas

1 An Interoperability Framework and Distributed Platform for Fast Data. . .

37

(XML, Web services) or of types (JSON, REST), providing a better support for
heterogeneity of services and data types.
The distributed platform should be based on nodes as universal and as independent from the network protocol as possible. This way, it can be overlaid on
different networks in a more transparent way.
The resource serialization format should be natively binary and adjust automatically the level of metadata included, to minimize message size.
The message protocol should be simple but efficient. Features that are more
complex should be built on top of the basic mechanisms, according to the levels
as shown in Table 1.1.

References
1. Zikopoulos P et al (2012) Understanding big data. McGraw-Hill, New York
2. White T (2012) Hadoop: the definitive guide. OReilly Media, Inc., Sebastopol
3. Zaharia M, Chowdhury M, Franklin M, Shenker S, Stoica I (2010) Spark: cluster computing
with working sets. In: Proceedings of the 2nd USENIX conference on Hot topics in cloud
computing, Boston, MA, pp 10
4. Mishne G, Dalton J, Li Z, Sharma A, Lin J (2013) Fast data in the era of big data: twitters realtime related query suggestion architecture. In: Proceedings of the international conference on
management of data, Ahmedabad, India, pp 11471158
5. Luigi A, Iera A, Morabito G (2010) The internet of things: a survey. Comput Netw
54:27872805
6. Gubbi J, Buyya R, Marusic S, Palaniswami M (2013) Internet of things (IoT): a vision,
architectural elements, and future directions. Futur Gener Comp Syst 29(7):16451660
7. Sundmaeker H, Guillemin P, Friess P, Woelffle S (2010) Vision and challenges for realising
the internet of things. European Commission Information Society and Media. http://bookshop.
europa.eu/en/vision-and-challenges-for-realising-the-internet-of-things-pbKK3110323/.
Accessed 20 Apr 2015
8. Schaffers H et al (2011) Smart cities and the future internet: towards cooperation frameworks
for open innovation. In: Domingue J et al (eds) The future internet. Springer, Berlin/Heidelberg, pp 431446
9. Delgado J (2014) The role of compliance and conformance in software engineering. In:
Ghani I, Kadir W, Ahmad M (eds) Handbook of research on emerging advancements and
technologies in software engineering. IGI Global, Hershey, pp 392420
10. Ohlhorst F (2012) Big data analytics: turning big data into big money. Wiley, Hoboken
11. Murdoch T, Detsky A (2013) The inevitable application of big data to health care. J Am Med
Assoc 309(13):13511352
12. Holzinger A (2014) Biomedical informatics: discovering knowledge in big data. Springer
International Publishing, Cham
13. Demchenko Y, Zhao Z, Grosso P, Wibisono A, de Laat C (2012) Addressing Big Data
challenges for scientific data infrastructure. In: Proceedings of the IEEE 4th international
conference on cloud computing technology and science, Taipei, Taiwan, pp 614617
14. Mayer-Schonberger V, Cukier K (2013) Big data: a revolution that will transform how we live,
work, and think. Houghton Mifflin Harcourt, New York
15. NIIST (2014) DRAFT NIST Big Data interoperability framework: volume 6, reference architecture.
http://bigdatawg.nist.gov/_uploadfiles/BD_Vol6-RefArchitecture_V1Draft_Prerelease.pdf. Accessed 20 Apr 2015

38

J.C.M. Delgado

16. Dean J, Ghemawat S (2008) MapReduce: simplified data processing on large clusters.
Commun ACM 51(1):107113
17. Heisterberg R, Verma A (2014) Creating business agility. Wiley, Hoboken
18. Liu X, Iftikhar N, Xie X (2014) Survey of real-time processing systems for big data. In:
Proceedings of the 18th international database engineering & applications symposium, Porto,
Portugal, pp 356361
19. Erl T (2007) SOA: principles of service design. Prentice Hall PTR, Upper Saddle River
20. Fielding R (2000) Architectural styles and the design of network-based software architectures.
Doctoral dissertation, University of California at Irvine. http://www.ics.uci.edu/~fielding/
pubs/dissertation/fielding_dissertation_2up.pdf. Accessed 20 Apr 2015
21. Chappell D (2004) Enterprise service bus. OReilly Media, Inc., Sebastopol
22. Armbrust M et al (2010) A view of cloud computing. Commun ACM 53(4):5058
23. Fernando N, Loke S, Rahayu W (2013) Mobile cloud computing: a survey. Futur Gener Comp
Syst 29(1):84106
24. Keyes J (2013) Bring your own devices (BYOD) survival guide. CRC Press, Boca Raton
25. Aggarwal C, Han J (2013) A survey of RFID data processing. In: Aggarwal C (ed) Managing
and mining sensor data. Springer, New York, pp 349382
26. Potdar V, Sharif A, Chang E (2009) Wireless sensor networks: a survey. In: Proceedings of the
international conference on advanced information networking and applications workshops,
Bradford, UK, pp 636641
27. Hartenstein H, Laberteaux K (eds) (2010) VANET: vehicular applications and internetworking technologies. Wiley, Chichester
28. Karen T (2013) How many internet connections are in the world? Right. Now. http://blogs.
cisco.com/news/cisco-connections-counter/. Accessed 20 Apr 2015
29. United Nations (2013) World population prospects: the 2012 revision, key findings and
advance tables. Working Paper No. ESA/P/WP.227. United Nations, Department of Economic
and Social Affairs, Population Division. http://esa.un.org/unpd/wpp/Documentation/pdf/
WPP2012_%20KEY%20FINDINGS.pdf. Accessed 20 Apr 2015
30. Liebowitz J (ed) (2013) Big data and business analytics. CRC Press, Boca Raton
31. Juric M, Pant K (2008) Business process driven SOA using BPMN and BPEL: from business
process modeling to orchestration and service oriented architecture. Packt Publishing,
Birmingham
32. Shadbolt N, Hall W, Berners-Lee T (2006) The semantic web revisited. IEEE Intell Syst 21
(3):96101
33. Palm J, Anderson K, Lieberherr K (2003) Investigating the relationship between violations of
the law of demeter and software maintainability. In: Proceedings of the workshop on softwareengineering properties of languages for aspect technologies. http://www.daimi.au.dk/~eernst/
splat03/papers/Jeffrey_Palm.pdf. Accessed 20 Apr 2015
34. Papazoglou M (2008) Web services: principles and technology. Pearson Education Limited,
Harlow
35. Webber J, Parastatidis S, Robinson I (2010) REST in practice: hypermedia and systems
architecture. OReilly Media, Sebastopol
36. Jacobsen R, Toftegaard T, Kjrgaard J (2012) IP connected low power wireless personal area
networks in the future internet. In: Vidyarthi D (ed) Technologies and protocols for the future
of internet design: reinventing the web. IGI Global, Hershey, pp 191213
37. Castellani A, Gheda M, Bui N, Rossi M, Zorzi M (2011) Web services for the internet of things
through CoAP and EXI. In: Proceedings of the international conference communications
workshops, pp 16. doi:10.1109/iccw.2011.5963563
38. Cugola G, Margara A (2012) Processing flows of information: from data stream to complex
event processing. ACM Comp Surv 44(3):article 15
39. Delgado J (2012) Bridging the SOA and REST architectural styles. In: Ionita A, Litoiu M,
Lewis G (eds) Migrating legacy applications: challenges in service oriented architecture and
cloud computing environments. IGI Global, Hershey, pp 276302

1 An Interoperability Framework and Distributed Platform for Fast Data. . .

39

40. Adamczyk P, Smith P, Johnson R, Hafiz M (2011) REST and web services: in theory and in
practice. In: Wilde E, Pautasso C (eds) REST: from research to practice. Springer, New York,
pp 3557
41. Berners-Lee T (1999) Weaving the web: the original design and ultimate destiny of the world
wide web by its inventor. HarperCollins Publishers, New York
42. Tolk A (2006) What comes after the semantic web PADS implications for the dynamic web.
In: Proceedings of the 20th workshop on principles of advanced and distributed simulation,
Beach Road, Singapore, pp 5562
43. Jarvi J, Freeman J (2010) C lambda expressions and closures. Sci Comp Program 75
(9):762772
44. Grau B et al (2008) OWL 2: the next step for OWL. Web Semant Sci Serv Agents World Wide
Web 6(4):309322
45. Baader F et al (eds) (2010) The description logic handbook: theory, implementation, and
applications, 2nd edn. Cambridge University Press, Cambridge
46. Jeong B, Lee D, Cho H, Lee J (2008) A novel method for measuring semantic similarity for
XML schema matching. Expert Syst Appl 34:16511658
47. Euzenat J, Shvaiko P (2007) Ontology matching. Springer, Berlin
48. Kokash N, Arbab F (2009) Formal behavioral modeling and compliance analysis for serviceoriented systems. In: de Boer F, Bonsangue M, Hallerstede S, Leuschel M (eds) Formal
methods for components and objects. Springer, Berlin/Heidelberg, pp 2141
49. Kim D, Shen W (2007) An approach to evaluating structural pattern conformance of UML
models. In: Proceedings of the ACM symposium on applied computing, Seoul, Korea, pp
14041408
50. Adriansyah A, van Dongen B, van der Aalst W (2010) Towards robust conformance checking.
In: Proceedings of the business process management workshops, Hoboken, NJ, pp 122133
51. Lubbers P, Albers B, Salim F (2010) Pro HTML5 programming: powerful APIs for richer
internet application development. Apress, New York
52. Schippers H (2009) Towards an actor-based concurrent machine model. In: Rogers I
(ed) Proceedings of the 4th workshop on the implementation, compilation, optimization of
object-oriented languages and programming systems, Genova, Italy, pp 49
53. Dubuisson O (2000) ASN.1 communication between heterogeneous systems. Academic, San
Diego
54. Parr T (2013) The definitive ANTLR 4 reference. The Pragmatic Bookshelf, Raleigh
55. Sumaray A, Makki S (2012) A comparison of data serialization formats for optimal efficiency
on a mobile platform. In: Proceedings 6th international conference on ubiquitous information
management and communication, article 48. doi:10.1145/2184751.2184810
56. Delgado J (2014) Structural services: a new approach to enterprise integration. In: Wadhwa M,
Harper A (eds) Technology, innovation, and enterprise transformation. IGI Global, Hershey,
pp 5091
57. Galiegue F, Zyp K (eds) (2013) JSON schema: core definitions and terminology. Internet
engineering task force. https://tools.ietf.org/html/draft-zyp-json-schema-04. Accessed 20 Apr
2015

Chapter 2

Complex Event Processing Framework

for Big Data Applications
Rentachintala Bhargavi

Abstract The fundamental requirement for modern IT systems is the ability to

detect and produce timely reaction to the occurrence of real-world situations in the
system environment. This applies to any of the Internet of Things (IoT) applications
where number of sensors and other smart devices are deployed. These sensors and
smart devices embedded in IoT networks continually produce huge amounts of
data. These data streams from heterogeneous sources arrive at high rates and need
to be processed in real time in order to detect more complex situations from the
low-level information embedded in the data. Complex event processing (CEP) has
emerged as an appropriate approach to tackle such scenarios. Complex event
processing is the technology used to process one or more streams of data/events
and identify patterns of interest from multiple streams of events to derive a
meaningful conclusion. This chapter proposes CEP-based solution to continuously
collect and analyze the data generated from multiple sources in real time. Two case
studies on intrusion detection in a heterogeneous sensor network and automated
healthcare monitoring of geriatric patient are also considered for experimenting and
validating the proposed solutions.
Keywords Complex event processing IoT Big data Data stream Event
Intrusion detection Geriatric health monitoring

2.1

Introduction

A wide range of IoT applications have been developed and deployed in recent
years. IoT has provided promising solution to several real-time applications by
leveraging the growing ubiquity of radio frequency identification (RFID) and
wireless mobile, sensors, and other smart devices [7]. Big data is a term
encompassing the use of techniques to capture, process, analyze, and visualize
potentially large datasets in a reasonable time frame not accessible to standard IT
technologies [11]. The platforms, tools, and software used for this purpose are
R. Bhargavi (*)
School of Computing Sciences and Engineering, VIT University, Chennai, India
e-mail: bhargavi.r@vit.ac.in
Springer International Publishing Switzerland 2016
Z. Mahmood (ed.), Data Science and Big Data Computing,
DOI 10.1007/978-3-319-31861-5_2

41

42

R. Bhargavi

collectively called big data technologies. These technologies refer to and present
the ability to crunch vast collections of information, analyze it instantly, and draw
conclusions. Big data technologies deal with petabytes of records, files, and transactional data either arriving as streams or in batches. The RFIDs, sensors, smart
devices, etc., embedded in IoT networks produce huge amounts of data/events
continuously [19]. This data has the characteristics like volume, variety, velocity,
variability, veracity, and complexity. Some of the characteristics of the data/event
streams and their implications are as follows:
Data streams are continuous and sequential and are ordered by a time stamp or
any other attribute value of the data item. Therefore, data items which belong to
the same stream are processed in the order they arrive.
Data streams are generated by external sources and are sent to a processing
system. Hence, the data stream processing system does not have any direct
control over the data sources.
The input characteristics and the rate of a data stream are unpredictable. The
input rate can be very irregular and, at times, bursty in nature. Also, the nature of
the input does not allow one to make multiple passes over the data while
processing.
The amount of data is very large and unbounded. Therefore, processing requirements may not permit persistence followed by processing. However, data or
summary of data can be stored for archival or other purposes.
The data types of the data items can be structured, semi-structured, or
unstructured.
Data items in a data stream are not error free because the data sources are
external. Some data may be corrupted or discarded due to network problems.
These data streams need to be processed and analyzed to identify some interesting patterns and take actions if necessary. Processing these continuous data/event
streams in real time to identify the patterns among them is a herculean task and has
raised new research challenges over the last few years [3, 5]. A large number of
solutions exist in terms of systems, middleware, applications, techniques, and
models proposed by researchers to solve different challenges [13]. Data generated
from multiple sources have logical and spatiotemporal relations among them. There
is a need for data fusion as the data from a single source may not be enough for
taking accurate decision. Conventional process-oriented control flow software
architectures do not explicitly target the efficient processing of continuous event
streams. Complex event processing is the technology used to process and analyze
one or more streams of data/events and identify patterns of interest from multiple
streams of events to derive a meaningful conclusion and respond by taking appropriate action [2].
The remaining parts of this chapter are organized as follows. Section 2.2 presents CEP preliminaries and event modeling. Semantic intrusion detection using
complex event processing is elaborated in Sect. 2.3. Section 2.4 discusses the
CEP-enabled geriatric health monitoring, and Sect. 2.5 concludes the chapter.

2 Complex Event Processing Framework for Big Data Applications

2.2

43

Complex Event Processing

Complex event processing (CEP) is a relatively new concept, but it has received
wider acceptance due to its systematic and multilevel architecture-driven concept
approach [9]. An event is an object that is a record of an activity; it signifies the
activity. A key stroke or the output reading produced by a sensor, etc., are examples
of an event. CEP allows one to set a request for an analysis or some query and then
have it executed continuously over a period of time against one or many streams of
events in a highly efficient manner. CEP is all about the processing of events that
combines data from many sources to infer events or patterns that represent more
complicated circumstances. In contrast to Hadoops two-stage disk-based
MapReduce paradigm, CEPs push-based paradigm supports faster processing of
data streams.

2.2.1

CEP Architectural Layers

The architectural layers of a CEP system are shown in Fig. 2.1. A typical system
consists of four layers, as briefly described below.

Application layer

GUI

External systems

Alert system

Output Adapters
Event Analysis layer
Rules

Queries

Stream mining Algorithms

Historic
Data

Formatted events
Data Collection layer

Device adapter

HTTP adapter

App adapter

Observed events
Event source layer

Sensors

Fig. 2.1 Architectural layers of CEP

RSS feed

logs

Files

app

44

R. Bhargavi

Event Source Layer This layer consists of event sources. These sources can be
sensors, RFID readers, RSS feed, feed from network monitoring systems, web
logs, etc.
Data Collection Layer Data collection layer is responsible for collecting the data
coming from various sensors and filtering the data. Data collection could be
distributed or centralized based on the requirement of the application. Basic
capabilities in WSN involve mechanisms like routing, tunneling, data aggregation,
and clustering to collect information from nodes and forward them to a sink node.
Event Analysis Layer The event analysis layer is responsible for data mining on
streams of data. An important feature of this layer is pattern matching and correlation across multiple event streams. In this layer, few fixed continuous queries act
on the incoming data streams to identify patterns of interest. Comparison of realtime data with historical or static data is also often required for event analysis. In
many instances, historical or static data is also often required along with the realtime data for analysis. Thus, the event analysis layer should contain facilities for
connecting to data base.
Application Layer This layer consists of event listeners, i.e., modules/systems that
receive the processed events. Examples of event listeners are event storage systems,
mobile phones and pagers, or other systems that can take actions based on the
results of event processing (e.g., GUI receives the processed events and displays the
information).
There are two underlying stages or steps involved in CEP. The first step is to
detect meaningful events or pattern of events which signifies either threats or
opportunities from the streams of events. The second step is to send alerts to the
responsible person/entity for the identified threat or opportunity for quick response.
CEP solutions and concepts can broadly be classified into two categories:
Computation-oriented CEP
Detection-oriented CEP
In computation-oriented CEP solutions, online algorithms are executed whenever an event or data enters the system. Simple example is to calculate the moving
average temperature sensed by a temperature sensor. Detection-oriented CEP
concentrates on detecting combinations of events or event patterns. Simple example
is to look for a sequence of events.
Complex event processing system is made up of a number of modules like input
adapters, output adapters, and event processing modules such as event filtering
modules, in-memory caching, aggregation, database lookup module, database write
module, correlation, joins, event pattern matching, state machines, dynamic
queries, etc, as shown in Fig. 2.2. In order to support more flexibility and adaptability for different use cases, more number of I/O adapters must be supported by
the CEP. The main component of CEP is the continuous queries which monitor
streams of simple/raw events for so-called complex events, that is, events that
manifest themselves in certain temporal, spatial, or logical combinations. Querying

2 Complex Event Processing Framework for Big Data Applications

45

CEP Engine

Event
Sources

Event
consumers

DB

Flat
file

Delivery Mode

Logs
Streaming Input

Input
Adapters

Sensor
Network

Social Feed

Sequencing
Filtering
Aggregation
Pattern matching
Classification

Output
Adapters

Prediction Tools
Tracking Tools
Logic Development

Feedback Events

Fig. 2.2 Logical view of CEP

events over data streams is different from traditional querying with database, in the
way that traditional database querying is pull based, whereas continuous querying
of events is push based.
Many of the real-time distributed applications require continuous monitoring
and processing and analysis of information or data in a timely fashion as it flows
from periphery to the system. Intrusion detection in surveillance and healthcare
monitoring are a couple of such applications. Traditional pull-based approaches can
hardly address the requirements of timeliness, response generation, etc. Hence, this
chapter proposes CEP-based solution to continuously collect and analyze the data
generated from multiple sources in real time. Two case studies: Intrusion detection
in a heterogeneous sensor network and automated healthcare monitoring of geriatric patient are considered for experimenting and validating the proposed solutions.

2.2.2

Event Modeling

An event is anything that happens, or is contemplated as happening [10] in the

real world, and normally it is of interest to some groups of people. A key stroke, a
sensor outputs a reading, etc., are couple of examples of an event. Sometimes these
events, in turn, may produce secondary events internally. Real-world occurrences
can be defined as events that happen over space and time. Events are of two types:
basic or primitive events and complex events. Events have event attributes. An
event attribute is a property of the event. For example, the entry of an identified
person in a restricted area could be treated as an activity. Then the form of the event

46

R. Bhargavi

instance could be composed by unique id of the person, time, and location (geographical coordinates). An event is an object that represents, encodes, or records an
event, generally for the purpose of computer processing [10].
A basic event is atomic and indivisible and occurs at a point in time. Attributes of
a basic or primitive event are the parameters of the activity that caused the event.
An event instance is a record of an activity, which has three features such as:
Significance, which gives its semantic
Form, which gives activity information that will be processed by the computer,
e.g., unique id, time, etc.
Relation with other event instances
Computer systems process the events by representing them as event objects.
From a software application perspective, an event is something that needs to be
monitored and may trigger a specific action. Specifying an event is therefore
providing a description of the happening. A common model of an event is a tuple
represented as
E Eid; a; t
where
id is the unique ID of an event.
a {a1, a2, . . .. . ., am}, m > 0, is a set of attributes.
t is the time of occurrence of the event.
For example, an RFID event can be described with some set of dimensions
which includes source of event, location of event, time at which the event occurred,
and a possible set of operations for combining events. An RFID event is denoted as
E e (o, r, t) where o is the tag EPC, r is the reader ID, and t is the time stamp of the
event.
Complex events are composed of basic events. Complex events are defined by
connecting basic events using temporal, spatial, or logical relations. A common
model for a complex event is as follows:
E Eid; a; c; tb ; te , tb < te
where
c {e1, e2, . . ...,en}, n > 0 is the vector that contains basic events and complex
events that cause this event to happen.
tb, te are starting and ending times of the complex event.
Attributes of complex events are derived from the attributes of the constituent
primitive events. Event constructors and event operators are used to express the
relationship among events and correlate events to form complex events.
Any basic event or a complex event is specified by an event expression. An event
expression is a mapping from histories (domain) to histories (range) [6], i.e.:

2 Complex Event Processing Framework for Big Data Applications

Fig. 2.3 Finite automata
representation of composite
event E1 ^ E2

E1

47

E2

F
E2

E1

E : histories ! histories:
Since event expressions are equivalent to regular expressions, it is possible to
implement event expressions using finite automata. For example, composite/complex event E1 ^ E2 can be represented using the finite automata as shown in
Fig. 2.3.
Event expression is formed by combining events with the event constructors.
Many of the event processing engines support different types of logical and
temporal constructs.

2.3

Semantic Intrusion Detection Using CEP

There is an increasing demand for security solutions in the society. This results in a
growing need for surveillance activities in many environments. Recent events, like
terrorist attacks, have resulted in an increased demand for security in society. There
is a growing interest in surveillance applications, because of the availability of
cheap sensors and processors at a reasonable cost. Intelligent remote monitoring
systems allow users to survey sites from remote location. Sensor networks bridge
the gap between the physical world and the virtual world of processing and
communication. It is envisioned that sensor networks can reduce or eliminate the
need for human involvement in information gathering and processing in surveillance applications if the ability of the sensor networks can be suitably harnessed.
Distributed sensors and smart devices in the surveillance application produce huge
data [14]. Processing, analyzing, and detecting abnormal patterns from this data are
very complex in nature. There is a need for multi-sensor data fusion in sensor
network applications as the data from homogeneous sensors may not be enough for
taking accurate decision [17]. Also the data generated from multiple sensors have
logical and spatiotemporal relations among them. Special processing and mining
algorithms are proposed in the literature for distributed and in-network processing.

48

R. Bhargavi

Existing solutions for surveillance and intrusion detection are based on pull-based
architecture where the data from sensors are stored, and later, the user runs queries
to retrieve the data. There are also solutions based on machine learning techniques,
but these solutions again work on the static data. Traditional approaches which are
pull based can hardly address the requirements of timeliness, response generation,
etc. The proposed CEP-based semantic intrusion detection system (CSIDS) [1]
addresses the abovementioned problems.
In the context of surveillance, the evaluation of available technologies shows
that the security bottleneck is not the hardware but rather the real-time analysis and
correlation of data provided by various sensors. The objective of proposed
CEP-based semantic intrusion detection system is early detection and prevention
of security compromises/risks by identifying abnormal situations and quickly
responding with appropriate action. To achieve this goal, sensors (such as cameras,
RFID readers, etc.) are installed in the places to be monitored, and a control center
receives the information transmitted by sensors. This information is processed and
analyzed by an event processing engine to detect the intrusive patterns. There are a
few factors that make this task rather difficult. First, the raw data to deal with is very
huge, and second, the information sources are heterogeneous. Therefore, there is a
need to fuse or aggregate the data coming from such sensors to construct a global
view of the situation and even more to go one step beyond to the intent assessment.
The proposed CSIDS allows fusion of information generated by heterogeneous
sensors that support the goal of providing a global situational view for intrusion
detection. The goal of this fusion is to transform lower-level data into higher-level
quality information and to improve the certainty of situation recognition, i.e.,
reduce the false alarm rate. The fusion is realized by taking into consideration the
semantics of the information. This means that there should be a representation or
model of the situation that needs to be detected (such as an intrusion in a restricted
area). In this context, a situation is viewed as a combination of several activity
elements (or smaller situations), each of them appearing at some place and time. As
a consequence, the finer the granularity of these elements, the more difficult the
situation modeling task will become. This task is further complicated if any of the
elements appear at different time instants and do not relate to the same situation.
Furthermore, the set of heterogeneous sensors capturing these sub-situations have
an asynchronous behavior, meaning each of the sensors computes the situation
independently, unconnected of each other.
CSIDS supports online detection of event patterns which represent anomalies.
The association of an event pattern, a constraint, and an action is referred to as a
rule. The approach consists of time, content, and context-based selection of a subset
of events described by their pattern. Finally, after the event pattern is matched, one
or more appropriate actions are executed such as creation and sending of a new
event or sending an alert message to the concern personnel, etc.
The architecture of the complex event processing system for semantic intrusion
detection system is shown in Fig. 2.4. CSIDS has multiple event receivers. Each
event receiver receives the data/events coming from a different data/event source.
The event receiver on receiving the data from the source converts them into event

2 Complex Event Processing Framework for Big Data Applications

Sensor Node 1

Sensor Node N

Sensor Node 2

Event
Receiver

49

Event
Receiver

Event
Receiver

Queue

Event Processing Engine

Notifier

Notifier

Notifier

Listener

Listener

Listener

Fig. 2.4 CEP architecture for semantic intrusion detection system

streams. Data generated by different sources follows a different format; hence,

event receivers also convert the data from different sources to the specific format
suitable for processing further by the event processing engine. The events generated
by the event receivers are inserted in to a FIFO or a queue. Events are organized in
the queue in the order of their detection time. To avoid the out of order arrival of the
events which is caused by the network delays, the events are stored internally in the
queue for some time T. Hence, the events are dequeued after a time T for
processing. Any event getting generated at time t will be processed after t+T time
by the CEP engine. Any event arriving at the queue with a time stamp smaller than
the time stamp of the already dequeued events is ignored. This leads to missing of
events. Missing events due to network delays can be avoided by having larger value
of T, i.e., by storing the events internally for longer duration. But this will delay the
event processing time. Hence, there is a trade-off between missing events due to
network delay and latency.
Event processing engine processes the event streams. The events generated from
heterogeneous sensors are collected and aggregated using logical and spatiotemporal relations to form complex events which model the intrusion patterns. All the
rules and patterns are to be registered initially. The listeners are intimated whenever

50

R. Bhargavi

the corresponding rule is hit or pattern is matched for which it is configured.

Modeling of the complex events using event expressions is explained later in this
section. Notifiers are used for intimating the listeners about the rule or pattern
occurrences. Listeners are the modules that take necessary action on notifications.
There are several supporting modules in semantic IDS to perform the activities
on the occurrence of certain events. These modules are Kalman tracking module,
person detection module, authentication module, etc. Person tracking is done using
Kalman filter. All these modules are listeners to the CEP engine.
Figure 2.5 shows how the physical events generated from different sensors can
be aggregated to generate a complex event which represents a pattern or a scenario
of interest. Sensed information from the sensors that interact with the physical
environment/world is collected by the event receivers. Various scenarios
representing simple and complex events have been modeled using event
expressions.
The following primitive events are considered for the present study:

Events generated during the interaction between the RFID readers and tags
PIR readings generated whenever a person crosses the sensor
GPS readings indicating the location of event occurrence
Time of event occurrence
Images captured by the camera

One of the complex event scenarios and its modeling using event expressions are
discussed below.

Complex Event Processing

EQL
Rule

Complex Event Processing Engine

Event Query Language

(EQL)

Pattern
Rule

Result
Event Processing
Agent (EPA)

EPA

Event
Listener 2

Event Stream/Event cloud

Result

Raw Events
Simple Event

Event Receiver

RFID

Event Receiver

PIR

GPS

Physical events (Generated from sensors)

Fig. 2.5 CSIDS

Event
Listener n

Complex Event

Event Receiver

Event
Listeners

Event Receiver

Timer

Event
Listener 1

Camera

2 Complex Event Processing Framework for Big Data Applications

51

Example: Unauthorized Entry

When a person who does not possess a valid RFID tag tries to enter a location just
behind an authorized person, it is called as tailgating. This scenario can be captured
as explained below.
CEP queries the wireless data for an event where PIR is present but RFID is zero.
Once such an event is identified, CEP gets the image from the database
corresponding to that events time stamp and gives the image to Haar face detection
module. The Haar face detection algorithm counts the number of persons present
and gives the count back to CEP. This scenario can be modeled as follows:
E1: Office entry
E1 (s1, o1, t1) type (s1) RFID event
E2: Office entry
E2 (s2, o1, t2) type (s2) PIR event
E3: Multiple people sensed by image sensor
The complex pattern can now be formulated as
everyE2 : E1 ^ E3
Now, the rule can be developed to raise an alert on the above complex event using
IF THEN condition:
IF (true)
Get the image with the same time stamp and node ID from the data base.
Send the location of the image in the data base to the image processing module for
checking number of people.
Receive the response from the image processing module.
Generate alert if needed.
End

2.3.1

Implementation and Validation

The proposed CSIDS is validated using a sensor network. A wireless sensor node is
devolved which consists of a passive infrared sensor, radio frequency identifier
reader, GPS receiver, and timer. Ten such sensor nodes are used for deploying the
wireless sensor network. Some of the nodes are connected with cameras. The
camera is triggered by the PIR sensor and is connected to a personal computer.
The images captured are stored in PC and transmitted to the server via wired
network. A unique ID is assigned to each sensor node. Data from the RFID and
PIR sensors are sent to the sink using wireless communication (using ZigBee
module). Data from the cameras is sent to the server using Ethernet. This is done

52

R. Bhargavi

to overcome the bandwidth, memory, processing, and power limitations of sensor

node. The complete CSIDS is developed using Java. Esper event processing engine
is used to implement and execute the rules. Patterns representing intrusion are
modeled as complex events which in turn are aggregated from base events and other
complex events using logical and spatiotemporal relations. Rules have been developed to identify the unauthorized entry of the people in to the security zones,
tailgating issue, and few other scenarios. It is observed that the proposed CSIDS
identifies the intrusion patterns efficiently in near real time. It is also observed that
the proposed system out performs the pull-based solutions in terms of detection
accuracy and detection time. The quality of results depends on the quality of sensors
and processes of their data. For example, if the sensors take more time to sense the
data, process, and communicate, then this introduces a delay between the occurrence of real situation and its recognition by the application which can become
considerable in sensitive situations.

2.4

CEP-Enabled Geriatric Health Monitoring

Healthcare monitoring is another application which can best exploit the advantages
of CEP. This section explains the proposed CEP-based geriatric health monitoring
system (CGHMS). Advancements in wireless body area networks have led
researchers to exploit the usage of these technologies in healthcare applications
[4, 15, 18]. Availability of wearable and cost-effective physiological and motion
sensors allow automated health monitoring of elderly people. There are a number of
healthcare applications using sensor networks. Within the hospital, there are three
applications: (1) to track people and objects around the hospital [16], (2) to safeguard use of equipment, and (3) to assist medical personnel with their jobs. Another
important application is elder care in home. Global increase in the ratio of elderly
people [8] in the world requires alternatives to the traditional homecare technologies that are available today. This in turn indicates that there will be challenges
related to giving proper care to the elderly people since care giving requires enough
people and resources. Hence, there is a need for alternatives that automate the home
care application domain.
The important requirement of healthcare monitoring is automatic anomaly
detection of vital parameters. Another important requirement is that when something dangerous or critical occurs, it is important that the system detects this
immediately and that the delay is minimal. There are several existing solutions
for monitoring the activity of a person or abnormality of a particular health
parameter. But there is no solution for identifying the abnormal situations like
fall of a person by combining the vital parameters, activities of the person, and the
context information. CEP-enabled geriatric healthcare monitoring system
(CGHMS) collects data generated from physiological and environmental sensors
and detects the abnormalities in vital parameters and fall of a person.

2 Complex Event Processing Framework for Big Data Applications

53

The main objective of the CEP-enabled geriatric healthcare monitoring is to

collect health parameters from the patient, detect the presence of an abnormality in
vital parameters, and provide feedback to the rules in decision-making about the
health condition and falls of a geriatric patient. In the geriatric healthcare monitoring domain, complex event processing involves analyzing raw sensor data and
recognizing next level of events like low BP, high BP, and high temperature to
complex events like person has fallen, etc. The data/event streams are filtered
correlated and aggregated to identify any abnormal patterns. Rules/patterns of
interest corresponding to various scenarios are stored in the knowledge base.
Rules are executed on the incoming data to detect the anomalies. Every rule has
associated action to be executed like sending SMS to caregiver or doctor or patient.
Figure 2.6 shows the architecture of the CEP-enabled geriatric health monitoring
system. The physical and semantic data flows are also shown. The proposed
CGHMS uses wearable BP sensor, pulse oximeter, and BioHarness 3 device, a
Zephyr product which contains both biosensors and triaxial accelerometer for
measuring the vital parameters like respiration/breathing rate, heart rate, ECG,
and movement along X-, Y-, and Z-axis, respectively. CGHMS also uses RFID
reader which can be connected and used with IPAQ PDA device. RFID tags and
environmental sensors are placed throughout the area to be monitored to get the
context information. Data generated by these sensors are collected and sent to a
central server where CEP engine. The knowledge base consists of CEP rules which
model the various abnormal health condition events. The CEP engine executes all
the rules whenever it receives the sensor data/facts to detect abnormality situations.

Physical Data Flow

Semantic
Data Flow

Basic
Events

Raw data

Physiological Sensors

Complex
Events

Alert
Generation

Working
Memory

CEP Engine

RFID tags RFID Reader

RFID
Middle ware

SVM
Accelerometer

Fig. 2.6 CGHMS architecture

Rule Base

CEP

Alert
Generation

54

R. Bhargavi

Every rule has associated action to be executed like sending SMS to the caregiver or
doctor or patient.

2.4.1

Implementation and Validation

The complete CGHMS is implemented in Java. Java-based Drools Expert engine is

used for rule development. Drools APIs are extended to support CGHMS. To
reduce the false alarms in fall detection, the proposed CGHMS uses vital parameter
data, activity of the person, RFID information which gives the contextual information, and camera. For activity classification using triaxial accelerometer and gyroscope, SVM classifier is used [12]. The proposed CGHMS provides continuous
accessibility to vital parameters and other relevant data and analysis of the data. The
CGHMS improves the quality of life of the elderly people by monitoring the wellbeing and alerting in case of emergency. The proposed CGHMS is tested and
validated in real-time environment using Zephyr BioHarness device; pulse oximeter; BP sensor, for measuring the vital parameters; RFID tags and reader to get the
context information; activity of a person using accelerometer and gyroscope; and
camera. It is observed that CGHMS detects the abnormalities in the vital parameters
and fall of a person more precisely compared to using the individual sensors. As
discussed earlier, the performance of the proposed CGHMS depends on the quality
of the sensors used. Also, the wearable sensor device must be worn by the person
correctly in the right position; otherwise, the captured data by the sensors may be
incorrect or sometimes the parameters may not be captured at all. Other problems
which affect the performance of the system include operating condition of the
device, battery charging, availability of the communication links, etc.

2.5

Conclusion

To conclude, this chapter discusses CEP framework for big data applications. Two
real-time case studies have also been discussed thoroughly. Early identification of
significant complex events provides situational awareness and better decisionmaking. Complex event processing enables sense-and-respond behavior, in which
incoming events or information is used to assess the current situation and generate a
response in a timely fashion. This chapter has proposed CEP-based solutions to
Internet of Things where early identification of significant complex events provides
situational awareness and better decision-making. Two real-time applications,
semantic intrusion detection and healthcare monitoring, are implemented and
validated to demonstrate the power of CEP. Patterns representing intrusion and
abnormal situations are modeled as complex events which are aggregated from base
events and other complex events using logical and spatiotemporal relations. The
CEP-based solutions provide the capabilities of heterogeneous information fusion

2 Complex Event Processing Framework for Big Data Applications

55

coming from several kinds of sensors to get a global situation view and take
necessary action in near real time. The proposed semantic intrusion detection
system is found to outperform the existing solutions in terms of detection latency
(i.e., time delay between the event occurrence and event detection) and detection
accuracy, i.e., event identification with less false positives and false negatives. The
proposed CEP-based solutions require domain knowledge of the application. All
the situations/scenarios of interest should be well understood and modeled as rules
using the events and available constructors. Failing to which, the pattern
representing the complex event (event of interest) may not be captured as expected.
Missing events is another common problem in case of wireless sensors. Missing
basic events will lead to the failure of getting the high-level complex situations
captured. The quality of results depends on the quality of sensors. For example, if
the sensors take more time to sense the data, process, and communicate, then this
introduces a delay between the occurrence of real situation and its recognition by
the application which can become considerable in sensitive situations.
In any of the complex event processing-based applications, the system is initially
configured with all the rules (defined by the domain experts) which define the
complex events or patterns to be identified and alerted. But there are several
applications in which new rules need to be added or deleted at run time. Hence,
the proposed solutions can be extended to support dynamic rule addition and rule
deletion. CEP is a rule-based technology for detecting known patterns of events and
reacting to the identified situations in real time. Incremental or dynamic learning is
used to learn the dynamic environments effectively. Hence, the proposed solutions
can be extended to incorporate dynamic learning, and the newly learned knowledge/concepts can be given as feed back to the CEP engine to add new rules and
identify new patterns or situations of interest. This makes the complete system more
adaptive and intelligent.

References
1. Bhargavi R, Vaidehi V (2013) Semantic intrusion detection with multisensor data fusion using
complex event processing. Sadhana Acad Proc Eng Sci 38(2):169185, ISSN: 02562499
2. Bastian Hobach and Bernhard Seeger (2013) Anomaly management using complex event
processing: extending data base technology paper. In: Proceedings of the 16th international
conference on extending database technology (EDBT 13). ACM, New York, pp 149154
3. Cantoni V, Lombardi L, Lombardi P (2006) Challenges for data mining in distributed sensor
networks. ICPR 1:10001007
4. Poon CYC, Liu Q, Gao H, Lin W-H, Zhang Y-T (2011) Wearable intelligent systems for
E-health. J Comput Sci and Eng 5(3):246256
5. Elnahrawy E (2003) Research directions in sensor data streams: solutions and challenges.
DCIS, Technical Report DCIS-TR-527, Rutgers University
6. Gehani NH, Jagadish HV, Shmueli O (1992) Composite event specification in active databases: model and implementation. In: VLDB 92: Proceedings of the 18th international
conference on very large data bases. Morgan Kaufmann Publishers Inc., San Francisco, pp
327338

56

R. Bhargavi

7. Jin J, Gubbi J, Marusic S, Palaniswami M (2014) An information framework for creating a

smart city through internet of things. Internet Things J IEEE 1(2):112121. doi:10.1109/JIOT.
2013.2296516
8. Kinsella K, He W (2009) An aging world: 2008. International population reports,
U.S. Department of Health and Human Services
9. Luckham DC (2010) The power of events: an introduction to complex event processing in
distributed enterprise systems. Addison Wesley Longman Publishing Co., Inc., Boston
10. Luckham DC, Schulte R (2008) Event processing glossary version 1.1.. Event processing
technical society. URL: http://www.ep-ts.com/component/option.com_docman/task.doc_
download/gid.66/Itemid.84/
11. NESSI White Paper, December 2012.
12. Palaniappan A, Bhargavi R, Vaidehi V (2012) Abnormal human activity recognition using
SVM based approach. In: Proceedings of IEEE international conference on recent trends in
information technology (ICRTIT 2012), Chennai, India, pp 97102, April 1921
13. Perera C, Zaslavsky A, Christen P, Georgakopoulos D (2014) Context aware computing for the
internet of things: a survey. Commun Surv Tutorials IEEE 16(1):414454
14. Tian He, Krishnamurthy S, Liqian Luo, Ting Yan, Lin Gu, Stoleru R, Gang Zhou, Qing Cao,
Vicaire P, John AS, Tarek FA, Jonathan Hui, Krogh B (2006) VigilNet: an integrated sensor
network system for energy-efficient surveillance. ACM Trans Sen Netw 2(1):1381
15. Vaidehi V, Bhargavi R, Ganapathy K, Sweetlin Hemalatha C (2012) Multi-sensor based
in-home health monitoring using complex event processing. In: Proceedings of IEEE international conference on recent trends in information technology (ICRTIT 2012), Chennai, India,
pp 570575, April 1921
16. Yao W, Chu C-H, Zang Li Yao W, Chu C, Li Z (2011) Leveraging complex event processing
for smart hospitals using RFID. J Netw and Comput Appl 34(3):799810
17. White Jr FE (1987) Data fusion lexicon. Data fusion subpanel of the joint directors of
laboratories, Technical Panel for C3, Naval Ocean Systems Centre, San Diego
18. Wood A, Virone G, Doan T, Cao Q, Selavo L, Wu Y, Fang L, He Z, Lin S, Stankovic S (2006)
ALARM-NET: wireless sensor networks for assisted-living and residential monitoring. Technical report, Department of Computer Science, University of Virginia, Wireless Sensor
Network Research Group
19. Zaslavsky A, Perera C, Georgakopoulos D (2012) Sensing as a service and big data. In:
International conference on advances in cloud computing (ACC-2012), Bangalore, India,
July 2012, pp 2129

Chapter 3

Agglomerative Approaches for Partitioning

of Networks in Big Data Scenarios
Anupam Biswas, Gourav Arora, Gaurav Tiwari, Srijan Khare,
Vyankatesh Agrawal, and Bhaskar Biswas

Abstract Big Data systems are often confronted with storage and processingrelated issues. Nowadays, data in various domains is growing so enormously and
so quickly that storage and processing are becoming the two key concerns in such
large systems of data. In addition to the size, complex relationship within the data is
making the system highly sophisticated. Such complex relationships are often
represented as network of data objects. Parallel processing, external memory
algorithms, and data partitioning are at the forefront of techniques to deal with
the Big Data issues. This chapter discusses these techniques in relation to storage
and processing of Big Data. The Big Data partitioning techniques, such as agglomerative approaches in particular, have been studied and reported. Network data
partitioning or clustering is common to most of the network-related applications
where the objective is to group similar objects based on the connectivity among
them. Application areas include social network analysis, World Wide Web, image
processing, biological networks, supply chain networks, and many others. In this
chapter, we discuss the relevant agglomerative approaches. Relative advantages
with respect to Big Data scenarios are also presented. The discussion also covers the
impact on Big Data scenarios with respect to strategic changes in the presented
agglomerative approaches. Tuning of various parameters of agglomerative
approaches is also addressed in this chapter.
Keywords Big Data Agglomerative approaches Clustering Partitioning
MapReduce Network partitioning

A. Biswas (*) G. Arora G. Tiwari S. Khare V. Agrawal B. Biswas

Department of Computer Science and Engineering, Indian Institute of Technology (BHU),
Varanasi, India
e-mail: anupam.rs.cse13@iitbhu.ac.in; gourav.arora.cse11@iitbhu.ac.in; gaurav.tiwari.
cse11@iitbhu.ac.in; srijan.khare.cse11@iitbhu.ac.in; v.agrawal.cse11@iitbhu.ac.in; bhaskar.
cse@iitbhu.ac.in
Springer International Publishing Switzerland 2016
Z. Mahmood (ed.), Data Science and Big Data Computing,
DOI 10.1007/978-3-319-31861-5_3

57

58

3.1

A. Biswas et al.

Introduction

Big Data maintenance and processing are the two major issues in modern-day data
mining. Complex relationships present within the data have resulted in data systems
becoming even more complicated. Often such relationships are represented as
networks in different domains [14]. In the network data representation, various
objects within the data are represented with nodes and relationships among those
objects that are represented with connections. The network representation of complex relationships present within the data is relatively easily interpretable. However, the main issues relating to memory and processing are retained due to huge
size of these networks. In this context, various parallel processing techniques
become highly useful for efficient processing. GPU techniques such as Pregel [5],
Giraph [6] and Seraph [7] are very popular for parallel processing of network data.
These techniques are based on the computation model called the bulk-synchronous
parallel (BSP) model [8]. On the other hand, to deal with memory-related issues
such as maintenance of larger network data than the memory, external memory
algorithms are often frequently utilized.
Both parallel processing and external memory techniques require partitioning of
networks [59]. In a parallel processing scheme, smaller parts of a network are
processed simultaneously in multiple processing units. Results of the multiple
processing units are accumulated to generate overall output. On the contrary,
external memory schemes partition the network to fit into the memory. Thus,
efficient partitioning of networks is very important from both perspectives. Meaningful partitions of networks can reduce computation cost. Moreover, such partitions can be easily handled with low-memory requirement. The objective of
meaningful partitioning is to divide the network into subnetworks such that nodes
within the subnetwork are close or similar. It means that nodes belonging to subnetworks have more connectivity than the connectivity among subnetworks. Such
subnetworks are often referred to as clusters or communities.
Numerous network data partitioning or clustering approaches have been developed, which can be broadly categorized as node centric [3, 10], group centric [11
14], network centric [15], and hierarchical [11, 1620]. Node-centric approaches
specify certain properties that needed to be satisfied by all the nodes of any cluster.
Group-centric approaches relax such restrictions from all nodes of clusters to some
members of the cluster. Instead of bounding individual nodes, group-centric
approaches consider properties for groups during partitioning of networks.
Network-centric approaches consider properties that are defined by covering the
entire network. Hierarchical approaches can be further subcategorized as divisive
and agglomerative. Divisive approaches follow top-down method, i.e., these
approaches keep partitioning the network until it maximizes prespecified criterion.
In contrast, agglomerative approaches follow natural way of group formation.
Starting from nodes, gradually groups grow by merging similar groups. In order
to ensure a meaningful clustering, several metrics [2123] are used to evaluate
clustering algorithms. In this context, agglomerative approaches are found better in

3 Agglomerative Approaches for Partitioning of Networks in Big Data Scenarios

59

generating accurate clustering than other approaches [11, 1620]. In this chapter,
we have presented agglomerative approaches in detail with their advantages and
limitations in relation to certain Big Data scenarios.
Remainder of the chapter is organized as follows. Section 3.2 explains various
Big Data scenarios and their inherent issues. Section 3.3 illustrates parallel
processing techniques of Big Data for the resolution of related issues. Section 3.4
explains various external memory operations to tackle memory-related issues.
Section 3.5 discusses how agglomeration can be helpful for Big Data scenarios in
parallel processing and external memory operation techniques. Section 3.6
describes various agglomerative approaches with a generic model. Section 3.7
discusses relative impact in handling large graphs with various strategic changes
in the agglomerative approaches. Section 3.8 lists the specific parameters of various
agglomerative algorithms and discusses beneficiary treatments of those parameter
in Big Data scenarios. In Sect. 3.9, we discuss various advantages, disadvantages,
and issues related to incorporation of agglomerative approaches in Big Data
scenarios. Finally, the conclusion is presented in Sect. 3.10.

3.2

Big Data Scenarios and Issues

Big Data involves three Vs (volume, variety, and velocity) [24] in its definition,
which resembles large, heterogeneous, and rapidly and continuously growing data
sets. In this section, we study applications of various domains that deal with large
data and associated issues. For example, social networks such as Facebook and
Twitter are growing enormously and are therefore generation huge volumes of user
data. Graph-based analysis of links and text analysis-related application to such
social networks are very common. Facebook has more than 802 million active users
and has to process over 1.26 billion queries every day [25]. Twitter also has almost
500 million active users who generate more than 58 million tweets and 2.1 billion
search queries every day [26]. Another instance is the searching of the World Wide
Web. Popular search engine Google processes 3.5 billion searches per day
[27]. Dealing with such large data is often confronted with the major issues mainly
related to handling of processing and storage of data.
Telecommunication traffic management system [28] receives billions of common channel signaling (CCS) messages and calls from multiple subnetworks that
require gigabytes of memory. Analysis of characteristics of both the telephone call
arrival process and the signaling message arrival process is key requisite for the
traffic system. Major concern is to analyze call arrival, call holding, message
arrival, message routing, etc. Such analysis requires processing of gigabytes of
data. The size and complexity of traffic data require sophisticated methodology for
data management, data translation and manipulation, and data analysis. Todays
Internet traffic management system [29] requires even more sophisticated and
efficient processing of such enormous amounts of data collected from the network.

60

A. Biswas et al.

In the image processing domain, content-based image retrieval (CBIR) systems

[30], duplicate image detection (DID) systems [31], and face recognition systems
[32] involve massive image data processing. Billions of images are available on
Facebook and Picasa. These image processing systems require processing each of
the images, which implies necessity of memory as well as efficient processing.
From the perspective of the data analysis specialist, the Big Data scenario infers
various specialized difficulties. The most evident of them is the computational
issue of needing to manage large data in memory (memory management) and to do
as such in a sensible amount of time (efficient processing). Efficient processing of
algorithms depends on super-linearity on size of the data and dimensions of data.
Increasing any one of these two variables requires memory optimization, parallelism, and new creative approaches. Other fundamental issues with aforementioned scenarios are to deal with dimensionality, variety, and heterogeneity
of data.
In practical applications of real-world problems, dealing with large data with
billions of instances and high dimensions focuses mainly on efficient processing
and creative memory management. In recent years, researchers have developed
brilliant models that optimize the processing part by executing in parallel. Prime
focus of this chapter is on large graph processing or processing of large-scale
linked data. MapReduce [33] and bulk-synchronous parallel (BSP) [8] are the
widely accepted models for processing of large graphs in parallel with multiple
machines. Similarly, for the storage issues, numerous solutions are proposed
which have proven to be very efficient in handling large data. External memory
operations are very popular for efficient memory management. In coming sections,
various BSP models for parallel processing and external memory operations are
elaborated.

3.3

Parallel Processing

This section explains the bulk-synchronous parallel (BSP) model [8] for the
processing of large graphs in parallel. In addition to this, techniques such as Pregel
[5] and Seraph [7] that incorporate BSP model for graph processing are also
illustrated.

3.3.1

Bulk-Synchronous Parallel (BSP)

Main objective of both MapReduce and BSP models is the same, i.e., processing of
graphs in parallel. The difference between the two models lies in the flow of data
while processing takes place. A typical BSP model adopted from [7] is presented in
Fig. 3.1. BSP partitions the input data once and for all during initialization. Nodes
process the part of data that are assigned to them independently by consuming

3 Agglomerative Approaches for Partitioning of Networks in Big Data Scenarios

61

Fig. 3.1 BSP model

messages from a private box. Data does not flow between the nodes; rather
instruction messages are exchanged between nodes that are smaller and efficiently
transmitted. This is an incarnation of message passing interface (MPI) pattern.
There is no restriction on the outcome of the processing phase. It can be the
modification of the local partitions data or output of a new data set or in particular
production of new messages, which will be shipped to the appropriate nodes and
consumed on the next iteration and so on.
Processing moves ahead with the completion of smaller units called supersteps.
It indicates the completion of processing in all nodes for the current iteration. Thus,
after each superstep, all nodes join a common synchronization barrier. The role of
such barriers is to decouple the production and consumption of messages. All nodes
pause their activity, while the BSP plumbing enqueues messages into each nodes
private mailbox and signals ready for consumption when processing resumes at the
next superstep. In this way, a message produced by node A for node B will not
interrupt the latter during its processing activity. Resources are mostly dedicated to
computing. In case of nonuniform processing of nodes, faster nodes will sit idle
while waiting for the other ones to complete a superstep. The superstep that
produces no more messages that will be shipped to any other nodes is the terminating step of the algorithm.

62

3.3.2

A. Biswas et al.

Overview of Pregel

The idea behind Pregel [5] is the exploring the graph along its edges. Starting from a
fixed set of vertices, one can hop from vertex to vertex and propagate the
execution of the algorithm, so to speak, across the set of vertices. This reminds
about how BSP model had executed an algorithm iteratively by passing messages.
In BSP, messages are exchanged between computation nodes, and in Pregel,
messages are exchanged between vertices of the graph. After knowing the
partitioning function of the graph, it is easy to know what node hosts each vertex
so that each node can dispatch incoming messages at the finest level. Thus, at a
logical level, vertices send messages to each other; at the physical level, the
underlying BSP foundation takes care of grouping those messages together and
delivering them to the appropriate nodes of the cluster. After each superstep, the
processing covers a bigger portion of the graph. Of course, messages are not
necessarily sent along edges of the graph, but it is the most common scenario.
The default partitioning function in Pregel does not take into account the shape of
the graph. It does not try to optimize the partitioning to minimize node-to-node
communication by keeping strongly connected vertices together.

3.3.3

Overview of Seraph

Seraph [7] is an efficient, low-cost system for concurrent graph processing. Existing
systems such as Pregel can process single graph job efficiently, but they incur high
cost while dealing with multiple concurrent jobs. These systems limit the sharing of
graphs in memory for multiple jobs. In other words, these systems strongly associate graph structure with job-specific vertex values. Therefore, each individual job
needs to maintain a separate graph data in memory that imply inefficient use of
memory. Seraph solves the above issues as follows. Concurrent jobs can share
graph structure, which reduces memory requirement even if the fact that graph
structure and job-specific data are strongly associated. Small amount of data is
required in Seraph; with that, the entire computation states can be recovered.
Seraph enables multiple jobs to share the same graph structure data in memory
[7] as illustrated in Fig. 3.2. To maximize job parallelism, Seraph incorporates three
new features as follows:
In order to isolate mutations in graph from concurrent jobs, Seraph adopts
copy-on-write semantic. When the graph structure needs to be modified by a
job, Seraph copies the local region that corresponds to that job, and mutations are
applied without affecting remaining jobs.
A lazy snapshot protocol is used. It maintains a continuous graph updates.
Further, consistent snapshot is generated for each new submitted job.
Delta graph check pointing and state regeneration are used to implement fault
tolerance. This is because it deals with both, job-level and worker-level failures.

3 Agglomerative Approaches for Partitioning of Networks in Big Data Scenarios

63

Fig. 3.2 The execution of concurrent jobs in Seraph

A job scheduler controls the execution progress and message flow of each
concurrent job. Therefore, avoiding aggressive resource competition requires a
job scheduler to be a part of Seraph.

3.4

External Memory Operations

In the case of large graph networks that cannot be stored completely in the main
memory, external disk is used. However, this incurs I/O interaction overheads.
External memory breadth-first search (EM BFS) algorithm is one of the competent
algorithms in efficiently handling such overheads. This section briefly discusses
different variants of the EM BFS algorithm.

3.4.1

External Memory BFS

EM BFS consists mainly of two phases: partitioning of graphs into disjoint subgraphs and breadth-first search (BFS) generation phase.
Partitioning phase: Partitioning of graph is done generally in parallel to speed up
the process. Adjacency list representation of graph is restructured in this phase.
With randomly selected nodes called master nodes, multiple connected subgraphs are prepared and stored in external file and spatial indexing is used to

64

A. Biswas et al.

access them. Boundaries of these subgraphs are actually the master nodes.
Therefore, the external files can be used to define the level of BFS.
BFS generation phase: External files are sorted in accordance with their identifiers. External file at level i contains adjacency list of subgraph of level i. Thus,
starting with level 0 through level n will generate levels of BFS from 0 to n.
There are several variants of EM BFS that have been proposed. The MM BFS
[34] and MR BFS [35] are the most efficient variants of EM BFS. These have
successfully managed to reduce the I/O-bound complexity in the following ways:
MR BFS: Worst-case I/O-bound complexity is reduced to On sortn m
by I/O efficient comparing the two most recently completed BFS levels with
current BFS.
MM BFS: It introduced a preprocessing step where nodes are grouped and
accessed orderly. Whenever the first node of a group is visited, then the
remaining nodes of that group will be accessed soon. Thus, it has reduced I/Obound complexity.

3.5

Agglomeration in Big Data Scenarios

Discussions in Sects. 3.3 and 3.4 indicate that the handling of big data issues with
both the parallel processing and external memory operation requires partitioning of
graphs or networks. This section illustrates how the agglomerative approaches are
becoming beneficial for the partitioning of large networks. In agglomerative
approaches, partitioning is started from arbitrary nodes and gradually enlarged to
obtain final clusters present in the entire network. Clusters are identified locally
with agglomeration approaches. Most of these approaches do not require access to
the entire network at the same time, which helps during processing large networks.
Large networks are processed in parallel by partitioning them into smaller subnetworks. Thus, clustering is very important for parallelization of processing.
However, most of the clustering algorithms other than agglomerative approaches
require processing of the entire network at the same time, which takes us on the
processing and memory-related issues. On the contrary, agglomerative approaches
partition the networks through local processing of parts of the networks.
Local nature of the agglomerative approaches gives a place for implication of
these approaches to execute and start clustering from multiple parts of the network
in parallel. For example, local nature of merging clusters based on incremental
modularity in fast unfolding algorithm [11] can be noticed easily. Processing of two
clusters for merging requires only information about those clusters. If it indicates
that the modularity increases with merging, then simply merge those clusters
without bothering about other parts of the network. Hence, merging operation in
fast unfolding algorithm can be treated as local operation and can be executed in
parallel. Similarly, visiting of nodes in random walk algorithm [20] during similarity matrix preparation requires processing only those nodes, which come under

3 Agglomerative Approaches for Partitioning of Networks in Big Data Scenarios

65

the path with prespecified length. Visiting of nodes in multiple random paths can be
performed in parallel. Thus, clustering process itself can be parallelized easily with
agglomerative approaches.
The local nature of agglomerative approaches can be parallelized if local
operations are defined explicitly. However, some of the algorithms are a little bit
complex, and the local nature of these algorithms is implicitly defined within the
process. Local operations cannot be extracted from the main process of the algorithm. These approaches limit the implication of the local operations in parallel. For
example, the evaluation of membership of nodes in LICOD algorithm [10] requires
the identification of the shortest path. The entire network has to be processed to find
the shortest path. Therefore, the algorithm cannot be parallelized. Similarly, leaderfollower algorithm [16] also requires the shortest path. Moreover, the merging of
clustering in LICOD is also not local, since it incorporates of Borda voting scheme.
Therefore, these algorithms cannot be parallelized. On the contrary, top leader
algorithm [18] can be parallelized though the local nature is implicit. Identification
of top k nodes requires access to the entire network. However, after identification of
top k nodes, the portion that finds associated members to these nodes can be easily
executed in parallel.
Partitioning of large networks and processing those parts in parallel are the
primary aspects to deal with most of the Big Data scenarios where linked information are presented in the form of a network. Therefore, efficient and proper
partitioning of network is prerequisite for processing the networks in parallel. If
partitioning scheme itself can be parallelized, then it avails additional advantage for
dealing with Big Data. Agglomerative approaches have privilege over other clustering or partitioning schemes, since these approaches can be parallelized. However, due to implicit definition of local operation or inseparable local operations,
parallelization of the algorithms cannot be performed for some cases. Nevertheless,
these approaches reduce significantly the processing time. Specially, the
approaches such as fast unfolding and SCAN [19] are very fast and overwhelmingly
accepted approaches for dealing with Big Data scenarios.

3.6

Agglomerative Approaches

This section elaborates on various agglomerative approaches with a generic model

for partitioning graphs. Agglomerative approaches follow bottom-up approach
where each observation starts in its own cluster and pairs of clusters are merged
as they move up in the hierarchy. Unlike divisive approaches, here clusters merge
instead of splitting. The generic agglomerative approach is briefly described, first,
as follows.

66

3.6.1

A. Biswas et al.

Generic Model

A generic model of agglomerative approach is shown in Fig. 3.3. Initially, each

node is considered as a cluster. At each level, these clusters are merged with other
clusters. Generally, merging of nodes is done in a greedy manner. Each level of
merging carries cost or benefit of merging. If merging is associated with any cost
function, then definitely that has to be minimized; otherwise, it has to be maximized. Generally, functions related to benefit of merging are considered to merge
clusters. The function associated with each level can be presented in the dendrogram. These functions are also used to determine when to stop merging clusters. For
example, in Fig. 3.3, the values of benefit of merging associated with each level are
fS1 , S2 , S3 , :::, Sn g, here n 4. Merging of clusters will stop at level 2 if S2 > S3 ,
i.e., the benefit of merging decreases with further merging of clusters. Hence, to
gain maximum benefit, algorithm will stop at level 2.
Apart from the cost and benefits, distance among clusters and number of clusters
are also used as stopping criteria. Each agglomeration occurs at a greater distance
between clusters than the previous agglomeration. One can decide to stop clustering
either when the clusters are too far apart to be merged (distance criterion) or when
there is a sufficiently small number of clusters (number criterion). Generally,
merging of clusters takes place in two ways depending on formation of clustering.
The model described in Fig. 3.3 is the global approach where clusters are formed on
processing of the entire graph. On the contrary, Fig. 3.4 presents generic model for
the local approach. Clusters are formed locally without processing the entire graph.
Clusters are formed sequentially one after another. However, each cluster is formed
by merging smaller clusters in successive levels until a maximum benefitted cluster
is obtained. Such merging is centric to a particular node that advocates benefit

Fig. 3.3 Generic model of agglomerative approach

3 Agglomerative Approaches for Partitioning of Networks in Big Data Scenarios

67

Fig. 3.4 Generic local model of agglomerative approach

associated with the cluster. Once the cluster associated with the advocating node
achieves maximum possible benefit, it starts new cluster associated with other
advocating node. In Fig. 3.4, S1i is the maximum benefit of the cluster 1. After
achieving the value S1i , processing for the cluster 1 ends and processing for cluster
2 begins. Most of the agglomerative algorithms follow local agglomeration
approach. In the next subsections, we illustrate strategies incorporated in different
agglomerative approaches in perspective of the generic model.

3.6.2

Fast Unfolding

Fast unfolding algorithm [11] uses the modularity function as benefit function to
obtain perfect clusters. The algorithm keeps on merging the clusters until maximum
modularity is reached. During intermediate states of the algorithm, the actual size of
graph reduces to the number of clusters formed at the current level, which are the
new nodes of the graph. These nodes are again sent back to the algorithm to reduce
it further to obtain maximum modularity. The algorithm has two phases. In the first
phase, a particular node is assigned to a neighboring node based on the change in
modularity, while in the second phase, the reduced graph is obtained from the first
phase. This process continues until there is no change in the graph obtained from
the first phase and the second phase.

68

A. Biswas et al.

Phase 1:
Initially each individual node is represented as a cluster.
Each node is processed with its neighbors for the value of Q that is the change
in modularity of the node in case neighbor is merged with node representing a
cluster.
The neighbor with maximum positive gain in modularity is chosen for merging
with the node to obtain clusters.
Intermediate clusters are obtained with the merging.
Phase 2:
A new graph is prepared with the clusters obtained from phase 1. Nodes of the
graph are the clusters. Edges of the new graph are prepared as follows. The edges
that lie within a cluster represented by self-loop. The weight of the edge between
two clusters is given by number of inter-cluster edges.
If the graph obtained from phase 2 is the same as the one used in phase 1, then
the algorithm stops. The nodes at this stage represent the final clusters. If two
graphs are not the same, the process continues.
The fast unfolding algorithm follows the generic agglomeration process
discussed in Fig. 3.3. Benefit function used in the algorithm is the modularity
function. Stopping criterion is the difference between two graphs that are processed
in two successive levels.

3.6.3

SCAN

SCAN [19] is another popular agglomerative clustering approach. In addition to

clusters, it also detects hubs and outliers. It follows local agglomerative approach as
presented in Fig. 3.4. At the initial stage, the core nodes are identified. All those
nodes that are structurally reachable from the core nodes are assigned to respective
core nodes that form a cluster. The process continues until all the nodes are either
classified or marked as nonmember. The nodes that are identified as nonmembers
are further classified as hubs and outliers. The algorithm uses the following special
terminologies.
Structural similarity: Represented by u; v where u and v are nodes and is
evaluated as number of common neighbors of u and v divided by geometric mean
of number of neighbors of u and v.
e-neighborhood: The number of neighbors of v whose structural similarity with v
is greater than e belongs to its e-neighborhood.
Core: A vertex v is defined as core if its e-neighborhood contains at least
vertices.
Direct structure reachability: The nodes that are neighbors of the core are
directly structure reachable from that core.

3 Agglomerative Approaches for Partitioning of Networks in Big Data Scenarios

69

Structure reachability: A node v is structure reachable from some node w, if there

exist a path such that all the intermediate nodes are cores.
Hub: A node is defined as a hub if some of its neighbors are in one cluster while
some in other.
Outliers: An isolated node is treated as an outlier if all its neighbors lie in one
cluster or does not belong to any as such.
Unlike the fast unfolding algorithm, SCAN does not have clear levels of
agglomeration. It merges nodes based on structural reachability and structural
similarity as follows.
Nodes are arbitrarily selected and find their neighbors to decide if the nodes
represent core or not.
If a node is not a core node, then it is marked as a nonmember. Otherwise, a new
cluster is formulated corresponding to the core node.
The neighbors of the core node that belong to its e-neighborhood are inserted
into a queue Q.
Nodes that are in Q are dequeued one by one until it becomes empty.
For each dequeued element from Q, direct structure reachable nodes are identified. If any of these nodes is a nonmember node, then assign that node to
current cluster and enqueue it into the queue.
After completing the processing of the entire graph, nonmembers that do not
belong to any cluster are marked as hubs or outliers.
One of the major advantages of SCAN algorithm is that it not only forms the
clusters but also identifies the nodes that are not formally in any cluster as hubs and
outliers. The algorithm is very fast. The time complexity of the algorithm is Om
for a graph with n vertices and m edges.

3.6.4

Leader-Follower

The leader-follower algorithm [16] is designed based on the idea of leaders and
their followers. Nodes are divided into either leader or follower by measuring
distance centrality. It measures the closeness of a node to all other nodes. Followers
are again subdivided as loyal or not loyal. A loyal follower must have its distance
centrality value at least equal to each of its neighbors.
Leaders are arranged in increasing order of their distance centrality. Initially, all
the leaders represent potential clusters. The representative cluster for a leader
expands as follows. The expansion process follows local agglomerative approach
as presented in Fig. 3.4. The neighbors that are loyal followers of the leader are
assigned to the representative cluster of that leader. Many of the leaders get no
followers and constitute cluster with single node. Such leaders are identified as fake
leaders. These fake leaders are removed by assigning them to other clusters. The

70

A. Biswas et al.

neighbors of the leader are considered for such assignment. The leader that has the
majority of its neighbors are loyal follower is selected as the leader for assigning
fake leaders. A special case occurs when the neighbors of a fake leader are
themselves leaders. This is handled by the same procedure of finding majority
leader and breaking ties randomly that results in random leader assignment. Major
steps of the algorithm are summarized as follows:
Initialization is done by calculating the distance centrality of each node. This
method utilizes Floyd-Warshall algorithm for computing the shortest path
among all pairs of nodes.
Identify leaders and followers by measuring distance centrality.
Each representative cluster for a leader is expanded by assigning loyal followers.
Single-node clusters, i.e., fake leaders are assigned to appropriate cluster as
explained in the above discussion.
Leader-follower algorithm can identify smaller clusters. The algorithm does not
need to specify any parameter as it naturally discovers the underlying structure.
Most importantly, it does not generate singleton clusters because all such clusters
are assigned to other clusters. The algorithm results unexpected clusters, as most of
the data sets do not follow the definition of cluster utilized in the algorithm. There
are numerous other leader-based algorithms in existence. These are covered in the
following sections in terms of their efficiency with respect to Big Data scenarios.

3.6.5

HC-PIN

HC-PIN [17] is another agglomerative algorithm that was designed to identify

clustering in protein-protein interaction networks. The algorithm follows local
approach as in Fig. 3.4 for expanding clusters. However, at the initial stage, it
processes the entire network for calculating the edge clustering value (ECV). In the
essence of agglomeration, the algorithm initially considers each individual node as
cluster. For each of the edges, ECV is calculated, and those are arranged in a list in
decreasing order. The ECV of an edge is the square of the product of a number of
common neighbors of the nodes divided by the product of a number of neighbors of
each node. Higher ECV for an edge indicates greater the tendency that both the
associated nodes of that edge to be included in the same cluster.
The edges are removed one by one from the ECV list. The weighted degree for
the clusters of each of the representative nodes is computed. Weighted degree is
defined as the sum of weights of all the incoming edges to a particular node. For a
cluster, it is the summation of weighted degrees for all the nodes belong that cluster.
A parameter is used for decision-making. The ratio of weighted in-degree to
out-degree for each of the clusters is compared with parameter. Another parameter
s, is used at the end to determine the minimum size of each cluster. Only clusters
that have minimum size as s are included in the result. The algorithm is summarized
as follows:

3 Agglomerative Approaches for Partitioning of Networks in Big Data Scenarios

71

All the vertices are initialized as a singleton cluster.

ECV value is calculated for each of the edges of the network. All the ECV values
along with associated edges are arranged in a list in decreasing order.
Edges are removed one by one from the list, and associated nodes of that edge
are compared. The comparison-based decision is made to merge both the nodes
either in the cluster or in the different clusters.
Weighted in-degree and out-degree of all clusters are computed. These weighted
values are compared with parameter in order to determine if it satisfies
necessary structural requirement of the cluster or not. If a cluster does not fulfill
this objective, then the nodes of that cluster are assigned to other cluster.
In addition, clusters are also refined with parameter s to ensure the sizes of
clusters that are identified.
HC-PIN is a successful algorithm in terms of accuracy of clusters it generates.

3.6.6

Other Approaches

Apart from the algorithms discussed in the above subsections, there are numerous
agglomerative approaches that have been developed which fall within either of the
two models. For instance, random walk algorithm [20] has been developed utilizing
very similar mechanism as SCAN and HC-PIN algorithms. The leader-based
algorithms such as top leader [18] and LICOD [10] have been developed along
the same lines as leader-follower algorithm. Recently, another agglomerative
algorithm, ENBC [36], has been proposed. The ENBC algorithm incorporates
both models discussed above. The algorithm is also very efficient in generating
accurate clusters. Details of the strategic changes of these algorithms and their
impact on Big Data scenarios are discussed in the next section.

3.7

Agglomerative Strategic Changes for Big Data

Scenarios

Strategies incorporated in handling Big Data problems have a vital role for making
processing efficient and effective. As part of large graph processing with parallel
processing and external memory schemes requires partitioning. Section 3.5 has
discussed several advantages of incorporating agglomerative clustering in
partitioning of large graphs. Prominent agglomerative approaches are studied in
Sect. 3.6 with generic models. This section is devoted to discussing the agglomerative approaches by accounting their strategic changes and relative impact in
handling large graphs with such strategic changes.
Numerous strategic changes in agglomerative approach have been proposed in
the quest for more effective and efficient clustering. Most promising strategic

72

A. Biswas et al.

change in agglomerative approach is the local processing of graphs. Several

algorithms have been proposed to make the local processing more efficient. One
common approach is to derive a mechanism which defines similarity among the
nodes of the entire graph and identifies clusters locally. Normally, clusters start
from any arbitrary node and gradually expand by utilizing similarities among
nodes. Most of such algorithms can be roughly viewed as two phases, i.e., similarity
matrix creation and cluster formation. For example, in random walk algorithm [20],
a similarity matrix is prepared that keeps track of frequencies of pairs of nodes that
appear in the random path. Initially, the algorithm traverses randomly in fixed
length path started from any random node. The strategy here is the nodes that are
in densely connected region will appear frequently in the random path. Thus, those
nodes will get higher similarity values. Similar kind of strategy is followed in
HC-PIN [17], where edge clustering value (ECV) is computed for all pairs of edges
and maintained in a list. Cluster formation part of such strategies is local, but the
similarity matrix preparation requires processing of the entire network. Therefore, it
requires special mechanism to generate similarities locally for large graphs.
An alternative of similarity computation is identification of leader nodes locally
among the representative groups of nodes within the graph. Leader-based algorithms are designed based on this principle. These approaches can also be viewed in
two phases, i.e., leader identification and cluster formation. In the LICOD algorithm
[10], leader nodes are identified utilizing the degree centrality. Since leader node
represents potential cluster, the group of identified leader nodes is optimized to get
minimized set of nodes. Followers of these leader nodes are identified and based on
their membership values assigned to the representative cluster of a leader node.
Leader-follower algorithm [16] divides followers further into two groups: loyal
followers and non-loyal followers. Distinction of follower results in more refined
clusters. In contrast to both LICOD and leader-follower algorithms, top leader
algorithm [18] identifies top k potential leaders instead of optimized set of leader
nodes. With this strategy, the optimization cost for obtaining leader nodes is
diminished, and as a result, top leader algorithm [18] requires comparatively less
time. On the other hand, SCAN algorithm [19] uses same mechanism as leaders
with new terminology called core nodes. It utilizes notion of structural reachability
instead of followers to define clusters represented for core nodes.
Both the similarity matrix-based approaches and leader-based approaches
require twofold processing of graphs. For large graphs, such processing costs
double. In that case, instead of two phases of clustering, merging both phases into
one will be more beneficial for large graphs. Fast unfolding algorithm [11] utilizes
such strategy with multilevel clustering. Clusters obtained in each level agglomeration can be treated as potential clusters with specific properties. The algorithm is
very efficient for large-scale graphs with billions of nodes and edges. On the other
hand, ENBC algorithm [36] utilizes single-level agglomeration as in the similarity
matrix-based and leader-based approaches. However, ENBC identifies clusters by
merging similarity computation phase and cluster formation phase into single
phase. It starts from any arbitrary node that represents potential cluster and that
expands based on structural similarity of neighbors. The process of cluster

3 Agglomerative Approaches for Partitioning of Networks in Big Data Scenarios

73

formation followed in ENBC is local and more efficient than both similarity matrixbased and leader-based approaches. Therefore, it is suitable for large graph
processing both in perspective of parallel processing and external memory
operations.

3.8

Parameter Tuning for Big Data Scenarios

Most of the agglomerative algorithms have parameters to control various aspects of

clusters such as size and quality. From the perspective of Big Data, these aspects are
very important. As mentioned above, graph clustering is a prominent intermediary
step for handling Big Data with both parallel processing and external memory
operations. Sizes of clusters have to be balanced in a sense that neither the large nor
the small clusters are suitable for parallel processing and external memory operations. Cluster sizes have to be as homogeneous as possible because too much
variance in clustering is not suitable or ineffective for both parallel processing
and external memory operations. Agglomerative algorithms used different parameters for controlling cluster sizes.
Controlling of cluster sizes is done in two ways as follows. First approach is to
define the limit to the total number of nodes that can belong in a cluster. In HC-PIN
algorithm, this approach is utilized very effectively with parameter s for controlling
cluster size. Second approach on the other hand sets limit on the total number of
clusters. Top leader algorithm has this approach by utilizing parameter k. It
indirectly sets the size of cluster as follows. If the value of k is very large, then
the sizes of cluster will become very small. Again, if the value of k is very small,
then the sizes of cluster will become huge. Therefore, depending on the size of the
network, the parameter k can be set with a value that may result in reasonable
cluster sizes.
Quality of clustering is very important for structurally feasible clustering. It
indicates how a cluster is confident to claim that it is a cluster. Generally, it is
defined as the ratio of number of edges within the cluster to the number of edges
outside the cluster. Structurally feasible clusters are easy to handle and require less
effort specially for parallel processing. Agglomerative approaches are very rich in
controlling quality of clustering with different parameters to ensure structurally
feasible clusters. The ways of controlling quality of clustering through parameters
vary with the strategy of the algorithm. Different algorithms ensure qualities
through different parameters as follows:
SCAN: The algorithm has two parameters e and for ensuring quality of
clustering. The parameter e is a threshold that represents the minimum structure
similarity for a node u to be in e-neighborhood of node v. Structure similarity
defines the closeness of two nodes. The parameter shows whether a node v is the
more central one by setting this threshold of representing minimum elements
in e-neighborhood of node v, thereby declaring v as core.

74

A. Biswas et al.

Top leader: It also has two parameters and for assuring quality of clustering.
With the parameter , the algorithm identifies outliers that do not fit in any
cluster. Higher value implies more connections, i.e., higher edge density for
a node to be not an outlier. The parameter defines the depth up to which the
common neighbors between a node and a leader can have. It represents the
required strength of connection between the node and a leader for the node to get
included in the cluster of leader.
HC-PIN: The parameter is used in comparison with the weighted degree of
clusters. It indicates the kind of cluster quality that is desired by the algorithm.
Varying the parameter, different hierarchical structures can be obtained with
different qualities. In addition, another parameter s determines the minimum size
of clusters. With parameter s, small and unrelated clusters easily distinguished
from large, significant, and dominant clusters and process further. Such
processing of insignificant clusters enhances overall qualities of clusters. Some
algorithms often report small group of nodes as a separate cluster that might have
been a part of larger cluster. HC-PIN has undermined such difficulty with proper
settings of values of parameter s.
LICOD: The algorithm has three parameters , , and . The parameter
determines of leaders, which actually represents the node that is surrounded by
densely connected region. The parameter is used for merging of the identified
leaders. It determines the closeness between two leaders. If two leaders are
closed, those are kept in same cluster rather than creating different clusters for
each. Therefore, the parameter has very important role in gaining better quality
clusters. The parameter is used to control the extent to which a node might
belong in more than one cluster. By varying parameter, the degree of
overlapping of the clusters can be controlled. For nonoverlapping clustering,
the parameter has no role in determining quality of clusters.
Due to a variety of parameters, most agglomerative algorithms provide large
degree of control to obtain variety of clusters with specific properties as desired.
The different characteristics of the clusters can be highlighted by changing these
parameters. Therefore, with agglomerative approaches, graphs can be partitioned
easily as per the requirements of both parallel processing and external memory
operations for dealing with large graphs.

3.9

Discussion

Big Data scenarios need to be handled in the context of both processing time and
memory requirements. Data clustering has a vital role in dealing with both these
issues. Representation of linked data in the form of network is another aspect of Big
Data scenarios. Parallel processing and external memory operations are prominent
solutions of these Big Data problems. Partitioning of such network is an intermediary step for parallel processing as well as memory management. For a meaningful

3 Agglomerative Approaches for Partitioning of Networks in Big Data Scenarios

75

and efficient partitioning of networks, graph clustering techniques are used. However, as regards implications of clustering in dealing with Big Data scenarios, the
following major issues must be tackled:
Efficient and proper utilization of partitions or clusters is very important in
parallel processing.
Dealing with size of clusters is a major issue. Smaller clusters may reduce
memory requirement but increases processing overhead. Too large clusters
may not fit into memory and increase time complexity.
Meaningful clusters are foremost important for efficient parallel processing.
Evaluation of clusters is also very important for measuring clustering and
algorithms.
Selection of clustering algorithm is another trivial problem especially with the
large numbers of algorithm available for clustering. Each of these algorithms
possesses specific advantages, which makes the selection process more difficult.
Agglomerative approaches are helpful in dealing with Big Data scenarios. These
approaches are local in nature and hence can be easily parallelized. The local nature
of these approaches can be utilized in efficient processing and memory management. Beside this, agglomerative approaches have numerous other advantages that
are summarized as follows:
Most of these approaches are very fast. Time complexities as well as space
complexity are low.
These approaches generate clusters that are more accurate than other
approaches.
Merging of clusters in level by level allows generation of structurally feasible
and meaningful clusters.
Most of the operations incorporated in agglomerative approaches are local,
which do not require accessing of the entire networks. Therefore, it reduces
average memory requirement during intermediate steps of algorithms.
Explicit local operations of agglomerative approaches can be executed in parallel that can reduce clustering time significantly.
Control parameters used in agglomerative approaches allow generating customized clusters. The algorithms are flexible to generate clusters with specific
properties.
Clearly, agglomerative approaches have several advantages in the perspective of
Big Data scenarios. However, incorporation of some of these approaches results in
the following disadvantages:
Too many parameters used in algorithms may cause difficulty in customization
of clustering.
Algorithm cannot be parallelized if operations of algorithms are not local or
local operations are implicit and inseparable from the main algorithm.
Local operations in some algorithms caused large numbers of smaller sized
clusters.

76

3.10

A. Biswas et al.

Conclusion

Big Data handling systems are often confronted with two major issues: efficient
processing and memory management. Parallel processing and external memory
operation techniques are the most favored processes to solve these issues where
partitioning of data is the vital prerequisite for these techniques. Both these
techniques have undertaken processing and memory-related issues to an extent.
However, the rapid growth of data and complex relationships within the data have
unleashed more difficulty in handling. For such Big Data scenarios, incorporation
of network representation and clustering has been addressed and discussed in this
chapter. The utilization of clustering is advantageous to get a meaningful and
efficient partitioning for parallel processing and memory-related issues. Nevertheless, this kind of utilization requires bearing extra overheads related to clustering.
Agglomerative approaches can be very helpful due to their local nature for clustering network data. Besides, several other advantages are also discussed while
dealing with Big Data scenarios. Despite numerous limitations and concerns,
agglomerative approaches have emerged as a core catalyst for dealing with
processing and memory-related issues, especially for Big Data scenarios of
modern-day data systems.

References
1. Wei T, Lu Y, Chang H, Zhou Q, Bao X (2015) A semantic approach for text clustering using
wordnet and lexical chains. Expert Sys with Appl 42:22642275
2. Li S, Wu D (2015) Modularity-based image segmentation. IEEE Trans Circuit Syst Video
Technol 25:570581
3. Nikolaev AG, Razib R, Kucheriya A (2015) On efficient use of entropy centrality for social
network analysis and community detection. Soc Networks 40:154162
4. Li S, Daie P (2014) Configuration of assembly supply chain using hierarchical cluster analysis.
Procedia fCIRPg 17:622627
5. Malewicz G, Austern MH, Bik AJ, Dehnert JC, Horn I, Leiser N, Czajkowski G (2010) Pregel:
a system for large-scale graph processing. In: SIGMOD10, 2010
6. The Apache Software Foundation (2014) http://giraph.apache.org/. Accessed 28 Apr 2015
7. Xue J, Yang Z, Qu Z, Hou S, Dai Y (2014) Seraph: an efficient, low-cost system for concurrent
graph processing. In: Proceedings of ACM HPDC2014, Vancouver, Canada, 2326 June
8. Vial T
(2012) http://blog.octo.com/en/introduction-to-large-scale-graph-processing/.
Accessed 28 Apr 2015.
9. Ajwani D, Dementiev R, Meyer U (2006) A computational study of external-memory BFS
algorithms, SODA 2006 ACM-SIAM Symposium on Discrete Algorithms, Miami, Florida,
USA, January 2006
10. Kanawati R (2011) Licod: leaders identification for community detection in complex networks. In: SocialCom/PASSAT, pp 577582
11. Blondel VD, Guillaume JL, Lambiotte R, Lefebvre E (2008) Fast unfolding of communities in
large networks. J Stat Mech: Theory Exp 10:P10008

3 Agglomerative Approaches for Partitioning of Networks in Big Data Scenarios

77

12. Fan W, Yeung K (2015) Similarity between community structures of different online social
networks and its impact on underlying community detection. Commun Nonlinear Sci Numer
Simul 20:10151025
13. Zhang T, Ramakrishnan R, Livny M (1996) BIRCH: an efficient data clustering method for
very large databases. In: ACM SIGMOD Record, vol. 25, No. 2, pp. 103114, ACM.
14. Wang W, Yang J, Muntz R (1997) STING: a statistical information grid approach to spatial
data mining. In: VLDB, vol 97, pp 186195
15. Shang R, Luo S, Li Y, Jiao L, Stolkin R (2015) Large-scale community detection based on
node membership grade and sub-communities integration. Physica A Stat Mech Appl
428:279294
16. Shah D, Zaman T (2010) Community detection in networks: the leaderfollower algorithm. In:
Workshop on networks across disciplines in theory and applications, NIPS, November 2010
17. Wang J, Li M, Chen J, Pan Y (2011) A fast hierarchical clustering algorithm for functional
modules discovery in protein interaction networks. IEEE/ACM Trans Comput Biol Bioinform
8(3):607620
18. Khorasgani RR, Chen J, Zaane OR (2010) Top leaders community detection approach in
information networks. In: Proceedings of the 4th workshop on social network mining and
analysis, ACM.
19. Xu X, Yuruk N, Feng Z, Schweiger TAJ (2007) SCAN: a structural clustering algorithm for
networks. In: Proceedings of the 13th ACM SIGKDD international conference on knowledge
discovery and data mining, San Jose, CA, USA, August 1215, 2007
20. Steinhaeuser K, Chawla NV (2010) Identifying and evaluating community structure in complex networks. Pattern Recogn Lett 31(5):413421
21. Hubert L, Arabie P (1985) Comparing partitions. J Classif 2(1):193218
22. Fred A, Jain A (2003) Robust data clustering. In: Proceedings of the CVPR, 2003
23. Wikipedia (2015) Precision and recall http://en.wikipedia.org/wiki/Precision_and_recall.
Accessed 28 Apr 2015
24. Laney D (2001) 3d data management: controlling data volume, velocity, and variety, application delivery strategies. META Group Inc, Stamford
25. Minas M, Subrahmanyam K, Dennis J (2015) Facebook use and academic performance among
college students: a mixed-methods study with a multi-ethnic sample. Comput Hum Behav
45:265272
26. Debra AG, Kullar R, Newland JG (2015) Review of Twitter for infectious diseases clinicians:
useful or a waste of time? Clin Infect Dis 60(10):15331540
27. Google Search Statistics (2015) http://www.internetlivestats.com/google-search-statistics/.
Accessed 28 Apr 2015
28. Duffy DE, McIntosh AA, Rosenstein M, Willinger W (1993) Analyzing telecommunications
traffic data from working common channel signaling subnetworks. Comput Sci
Stat 1993:156156
29. Joo H, Hong B, Choi H (2015) A study on the monitoring model development for quality
measurement of internet traffic. Inf Syst 48:236240
30. Joan BE (2015) Content-based image retrieval methods and professional image users. J Assoc
Inform Sci Technol 67(2):23301643
31. Sebastiano B, Farinella GM, Puglisi G, Rav` D (2014) Aligning codebooks for near duplicate
image detection. Multimedia Tools Appl 72(2):14831506
32. Karmakar D, Murthy CA,(2015) Face recognition using face-autocropping and facial feature
points extraction. In: Proceedings of the 2nd international conference on perception and
machine Intelligence, Kolkata, West Bengal, India, pp 116122
33. Sabeur A, Lacomme P, Ren L, Vincent B (2015) A MapReduce-based approach for shortest
path problem in large-scale networks. Eng Appl Artif Intell 41:151165
34. Mehlhorn K, Meyer U (2002) External-memory breadth-first search with sublinear I/O. In:
Proceedings 10th annual European Symposium on Algorithms (ESA), vol 2461 of LNCS, pp
723735. Springer

78

A. Biswas et al.

35. Munagala K, Ranade A (1999) I/O-complexity of graph algorithms. In: Proceedings 10th
symposium on discrete algorithms, ACM-SIAM, pp 687694
36. Biswas A, Biswas B (2015) Investigating community structure in perspective of ego network.
Expert Sys Appl 42(20):69136934

Chapter 4

Identifying Minimum-Sized Influential

Vertices on Large-Scale Weighted Graphs:
A Big Data Perspective
Ying Xie, Jing (Selena) He, and Vijay V. Raghavan

Abstract Weighted graphs can be used to model any data sets composed of entities
and relationships. Social networks, concept networks, and document networks are
among the types of data that can be abstracted as weighted graphs. Identifying
minimum-sized influential vertices (MIV) in a weighted graph is an important task
in graph mining that gains valuable commercial applications. Although different
algorithms for this task have been proposed, it remains challenging for processing
web-scale weighted graph. In this chapter, we propose a highly scalable algorithm
for identifying MIV on large-scale weighted graph using the MapReduce framework. The proposed algorithm starts with identifying an individual zone for every
vertex in the graph using an -cut fuzzy set. This approximation allows to divide the
whole graph into multiple subgraphs that can be processed independently. Then, for
each subgraph, a MapReduce-based greedy algorithm can be designed to identify
the minimum-sized influential vertices for the whole graph.
Keywords MapReduce framework Minimum-sized influential vertices Social
influences Data mining Large-scale weighed graph, big data Big data analytics

4.1

Introduction

Weighted graphs can be used to model any data sets composed of entities and
relationships. Social networks, concept networks, and document networks are
among the types of data that can be abstracted as weighted graphs. Identifying
minimum-sized influential vertices (MIV) in a weighted graph is an important task
in graph mining that gains valuable commercial applications. Consider the
Y. Xie J. He (*)
Department of Computer Science, Kennesaw State University, Marietta, GA, USA
e-mail: yxie2@kennesaw.edu; jhe4@kennesaw.edu
V.V. Raghavan
The Center for Advanced Computer Studies, University of Louisiana at Lafayette, Lafayette,
LA, USA
e-mail: raghavan@louisuana.edu
Springer International Publishing Switzerland 2016
Z. Mahmood (ed.), Data Science and Big Data Computing,
DOI 10.1007/978-3-319-31861-5_4

79

80

Y. Xie et al.

following hypothetical scenario as a motivating example. A small company

develops a new online application and would like to market it through online social
networks. It is worth to mention that word-of-mouth or viral marketing through
online social networks differentiates itself from other marketing strategies because
it is based on trust among individuals close social circle of families, friends, and
co-workers. Research shows that people trust the information obtained from their
close social circle far more than the information obtained from general advertisement channels such as TV, newspaper, and online advertisements [1]. The company
has a limited budget such that it can only select a small number of initial users to use
it (by giving the initial users gifts or payments). The company wishes that these
initial users would like the application and start influencing their friends on the
social networks to use it. And their friends would influence their friends friends
and so on, and thus through the word-of-mouth effect a large population in the
social network would adopt this application. Using the social network shown in
Fig. 4.1 as the example, the links between each individual represent the interactions
between the pair of the individuals, while the numbers on the links mean the social
influence between the pair of the individual. It is obvious that the company may first
choose Andy as the initial user, since Andy can influence Tony, John, and Ricky
with some influences. In sum, the MIV problem is who to be selected as the initial
users so that they eventually influence the largest number of people in the network.
The objective of the MIV problem is to minimize the size and the set of the initial
users because of budget limitations.
The problem is first introduced for social networks by Domingos and Richardson
in [2] and [3]. Subsequently, Kempe et al. [4] proved this problem to be NP-hard
and propose a basic greedy algorithm that provides good approximation to the
optimal solution. However, the greedy algorithm is seriously limited in efficiency
because it needs to run Monte-Carlo simulation for considerably long time period to
guarantee an accurate estimate. Although a number of successive efforts have been
Fig. 4.1 A sample of a
social network

4 Identifying Minimum-Sized Influential Vertices on Large-Scale Weighted. . .

81

made to improve the efficiency, the state-of-the-art approaches still suffer from
excessively long execution time due to the high-computational complexity for
large-scale weighted graph. Furthermore, the graph structures of real-world social
networks are highly irregular, making MapReduce acceleration a nontrivial task.
For example, Barack Obama, the US president, has more than 11 million followers
in Twitter, while for more than 90 % of Twitter users, the follower number is under
100 [5]. Such irregularities may lead to severe performance degradation.
On the other hand, MapReduce framework has recently been widely used as a
popular general-purpose computing framework and has also been shown promising
potential in accelerating computation of graph problems such as breadth-first
searching and minimum spanning tree [6-9] due to its parallel processing capacity
and ample memory bandwidth. Therefore, in this chapter, we explore the use of
MapReduce framework to accelerate the computation of MIV in large-scale
weighted graphs.
The proposed framework starts with identifying an individual zone for every
vertex in the graph. The individual zone of a given vertex is the set of vertices that
the given vertex can influence. To design a scalable algorithm to address this, we
approximate individual zone by using the concept of -cut fuzzy set. This approximation allows us to reduce the complexity of multi-hop influence propagation to
the level of single-hop propagation. Subsequently, we aim to find a minimum-sized
set of vertices whose influence (the formal definition will be presented in Sect.4.3)
reaches a predefined threshold. To reach this goal, a MapReduce-based greedy
algorithm is designed by processing individual zones (the formal definition will be
given in Sect. 4.3) for all vertices.
The contribution of this work can be summarized as follows:
A fuzzy propagation model was proposed to describe multi-hop influence
propagation along social links in weighted social networks.
An -cut fuzzy set called individual zone was defined to approximate multi-hop
influence propagation from each vertex.
MapReduce algorithms were designed to locate each individual zone and then
identify minimum-sized influential vertices (MIV) using a greedy strategy.
This remainder of this chapter is organized as follows: Sect. 4.2 reviews related
literatures of the MIV problem. Network model and the formal definition of the
MIV problem are given in Sect. 4.3. The MapReduce-accelerated framework is
presented in Sects. 4.4 and 4.5. Finally, the work is concluded and future working
direction is given in Sect. 4.6.

4.2

Related Works

In this section, we first summarize the current research status of influence maximization problem in social network and then summarize some works; try to use GPU
framework to accelerate the process to find the solution of the influence

82

Y. Xie et al.

maximization problem. Finally, we remark on the differences of the work described

in the chapter and emphasize the contribution of this work.

4.2.1

Influence Maximization Problem in Social Network

Finding the influential vertices and then eventually influencing most of the population in the network are first proposed by Domingos et al. in [2, 3]. They model the
interaction of users as a Markov random field and provide heuristics to choose users
who have large influence in the network. Kempe et al. [4] formulate the problem as
a discrete optimization problem and propose a greedy algorithm. However, the
greedy algorithm is time-consuming. Hence, recently huge amount of researchers
try to improve the greedy algorithm in two ways. One is reduce the number of
individual searched in the graph. The other is improving the efficiency of calculating the influence of each individual. Leskovec et al. [10] propose an improved
approach which is called CELF to reduce the number of individual searched in the
graph. Later, Goyal et al. [11] propose an extension to CELF called CELF,
which can further reduce the number. Kimura et al. [12] utilize the Strong
Connected Component (SCC) to improve the efficiency of the greedy algorithm.
Although many algorithms are proposed to improve the greedy algorithm, they
are not efficient enough for the large scale of current social networks. Hence, some
works are proposed to fit for large-scale networks. Chen et al. proposed a method
called MixGreedy [13] that reduces the computational complexity by computing
the marginal influence spread for each node and then selects the nodes that offer the
maximum influence spread. Subsequently, Chen et al. [14] use local arborescence
of the most probable influence path between two individuals to further improve the
efficiency of the algorithm. However, both of the algorithms provide no accuracy
guarantee. In [15], Liu et al. propose ESMCE, a power-law exponent supervised
Monte-Carlo method that efficiently estimates the influence spread by randomly
sampling only a portion of the nodes. There have been also many other algorithm
and heuristics proposed for improving the efficiency issues for large-scale social
networks, such as [1619]. However, all of the aforementioned improvements are
not effective enough to reduce execution time to an acceptable range especially for
large-scale networks.

4.2.2

GPU Framework

Completely different from the previously mentioned work, Liu et al. [20] present a
GPU framework to accelerate influence maximization in large-scale social networks called IMGPU, which leveraging the parallel processing capability of

4 Identifying Minimum-Sized Influential Vertices on Large-Scale Weighted. . .

83

graphics processing unit (GPU). The authors first design a bottom-up traversal
algorithm with GPU implementation to improve the existing greedy algorithm.
To best fit the bottom-up algorithm with the GPU architecture, the authors further
develop an adaptive K-level combination method to maximize the parallelism and
reorganize the influence graph to minimize the potential divergence. Comprehensive experiments with both real-world and synthetic social network traces demonstrate that the proposed IMGPU framework outperforms the state-of-the-art
influence maximization algorithm up to a factor of 60.

4.2.3

Remarks

In this chapter, we focus on addressing the MIV problem on large-scale weighted

graphs using the MapReduce framework. The proposed method first improves the
algorithm efficiency by dividing the whole graph into some subgraphs using fuzzy
propagation model. Subsequently, a MapReduce greedy algorithm is presented to
search the best candidates in each subgraph to achieve high parallelism. The
proposed framework shows potential to scale up to extraordinarily large-scale
graphs especially from big data perspective.

4.3

Graph Model and Problem Definition

In this section, we first introduce how to model a weighted graph and then formally
define the minimum-sized influential vertices (MIV) problem.

4.3.1

Graph Model

We model a weighted graph by an undirected graph G(V,E,W(E)), where V is the set

of N vertices, denoted by vi, and 0 < i < N. i is called the vertex ID of vi. An
undirected edge eij (vi, vj) E represents weights between the pair of vertices. W
(E) {pij, if (vi, vj) E, then 0 < pij < 1, else pij 0}, where pij indicates the
weights between vertices vi and vj. For simplicity, we assume the links are undirected (bidirectional), which means two linked vertices have the same weight (i.e.,
pij value) on each other. Figure 4.2 shows an example of a weighted graph. There
are nine vertices in the graph. The weights over edges represent the social influences between the pair of vertices.

84

Y. Xie et al.

0.8
2
1

0.7

0.9

0.6
0.8

0.6

0.8
0.7

0.6
0.6

0.7

0.5

0.6

0.7
9

Fig. 4.2 A sample of a weighted graph

4.3.2

Problem Definition

The objective of the MIV problem is to identify a minimum-sized subset of

influential vertices in the weighted graph. Such that, eventually large number of
vertices in the graph can be influenced by these initially selected vertices. As we
mentioned in Sect. 4.1, we first partition the whole graph into subgraphs by
applying fuzzy propagation modes. The subgraph is called individual zone for
each vertex. Subsequently, we formally define individual zone as follows:
Definition 1. Individual zone (Zonev). For weighted graph G(V;E;W(E)), the individual zone is a fuzzy set (Uv,Mv), where Uv is the set of vertices and Mv is a
function Uv! [0; 1], such that for any vertex v Uv, we have:
8
< 1,Y  
ifx v;
M v x
W eij ,
otherwise;
: e 2P
ij

xv

where Pxv is the path from vertex x to vertex v and eij is an edge in the path. We
further define W(eii) 1.
Another important terminology influence must be formally defined before the
problem definition.
Definition 2. Influence (v). For weighted graph G(V;E;W(E)), the influence of
vertex v is denoted by v, which is the sum of the membership value of all vertices in
Zonev.
Now, we are ready to define the minimum-sized influential vertices (MIV)
problem as follows:
Definition 3. Minimum-sized influential vertices (MIV). For weighted graph G(V;
E;W(E)), the MIV problem is to find minimum-sized influential vertices  V,
such that 8v 2 , v > N  x%, where x% is a predefined threshold.

4 Identifying Minimum-Sized Influential Vertices on Large-Scale Weighted. . .

4.4

85

MapReduce Algorithm for Identifying Individual

Zones

To scale the MIV problem to a large-scale weighted graph, we approximate

individual zones by using -cut fuzzy sets. That is, given a vertex v, the -cut
individual zone of v contains and only contains all vertices whose membership
value toward v is greater than or equal to the given parameter . For simplicity, in
the description of the MapReduce algorithms shown in Sects. 4.4 and 4.5, individual zone actually means -cut individual zone.
A given weighted graph will be represented by using adjacency lists, which are
similar representations used in MapReduce-based algorithms for breath-first
searching and minimum spanning tree [6-8]. For instance, the adjacency lists for
the weighted graph shown in Fig. 4.2 are described as follows:
1.
2.
3.
4.
5.
6.
7.
8.
9.

2(0.7), 7(0.6)
1(0.7), 3(0.8), 6(0.6), 8(0.8)
2(0.8), 4(0.7), 5(0.8), 8(0.6), 6(0.9)
3(0.7), 5(0.5), 9(0.6)
3(0.8), 4(0.7), 8(0.7)
2(0.6), 3(0.9)
1(0.6), 8(0.6)
2(0.8), 3(0.6), 5(0.7), 7(0.6), 9(0.7)
8(0.7), 4(0.6)

As already shown, each line in the adjacency lists represents the one-hop
neighbors for each vertex. The values in the parenthesis are the weights over
each link.
For a big weighted graph, we divide its adjacency lists into k equal-sized files,
where k total size/block size (total size is the total size of the adjacency lists of the
graph data, and block size is the block size of the HDFS of the Hadoop cluster).
Assume there are n vertices for each of the k files. Then we select m vertices from
each file, respectively, to form a set of target vertices for which we identify
individual zones. In other words, the execution of the following MapReduce
algorithm is able to identify individual zones for k*m target vertices. Therefore,
we just need to run n/m times of this algorithm in order to identify individual zones
for all vertices. Fortunately, the n/m times of executing this algorithm are totally
independent to each other and, thus, can run in a completely parallel manner.
As an illustration, we assume k 2 and n 4 for the weighted graph given in
Fig. 4.2. Then, let m 2, i.e., for each of the 2 files, we select 2 vertices as the target
vertices. Assume for the first run of the algorithm, we select vertices 1 and 2 from
file 1 and vertices 5 and 6 from file 2. Then we will have the following two input
files to identify individual zones for vertices 1, 2, 5, and 6.
File 1:
1. 2(0.7), 7(0.6) | 1, 0, 0, 0 | G, W, W, W| 0, 1, 1, 1
2. 1(0.7), 3(0.8), 6(0.6), 8(0.8) |0, 1, 0, 0 | W, G, W, W| 1, 0, 1, 1

86

Y. Xie et al.

3. 2(0.8), 4(0.7), 5(0.8), 8(0.6), 6(0.9) |0, 0, 0, 0| W, W, W, W| 1, 1, 1, 1

4. 3(0.7), 5(0.5), 9(0.6) |0, 0, 0, 0| W, W, W, W| 1, 1, 1, 1
File 2:
5.
6.
7.
8.
9.

3(0.8), 4(0.7), 8(0.7)|0,0,1,0|W,W,G,W| 1, 1, 0, 1

2(0.6), 3(0.9)|0,0,0,1|W,W,W,G| 1, 1, 1, 0
1(0.6), 8(0.6)| 0,0,0,0|W,W,W,W| 1, 1, 1, 1
7(0.6), 2(0.8), 3(0.6), 5(0.7), 9(0.7)| 0,0,0,0|W,W,W,W| 1, 1, 1, 1
8(0.7), 4(0.6) | 0,0,0,0|W,W,W,W| 1, 1, 1, 1

In the two input files shown above, each vertex is represented in the following
format:
VertexID list of adjacent vertices and weights | membership
values target vertices | color decorations towards target
vertices| immediate parent towards target vertices
Take vertex 1 as an example; its membership values to the four target vertices
(vertex 1, 2, 5, and 6) are initialized to be 1, 0, 0, 0, respectively. Its color decoration
is set to be gray (G), white (W), white (W), white (W), respectively, where the first
G means that more vertices belonging to the individual zone of the first target vertex
need to be located starting from this vertex; the rest W means that, for other target
vertices, no immediate action is needed from this vertex; and the other possible color
value is black (B), which means no further development from the current vertex is
needed for the corresponding target vertex. We further assume the ordinal among the
three color values are B > G > W. Its immediate parent vertices toward each of the
target vertices are initialized to be 0, 1, 1, and 1, respectively, where the first
0 means that this vertex itself is the first target vertex and the rest 1 mean that its
parent vertex to the rest of the target vertices remains unknown for right now.
We further use the following data structure vertex to hold the information on
each individual vertex:
Vertex
ID: int
NeighborsID: List<Integer>
MembershipOfNeighbors: List<Double>
MembershipToTargets: Array<Double>
ColorToTargets: Array<Char>
ParentToTargets: Array<Integer>
Then, the MapReduce algorithm can be described in Algorithm 1.
When this MapReduce job is executed on input File 1 and File 2, it will generate
the following output, if 0:5:
1.
2.
3.
4.
5.

2(0.7), 7(0.6) | 1, 0.7, 0, 0 | B, G, W, W | 0, 2, 1, 1

1(0.7), 3(0.8), 6(0.6), 8(0.8) | 0.7, 1, 0, 0.6 | G, B, W, G | 1, 0, 1, 6
2(0.8), 4(0.7), 5(0.8), 8(0.6), 6(0.9) | 0, 0.8, 0.8, 0.9 | W, G, G, G | 1, 2, 5, 6
3(0.7), 5(0.5), 9(0.6) | 0, 0, 0.7, 0 | W, W, G, W | 1, 1, 5, 1
3(0.8), 4(0.7), 8(0.7) | 0,0,1,0 | W,W,B,W | 1, 1, 0, 1

4 Identifying Minimum-Sized Influential Vertices on Large-Scale Weighted. . .

6.
7.
8.
9.

87

2(0.6), 3(0.9) | 0,0.6,0,1 | W,G,W,B | 1, 2, 1, 0

1(0.6), 8(0.6)| 0.6,0,0,0 | G,W,W,W | 1, 1, 1, 1
7(0.6), 2(0.8), 3(0.6), 5(0.7), 9(0.7) | 0,0.8,0.7,0 | W,G,G,W | 1, 2, 5, 1
8(0.7), 4(0.6) | 0,0,0,0 | W,W,W,W | 1, 1, 1, 1

Since the output records contain G color, the job counter numberOfIterations
will be greater than 0. So we run above MapReduce job for another iteration by
using the output of the first run as input. This process will continue until no record
in output contains any G color. Then the output contains information on individual
zones for the target vertices 1, 2, 5, and 6. In the same way, we are able to obtain
individual zones for vertices 3, 4, 7, 8, and 9.

4.4.1

Algorithm 1: Mapper Part

Method Map(vertexID id, vertexRecord: r)

Instantiate Vertex v from id and r
for (int i 0; i&amp;amp;lt;v.ColorToTargets.size(); i)
{
if (v.ColorToTargets[i] G) {
for(int j0; j&amp;amp;lt;v.NeighborsID.size(); j){
double membershipToTarget
v.MembershipToTarges[i] *
v.MembershipOfNeighbors[j];
if (membershipToTarget > ){
Instantiate Vertex vv for Neighbors[j], such
that
vv.ID NeighborsID[j];
vv.NeighborsID Null;
vv.MembershipOfNeighbors Null;
for (int k0; k< v.ColorToTargets.size();
k){
if (ik)
vv.MembershipToTargets[k]
membershipToTarget;
vv.ColorToTargets[k] G;
vv.parentToTargets[k] v.VertexID;
else
vv.MembershipToTargets[k]0;
vv.ColorToTargets[k] W;
vv.ColorToTargets[k] -1;
}
Create record rr from vv in following format:
list of adjacent vertices and weights | membership

88

Y. Xie et al.

values target vertices color decorations towards

target vertices | immediate parent towards
target vertices;
EMIT (vv.VertextID, rr);
}
}
v.ColorToTargets[i] B.
}
}
Create record r from v in the following format:
list of adjacent vertices and weights | membership values
target vertices | color decorations towards target vertices| immediate parent towards target vertices;
EMIT(id, r);

4.4.2

Algorithm 1: Reducer Part

Method Reduce(vertexID id, [r1, r2, r3, . . ., rl])

Instantiate Vertex v from id;
v.VertexID id
v.NeighborsID Null
v.MembersOfNeighbors[0,0,0,. . .]
v.ColorToTargets[W, W,W,. . .]
v.ParentToTargets[-1,-1,-1,. . .]
for each ri in [r1, r2, r3, . . ., rl] {
Instantiate Vertex vv from id, and ri
if (vv.NeighborsID ! Null){
v.NeighborsID vv.NeighborsID;
v.MembershipOfNeighborsvv.MembershipOfNeighbors;
}
for (int j0; j&amp;amp;lt;vv.ColorToTargets.size();
j){
if (vv.ColorToTargets[j] > v.ColorToTargets[j]){
v.ColorToTargets[j] vv.ColorToTargets[j]
}
}
for (int j0; j&amp;amp;lt;vv.MembershipToTargets.size
(); j){
if (vv.MembershipToTargets[j] >
v.MembershipToTargets[j]){
v.MembershipToTargets[j]
vv.MembershipToTargets[j]

4 Identifying Minimum-Sized Influential Vertices on Large-Scale Weighted. . .

89

}
}
}
for (int i0; i&amp;amp;lt;v.ColorToTargets.size(); i)
{
if (v.ColorToTargets[i] G){
Increment
a
predefined
job
counter
called
numberOfIterations;
}
}
Create record r from v in the following format:
list of adjacent vertices and weights | membership values
target vertices | color decorations towards target vertices
| immediate parent towards target vertices;
EMIT(id, r);

4.5

MapReduce Algorithm for Solving MIV

Using the graph shown in Fig. 4.2 as an illustration again, the output of algorithm
1 on this graph can be easily converted to the following format by a MapReduce
process. Each record represents the -cut individual zone for a vertex. Given a
vertex, its influence is the sum of all the membership values in its -cut individual
zone. For example, the influence of vertex 1 is 0.7 0.6 0.56 0.56 2.42. Now,
the task is to find minimum-sized influential vertices (MIV) whose influence
reaches N*x%:
1.
2.
3.
4.
5.
6.
7.
8.
9.

2(0.7), 7(0.6), 8(0.56), 3(0.56)

1(0.7), 8(0.8), 6(0.6), 3(0.8), 4(0.56), 5(0.64), 9(0.56)
2(0.8), 6(0.9), 8(0.64), 5(0.8), 4(0.7), 1(0.56)
3(0.7), 5(0.5), 9(0.6), 2(0.56), 6(0.63)
8(0.7), 3(0.8), 4(0.56), 2(0.64), 6(0.72)
2(0.6), 3(0.9),8(0.54),5(0.72), 4(0.63)
1(0.6), 8(0.6)
7(0.6), 2(0.8), 3(0.6), 5(0.7), 9(0.7), 1(0.56), 6(0.54)
8(0.7), 4(0.6)

We design a MapReduce-based greedy algorithm for computing this task. Let

the minimum-sized set of influential vertices be denoted as S. We also introduce
another set denoted as I, which includes all the vertices that are influenced by
vertices in S as well as their maximum membership values toward all influential
vertices in S. For instance, for the graph shown in Fig. 4.2, if S {2, 8}, then I {2
(1), 8(1), 1(0.7), 6(0.6), 3(0.8), 4(0.56), 5(0.7), 9(0.7), 7(0.6)}, and the influence of
S is 1 1 0.7 0.6 0.8 0.56 0.7 0.7 0.6 6.66.

90

Y. Xie et al.

A MapReduce-based greedy algorithm for identifying minimum-sized influential vertices can be described in Algorithm 2:

4.5.1

Algorithm 2: Mapper Part

4.5.2

Algorithm 2: Reducer Part

As shown in Algorithm 2, in the mapper part, the sum of the influences of each
vertex is calculated and stored. In the reducer part, if the job counter
numberOfIterations is greater than 0, then the algorithm 2 will run again to add

4 Identifying Minimum-Sized Influential Vertices on Large-Scale Weighted. . .

91

the next most influential vertex to the set S, which is the influential set we are trying
to construct.

4.6

Conclusion and Future Work

In this chapter, we proposed a fuzzy propagation model to simplify the description

of how a vertex influences others in a large-scale weighted social network from big
data perspective. A MapReduce-based algorithm was then designed to locate
individual zone for each vertex of the network. The concept of individual zone
approximates the influence propagated from a vertex by using an -cut fuzzy set.
Then, a MapReduce-based greedy algorithm was designed to identify MIV from all
individual zones.
The MapReduce algorithms are well designed in the chapter. However, how to
validate the proposed algorithms needs to be investigated in the future work. In the
future, we plan to conduct both simulations and experiments (crawl real social data
online) to validate and evaluate the performance of the proposed framework.
Another research direction is to extend the influence diffusion model. In the
chapter, we assume linear threshold model which is adding neighbors social
influences together to compare with the preset threshold. More practical diffusion
model can be designed and proposed to solve the MIV problem, such as independent cascade model [21], susceptible/infected/susceptible (SIS) model [22], or voter
model [23]. One interesting direction could be model social influences as positive
and negative influences [23], since there are difference campaigns and opposite
ideas competing for their influence in the social network.

References
1. Nail J (2004) The consumer advertising. Forrester Research and Intelliseek Market Research
Report
2. Domingos P, Richardson M (2001) Mining the network value of customers. ACM SIGKDD,
pp 5766
3. Richardson M, Domingos P (2002) Mining knowledge-sharing sites for viral marketing. ACM
SIGKDD, pp 6170
4. Kempe D, Kleinberg J, Tardos E (2003) Maximizing the spread of influence through a social
network. ACM SIGKDD, pp 137146
5. Kwak H, Lee C, Park H, Moon S (2010) What is Twitter, a social network or a news media?
WWW, pp 591600
6. Goodrich MT (2010) Simulating parallel algorithms in the MapReduce framework with
applications to parallel computational geometry, CORR
7. Lattanzi S, Moseley B, Suri S, Vassilvitskii S (2011) Filtering: a method for solving graph
problems in MapReduce. SPAA, pp 8594
8. Qin L, Yu J, Chang L, Cheng H, Zhang C, Lin X (2014) Scalable big graph processing in
MapReduce. ACM SIGMOD, pp 827838

92

Y. Xie et al.

9. White T (2012) Hadoop: the definitive guide, 3rd edn. OReilly, Beijing
10. Leskovec J, Krause A, Guestrin C, Faloutsos C, VanBriesen J, Glance N (2007) Cost-effective
outbreak detection in networks. ACM SIGKDD, pp 420429
11. Goyal A, Lu W, Lakshmanan LV (2011) Celf: optimizing the greedy algorithm for
influence maximization in social networks. WWW, pp 4758
12. Kimura M, Saito K, Nakano R (2007), Extracting influential nodes for information diffusion
on a social network. NCAI, vol 22, 13711380
13. Chen W, Wang Y, Yang S (2009) Efficient influence maximization in social networks. ACM
SIGKDD
14. Chen W, Wang C, Wang Y (2010), Scalable influence maximization for prevalent viral
marketing in large scale social networks. ACM SIGKDD
15. Liu X, Li S, Liao X, Wang L, Wu Q (2012) In-time estimation for influence maximization in
large-scale social networks. ACM proceedings of EuroSys workshop social network systems,
pp 16
16. Hu J, Meng K, Chen X, Lin C, Huang J (2013) Analysis of influence maximization in largescale social networks. ACM sigmetrics Big Data analytics workshop
17. Kim J, Kim SK, Yu H (2013) Scalable and parallelizable processing of influence maximization
for large-scale social networks. ICDCS
18. He J, Ji S, Beyah R, Cai Z (2014) Minimum-sized influential node set selection for social
networks under the independent cascade model. ACM MOBIHOC 2014, Philadelphia, PA,
USA, August 1114 2014
19. Shi Q, Wang H, Li D, Shi X, Ye C, Gao H (2015) Maximal influence spread for social network
based on MapReduce. Commun Comput Inf Sci 503:128136
20. Liu X, Li M, Li S, Penng S, Liao X, Lu X (2014) IMGPU: GPU accelerated influence
maximization in large-scale social networks. Trans Parallel Distrib Syst 25(1):136145
21. He S Ji, Beyah R, Cai Z (2014) Minimum-sized influential node set selection for social
networks under the independent cascade model. ACM MOBIHOC, August
22. Saito K, Kimura M, Motoda H (2009) Discovering influential nodes for SIS models in social
networks. Discov Science 5808:302316
23. Li Y, Chen W, Wang Y, Zhang ZL (2013) Influence diffusion dynamics and influence
maximization in social networks with friend and foe relationships. WSDM

Part II

Big Data Modelling and Frameworks

Chapter 5

A Unified Approach to Data Modeling

and Management in Big Data Era
Catalin Negru, Florin Pop, Mariana Mocanu, and Valentin Cristea

Abstract The emergence of big data paradigm and developments in other areas,
such as cyber-infrastructures, smart cities, e-health, social media, Web 3.0, etc., has
led to the production of huge volumes of data. Moreover, these data are often
unstructured or semi-structured, with a high level of heterogeneity. Nowadays,
information represents an essential factor in the process for supporting decisionmaking, and that is the reason that heterogeneous data must be integrated and
analyzed to provide a unique view of information for many types of application.
This chapter addresses the problem of modeling and integration of heterogeneous
data that comes from multiple heterogeneous sources in the context of cyberinfrastructure systems and big data platforms. Furthermore, this chapter analyzes
different heterogeneous data models in relation to heterogeneous sources such as
the following: sensors, mobile users, web, and public open data sources (e.g.,
regulatory institutions). A CyberWater case study is also presented for the purposes
of modeling, integration, and operation of these data in order to provide a unified
approach and a unique view. The case study aims to offer support for different
processes inside the CyberWater platform such as monitoring, analysis, and control
of natural water resources, with the scope to preserve the water quality.
Keywords Big data Data modeling Data management Cloud computing
Heterogeneous distributed systems Cyber-infrastructure Natural resources

5.1

Introduction

Big data paradigm represents an interesting and challenging research topic. Many
scientific fields, such as cyber-infrastructures, smart cities, e-health, social media,
Web 3.0 etc., try to extract valuable information from huge amounts of data,
generated on a daily basis. Moreover these data are unstructured or semi-structured,
having a high level of heterogeneity [1].

C. Negru F. Pop (*) M. Mocanu V. Cristea

Computer Science Department, Faculty of Automatic Control and Computers, University
Politehnica of Bucharest, Bucharest, Romania
e-mail: florin.pop@cs.pub.ro
Springer International Publishing Switzerland 2016
Z. Mahmood (ed.), Data Science and Big Data Computing,
DOI 10.1007/978-3-319-31861-5_5

95

96

C. Negru et al.

In the case of smart cities, big data relates to urban data, basically referring to
space and time perspective, which is gathered mostly from different sensors.
Furthermore, the growth of big data changes the planning strategies from longterm thinking to short-term thinking as the management of the city can be made
more efficient [2]. Moreover, this data can open the possibility for real-time
analysis of city life and new modes of city administration and offer also the
possibility for more efficient, sustainable, competitive, productive, open, and transparent city [3].
Healthcare systems are also transformed by big data paradigm as data is generated from different sources such as electronic medical records systems, mobilized
health records, personal health records, mobile healthcare monitors, genetic
sequencing, and predictive analytics as well as a large array of biomedical sensors
and smart devices that rise up to 1000 petabyte [4].
People and the interaction between them produce, regarding the social media
and Web 3.0, massive quantities of information. A huge effort is made to understand social media interactions, online communities, human communication, and
culture [5].
The motivation for this chapter comes from the necessity of a unified approach
of data processing in large-scale cyber-infrastructure systems as the characteristics
of nontrivial scale cyber-physical systems (e.g., data sources, communication, and
computing) exhibit significant heterogeneity. The main contribution of the chapter
is the analysis of different heterogeneous sources and format of data collected from
sensors, users, and the web from public open data sources in the context of the big
data and cyber-infrastructure systems. First, we address the problem of data modeling for the water resource monitoring and management processes. Next, this
analysis is extended considering the big data characteristics of the collected data.
Furthermore, describe a unified approach for the water resource data models.
Finally, we illustrate the applicability of the presented analysis on a real research
project. The chapter is structured as follows.
The first part of the chapter analyzes existing data models for water resource
monitoring and management considering computational methods (e.g., neural networks) for handling environmental data, using software engineering to deliver the
computational solutions to the end user and developing methods for continuous
environmental monitoring based on sensor networks as part of Internet of things.
In the second part, based on this analysis, we address the problem of data
modeling considering a huge volume, high velocity, and variety of data collected
from environments (another big data problem) and process in real time to be used
for water quality support.
The third part presents the unified approach for water resource data models
considering the context of observation and the possibility of usage to enhance the
organization, publication, and analysis of world-distributed point observation data
while retaining a simple relational format. This will enable development of new
applications and services for water management that are increasingly aware of and
adapt to their changing contexts in any dynamic environment. A context-aware
approach requires an appropriate model to aggregate, semantically organize, and

5 A Unified Approach to Data Modeling and Management in Big Data Era

97

access large amounts of data, in various formats, collected from sensors or users,
from public open data sources.
Finally, we present the case study of CyberWater project, a prototype cyberinfrastructure-based system for decision-making support in water resource management, as data modeling of water resources is a fundamental requirement. Cyberinfrastructure basically refers to a mix of advanced data acquisition through realtime sensor networks, big data, visualization tools, high-performance computational platforms, analytics, data integration, and web services.

5.2

Big Data: Heterogeneous Data

Big data refers to dealing with huge amounts of heterogeneous data. This requires a
pipeline of processing operations in order to accomplish efficient analytics. The
overall scope is to offer support in the decision-making process. An important
challenge, besides the large volumes and high-speed production rates (e.g., velocity), is raised by the enormous heterogeneity (e.g., variety) of such data.
Research fields, such as environmental research, disaster management, decision
support systems, information management in relief operations, crowdsourcing,
citizen sensing, and sensor web technologies, need to make use of new and
innovative tools and methods for big data. Ongoing research, challenges, and
possible solutions have been presented through several important projects and in
research papers that face with the previous presented challenges and try to
overcome them.
A new and innovative approach is presented in Dancer FP7 project being
oriented on developing new instruments and tools that will enhance environmental
research and promote innovation in Danube Region, including the Danube Delta
and the Black Sea. The project will undertake a critical analysis of what has been
achieved so far in the region and will build upon results of achievements to date, to
design innovative solutions and to strengthen knowledge transfer in this area.
In [6], the authors address the use of Geo-ICT in two stages of handling disasters,
before and during occurrence, giving special attention to the real utilization of
geo-information, e.g., risk maps, topographic maps, etc. They conclude that in both
risk management and disaster management is a growing awareness of the importance of spatial information and the increasing types of spatial data are used for
performing tasks within risk and disaster management; secondly a general understanding is building up about sharing information between the two domains.
A solution, called Sahana, is presented in [7] of a free and open-source software
(FOSS) application. It aims to be a comprehensive solution for information management in relief operations, recovery, and rehabilitation. After presenting the
architecture of the system, they conclude that there is much left to be done in the
technical development of FOSS disaster management software, and of primary
importance is the need to deal with a heterogeneity of data types (text, semistructured, Web HTML and XML, GIS, tabular, and DBMS) and to develop

98

C. Negru et al.

standards and protocols for data sharing. The purpose of these solutions is to assist
saving procedures for environment (by improving the quality of air, water, sediments, and soil) and people in the same time during a disaster. In periods without
disasters, prevention and improvement actions, like people training, will be
considered.
The authors of [8] survey the crowdsourcing, citizen sensing, and sensor web
technologies for public and environmental health surveillance and crisis management and reinforce the roles of these technologies in environmental and public
health surveillance and crisis/disaster. The data used by these services, collected
from different sources (mobile smart devices, cameras, dedicated equipment, sensors, etc.), must be stored in an aggregated way in the repository.
The authors of [9] propose a cloud-based natural disaster management system.
They conclude that it is essential to have a scalable environment with flexible
information access, easy communication, and real-time collaboration from all types
of computing devices, including mobile handheld devices (e.g., smartphones,
PDAs, and tablets (iPads)). Also, it is mandatory that the system must be accessible,
scalable, and transparent from location, migration, and resource perspectives.
An overview of all the extreme events that threaten people and what they value
in the twenty-first century is presented by Keith Smith in the book named Environmental Hazards: Assessing Risk and Reducing Disaster [10]. It integrates
cutting-edge material from the physical and social sciences to illustrate how natural
and human systems interact to place communities of all sizes, and at all stages of
economic development, at risk. It also explains in detail the various measures
available to reduce the ongoing losses to life and property.

5.2.1

Characteristics, Promise, and Benefits

Big data is a business term that can also represent a buzzword for the advancing
trends in the latest technologies. It describes a new approach in order to understand
the process of data analysis for decision-making. Currently, data is growing at a rate
of 50 % per year through new streams such as digital sensors, industrial equipment,
automobiles, and electric meters (measuring and communicating all sorts of parameters like location, temperature, movement) [11].
In this section, some of the most important big data characteristics such as
volume, variety, velocity, value and veracity, volatile, and vicissitude are analyzed.
Volume. represents the main challenge, as traditional relational database management systems did not succeed to handle volumes of data in terms of terabyte and
petabyte levels. According to [12], it is estimated that 2.5 quintillion bytes of data
are created each day, six billion people have cell phones, and 40 zettabytes will be
created by 2020, which represents an increase of 300 times from 2005.
Variety. characteristic refers to different data formats and sources such as data
from sensors, documents, emails, social media texts, mobile devices, etc.

5 A Unified Approach to Data Modeling and Management in Big Data Era

99

According to [12], 400 million tweets are sent per day by about 200 million
monthly active users, four billion hours of video are watched on YouTube each
month, and more than 420 million wearable wireless health monitors exist.
Velocity. refers to the data acquisition rate, as data can be acquired at different
speeds. For instance, in the case of a natural disaster or pollution accident, the data
acquisition rate from the wireless sensor network rises exponentially, compared
with normal circumstances. According to [12], the New York Stock Exchange
captures 1 TB of trade information during each session, modern cars have close
to 100 sensors that monitor items such as fuel level and tire pressure, and by 2016 it
is projected that there will be 18.9 billion network connections, which represent
almost 2.5 connections per person on Earth.
Value property illustrates the potential gain of data, obtained after some
processing operations. Big data processing operations can help to uncover the
fine interactions in data, allowing to manipulate heretofore hidden often counterintuitive levels that directly impact different domains and activities.
Veracity. characteristic describes how accurate is data collected from different
sources. For example, data gathered from a social media website have a specific
degree of accuracy. In other words, veracity represents the degree of uncertainty in
the data. According to [12], the poor data quality costs the US economy around $1.3
trillion per year, one of three business managers dont trust the information they use
to make decisions, and 27 % of respondents in one survey were unsure about how
much of their data was inaccurate.
Volatility. propriety of big data refers to the time of storage of the data after the
processing step. Volatility has a direct impact on the other aspects of big data such
as volume and veracity. So, data volatility policies have to be defined, in order to
avoid the interference with other properties of even the damage of those.
Vicissitude. property refers to the challenge of scaling big data complex
workflows. This property signifies a combination between the large volume of
data and the complexity of the processing workflow, which prevent to gather useful
insights for data [13].
There are three categories of promises related to big data analytics identified by
the authors of [14]:
Cost reduction technologies like Hadoop and cloud-based analytics can provide significant cost advantages.
Faster and better decision-making this is obtained with the aid of frameworks
such as Apache Storm and Apache Spark designed to run programs up to 100

faster than Hadoop. Also the possibility to analyze new sources of data can help
in decision-making process. For example, healthcare companies try to use
natural language processing tools to better understand customers satisfaction.
New products and services this is possible due to new sources of data such as
mobile phones. For example, a company called Verizon sells data about how
often mobile phone users are in certain locations.

100

5.2.2

C. Negru et al.

Data Models

Data models represent the building blocks of big data applications, having a major
impact on the performance and capabilities of those applications. Moreover, the
different tools that are used for data processing impose these data models. Next, we
will discuss the different data models that exist in the context of big data: structured
data, text file data, semi-structured data, key-value pair data, and XML data [15].
Structured data models refer to data that is contained in database or spreadsheet
files. The sources of structured data in a cyber-infrastructure are water quality
sensors (e.g., water parameters such as conductivity, salinity, total dissolved solids,
resistivity, density, dissolved oxygen, pH, temperature, and so on), mobile phone
sensors (e.g., location data), geographical information systems (e.g., geo-database
that contains spatial data such as points, polygons, raster, annotations, and so on),
and click-stream sources (e.g., data generated by human intervention in the case of
reporting an incident related to a pollution event). Figure 5.1 presents a database
table that contains information about a spatial dataset which has different attributes.
Data can be aggregated and queried across the entire database. Things get more
complicated when we want to aggregate data from many tables because the
problem becomes exponentially more complex. The reason behind this complexity
is that each query requires the read of the entire dataset.
Text data models, on the other hand, are at the opposite end of structured data, as
this type of data has no well-defined structure and meaning. The sources for this
data are represented by different documents related to different regulations released
by regulatory institutions in the field of cyber-infrastructure domains such as water
management [15].
Semi-structured data represents data that has a structure but is not relational.
Instrumentation equipment such as sensors generates this data. In order to be stored
in a relational database, data must be transformed. A major advantage of semistructured data is represented by the fact that can be loaded directly into a Hadoop
HDFS file system and processed in raw from there [16].
Key-value pair data represents the driver for performance in MapReduce programming model. This model has a single key-value index for all data being similar
to memcached distributed in memory cache. This type of data is stored in key-value
stores and in general provides a persistence mechanism and additional functionality
as well: replication, versioning, locking, transactions, sorting, and/or other
features [17].
OBJECT_ID
1
2
3
5
6

Zone
RightBank
MainStream
LeftBank
MainStream
LeftBank

Fig. 5.1 Geographical data example

SHAPE Length (m)

54.021,952344830665
37.099,640979200485
40.974,860305322997
21.167,611441057656
23.707,147651820300

5 A Unified Approach to Data Modeling and Management in Big Data Era

5.2.3

101

Data Gathering

Environmental data gathering and analysis, resulting from continuous monitoring,

pose unprecedented challenges to the design and analysis of methods, workflows,
and interaction with datasets. A big part of these data that resulted from monitoring
is made available by public institutions, private companies, and scientists [18]. Furthermore, data collected from social media sites such as geo-tagged photos, video
clips, and social interactions may contain information related to environmental
conditions. Moreover this process of data gathering must be energy efficient and
cost aware in the context of cloud computing services. New methods for gathering
and integration for these types of data for further processing stages must be
designed.
Geospatial sensor web [19] has been used widely for environmental monitoring.
Different from a sensor network, in a sensor web infrastructure, the device layers,
the network communication details, and the heterogeneous sensor hardware are
hidden [20]. The Sensor Web Enablement (SWE) of the Open Geospatial Consortium (OGC) defines a sensor web as an infrastructure that enables access to sensor
networks and archived sensor data that can be discovered and accessed using
standard protocols and interfaces [21]. In [22] the authors propose a sensor web
heterogeneous node meta-model discussing development of five basic metadata
components, the design of a nine-tuple node information description structure as
shown in Fig. 5.2.
In [23], the authors describe the development of a service-oriented multipurpose
SOS framework in order to access data in a single method approach, integrating the
sensor observation service with other OGC services. The proposed solution
includes few components such as extensible sensor data adapter, OGC-compliant
Level 0 Instance

Level 2 Meta-model

Information
Model

Formalization
Metamodel

XML

Modelling
Facility
Metamodel

Processing
Node

Application
Node

Sensor web
nodes

Level 3 Meta-Metamodel

Information
Describing
Metamodel
Identification

Characteristic

Capability

Quality

Service

Administration

Geolocation

Working

Constraint
Metadata Structure

Fig. 5.2 Sensor web meta-model

Metadata Components

Tag
Performance

State

Accessibility
Service

Sensor Web node Meta-Metamodel

Sensing
Node

Level 1 Model

102

C. Negru et al.

geospatial SOS, geospatial catalogue service, a WFST, and a WCS-T for the SOS,
and a geospatial sensor client. Data from live sensors, models, and simulation is
stored and managed with the aid of extensible sensor data adapter.
There is an important need for data gathering and content analysis techniques
from social media streams, such as Twitter and Facebook. These have become
essential real-time information resources with a wide range of users and applications [24]. Gathering valuable data from social media represents a good opportunity
for environmental monitoring. Also social media can play an important role in
pollution accidents or natural disasters as an information propagator.

5.2.4

Open Issues and Challenges

The challenges addressed in the chapter were also presented in the literature in
various forms: methodologies, applications, etc. The authors of [25] review stateof-the-art research issues in the field of analytics over big data. Among the open
problems and research trends in big data analytics identified by the authors are the
following:

Data source heterogeneity and incongruence

Filtering-out uncorrelated data
Strongly unstructured nature of data sources
High scalability
Combining the benefits of RDBMS and NoSQL database systems
Query optimization

Integration of distributed heterogeneous web data sources is another important

challenge. In [26], the authors propose DeXIN (Distributed extended XQuery for
heterogeneous data INtegration), which is an extensible framework for distributed
query processing over heterogeneous, distributed, and autonomous data sources. In
this approach the base idea is that one data format is considered as the basis or the
aggregation model. It extends the corresponding query language to executing
queries over heterogeneous data sources in their respective query languages.
Disasters, such as water pollution accidents, bring together institutions, organizations, and people. Moreover, these produce large amounts of data with a high
level of heterogeneity that must be handled in order to respond efficiently in disaster
management operations (a special attention can be offered for specific scenarios in
agriculture disasters [27] and land degradation assessment [28]).
In the case of a disaster, it is usually formed a crisis commandment which has to
coordinate a large number of people and resources involved in disaster management
operations. This often led to a lack of coordination and chaos among these institutions, as they track the needs of the people affected by the disaster. Many autonomous actors control information and may work together for the first time, so the
development of a system that enable the information to be shared and analyzed to
target resources is fundamental to build a better response capacity [29].

5 A Unified Approach to Data Modeling and Management in Big Data Era

103

With the aid of big data solutions, we can acquire better results in response time,
coordination, and distribution of efforts, tracking displaced and vulnerable
populations; logging the damage to housing, infrastructure, and services; dealing
with the sudden influx of humanitarian supplies; and coordinating the work of
dozens and even hundreds of responding agencies [29]. However, there is no
current solution that provides information system or data management support for
the basic functionality of disaster management such as registering organizations,
locating missing persons, and requesting assistance [30].
So, a major challenge is to provide an integrated and interoperable framework
for environmental monitoring and data processing that include development of
sound solutions for a sustainable crisis management in order to support human
life in the case of disasters. For instance, the authors of [31] propose an approach to
integrate heterogeneous data with uncertainty in emergency systems.
Based on ICT (information and communications technology) support and big
data technologies, the response to natural disasters needs to be both rapid and
carefully coordinated. Saving lives and providing shelter for those displaced are
two key priorities for first responders. Helping citizens post disaster is a key part of
the mandate for damage assessors. This is by measuring and quantifying a disasters
impact on the community and then providing assistance to individuals. Mobile and
cloud-based GIS offers many potential benefits to improve disaster management.
This is the position of WebMapSolutions Company, which is implementing stateof-the-art disaster management GIS solutions.
Another important challenge is to combine mobile applications, existing electronic services, and data repositories in an architecture based on cloud solutions and
existing big data approaches. Also we have to contribute to the development of
INSPIRE compliant solutions [32], based on the ISO 19156 standard on Observations and Measurements (O&M) and SOS (Sensor Observation Service), SES
(Sensor Event Service), and SAS (Sensor Alert Service) as a main contribution to
standardization.
Incorporating and extending the technology presented in the literature to create a
new one that will transform data into valuable information is another big challenge.
Even when faced with the big data problem, solutions based on cloud computing
and HPC infrastructure to process data must be used.
The main challenges when dealing with systems for water quality management
are the following: storage, processing, and management of data. For the storage
part, scale-out architectures have been developed to store large amount of data, and
also purpose-built appliance has improved the processing capability.
The major challenge is represented by the management of big data throughout its
entire life cycle, from acquisition to visualization in order to get valuable information that helps in decision-making processes. Further challenges are in the field of
real-time control and action prediction for people and authoritys assistance and
setting up correlations between related phenomena, based on the available
information.

104

5.3

C. Negru et al.

Unified Approach to Big Data Modeling

In this part of the chapter, we present a unified approach for water resource data
models considering the context of observation and the possibility of usage to
enhance the organization, publication, and analysis of world-distributed point
observation data while retaining a simple relational format. This will enable
development of new applications and services for water management that are
increasingly aware of and adapt to their changing contexts in any dynamic
environment.
According to the National Oceanic and Atmospheric Administration (NOAA)
that collects, manages, and disseminates a wide range of climate, weather, ecosystem, and other environmental data, there are nine principles for effective environmental data management:
Environmental data should be archived and made accessible.
Data-generating activities should include adequate resources to support end-toend data management.
Environmental data management activities should recognize users needs.
Effective interagency and international partnerships are essential.
Metadata are essential for data management.
Data and metadata require expert stewardship.
A formal, ongoing process, with broad community input, is needed to decide
what data to archive and what data not to archive.
An effective data archive should provide for discovery, access, and integration.
Effective data management requires a formal, ongoing planning process.
A matrix represents the basic structure of an environmental dataset, where
usually the rows correspond to the individual objects (measurements, time units,
or measuring spots) and the columns contain the series of reading for the
corresponding variable. The units in the columns may be logical characters
(true 1, false 0), ordered or unordered categories, integers (count data), or
reals (measurements); they may also contain a time information or a coding of
the measurement spots. The coding of missing data and of censored data (for
extreme values) has to be fixed. Of course some describing or classifying text
may be contained as well in the rows.
In a geographical information system (GIS), there are four basic data structures,
viz., vectors, rasters, triangulated irregular networks, and tabular information (table
of attributes). For example, a virtual representation of the Earth mostly contains
data values that are observed within the physical Earth system. On the other side,
data models are required to allow the integration of data across the silos of various
Earth and environmental science domains. Creating a mapping between the welldefined terminologies of these silos is a challenging problem. A generalized
ontology for use within Web 3.0 services, which builds on European Commission
spatial data infrastructure models, that acknowledge that there are many complexities to the description of environmental properties which can be observed within

5 A Unified Approach to Data Modeling and Management in Big Data Era

105

the physical Earth system is presented in [33]. The ontology is shown to be flexible
and robust enough to describe concepts drawn from a range of Earth science
disciplines, including ecology, geochemistry, hydrology, and oceanography.

5.3.1

Unified Data Representation and Aggregation

When it comes down to big data representation and aggregation, the most important
question that has to be answered is how to represent and aggregate relational and
non-relational data in the same storage engine. Moreover, this data must be queried
in an efficient way so that it offers relevant results across all data types.
An important aspect in wireless sensor networks is end-to-end data aggregation
without degrading sensing accuracy that can prevent network congestion to occur
[34]. With the aid of sensor web technologies, environmental monitoring observations are published. The integration of heterogeneous observations in different
applications poses important challenges as these differ in spatial-temporal coverage
and resolution. In [35], the authors present an approach for spatial-temporal aggregation in the sensor web using the geo-processing web by defining a tailored
observation model for different aggregation levels, a process model for aggregation
processes, and a spatial-temporal aggregation service.
Also, context awareness represents a core function for the development of
modern ubiquitous systems. This offers the capacity to gather and deliver to the
next level any relevant information that can characterize service-provisioning
environment, such as computing resources/capabilities, physical device location,
user preferences, time constraints, and so on [36].
There are numerous applications that need to query multiple databases, with
heterogeneous schemas. The semantic integration approaches focus on heterogeneous schemas, with homogenous data sources. In [37], the authors propose a new
query type, called decomposition aggregate query, to integrate heterogeneous data
source domains. This approach is based on three-role structure. Decomposing
compounds into components and translating non-aggregate queries over compounds into aggregate queries answerable by other data sources are achieved by a
type of data sources called dnodes. It is in order to mention that this is a solution
mainly designed for database management systems.

5.3.2

Data Access and Real-Time Processing

A monitoring platform for water data management needs to access distributed data
sources (e.g., sensor networks, mobile systems, data repository, social web, and so
on). Next, this data has to be processed in real time in order to prevent natural
disasters such as water pollution and more importantly to alert the possible affected
people.

106

C. Negru et al.

Apache Kafka is a solution that proposes a unified approach to offline and

online processing by providing a mechanism for parallel load in Hadoop systems
as well as the ability to partition real-time consumption over a cluster of machines.
Also it provides a real-time publish-subscribe solution, which overcomes the
challenges of real-time data usage for consumption, for data volumes that may
grow in order of magnitude, larger than the real data [38]. Figure 5.3 presents big
data aggregation-and-analysis scenario supported by the Apache Kafka messaging
system.
There are two main categories of data: one which has a value at the given
moment in time such as prediction data and the other whose value remains forever
such as the maximum value possible for pollutants or sensor data which represents
historical data. Mining the instantaneously valued data requires a real-time platform. In [39], the authors propose a method of dynamic pattern identification for
logically clustering log data. The method is a real-time and generalized solution to
the process of log file management and analysis.
Computing frameworks such as MapReduce [40] or Dryad [41] are used for
large-scale data processing. In this paradigm users write parallel computations
with the aid of high-level operators, without paying attention to data distribution
or fault tolerance. The main drawback of these systems is represented by the fact
that these are batch-processing systems and are not designed for real-time
processing.
Storm [42] and Spark [43] represent a possible solution for real-time data
streaming processing. Storm is used currently at Twitter for real-time distributed
processing of stream data. The most important proprieties of Storm are scalability,
resiliency, extensibility, efficiency, and easy administration [44].

Front End

Real-Time

Services

NoSQL

Proxies

Kafka Broker

Hadoop

Adapters

Warehouses

Other

Other

Producers

Consumers

Fig. 5.3 Big data aggregation and analysis in Kafka

5 A Unified Approach to Data Modeling and Management in Big Data Era

5.4

107

Uniform Data Management

In relation to the uniform data management, a context-aware approach requires an

appropriate model to aggregate, semantically organize, and access large amounts of
data, in various formats, collected from sensors or users, from public open data
sources. When it comes down to uniform data management, the most important
question that has to be answered is how to handle relational and non-relational data
in the same storage engine. Moreover, this data must be queried in an efficient way
so that it offers relevant results across all data types.
Data handling methods have been applied in several areas including water
network data analysis and modeling, water quality analysis, energy-water relationship, efficiency modeling of regional water resource under disaster constrains, etc.
Relational databases can handle various types of data, for example, sensor data or
GIS data. Every query has the same kind of data location, water parameters, map,
and so on. These are all stored in a table with one column for each piece of data.
On the other hand, we have multimedia files such as images or videos that can be
attached to an event reporting action. Multimedia files cannot be represented in the
same way as a series of columns. What can be stored in a relational way is
represented by the data about the files, which is in fact metadata. Alongside with
multimedia files, we have social media objects such as blogs, tweets, and emails,
which can be categorized in the category of non-relational data.
One approach is to modify existing relational database management systems in a
way that can store non-relational data. For example, multimedia files can be stored
in Blob data type. In this way, data can be retrieved. The worst part is represented
by the fact that data stored in this way cannot be processed, for example, the image
cannot be scanned in order to find useful information.
Another approach is to design new engines for database systems that can handle
all these big data challenges. So, for example, the new engine can have functions for
parsing multimedia files in order to find information about a pollution event.
The authors in [45] propose a hybrid system called HadoopDB, which combines
the parallel database systems with MapReduce-based systems, in order to benefit
from the performance and efficiency of the first type of systems and scalability,
fault tolerance, and flexibility of the second ones. The main idea of this solution is
to connect multiple single-node database systems using Hadoop as the task coordinator and to communicate through network layer. In this way queries can be
parallelized with MapReduce. The drawback of this approach is that HadoopDB
does not match the performance of parallel database systems.
Another approach is presented in [46]. This proposes a novel definition of a
declarative language that has to be able to map, with precision, an ontology into
queries for a set of data sources. This method is mainly designed for the case of
integration of multiple heterogeneous relational database systems. The authors
introduce the concepts of semantic identifier and semantic join. The first represents
a solution for the problem of entity resolution, and the second one is designed to
help in the problem of record linkage. Although, it is an interesting approach, this

108

C. Negru et al.

has to be modified in order to be used for multiple heterogeneous data sources in the
context of big data. Still, we cannot know for sure how accurate the mapping phase
can be when dealing with these types of data sources.
Tasso Argyros and Mayank Bawa from Teradata propose also an interesting
approach. This supposes the design of a management system composed of three
elements, storage engine, processing layer, and function library [47]. In the storage
engine, relational data is stored in database tables and non-relational data is stored
as de-serialized objects that are similar to Blobs. An extended SQL engine that
includes MapReduce functions represents the processing layer. In this way relational data is queried with SQL and non-relational data is queried with MapReduce
functions. The layer of function library is the core element. In this layer functions
written by users permit the manipulation and query non-relational data. Next these
functions are stored in a library and also the results of the functions are stored in
database tables. This is due to the closure principle from relational database model.
Basically this principle states that any query against a table or tables of data must
return the answer in the form of a table, permitting in this way the chaining of
queries.
So, in order to overcome the data heterogeneity in big data platforms and to
provide a unified and unique view of heterogeneous data, a layer on top of the
different data management systems, with aggregation and integration functions,
must be created.

5.5

CyberWater Case Study

Water resource monitoring implies a huge amount of information with different

levels of heterogeneity (e.g., spatial data, sensor data, multimedia data), availability
(e.g., data must have a minimum degree of redundancy), and accessibility (e.g.,
methods for data access such as REST, SOAP, WSDL) [48]. It is very important to
acquire, store, transmit, and analyze data in order to respond in real time to possible
threads or pollution accidents, for instance.
In this section, we present the case study of CyberWater project. This is a
prototype cyber-infrastructure-based system for decision-making support in water
resource management. The main reason we present this case study is to highlight
the need for heterogeneous data modeling in big data era, especially the context of
cyber-infrastructure-based systems.
Water resource data modeling is a fundamental process for cyber-infrastructures.
These are characterized by a mix of technologies such as advanced data acquisition
through real-time sensor networks, big data, visualization tools, high-performance
computational platforms, analytics, data integration, and web services.
Prototype Cyber Infrastructure-based System for Decision-Making Support in
Water Resources Management (CyberWater) project [49], a national project, aims
to create a prototype platform using advanced computational and communication
technology for implementation of new frameworks for managing water and land

5 A Unified Approach to Data Modeling and Management in Big Data Era

109

resources in a sustainable and integrative manner. The main focus of this effort will
be on acquiring diverse data from various disciplines in a common digital platform
that is subsequently used for routine decision-making in normal conditions and for
providing assistance in critical situations related to water, such as accidental
pollution flooding. So, one of the main challenges of this project will be to integrate
the heterogeneous data sources into a unique and unified view.
In Fig. 5.4, we present a layered architecture of CyberWater monitoring platform. This is focused on three levels: data level, storage level, and level of access,
management, and the data processing. Between data and storage levels, there exists
the interface for monitoring, analysis, and processing rules. Further, between
storage and management level is placed the data access interface.
The data level consists of various heterogeneous data sources used by the
platform, such as sensor network, data suppliers (e.g., GIS, water treatments plants,
etc.), and third-party services (e.g., ANAR, ApaNova, and other Romanian
institutions).
At the storage level, data collected from heterogeneous data sources is stored,
in order to offer aggregation services, such as workflow execution, pollution
propagation modeling, pollution prediction methods, and platform configuration.
Also, the aggregation service is connected with a knowledge management service
and model parameter estimator service, in order to offer support for the offered
services.

Fig. 5.4 CyberWater-layered architecture

110

C. Negru et al.

The top layer of the architecture is the level of access, management, and the data
processing. Here, we have services such as spatial and temporal query services,
customized view of data, data validation, or specific preferences providing functionality for applications like decision support systems, data analysis, real-time
alerts, video, mobile access, and online support.
The interface for monitoring, analysis, and processing rules has the role to store
data from data level (e.g., sensor network, data supplier, external events, and thirdparty services) in a flexible and efficient manner.
Data access interface ensures the access to data for CyberWater services and
applications and also for third-party services which need access to data, offering
access methods such as REST, SOAP, and WSDL.
A sensor gathers the following parameters from water resources: temperature,
pH, specific conductivity, turbidity, and dissolved oxygen.
Data obtained from third-party suppliers is received in heterogeneous formats, in
a form of files such as pdf, csv, txt, etc. For example, data obtained from ApaNova
is in a form of a pdf file, called analysis bulletin, and contains information about
taste, color, odor, turbidity, pH, conductivity, free residual chlorine, ammonium,
nitrites, nitrates, iron, total hardness, and aluminum. From this format, we can
identify three main components:
Limitations specification in a fixed format
Measured values
Semantics which refers to explanation of the measures
Based on the presented architecture, we modeled few conceptual workflows for
decision support module, real-time alerts and visualization of the maps, and informations presented on them. The applications offered by the platform need a unique
view and integrated approach of the data.
Figure 5.5 presents the visualization workflow. A user accesses the application
and can do one of the following actions: can do a zoom in action, click on a sensor,

Fig. 5.5 Visualization workflow

5 A Unified Approach to Data Modeling and Management in Big Data Era

111

or apply a filter. In the case of the first action, a detailed portion of a map is
visualized and data are brought from storage system by the aggregation service. In
this module can be identified the following components: analytics component, data
filtering component, and access to distributed databases. Analytics component
refers to all the activities in the workflow, such as zoom in, click on a sensor, and
apply a filter, and is important because it is the most accessed and utilized part of the
platform, all the users having access to this component.
Data filtering component occurs when users want to find specific data about an
event. So, in order to get data, the user must apply a series of filters, which will be
used as an input to query the storage system. Access to distributed storage supposes
that data gathered from sensors are stored in the near storage site and all the data
must be synchronized and processed.
Figure 5.6 presents the conceptual model for decision support module. First, the
client accesses decision support module from the web application, by making one
of the two requests: data validation or temporal and spatial queries. Next, the
aggregation service will interrogate the storage and service repositories in order
to calculate the propagation model or prediction methods. There are two important
components: analytics component and automatic control component.
Analytics component can be identified in the sub-workflow, which is responsible
with temporal and spatial queries. Data is aggregated from the storage system in
order to provide the application with historical data. In this case the cost for data
aggregation is very important in order to optimize the overall cost of the entire
system.
Another important service offered by the CyberWater project is real-time alerts.
Figure 5.7 presents the workflow of the service. The clients or administrators will
make preferences for platform configuration, and the service will perform spatial
and temporal queries, aggregating data from storage and service repositories. After
this step data is validated against third-party data suppliers and passed to alert

Fig. 5.6 Decision support workflow

112

C. Negru et al.

Fig. 5.7 Real-time alert workflow

Fig. 5.8 CyberWater front-end

generation module, which will classify the alert and take necessary steps for
sending the alert.
All workflows presented above have been encapsulated in a web mapping
application presented in Fig. 5.8. Users are informed about the water quality on a
certain river in a certain point by clicking on a sensor marked on river course. Also
the watercourse is colored based on water quality (blue is a good quality and red bad
quality of the water). The evolution of a certain parameter can be viewed in a graph
displayed on top of the map. Our system also offers a prediction and alert service for
water quality in the case of a pollution event.

5 A Unified Approach to Data Modeling and Management in Big Data Era

113

Users can report an incident by clicking on the map and filling a web form that
describes the reported incident, can upload multimedia files related to the incident,
and can share the reported event on social networks.

5.6

Conclusion

In this chapter, we analyzed the problem of modeling and integration of heterogeneous data sources in the context of cyber-infrastructure systems and big data as it is
a significant research issue. In this type of systems, information represents a key
factor in the process of decision-making. Hence, there is an urgent need to integrate
and operate heterogeneous data, in order to provide a unique and unified view of
information. Benefits such as cost reduction, faster and better decision-making, and
the development of new products and services can be obtained.
The unified approach for big data in environmental monitoring relies on a
separation of heterogeneous information into data, metadata, and semantics. In
terms of data management, operations such as integration, reduction, querying,
indexing analysis, and mining, must be performed. Every operation corresponds to
a specific layer and represents an important stage in the processing pipeline.
Data integrity check and validation also play an important role as there are often
many errors and missing data in environmental datasets. Visualization tools also
play a central role as front-end interface can offer a unified and uniform view of the
entire system permitting users to get useful insights from data and take the necessary actions.
A key research issue nowadays is to find an approach that combines the
relational database management systems and NoSQL database systems in order to
benefit from the two paradigms. Equally important is to have new methods for the
query optimization.
The CyberWater case study presented in the chapter highlights the need for
methods and tools that integrate heterogeneous data sources. These need to be
integrated in order to build a robust and resilient system that offers support in the
decision-making process in the case of water resource management.
Although there are many tools for integrating various heterogeneous data
sources, these do not provide the performance needed by a real-time system. New
scalable and resilient methods and tools to integrate data source heterogeneity to
filter our uncorrelated data need to be developed.

References
1. Jagadish HV, Gehrke J, Labrinidis A, Papakonstantinou Y, Patel JM, Ramakrishnan R,
Shahabi C (2014) Big Data and its technical challenges. Commun ACM 57(7):8694.
doi:10.1145/2611567

114

C. Negru et al.

2. Kitchin R (2013) Big Data and human geography opportunities, challenges and risks. Dialogues in Hum Geogr 3(3):262267. doi:10.1177/2043820613513388
3. Kitchin R (2014) The real-time city? Big Data and smart urbanism. GeoJournal 79(1):114.
doi:10.1007/s10708-013-9516-8
4. Liu W, Park EK (2014) Big Data as an e-health service. In: International conference on
Computing, Networking and Communications (ICNC), February 2014, IEEE, pp 982988.
doi:10.1109/ICCNC.2014.6785471
5. Boyd D, Crawford K (2012) Critical questions for Big Data: provocations for a cultural,
technological, and scholarly phenomenon. Inf Commun Soc 15(5):662679. doi:10.1080/
1369118X.2012.678878
6. Zlatanova S, Fabbri AG (2009) Geo-ICT for risk and disaster management. In: Geospatial
technology and the role of location in science, Springer, Dordrecht, pp 239266. doi:10.1007/
978-90-481-2620-0_13
7. Careem M, De Silva C, De Silva R, Raschid L, Weerawarana S (2006) Sahana: overview of a
disaster management system. In: International conference on Information and Automation,
2006. ICIA 2006, IEEE, pp 361366. doi:10.1109/ICINFA.2006.374152
8. Boulos MNK, Resch B, Crowley DN, Breslin JG, Sohn G, Burtner R, Chuang KYS (2011)
Crowdsourcing, citizen sensing and sensor web technologies for public and environmental
health surveillance and crisis management: trends, OGC standards and application examples.
Int J Health Geogr 10(1):67. doi:10.1186/1476-072X-10-67
9. Habiba M, Akhter S (2013) A cloud based natural disaster management system. In: Grid and
pervasive computing. Springer, Berlin/Heidelberg, pp 152161. doi:10.1007/978-3-64238027-3_16
10. Smith K (2013) Environmental hazards: assessing risk and reducing disaster. Routledge,
New York
11. Lohr S (2012) The age of Big Data. New York Times, http://www.nytimes.com/2012/02/12/
sunday-review/big-datas-impact-in-the-world.html?_r0. Accessed 25 Mar 2015
12. Bizer C, Boncz P, Brodie ML, Erling O (2012) The meaningful use of Big Data: four
perspectives-four challenges. ACM SIGMOD Rec 40(4):5660. doi:10.1145/2094114.
2094129
13. Ghit B, Capota M, Hegeman T, Hidders J, Epema D, Iosup A (2014) V for Vicissitude: the
challenge of scaling complex Big Data workflows. In: 2014 14th IEEE/ACM international
symposium on Cluster, Cloud and Grid Computing (CCGrid), IEEE, May, pp 927932. doi:10.
1109/CCGrid.2014.97
14. Davenport T (2014) Three big benefits of Big Data analytics. http://www.sas.Com/tr_tr/news/
sascom/2014q3/Big-data-davenport.html. Accessed 25 Mar 2015
15. Pop F, Cristea V (2015) The art of scheduling for big data science. In: Kuan-Ching Li, Hai
Jiang, Yang LT, Cuzzocrea A (eds) Big data: algorithms, analytics, and applications. Chapman
& Hall/CRC Big Data Series, pp 105120, ISBN 978-1482240559
16. Freitas A, Curry E, Oliveira JG, ORiain S (2012) Querying heterogeneous datasets on the
linked data web: challenges, approaches, and trends. IEEE Internet Comput 16(1):2433.
doi:10.1109/MIC.2011.141
17. Cal` A, Calvanese D, De Giacomo G, Lenzerini M (2013) Data integration under integrity
constraints. In: Seminal contributions to information systems engineering. Springer, Berlin/
Heidelberg, pp 335352. doi:10.1007/3-540-47961-9_20
18. Buytaert W, Vitolo C, Reaney SM, Beven K (2012) Hydrological models as web services:
experiences from the environmental virtual observatory project. In: AGU fall meeting
abstracts, vol 1, p 1491
19. The Open Geospatial Consortium (OGC) Why is the OGC involved in sensor webs? http://
www.opengeospatial.org/domain/swe. Accessed: 15 Apr 2015
20. Broring A, Echterhoff J, Jirka S, Simonis I, Everding T, Stasch C, Lemmens R (2011) New
generation sensor web enablement. Sensors 11(3):26522699. doi:10.3390/s110302652

5 A Unified Approach to Data Modeling and Management in Big Data Era

115

21. Reed C, Botts M, Davidson J, Percivall G (2007) OGC sensor web enablement: overview and
high level architecture. In: Autotestcon, 2007 IEEE, IEEE, pp 372380. doi:10.1007/978-3540-79996-2_10
22. Chen N, Wang K, Xiao C, Gong J (2014) A heterogeneous sensor web node meta-model for
the management of a flood monitoring system. Environ Model Softw 54:222237. doi:10.
1016/j.envsoft.2014.01.014
23. Chen N, Di L, Yu G, Min M (2009) A flexible geospatial sensor observation service for diverse
sensor data based on web service. ISPRS J Photogramm Remote Sens 64(2):234242. doi:10.
1016/j.isprsjprs.2008.12.001
24. Gao Y, Wang F, Luan H, Chua TS (2014) Brand data gathering from live social media streams.
In: Proceedings of international conference on multimedia retrieval, ACM, April, p 169.
doi:10.1145/2578726.2578748
25. Cuzzocrea A, Song IY, Davis KC (2011) Analytics over large-scale multidimensional data: the
Big Data revolution! In: Proceedings of the ACM 14th international workshop on Data
Warehousing and OLAP, ACM, October, pp 101104. doi:10.1145/2064676.2064695
26. Ali MI (2011) Distributed heterogeneous web data sources integration: DeXIN approach. LAP
Lambert Academic Publishing, Saarbr
ucken
27. Riley J (2001) The indicator explosion: local needs and international challenges. Agric Ecosyst
Environ 87:119120. doi:10.1016/S0167-8809(01)00271-7
28. Wessels KJ, Van Den Bergh F, Scholes RJ (2012) Limits to detectability of land degradation
by trend analysis of vegetation index data. Remote Sens Environ 125:1022. doi:10.1016/j.rse.
2012.06.022
29. Amin S, Goldstein MP (eds) (2008) Data against natural disasters: establishing effective
systems for relief, recovery, and reconstruction. World Bank-free PDF
30. Alazawi Z, Altowaijri S, Mehmood R, Abdljabar MB (2011) Intelligent disaster management
system based on cloud-enabled vehicular networks. In: 2011 11th international conference on
ITS Telecommunications (ITST), IEEE, August, pp 361368. doi:10.1109/ITST.2011.
6060083
31. Huang W, Chen KW, Xiao C (2014) Integration on heterogeneous data with uncertainty in
emergency system. In: Fuzzy information & engineering and operations research & management. Springer, Berlin/Heidelberg, pp 483490. doi:10.1007/978-3-642-38667-1_48
32. van Loenen B, Grothe M (2014) INSPIRE as enabler of open data objectives. In: INSPIRE
conference: INSPIRE for good governance
33. Leadbetter AM, Vodden PN (2015) Semantic linking of complex properties, monitoring
processes and facilities in web-based representations of the environment. Int J of Digital
Earth 9(3):138. doi:10.1080/17538947.2015.1033483
34. Sicari S, Grieco LA, Boggia G, Coen-Porisini A (2012) DyDAP: a dynamic data aggregation
scheme for privacy aware wireless sensor networks. J Syst Softw 85(1):152166. doi:10.1016/
j.jss.2011.07.043
35. Stasch C, Foerster T, Autermann C, Pebesma E (2012) Spatio-temporal aggregation of
European air quality observations in the sensor web. Comput Geosci 47:111118. doi:10.
1016/j.cageo.2011.11.008
36. Bellavista P, Corradi A, Fanelli M, Foschini L (2012) A survey of context data distribution for
mobile ubiquitous systems. ACM Comput Surv (CSUR) 44(4):24. doi:10.1145/2333112.
2333119
37. Xu J, Pottinger R (2014) Integrating domain heterogeneous data sources using decomposition
aggregation queries. Inf Syst 39:80107. doi:10.1016/j.is.2013.06.003
38. Garg N (2013) Apache Kafka. Packt Publishing Ltd., Birmingham
39. Moharil B, Gokhale C, Ghadge V, Tambvekar P, Pundlik S, Rai G (2014) Real time generalized log file management and analysis using pattern matching and dynamic clustering. Int J
Comput Appl 91(16):16. doi:10.5120/15962-5320
40. Dean J, Ghemawat S (2008) MapReduce: simplified data processing on large clusters.
Commun ACM 51(1):107113. doi:10.1145/1327452.1327492

116

C. Negru et al.

41. Isard M, Budiu M, Yu Y, Birrell A, Fetterly D (2007) Dryad: distributed data-parallel programs from sequential building blocks. In: ACM SIGOPS operating systems review, vol
41, no. 3, pp 5972, March ACM. doi:10.1145/1272998.1273005
42. Ankit T, Siddarth T, Amit S, Karthik R, Jignesh MP, Sanjeev K, Jason J, Krishna G,
Maosong F, Jake D, Nikunj B, Sailesh M, Dmitriy R (2014) Storm@twitter. In: Proceedings
of the 2014 ACM SIGMOD international conference on Management of data (SIGMOD 14).
ACM, New York, pp 147156. doi:10.1145/2588555.2595641
43. Matei Z, Mosharaf C, Michael JF, Scott S, Ion S (2010) Spark: cluster computing with working
sets. In: Proceedings of the 2nd USENIX conference on Hot topics in Cloud Computing
(HotCloud10). USENIX Association, Berkeley, CA, USA, pp 1010
44. Toshniwal A, Taneja S, Shukla A, Ramasamy K, Patel JM, Kulkarni S, Ryaboy D (2014)
Storm@twitter. In: Proceedings of the 2014 ACM SIGMOD international conference on
Management of data, ACM, June, pp 147156. doi:10.1145/2588555.2595641
45. Abouzeid A, Bajda-Pawlikowski K, Abadi D, Silberschatz A, Rasin A (2009) HadoopDB: an
architectural hybrid of MapReduce and DBMS technologies for analytical workloads. Proceedings of the VLDB Endowment 2(1):922933. doi:10.14778/1687627.1687731
46. Leida M, Gusmini A, Davies J (2013) Semantics-aware data integration for heterogeneous data
sources. J Ambient Intell Humaniz Comput 4(4):471491. doi:10.1007/s12652-012-0165-4
47. Whitehorn M. Aster Data founders explain unified approach to data big and small. http://www.
computerweekly.com/feature/Aster-Data-founders-explain-unified-approach-to-data-big-andsmall. Accessed: 15 Apr 2015
48. Singh VK, Gao M, Jain R (2012) Situation recognition: an evolving problem for heterogeneous
dynamic big multimedia data. In: Proceedings of the 20th ACM international conference on
Multimedia, ACM, October, pp 12091218. doi:10.1145/2393347.2396421
49. Ciolofan SN, Mocanu M, Pop F, Cristea V (2014) Improving quality of water related data in a
cyberinfrastructure. In: Third international workshop on cyber physical systems. doi:10.
13140/2.1.1380.4803

Chapter 6

Interfacing Physical and Cyber Worlds:

A Big Data Perspective
Zartasha Baloch, Faisal Karim Shaikh, and Mukhtiar A. Unar

Abstract With the increase in utilization and pervasiveness of smart gadgets, there
is a rise in new application domains. For that reason, computational technologies
are progressing very rapidly, and computations are becoming an essential part of
our life. Cyber-physical systems (CPSs) are a new evolution in computing that are
integrated with the real world along with the physical devices to provide control in
real-time environments. CPS generally takes input through sensors and controls the
physical system through cyber systems using actuators. Such systems are really
complex and challenging as they control real environments. This necessitates a
proper interfacing of physical and cyber domains. To this end, the data generated by
physical devices is getting bigger and bigger that is collectively acknowledged as
big data. The real challenge in interfacing cyber and physical domains is the
efficient management of big data. Accordingly, this chapter discusses big data
sources and the relevant computing paradigms. It also classifies and discusses the
main phases of data management for interfacing CPS, viz., data acquisition, data
preprocessing, storage, query processing, data analysis, and actuation.
Keywords Big Data Cyber-physical systems Cloud computing Data
analytics Decision support systems Data management Big data sources

6.1

Introduction

The computing paradigm has evolved in line with the development of the latest and
newer technologies. With these advancements, there is a perception that 1 day
computing will become the fifth utility (after water, gas, electricity, and telephone)
which will be essential for everyday needs of the society [1]. Cyber-physical

Z. Baloch (*) M.A. Unar

IICT, Mehran University of Engineering and Technology, Jamshoro, Pakistan
e-mail: zartasha.baloch@faculty.muet.edu.pk
F.K. Shaikh
IICT, Mehran University of Engineering and Technology, Jamshoro, Pakistan
TCMCORE, STU, University of Umm Al-Qura, Mecca, Saudi Arabia
Springer International Publishing Switzerland 2016
Z. Mahmood (ed.), Data Science and Big Data Computing,
DOI 10.1007/978-3-319-31861-5_6

117

118

Z. Baloch et al.

Fig. 6.1 Generic cyber-physical system

systems (CPSs) may support a new wave of computing by actively engaging it with
the real world in real time [2]. Cyber-physical system is a new generation of systems
with integrated computational and physical capabilities that can interact with
humans through many new modalities [3]. It is a bridge between the cyber world
and the physical world [4], where the physical world is simply the real world and
the cyber world comprised of computing paradigms.
A cyber-physical system is the integration of the physical world with the cyber
world to monitor and control physical entities by using feedback loops. It is an
emerging technology which provides computing and communication facilities to
the real-world systems and adds intelligence to the physical entities (see Fig. 6.1).
CPS uses digital capabilities of computing to control analog physical systems.
In cyber-physical systems, multiple static or mobile sensors and actuators may
be used that are integrated with intelligent decision support systems [5, 6]. The

6 Interfacing Physical and Cyber Worlds: A Big Data Perspective

119

sensors are constrained due to low energy, low computational power, and less
storage capacity. Also a sensor does not have enough storage capacity to
accommodate huge datasets. Cloud computing is a solution to some of these
issues related to sensors. The combination of sensors and cloud is known as
sensor cloud [7]. The sensor cloud infrastructure is a vital part of CPS, where the
cloud performs computing (cyber) activities and sensor supports physical
activities [7].
Kim et al. [8] proposed a generic framework for design, modeling, and simulation of CPS. The paper highlights many important features that need to be part of
that framework. The features include heterogeneous application support, physical
modeling environments that support mathematical expressions, scalability that
helps to increase in number of sensors deployed, support to connect with existing
simulation tools, and software reusability, and all the proprietary solutions and open
standards should be integrated into generic framework [8].
Due to the increase in the use of smart computing devices, a huge amount of data
is generated across the physical world. The term big data is used for those huge
datasets and is defined as a massive volume of both structured and unstructured
data that is so large that it is difficult to process using traditional database and
software techniques [9]. There are many sources of big data at the physical world;
there may be wireless sensor networks, social networks, wireless body area networks, mobile networks and vehicular ad hoc networks, etc. The data are physically
managed through some data management frameworks, and then that captured data
is sent to the cyber world for analytics.
The cyber world may include big data, big data management, cloud storage, data
analytics, and decision support systems. The continuous data growth poses many
challenges. The major issues are storing that data and extracting valuable information from such a large amount of data. The data is not limited, but it is increasing
exponentially so there is a major issue to store this data efficiently and in a costeffective manner. The cloud storage provides a cost-effective way to facilitate the
users with ease of computing, storing, and networking resources. As the big data is
in large quantity and all of that data is not important, there is a need to extract
valuable data through data analytics. Big data analytics is the process of capturing,
arranging, and analyzing huge sets of data to identify patterns and valuable information [10]. The analyzed data will be sent back to the physical world. Big data is a
buzz word today, so it provides wide space for research in this field. This chapter
presents the review of various technical aspects of big data for cyber-physical
systems.
The remaining chapter is organized as follows. Section 6.2 discusses various
sources of big data. Section 6.3 briefly describes data management at cyberspace
that includes cloud computing and decision support systems. Interfacing cyber
and physical worlds is discussed in Sect. 6.4. Section 6.5 identifies the main
challenges of cyber physical systems in terms of big data, and Sect. 6.6 concludes
the chapter.

120

6.2

Z. Baloch et al.

Data Generation by Physical Systems: Big Data

Sources

The first step in big data scenario is data generation. There are many sources of big
data which are generating highly diverse and complex datasets. These sources
include wireless sensor networks, mobile ad hoc networks, social networks, vehicular networks, RFIDs, web servers, online transactions, etc.
Big data can be structured, unstructured, and semistructured. The data, which are
well organized and are based on some data model, are referred to as structured data.
On the other hand, the unstructured data does not follow any data model. The
semistructured data is the combination of structured and unstructured. It is a type of
structured data, but somehow it lacks the data model structure and uses markers or
tags to mark specific data elements. For example, emails contain unstructured data,
but it has some fields like date, time, sender, recipient, etc. which are considered to
be as structured data. Generally, big data is considered as unstructured.
There are three main characteristics of big data: volume, variety, and velocity
[11]. The volume characteristic is defined as the amount of data, variety as different
formats of data/data sources, and velocity is the speed at which the data is growing
[12]. The data is not just large in volume, but there is variety of complex datasets. The
real challenge is to handle that diversity and variety. We can categorize the data
growth as business application data, personal data, and machine data [13]. The data
generated by business applications is moderate in volume, variety, and velocity. This
type of data is highly structured data. It includes online transactions. The personal
data includes web logs, documents, emails, social media, etc. It is highly unstructured
data, and it is moderate in variety but high in volume and velocity. The data growth is
two times more than business application data. The third category is machine data
which include sensors, machine logs data, audio and video recordings, bio-informatics, etc. This type of data is highly structured, and it is high in volume, variety, and
velocity [14]. The growth is three times more than business application data.
In this section, we discuss a number of common data sources such as wireless
sensor networks, social networks, body area networks, and vehicular ad hoc
networks.

6.2.1

Wireless Sensor Networks

In the past few years, the applications of wireless sensor networks (WSNs) have
been increasing rapidly, such as monitoring, event detection, surveillance, etc.
Wireless sensor network is a wireless network of many small devices which are
capable of sensing, computation, and communication. A sensor network consists of
multiple sensor nodes, which are small and lightweight. The sensor nodes are
generally dispersed in a sensor field as shown in Fig. 6.2. Every sensor node
contains a transducer, microcomputer, transceiver, and a power source [15, 16].
When a sensor node senses a physical phenomenon, an electrical signal is

6 Interfacing Physical and Cyber Worlds: A Big Data Perspective

121

Fig. 6.2 Generic wireless sensor network model

generated by the transducer, which is processed and stored by microcomputer. The

collected data will then be sent to the sink/gateway which sends it back to the end user
via Internet/satellite or any other means of communication [17].
The traditional technologies for data processing, storing, and reporting provide
limited support for analyzing WSN data. These technologies have become prohibitively expensive, while dealing with sensor-generated big data [18]. Even then,
they cannot handle the processing requirements for real-time processes such as fire
detection, natural disasters, and traffic control [18]. Thus, the research is directed
toward new technologies for processing big data. There are many attempts in
combining big data and WSN. Jardak et al. [19] proposed a data model for
structuring the stored data by allowing a wide range of analysis by using Bigtable,
Hadoop, and MapReduce algorithms. A Hadoop-based cloud storage solution for
WSN data is presented by Fan et al. [20]. Similarly, Ahmed et al. [21] proposed an
infrastructure for integrating cloud computing with WSNs.

6.2.2

Social Networks

A social network is a Web-based service that connects people with each other. The
user can make their profile and share information, experiences, ideas, etc. The most
popular social networking sites are Twitter [22], Facebook [23], Pinterest [24],
Google+ [25], Instagram [26], and many more. These social sites are very popular
among youngsters. Data is being shared on these networks, which include billions
of photos, videos, and other information. Most of the data is unstructured with high
volume and velocity.

122

Z. Baloch et al.

The social networks can have enormous benefits for the society as it can help in
disasters. Through the social networks, the information about disasters can quickly
be disseminated among the people. The most prominent and widely used social
media networks like Twitter and Facebook are playing an important role in the
propagation of information which could be of different genres. The widespread use
of hashtag trends can help with easy access of the latest trends going on. In case of
any disaster or catastrophic crisis, the faster spread of information through these
sites could be an epidemic in saving lives and providing assistance for the further
course of action. One of the crisis situation examples could be the current deadliest
earthquake which struck Nepal in April 25, 2015 [27, 28], leaving behind thousands
of people dead and other severe casualties. It was within minutes that this news
broke through the whole social networks and spread throughout the whole world.
Immediate actions were taken to help the people affected by this devastating
tragedy. This was due to social networking sites which showed the world how
severe the situation was, and because of it various rescue and relief aids were
instantly sent from around the world to Nepal. Social networks have become a
binding force in the world where within seconds information could be propagated
from one corner of the world to the other. The only disadvantage is that we cannot
verify the credibility of the information being generated on the social networking
sites. Furthermore, social media analysis can be helpful for the organizations to
redesign their policies to address the public issues [29]. Social networks are leading
toward a new generation of crowd sourcing applications [30], which will help in
analyzing in-depth physical environments.

6.2.3

Vehicular Ad Hoc Networks

The research on integrating communication technologies with vehicles has begun

since long ago. The communication between vehicles by using ad hoc networks is
known as vehicular ad hoc networks (VANETs) [31, 32] as shown in Fig. 6.3. It is a
subcategory of intelligent transport systems and mobile ad hoc networks. The
vehicles can share necessary information with each other (referred to as vehicle

Fig. 6.3 Generic vehicular ad hoc network

6 Interfacing Physical and Cyber Worlds: A Big Data Perspective

123

to vehicle or V2V), such as traffic information (traffic jam or accident), emergency

warnings, weather information, road condition warnings, etc. Furthermore, the data
can also be shared with the data center to the passing vehicles (referred as infrastructure to vehicle or I2V and vice versa). It is not only important to pass along the
latest information but also to remove the outdated data [33, 34]. Since there are
many vehicles passing on the roads, they may consume the total bandwidth in data
dissemination. Therefore, it is important to efficiently transmit the data by using
limited bandwidth [35].

6.2.4

Wireless Body Area Networks

The recent development in wireless networks and microelectronics has resulted in

wireless body area networks (WBANs). A WBAN may consist of miniature
lightweight sensor nodes with low power and is used for healthcare application to
monitor physiological status of the human body, such as blood pressure, blood
sugar, ECG, pulse rate, etc. [36].
Figure 6.4 shows a generic architecture for WBAN-based health monitoring
system where sensor nodes senses medical data and sends it to the base station (BS).

Fig. 6.4 Generic WBAN scenario

124

Z. Baloch et al.

Table 6.1 Comparison of big data sources

Big data sources
WSN
Social networks
VANET
WBAN

Volume
High
Moderate to high
Moderate to high
High

Velocity
High
High
Moderate
High

Variety
Low
Low to moderate
Low to moderate
High

Structured/unstructured
Structured
Unstructured
Structured
Unstructured

The BS then transmits the data to the doctor for real-time diagnosis, to a database
for keeping medical records or to a particular device that generates emergency
alerts via medical health service (MHS) center [37, 38].
The common factor in all the emerging data sources is that they all are
continuously generating data. The data is high in volume, with lots of variety
and with dynamic velocities [39]. In general, the characteristics of big data for
these data sources are summarized in Table 6.1. The data from these four sources
are mostly high in volume and velocity. As the WSN, VANET, and WBAN data
are sensor-generated data, so they have a vast data variety depending on the nature
of sensors used and are mostly highly structured. The social network data is
moderate to high in volume but does not have much data variety. This type of
data is mostly unstructured, as it mostly contains images and audio and video
streams.

6.3

Data in Cyber Systems: Big Data Management

Once the raw data is collected from the physical world, it is passed to the cyber
world for further processing. In this section, we will mainly consider how the data is
managed in the cyber world and what the prerequisites are. This section also
highlights the role of existing cloud computing paradigm and decision support
systems that solves the problem of storage and provides other computational
facilities.

6.3.1

Cloud Computing Paradigms

For the past few years, cloud computing and big data have been the two key fields
which gained significant attention of the researchers [40, 41]. A few years back,
large datasets assumed to be of a few terabytes, but nowadays this concept has
been changed, and individual applications are producing more than that: new units
being used as terabytes and petabytes. With this continuous growth of data, it is
difficult for an organization to handle the data which is too big, too versatile, and

6 Interfacing Physical and Cyber Worlds: A Big Data Perspective

125

too fast because the traditional storage methods are not designed for such a huge
data. One solution is cloud storage. Cloud computing is the emerging technology
that provides users to perform complex computations without maintaining expensive hardware and software. Although cloud computing is currently used by
almost all the leading companies, still there is no universally agreed definition
[42]. Gartner [43] defines cloud computing as a computing style which provides
scalable and elastic IT-enabled capabilities as a service by using Internet technologies. It provides many computational services to the users such as infrastructure as a service (IaaS) [44], platform as a service (PaaS) [44], and software as a
service (SaaS) [44]. The IaaS offers storage and processing infrastructure as a
service. The user is provided virtualized infrastructure without worrying about
hardware resources [45]. The PaaS provides a platform for the software developers to write and upload their application code [45]. The SaaS is the most
common layer of cloud computing. It provides software application as a service
to users on pay-as-per-use basis [45]. Cloud computing has many advantages over
other computational services which includes parallel processing, security, scalable
data storage, and resource virtualization [46]. It also reduces maintenance of
infrastructure. Cloud computing supports virtualization; through virtualization
software, a simple computer can behave like a supercomputer at an affordable
cost [46].
There is a big challenge for researchers to design an appropriate platform for
cloud computing that handles it and performs data analytics. There are many cloud
service providers, whenever an enterprise tend to migrate from IT system to the
cloud, the decision can be difficult, as the enterprise want to evaluate cost, benefits,
and risks of using cloud computing [47]. Hashem et al. [47] presented two decision
support tools for migration to the public IaaS cloud. These tools help an enterprise
to make cloud migration decisions. The first tool is a cost modeling tool [47], which
can be used in modeling the requirements of enterprise data, applications, and
infrastructure along with the usage patterns of computational resource. This tool
can also be used to compare the cost of cloud services from different cloud service
providers with different deployment options and usage. The second tool, which is a
spreadsheet, shows the usage benefits of IaaS cloud, and it also provides beginning
point for risk assessment [47].
The data management applications in the cloud include two main data management components: transactional data management and analytical data management
[48]. Transactional data management is the database related with transactions like
banking and online reservations [49]. Shared nothing is simply a distributed
architecture in which each node consists of a processor, main memory, and disk,
and the nodes communicate with each other via interconnecting network [50]. For
implementation of transactional data management, the usage of shared nothing
architecture may result in the complex distributed locking and commit protocols
[48]. There is an extensive risk of storing transactional data on untrusted host
because of the sensitive information which includes credit card numbers and pin
codes. Analytical data management is for the applications that query a data store for

126

Z. Baloch et al.

business intelligence and decision making purposes [48]. Its scale is larger than
transactional database management. The analytical data management systems are
best suited for execution in a cloud environment. Generally, for implementation of
analytical data management, the use of shared nothing architecture is well suited.
The atomicity, consistency, and isolation are easy to obtain as compared to
transactional data.

6.3.2

Service-Oriented Decision Support Systems

Decision support system (DSS) is a computer application that analyzes business

data and presents it in a way so that users can make business decisions more easily
[51]. It may use artificial intelligence for analyzing data. DSS finds certain patterns
in data which helps humans to take decisions. For example, DSS helps doctors to
diagnose the disease on the basis of symptoms.
There are three service models for service-oriented DSS [52], namely, data as a
service (DaaS), information as a service (IaaS), and analytics as a service (AaaS).
These are discussed in the following subsections.

6.3.2.1

Data as a Service

The service-oriented architecture provides access to the data from anywhere,

independent of the platform. The data as a service provides the business applications, a facility to access the data wherever it resides [53]. With the provision of
DaaS, the data quality can be maintained at central place, i.e., at cloud. According
to Demirkan et al. [52], the data cleansing and data enriching can be done by two
solutions, namely, master data management (MDM) and customer data integration
(CDI), where customer data can be placed anywhere and can be accessed as a
service through any application that has service provision.

6.3.2.2

Information as a Service

Sometimes the information repositories at organizations are not efficiently designed

to transmit information to the required destinations; this is due to the increased
complexity of processes and architectures. Demirkan et al. [52] defines information
as a service as an idea to make information quickly available to users, processes,
and applications in an organization. This shares real-time information with emerging applications and hides complexities. It increases availability with virtualization.
It also provides master data management (MDM), content management services,
and business intelligence services [52].

6 Interfacing Physical and Cyber Worlds: A Big Data Perspective

6.3.2.3

127

Analytics as a Service

Analytics as a Service can be defined as the combination of cloud computing and

big data analytics [53]. It enables data scientists to access datasets that are centrally
managed by cloud providers. The business analysts can make decisions effectively
and delivers successful outcomes. AaaS is a cloud-based analytical platform, where
several data analytical tools are available, which can be configured by end users to
process and analyze large amounts of heterogeneous data.

6.4

Interfacing Cyber World with Physical World

The error-free interaction between cyber and physical worlds is not an easy task
because the physical world is mostly unpredictable. The most critical element is
that human lives are dependent on the system. Thus, availability, connectivity,
predictability, and repeatability are very much important for the cyber-physical
interface [54].
In general, CPS can be divided into three parts: physical world, cyber world, and
interfacing physical and cyber worlds. The sensors sense physical characteristics,
then the sensed data is passed to cyber world to perform computations, and finally,
response is generated through some actuators (see Fig. 6.1).
The physical components may include power sources, energy storage, and
physical transducers that perform energy conversions in physical domains
[55]. The cyber components may include data stores, computation, and I/O interfaces. The interface contains both physical and cyber components and adds a few
more components to connect them. Rajhans et al. [56] presented two connector
types for modeling the interface between cyber and physical worlds. The connectors are physical-to-cyber (P2C) and cyber-to-physical (C2P) connectors. Simple
sensors can be used for physical-to-cyber connector type and actuators can be used
for cyber-to-physical connector type. For the complex interfaces in CPS, physicalto-cyber transducer and cyber-to-physical transducer may also be used, which have
ports to cyber and physical components on each side [56].
Applications of CPS apparently have potential to overshadow the twentiethcentury IT advancements. CPS applications include many components that cooperate through an unpredictable physical environment. In this regard, reliability and
security are major issues to be resolved. The CPS applications include transportation, defense, energy and industrial automation, health and biomedical, agriculture,
and other critical infrastructures [57].
Cyber-physical cloud computing is the integration of CPS and cloud computing.
The CPCC architectural framework is a system environment that can rapidly build,
modify, and provision cyber physical systems composed of a set of cloud computing based sensor, processing, control, and data services [58]. The customers can
access available resources through Internet independent of location and devices.

128

Z. Baloch et al.

CPCC automatically manages all the resources. CPCC can benefit many systems
such as traffic management, intelligent power grids, disaster management systems,
healthcare, etc. [58].
Despite the fact that data management is the focal point of interest for many
researchers, there is still lack of an agreed-upon definition for data management. It
includes many phases. The data management phases are defined by many
researchers; these steps depend on the nature of data to be managed. Some phases
can be added or removed accordingly. We will discuss some of them in this section.
In TDWI report [39], the data management is defined as data collection, storage,
processing, and delivery, and it considers data management as a broad practice that
includes many data disciplines such as data quality, data integration, data
warehousing, event processing, database administration and content management,
etc. According to Mokashi et al. [59], data management includes data collection,
data storage, and query processing. Padgavankar et al. [60] consider data management as a four-step process, i.e., data generation, big data acquisition, big data
storage, and data analysis. Furthermore, Sathe et al. [61] use four tasks for WSN
data management, i.e., data acquisition, data cleaning, query processing, and data
compression.
Accordingly, we classify the main phases of data management for interfacing
CPS as data acquisition, data preprocessing, storage, query processing, data analysis, and actuation.

6.4.1

Data Acquisition

Data acquisition is simply data gathering or data collection. As the physical world is
generating huge datasets, the data acquisition process determines which data should
be collected along with minimal energy consumption. This is challenging task,
because of the data uncertainty due to natural errors, noise, and missed readings in
sensor data [62]. It is responsible for efficiently collecting samples from sensors in
CPS. In sensor data acquisition, the main objective is to achieve energy efficiency
because sensors are battery powered and located mostly at unreachable locations.
Sathe et al. [61] presented model-based data acquisition techniques that are
designed to handle challenges such as minimal energy consumption and
communication cost.

6.4.2

Data Preprocessing

In the data acquisition phase, the raw data gets collected. The acquired sensor
datasets may sometimes contain erroneous or redundant data, which will definitely
occupy more storage space and will affect data analysis [60]. Therefore, before data

6 Interfacing Physical and Cyber Worlds: A Big Data Perspective

129

storage, preprocessing may be applied which may include data cleaning, data
fusion, and data compression [60].

6.4.2.1

Data Cleaning

It is a process of finding incomplete, inaccurate, and unreasonable data and then

correcting the errors to improve data quality [63]. In the data cleaning process, the
errors will be removed from raw sensor data. For incomplete datasets, regression or
interpolation models can be used to reconstruct missing data. Alonso et al. [64]
proposed an extensible receptor stream processing (ESP) framework for online data
cleaning of the acquired sensor data streams.

6.4.2.2

Data Fusion

As the name shows, data fusion combines data from various data sources. Sensor
fusion is a technique to merge data from many sources to provide accurate and
comprehensive information [65]. It is a technique to address sensor impairments.
Some other terms are also used in the literature that are related to data fusion such as
decision fusion, multisensor fusion, and information fusion.
Data fusion is a technique of combining data from various sensors and information from related databases to attain accuracy and more specific inferences than by
using a single sensor [66]. The data fusion techniques can be categorized into three
categories, that is, data association, state estimation, and decision fusion [67].

6.4.2.3

Data Compression

As the sensors continuously generate huge datasets, and sometimes the collected
data contains redundant data, this redundancy is common in environmental monitoring. The data compression techniques can help to reduce the redundancy which
helps in reducing storage space [60]. Data transmission is more energy consuming
than computation; thus, reduced data size, before data transmission, will minimize
the overall energy consumption [68]. Different data compression schemes have
been discussed by Kimura et al. [68]. Sathe et al. [61] also discussed many data
compression techniques such as linear approximation, model approximation, and
orthogonal transformation. Marcelloni et al. [69] introduces a new lossless compression algorithm that is suitable for reduced computational and storage resources
of a WSN node. Many other techniques are proposed in the literature, some of them
are derived from signal processing [70] and some has used correlations between
sensor data to compress the data streams [7173].

130

6.4.3

Z. Baloch et al.

Data Storage

Due to the limitations of sensors, it is important to store the sensor data efficiently
elsewhere [74], to improve the data retrieval and analytics processes. As the sensors
generate huge datasets, the question arise: do all the generated data is required to be
stored? For different applications, the answer could be different. For example, in
real-time applications, mostly, the recent data is important [75], so there is no need
to keep all the data for long periods. In such cases, live data streaming will be an
appropriate approach. In some applications, the historical data need to be stored for
future analysis, in those cases; the historical storage approach will be appropriate
[76]. For a variety of applications, both the storage approaches can be combined to
make an efficient storage system, but this could be very challenging [75].
The second important issue is determine where to store that data. Many
researchers have done work in this direction. Three data storage methods for
WSN have been discussed by Xing et al. [74]; these are local storage, external
storage, and data centric storage (DCS). In local storage, only short-lived data is
stored in the sensor node. In external storage, data is stored at an external point for
further processing. While in data-centric storage, data is stored along with the name
or location. In DCS, the related data is classified and named according to its
meaning. The data with the same name will be stored in the same sensor node.
For a particular name, the user queries will be sent directly to that particular node
which holds that named data [59].
For live querying data, many data management techniques have been proposed,
whereas for querying historical data, only a few data management solutions have
been developed. Those techniques are discussed by Diao et al. [76].
With the technology advancements, the storage devices are becoming more
energy efficient and cheaper in price. Thus the sensor networks are transforming
from communication-centric to storage-centric perspective which provides a network that efficiently stores data from sensors [76, 77]. The data can be batched or
accessed later. Energy efficiency can be improved by batching sensed data. In
storage-centric sensor network, the applications must be delay tolerant because
the data is not transmitted immediately. For the applications where immediate
response is needed, delay cannot be tolerated, in such cases communication-centric
approach is appropriate [77].

6.4.4

Query Processing

Another important component of data management is data retrieval or query

processing. Many important model-based query processing techniques, which aim
to process queries by retrieving minimum amount of data, are presented by Sathe
et al. [61]. Apart from that, these techniques also handle missing data and create an
abstraction layer on sensor network by using these models [78, 79]. Some of the

6 Interfacing Physical and Cyber Worlds: A Big Data Perspective

131

techniques are based on hidden Markov model (HMM) [80] or dynamic probabilistic model which is for spatiotemporal evolution of the data from sensors [79].
For querying the real-time applications of CPS, different researchers have
developed many tools that can be named as information flow processing (IFP)
systems [81]. Information flow processing is an application domain where users are
required to collect data from various data sources to process it within due time
[81]. After processing data, the collected data is generally discarded, except some
critical applications where historical analysis is important. This is done using two
popular models that are data stream processing [82] and complex event processing
models [83]. The data stream processing model is processing data from various
sources to produce output data streams. The data stream management system
(DSMS) is also based on database management systems (DBMS) with a few
differences such as DBMSs deals with data that is not updated constantly, whereas
DSMSs are specially designed to deal with data that is updated continuously
[81]. Apart from few differences, there are more similarities in between both
of them.
The recent developments in DSMSs are reviewed by Golab et al. [84]. The
complex event processing model considers information flow items as notification of
events of the physical world, which will be filtered and combined to visualize what
is happening in the form of high-level events [81]. The approach mainly focuses to
detect patterns of low-level events that will eventually be combined to represent the
high-level events that will be notified to the parties that are interested. An architecture for real-time analysis and processing of complex-event streams of sensor
networks, which is based on semantically rich event models, is presented by Dunkel
et al. [85].

6.4.5

Data Analysis

The data analysis is the most significant phase of data management for CPS
interfacing. As CPS data is larger in magnitude, the real challenge is to extract
the insight value from it which is valuable. The purpose of data analysis is to sift
valuable information. It helps organizations to better cope with the needs of their
customers and to make better decisions [86].
The traditional analytical methods, based on statistics and computer science,
may still be used for big data analysis, such as cluster analysis, factor analysis,
correlation analysis, regression analysis, real-time analysis, offline analysis, memory level analysis, business intelligence (BI) analysis, and massive analysis
[60]. Some advanced analysis techniques are also required to handle the complexity
of real-world heterogeneous datasets. Big data analytics is the set of modern
techniques that are designed to operate on heterogeneous data with large magnitudes [87]. The intelligent quantitative methods, such as artificial intelligence,
robotics, artificial neural networks, or machine learning, can be used to explore
and to identify hidden patterns and their relationships [87].

132

Z. Baloch et al.

In a typical CPS for environment monitoring, most of the collected data is

considered as regular, but some of them may be irregular; such data is known as
atypical data [88]. Atypical data is extremely crucial as it identifies a change in
environmental condition; therefore, such data need to be analyzed. Different
approaches have been discussed in the literature [8890] to analyze atypical data
in the CPS. Tang et al. [91] proposed a method named as Tru-alarm, which finds out
trustworthy alarms for the cyber-physical systems. It uses data analysis to eliminate
noisy data that can cause false emergency alarms.

6.4.6

Actuation

Actuation is the most crucial element in CPS because it controls the environment.
In many CPS applications, sensing data is not just sufficient, but a response is also
required to show how the system reacts in a particular situation [92]. For example,
in fire alarm systems, the actuators may be deployed to shower water on the fire.
Another example could be of an agricultural environment where crops can be
monitored and pesticides can be sprinkled by an actuation process if needed.
Data actuation is the process in which the processed data is sent to actuators to
perform some action. It transfers data back to the physical systems. Thouin
et al. [93] discussed different actuation strategies to acquire desired actions to be
performed on physical devices. A dynamic actuation strategy is group of decision
rules to find the actuation nature which will be executed throughout the course of
operations in a wireless sensor and actuator networks [93].

6.5

Future Challenges and Opportunities

The CPS is a multidisciplinary technology, which involves communication and

networks, embedded systems, and semantic technologies. To take the maximum
benefit from CPS and to handle big data that flows in between the cyber and
physical worlds, there are many challenges to be addressed. A few challenges of
CPS in big data perspective are given below.
Volume As discussed earlier, CPS data is enormous and keeps growing continuously, processing that huge dataset can lead to many challenges. Data abstraction,
i.e., summarizing the data and making it human comprehensible, is one of the
biggest challenge for big data generated across CPS. Another challenge could be
efficient use of distributed processing to scale the CPS computations. The simple
computations can become complex when scaling from terabytes to petabytes. Even
sequential scans to petabytes data takes too much time. The indexing techniques are
also very challenging while scaling to huge volume.

6 Interfacing Physical and Cyber Worlds: A Big Data Perspective

133

Variety There is a huge variety of datasets with different data formats, which need
to be integrated together. As the data is collected from distinct sources, the structure
of data can be complex and data processing can also be very complex. Thus,
efficient techniques are needed to cope with the increasing variety of data.
Velocity With these fast growing datasets, it is challenging to focus on the data
trends and the correlations between data. There is a great need of robust and realtime techniques to cope with velocity of the data generation and processing.
Veracity Sometimes, the sensors in a CPS generate erroneous data, or some data is
missed due to erroneous communication. Therefore, it is challenging to find
trustworthiness of the data.
Value The main challenge of interfacing CPS is to transform the collected raw data
into useful information in order to facilitate the decision making process. The
efficient transformation techniques are needed to provide the accurate value of
the information.
Query Load Generally, the query loads vary and are unpredictable. Due to lack of
flexibility, it is complex to handle these variations. Conti et al. [94] proposed a new
term data vitalization to sense the query load variations. There is still lot of work
needed to optimize the query load in accordance with the needed information and
available resources.
Quality of Service The methodologies to precisely capture and communicate
information and the quality needs of an application should be researched. Due
to increase in scalability and complexity of data, the computational techniques and
their results are very complex to reproduce. Thus, the relationship between data
from information producing systems and the operational systems need to be
studied such that applications quality of service requirements are fulfilled
efficiently.
Knowledge Association The constant sensor data streams are required to be
processed by CPS. These streams need to be efficiently associated with the existing
knowledge [6, 58]. For the complex and uncertain data, the temporal and spatial
correlations must be used with data mining tools to retrieve valuable knowledge
[6]. There is very little work in this direction, and more research is needed for
efficient knowledge association across CPS.
Open CPS Architecture A new open architecture is required, which can be
customized in different situations by different application scenarios. The physical
components are mostly unreliable; tools are needed to build a reliable CPS that
should be resilient to tolerate malicious attacks on the data [57]. For the complex
design of CPS, new modeling and analytical tools are essential to be
utilized [95].

134

6.6

Z. Baloch et al.

Conclusion

In the era of advanced computing, there is an emergent and rapid technological

enhancements in the fields of embedded systems, human computer interaction,
cloud computing, data analysis, cyber-physical systems, and many other computing
aspects. Cyber-physical systems, a new wave of computing, have enabled many
applications that were not practical before. The data from cyber-physical systems is
enormous and growing constantly which poses many challenges in this field. This
chapter discusses the state-of-the-art of the cyber-physical systems from big data
perspective. The data generation sources, cyberspace paradigms, and interfacing
them with the physical and cyber world have been discussed. From data generation
to its storage, different phases of data management for interfacing the two worlds
have also been elaborated. The main issues are efficient storage and processing of
cyber-physical systems big data. The cyber-physical system cloud computing
infrastructure has also been discussed which provides the framework to interface
with the computing devices. Furthermore, the research issues related to big data in
cyber-physical systems have been highlighted. The cyber-physical systems are in
its way of development; therefore, significant issues and challenges must be
addressed by researchers for long-term success.

References
1. Buyya R, Yeo CS, Venugopal S, Broberg J, Brandic I (2009) Cloud computing and emerging
IT platforms: vision, hype, and reality for delivering computing as the 5th utility. Futur Gener
Comput Syst 25(6):599616
2. Wolf W (2009) Cyber-physical systems. Computer 3:8889
3. Baheti R, Gill H (2011) Cyber-physical systems. Impact Control Technol 12:161166
4. Rajkumar R, Lee I, Sha L, Stankovic J (2010) Cyber-physical systems: the next computing
revolution. In: Proceedings of the 47th design automation conference. ACM, pp 731736
5. Shaikh FK, Zeadally S (2015) Mobile sensors in cyber-physical systems. Book Chapter in
cyber physical system design with sensor networking technologies, IET, 2015 (to appear)
6. Wu FJ, Kao YF, Tseng YC (2011) From wireless sensor networks towards cyber physical
systems. Pervasive Mob Comput 7(4):397413
7. Haque AS, Aziz SM, Rahman M (2014) Review of cyber-physical system in healthcare. Int J
Distrib Sens Netw 2014:120
8. Kim JE, Mosse D (2008) Generic framework for design, modeling and simulation of cyber
physical systems. ACM SIGBED Rev 5:1
9. Bloomberg J (2013) The big data long tail. http://www.devx.com/blog/the-big-data-long-tail.
html. Accessed 17 Jan 2015
10. Kambatla K, Kollias G, Kumar V, Grama A (2014) Trends in big data analytics. J Parallel Distr
Com 74(7):25612573
11. Madden S (2012) From databases to big data. IEEE Internet Comput 3:46
12. Rouse M (2015) 3Vs (volume, velocity & variety). http://whatis.techtarget.com/definition/
3Vs. Accessed Apr 2015
13. Hitachi Data Systems (2015) Capitalize on big data. http://www.hds.com/assets/pdf/hitachiwebtech-educational-series-capitalize-on-big-data.pdf. Accessed 20 Mar 2015

6 Interfacing Physical and Cyber Worlds: A Big Data Perspective

135

14. Hurwitz J, Nugent A, Halper F, Kaufman M (2015) Structured data in a big data environment. www.dummies.com/howto/ content/structured-data-in-a-big-data-environment.html.
Accessed 2 Apr 2015
15. Shaikh FK, Zeadally S, Siddiqui F (2013) Energy efficient routing in wireless sensor networks.
In: Next-generation wireless technologies. Springer, London, pp 131157
16. Rouse M (2006) Wireless sensor networks. http://searchdatacenter.techtarget.com/definition/
sensor-network. Accessed 20 Feb 2015
17. Akyildiz IF, Vuran MC (2010) Wireless sensor networks, 4th edn. Wiley, New York
18. Rios LG, Diguez JEAI (2014) Big data infrastructure for analyzing data generated by wireless
sensor networks. In: IEEE international congress on big data (BigData Congress), 2014. IEEE,
pp 816823
19. Jardak C, Riihijarvi J, Oldewurtel F, Mah
onen P (2010) Parallel processing of data from very
large-scale wireless sensor networks. In: Proceedings of the 19th ACM international symposium on high performance distributed computing. ACM, pp 787794
20. Fan T, Zhang X, Gao F (2013) Cloud storage solution for WSN in internet innovation union.
Int J Database Theory Appl 6(3):4958
21. Ahmed K, Gregory M (2011) Integrating wireless sensor networks with cloud computing. In:
Seventh international conference on Mobile Ad-hoc and Sensor Networks (MSN), 2011. IEEE,
pp 364366
22. Kwak H, Lee C, Park H, Moon S (2010) What is Twitter, a social network or a news media?.
In: Proceedings of the 19th international conference on world wide web. ACM, pp 591600
23. Ellison NB, Steinfield C, Lampe C (2007) The benefits of Facebook friends: social capital
and college students use of online social network sites. J Comput-Mediat Commun 12
(4):11431168
24. Gilbert E, Bakhshi S, Chang S, Terveen L (2013) I need to try this?: a statistical overview of
pinterest. In: Proceedings of the SIGCHI conference on human factors in computing systems.
ACM, pp 24272436
25. Shervington M (2015) What is google Plus? A complete user guide. http://www.
martinshervington.com/what-is-google-plus/. Accessed 20 Apr 2015
26. Hochman N, Schwartz R (2012) Visualizing instagram: tracing cultural visual rhythms. In:
Proceedings of the workshop on Social Media Visualization (SocMedVis) in conjunction with
the sixth international AAAI conference on Weblogs and Social Media (ICWSM12), pp 69
27. Watson I, Mullen J, Smith-Spark L (2015) CNN. Nepal earthquake: death toll passes 4,800 as
rescuers face challenges. http://edition.cnn.com/2015/04/28/asia/nepal-earthquake/. Accessed
on 05 May 2015
28. Ravilious K (2015) Nepal quake followed historic pattern. http://www.bbc.com/news/sci
ence-environment-32472310. Accessed on 28 Apr 2015
29. Garg Y, Chatterjee N (2014) Sentiment analysis of Twitter feeds. In: Big data analytics.
Springer International Publishing Switzerland, pp 3352
30. Felemban E, Sheikh AA, Shaikh FK (2014) MMaPFlow: a crowd-sourcing based approach for
mapping mass pedestrian flow. In: Proceedings of the 11th international conference on Mobile
and Ubiquitous Systems: Computing, Networking and Services (MOBIQUITOUS 14)
31. Yousefi S, Mousavi MS, Fathy M (2006) Vehicular ad hoc networks (VANETs): challenges
and perspectives. In: 6th international conference on ITS telecommunications proceedings,
2006. IEEE, pp 761766
32. Ali F, Shaikh FK, Ansari AQ, Mahoto NA, Felemban E (2015) Comparative analysis of
VANET routing protocols- on placement of road side units. Int J Wirel Pers Commun,
Springer, pp 114, 2015. doi:10.1007/s11277-015-2745-z
33. Zhang Y, Zhao J, Cao G (2010) Roadcast: a popularity aware content sharing scheme in
vanets. ACM SIGMOBILE Mobile Comput Commun Rev 13(4):114
34. Sutariya D, Pradhan SN (2010) Data dissemination techniques in vehicular ad hoc network. Int
J Comput Appl 8(10):3539

136

Z. Baloch et al.

35. Dubey BB, Chauhan N, Kumar P (2010) A survey on data dissemination techniques used in
VANETs. Int J Comput Appl 10(7):510
36. Talpur A, Baloch N, Bohra N, Shaikh FK, Felemban E (2014) Analyzing the impact of body
postures and power on communication in WBAN. Procedia Comput Sci 32:894899
37. Khelil A, Shaikh FK, Sheikh AA, Felemban E, Bojan H (2014) DigiAID: a wearable health
platform for automated self-tagging in emergency cases, In: 4th international conference on
wireless Mobile Communication and Healthcare (Mobihealth), 2014 EAI, pp 296,299
38. Aziz Z, Qureshi UM, Shaikh FK, Bohra N, Khelil A, Felemban E (2015) Revisiting routing in
wireless body area networks. In: Emerging communication technologies based on wireless
sensor networks: current research and future applications. CRC Press (to appear)
39. TDWI Best Practices Report (2015) Managing big data. http://tdwi.org/research/2013/10/
tdwi-best-practices-report-managing-big-data.aspx?tcpage0. Accessed 01 Mar 2015
40. Dinh HT, Lee C, Niyato D, Wang P (2013) A survey of mobile cloud computing: architecture,
applications, and approaches. Wirel Commun Mob Comput 13(18):15871611
41. Agrawal D, Das S, El Abbadi A (2010) Big data and cloud computing: new wine or just new
bottles? Proc VLDB Endowment 3(12):16471648
42. Elazhary H (2014) Cloud computing for big data. MAGNT Res Rep 2(4):135144
43. Gartner IT glossary (2013) Cloud computing. http://www.gartner.com/it-glossary/cloud-com
puting. Accessed 01 Apr 2015
44. Rodero-Merino L, Vaquero LM, Gil V, Galan F, Fontan J, Montero RS, Llorente IM (2010)
From infrastructure delivery to service management in clouds. Futur Gener Comput Syst 26
(8):12261240
45. Patidar S, Rane D, Jain P (2012) A survey paper on cloud computing. In: Second international
conference on Advanced Computing & Communication Technologies (ACCT), 2012. IEEE,
pp 394398
46. Khajeh-Hosseini A, Sommerville I, Bogaerts J, Teregowda P (2011) Decision support tools for
cloud migration in the enterprise. In: IEEE international conference on Cloud Computing
(CLOUD), 2011. IEEE, pp 541548
47. Hashem IAT, Yaqoob I, Anuar NB, Mokhtar S, Gani A, Khan SU (2015) The rise of big data
on cloud computing: review and open research issues. Inf Syst 47:98115
48. Abadi DJ (2009) Data management in the cloud: limitations and opportunities. IEEE Data Eng
Bull 32(1):312
49. Das S, Agrawal D, El Abbadi A (2009) Elastras: an elastic transactional data store in the cloud.
USENIX HotCloud 2:7
50. Valduriez P (2009) Shared-memory architecture. In: Encyclopedia of database systems.
Springer US, New York, pp 26382638
51. Jill Dyche (2015) Data as a service explained and defined. http://searchdatamanagement.
techtarget.com/answer/Data-as-a-service-explained-and-defined Accessed on 20 Mar 2015
52. Demirkan H, Delen D (2013) Leveraging the capabilities of service-oriented decision support
systems: putting analytics and big data in cloud. Decis Support Syst 55(1):412421
53. Mathiprakasam M (2015) The road to analytics as a service. http://www.forbes.com/sites/
oracle/2014/09/26/the-road-to-analytics-as-a-service/. Accessed on 20 Mar 2015
54. Poovendran R (2010) Cyberphysical systems: close encounters between two parallel worlds
[point of view]. Proc IEEE 98(8):13631366
55. Shaikh FK, Zeadally S, Exposito E (2015) Enabling technologies for green internet of things.
IEEE Syst J 99:112
56. Rajhans A, Cheng SW, Schmerl B, Garlan D, Krogh BH, Agbi C, Bhave A (2009) An
architectural approach to the design and analysis of cyber-physical systems. Electronic Communications of the EASST, 21:110
57. CPS Steering Group (2008) Cyber-physical systems executive summary. CPS Summit
58. Simmon E, Kim KS, Subrahmanian E, Lee R, de Vaulx F, Murakami Y, Zettsu K, Sriram RD
(2013) A vision of cyber-physical cloud computing for smart networked systems. NIST,
Gaithersburg

6 Interfacing Physical and Cyber Worlds: A Big Data Perspective

137

59. Mokashi M, Alvi AS (2013) Data management in wireless sensor network: a survey. Int J Adv
Res Comput Commun Eng 2:13801383
60. Padgavankar MH, Gupta SR (2014) Big data storage and challenges. Int J Comput Sci Inf
Technol 5:2
61. Sathe S, Papaioannou TG, Jeung H, Aberer K (2013) A survey of model-based sensor data
acquisition and management. In: Managing and mining sensor data. Springer US, New York,
pp 950
62. Aggarwal CC (2013) Managing and mining sensor data. Springer Science & Business Media,
New York
63. Chapman AD (2005) Principles and methods of data cleaning. GBIF, Copenhagen
64. Jeffery SR, Alonso G, Franklin MJ, Hong W, Widom J (2006) A pipelined framework for
online cleaning of sensor data streams. IEEE, p 140
65. Elmenreich W (2002) Sensor fusion in time-triggered systems, Ph.D. thesis, Faculty of
Informatics at the Vienna University of Technology, Austria. http://www.vmars.tuwien.ac.
at/~wilfried/papers/elmenreich_Dissertation_sensorFusionInTimeTriggeredSystems.pdf
66. Hall David L, Llinas J (1997) An introduction to multisensor data fusion. Proc IEEE 85
(1):623
67. Castanedo F (2013) A review of data fusion techniques. Sci World J 2013:119
68. Kimura N, Latifi S (2005) A survey on data compression in wireless sensor networks. In:
International conference on Information Technology: Coding and Computing (ITCC), 2005,
vol. 2. IEEE, pp 813
69. Marcelloni F, Vecchio M (2008) A simple algorithm for data compression in wireless sensor
networks. Commun Lett IEEE 12(6):411413
70. Agrawal R, Faloutsos C, Swami A (1993) Efficient similarity search in sequence databases.
Springer, Berlin/Heidelberg, pp 6984
71. Gandhi S, Nath S, Suri S, Liu J (2009) Gamps: compressing multi sensor data by grouping and
amplitude scaling. In: Proceedings of the 2009 ACM SIGMOD international conference on
management of data. ACM, pp 771784
72. Wang L, Deshpande A (2008) Predictive modeling-based data collection in wireless sensor
networks. In: Wireless sensor networks. Springer, Berlin/Heidelberg, pp 3451
73. Arion A, Jeung H, Aberer K (2011) Efficiently maintaining distributed model-based views on
real-time data streams. In: Global Telecommunications Conference (GLOBECOM 2011).
IEEE, pp 16
74. Xing K, Cheng X, Li J (2005) Location-centric storage for sensor networks. In: IEEE
international conference on mobile adhoc and sensor systems conference. IEEE, p 10
75. Petit L, Nafaa A, Jurdak R (2009) Historical data storage for large scale sensor networks. In:
Proceedings of the 5th French-speaking conference on mobility and ubiquity computing.
ACM, pp 4552
76. Diao Y, Ganesan D, Mathur G, Shenoy PJ (2007) Rethinking data management for storagecentric sensor networks. In: CIDR, vol. 7, pp 2231
77. Dutta P, Culler DE, Shenker S (2007) Procrastination might lead to a longer and more useful
life. In: The sixth workshop on Hot Topics in Networks (HotNets-VI) pp 17
78. Deshpande A, Madden S (2006) MauveDB: supporting model-based user views in database
systems. In: Proceedings of the 2006 ACM SIGMOD international conference on management
of data. ACM, pp 7384
79. Kanagal B, Deshpande A (2008) Online filtering, smoothing and probabilistic modeling of
streaming data. In: IEEE 24th international conference on Data Engineering, ICDE 2008.
IEEE, pp 11601169
80. Bhattacharya A, Meka A, Singh AK (2007) Mist: distributed indexing and querying in sensor
networks using statistical models. In: Proceedings of the 33rd international conference on very
large data bases. VLDB Endowment, pp 854865
81. Cugola G, Margara A (2012) Processing flows of information: from data stream to complex
event processing. ACM Comput Surv (CSUR) 44(3):15

138

Z. Baloch et al.

82. Babcock B, Babu S, Datar M, Motwani R, Widom J (2002) Models and issues in data stream
systems. In: Proceedings of the twenty-first ACM SIGMOD-SIGACT-SIGART symposium on
principles of database systems. ACM, pp 116
83. Luckham D (2002) The power of events, vol 204. Addison-Wesley, Reading
zsu MT (2003) Issues in data stream management. ACM Sigmod Rec 32(2):514
84. Golab L, O
85. Dunkel J (2009) On complex event processing for sensor networks. In: International symposium on autonomous decentralized systems, 2009. ISADS09. IEEE, pp 16
86. Miller S (2013) Big data analytics. Podcasts at Singapore Management University, Available
at: http://ink.library.smu.edu.sg/podcasts/8
87. Big Data in the Cloud Converging Technologies-Intel (2014) http://www.intel.com/content/
www/us/en/big-data/big-data-cloud-technologies-brief.html. Accessed on Apr 2015
88. Tang LA, Yu X, Kim S, Han J, Peng WC, Sun Y, Gonzalez H, Seith S (2012)
Multidimensional analysis of atypical events in cyber-physical data. In: IEEE 28th international conference on Data Engineering (ICDE), 2012. IEEE, pp 10251036
89. Tang LA, Yu X, Kim S, Han J, Peng WC, Sun Y, Leung A, La Porta T (2012)
Multidimensional sensor data analysis in cyber-physical system: an atypical cube approach.
Int J Distrib Sens Netw 2012:119
90. Yu X, Tang LA, Han J (2009) Filtering and refinement: a two-stage approach for efficient and
effective anomaly detection. In: Ninth IEEE international conference on Data Mining, 2009.
ICDM09. IEEE, pp 617626
91. Tang LA, Yu X, Kim S, Han J, Hung CC, Peng WC (2010) Tru-alarm: trustworthiness analysis
of sensor networks in cyber-physical systems. In: IEEE 10th international conference on Data
Mining (ICDM), 2010. IEEE, pp 10791084
92. Xia F, Kong X, Xu Z (2011) Cyber-physical control over wireless sensor and actuator
networks with packet loss. In: Wireless networking based control. Springer, New York, pp
85102
93. Thouin F, Thommes R, Coates MJ (2006) Optimal actuation strategies for sensor/actuator
networks. In: 3rd annual international conference on mobile and ubiquitous systems: networking & services, 2006. IEEE, pp 18
94. Conti M, Das SK, Bisdikian C, Kumar M, Ni LM, Passarella A, Roussos G, Tr
oster G,
Tsudik G, Zambonelli F (2012) Looking ahead in pervasive computing: challenges and
opportunities in the era of cyberphysical convergence. Pervasive Mob Comput 8(1):221
95. Guturu P, Bhargava B (2011) Cyber-physical systems: a confluence of cutting edge technological streams. International conference on advances in computing and communication

Chapter 7

Distributed Platforms and Cloud Services:

Enabling Machine Learning for Big Data
Daniel Pop, Gabriel Iuhasz, and Dana Petcu

Abstract Applying popular machine learning algorithms to large amounts of data

has raised new challenges for machine learning practitioners. Traditional libraries
do not support properly the processing of huge data sets, so the new approaches are
needed. Using modern distributed computing paradigms, such as MapReduce or
in-memory processing, novel machine learning libraries have been developed. At
the same time, the advance of cloud computing in the past 10 years could not be
ignored by the machine learning community. Thus, a rise of cloud-based platforms
has been of significance. This chapter aims at presenting an overview of novel
platforms, libraries, and cloud services that can be used by data scientists to extract
knowledge from unstructured and semi-structured, large data sets. The overview
covers several popular packages to enable distributed computing in popular
machine learning environments, distributed platforms for machine learning, and
cloud services for machine learning, known as machine-learning-as-a-service
approach. We also provide a number of recommendations for data scientists
when considering machine learning approach for their problem.
Keywords Machine learning Data mining Cloud computing Big data Data
scientist Distributed computing Distributed platforms

7.1

Introduction

Analyzing large amounts of data collected by companies, industries, and scientific

domains is becoming increasingly important for all impacted domains. The data to
be analyzed is no longer restricted to sensor data and classical databases, but it often
includes textual documents and Web pages (text mining, Web mining), spatial data,
multimedia data, or graph-like data (e.g., molecule configuration and social
networks).

D. Pop (*) G. Iuhasz D. Petcu

Institute e-Austria Timisoara, West University of Timisoara, Blvd. Vasile Parvan, nr. 4,
300223 Timioara, Romania
e-mail: daniel.pop@e-uvt.ro
Springer International Publishing Switzerland 2016
Z. Mahmood (ed.), Data Science and Big Data Computing,
DOI 10.1007/978-3-319-31861-5_7

139

140

D. Pop et al.

Although, for more than two decades, parallel database products such as
Teradata, Oracle, and Netezza have provided means to realize a parallel implementation of machine learning algorithms, expressing these algorithms in SQL
code is a complex and difficult-to-maintain task. On the other side, large-scale
installations of these products are expensive. Another reason for moving away
from relational databases is the exponential growth of the unstructured data (e.g.,
audio and video) and semi-structured data (e.g., Web traffic data, social media
content, sensor-generated data) in recent years. The needs of data science practitioners with respect to data analysis tools vary greatly across different domains,
from medical statistics and bioinformatics to social network analysis or even in
physics. This diversity is equally important for the advancement of machine
learning tools and platforms. Consequently, in the past decade, researchers
moved from the parallelization of machine learning algorithms and support in
relational databases toward the design and implementation on top of novel
distributed storage (e.g., NoSQL data stores, distributed file systems) and
processing paradigms (e.g., MapReduce). From the business perspective, Software-as-a-Service (SaaS) model opened up new opportunities for machine learning providers, who moved the stand-alone tools toward cloud-based machine
learning services.
In this chapter, we survey how distributed storage and processing platforms help
data scientists to process large, heterogeneous sets of data. The tools, frameworks,
and services included in this chapter share a common characteristic: all run on top
of distributed platforms. Thus, parallelization of machine learning algorithms,
either using multiple-core CPU or GPU, was not included here. The reader is
referred to [42], a recent comprehensive study covering that topic. We also avoided
commercial solution providers, small or big players, since their offerings are either
based on distributed open-source packages or they do not disclose the implementation details.
In the first section, we briefly introduce the reader to the machine learning field,
describing and classifying the types of problems and overviewing the challenges
of applying traditional algorithms to large, unstructured data sets. The first category of tools considered in this survey covers tools, packages, and libraries that
enable data scientists to use traditional environments for data analysis such as R
Systems, Python, or statistics applications, in order to deal with large data sets. We
survey, next, the distributed platforms for big data processing, either based on
Apache Hadoop or Spark, as well as platforms specifically designed for distributed
machine learning. We also include a section on scalable machine learning services
delivered using Software-as-a-Service business model since they offer easy-to-use,
user-friendly graphical interfaces supporting users in quickly getting and
deploying models. The last section of the chapter summarizes our findings and
provides readers with a collection of best practices in applying machine learning
algorithms.

7 Distributed Platforms and Cloud Services: Enabling Machine Learning for Big Data

7.2

141

Machine Learning for Data Science

The broadest and simplest definition of machine learning is that it is a collection of

computational methods that use experience, i.e., information available to the
system to improve performance or to make predictions [28]. This information
usually takes the form of electronic records collected and made available for
analytical purposes. These records can take the form of pre-labeled training sets
(usually by a human operator although this is not always the case). Another
important source of data is that resulting from direct interactions with a given
environment, either virtual, such as software interactions, network data, etc., or
relying on real-world natural scenarios, such as weather phenomena, water level,
etc. Data quality and quantity are extremely important in order to obtain an
acceptable learned model. Machine learning relies on data-driven methods that
combine fundamental concepts in the field of computer science with optimization,
probability, and statistics [28].
There is a wide array of applications to which machine learning can and is being
applied, such as taming (text mining and document classification), spam detection,
keyword extraction, emotion extraction, natural language processing (NPL),
unstructured text understanding, morphological analysis, speech synthesis and
recognition, optical character recognition (OCR), computational biology, face
detection, image segmentation, image recognition, fraud detection, network intrusion detection, board and video games, navigation in self-driving vehicles, planning, medical diagnosis, recommendation systems, or search engines. In all these
applications, we can identify several types of learning-related issues, which are:
Classification to assign each item from a data set to a specific category, e.g.,
given a document, assign a domain (history, biology, mathematics) to which it
belongs.
Regression and time series analysis to predict a real value for each item, e.g.,
future stock market values, rainfall runoff, etc.
Ranking to return an ordered set of features based on some user-defined
criterion (e.g., Web search).
Dimensionality reduction (feature selection) to use for transforming initial
large feature spaces into a lower-dimensional representation so that it preserves
the properties of the initial representation.
Clustering to group items based on some predefined distance measure. It is
usually used on very large data sets. In sociology, it can be used to group
individuals into communities [30].
Anomaly detection to conduct observation or series of observations which do
not resemble any pattern or data item in a data set [6, 37].
The most common classification techniques are called linear classifiers. In this
case, classification is expressed in the form of a linear function. This function
assigns scores to each possible category. Among the linear classifiers, we have
linear regression, perceptron, and support vector machines (SVM) [28]. Another

142

D. Pop et al.

form of classification is based on kernel estimation in the form of the k-nearest

neighbor (k-NN) algorithms. Decision trees such as C4.5 [28] are also used for this
type of problem and are based on information theory (difference in entropy), which
is used as the splitting criterion. Ensemble meta-algorithm-based techniques such
as AdaBoost [28] are also used although they have questionable performance on
noisy data sets. Some methods such as Classification and Regression Tree (CART)
algorithm can be used for both regression and classification problems. Nave Bayes,
for example, can also be used for both types of problems.
Clustering algorithms are largely split according to their particular definition of
what cluster model they use. Connectivity models (hierarchical clustering) are
based on distance connectivity. Centroid models, such as k-means (k-M), represent
each cluster with a single mean vector. Density models consider clusters as
connected dense regions from the data space. DBSCAN and OPTICS [28] are
two algorithms using this model. Statistical distribution-based models are also used.
Anomaly detection is a special case of either classification or clustering; thus, it
uses mostly the same algorithms and methodologies. Feature selections main goal
is the reduction of the amount of recourses required to analyze big data set. They are
extremely useful when no domain expert is available that could help in the
reduction of the dimensionality of the available data. There are a number of general
dimensionality reduction techniques such as principal component analysis (PCA),
kernel PCA, multilinear PCA, and wrapping methods [25].
In machine learning, there are different types of training scenarios [28]. Arguably
the most widely used type of training is called supervised learning. In this scenario,
the learner receives a set of labeled data for training and validation. The learned
prediction model can be then applied to a larger data set and identify all unseen data
points. This type of learning is used for classification and regression (time series
analysis). Supervised methods rely on the availability and accuracy of labeled data
sets. In unsupervised learning, the learner receives unlabeled data that it has to
group based on a distance measurement. In some scenarios, labeled data is
extremely hard to come by; thus, training a classification model is often unfeasible.
This type of learning is used for clustering, anomaly detections (a type of clustering), and dimensionality reduction.
In some cases, labeled data is only a small fraction of the overall training data
set. This is called semi-supervised learning. The idea is that the distribution of
unlabeled data can help the learner achieve a much better performance [11].
In reinforcement learning, the training is done using an evaluation function.
This means that training and testing are much more interlaced than in other
learning scenarios. The performance of an algorithm in a problem environment
is continuously evaluated through the monitoring and evaluation of its performance. Favorable outcomes are rewarded, while unfavorable ones are punished.
Reinforcement learning is used in genetic algorithm, neural networks, etc. Online
learning is used when data is available in a sequential way. This means that the
mapping between data sets and labels is established each time a new data point is
received.

7 Distributed Platforms and Cloud Services: Enabling Machine Learning for Big Data

143

Due to the popularity of data analytics, machine learning techniques are being
pursued by teams with complementary skills across very different businesses
(finance, telecommunications, life sciences, etc.). This section aims to classify the
diversity of groups of interests with respect to machine for big data. We must state
upfront that there is no clear line between these perspectives, as competencies and
expectations blur the edges and multidisciplinary teams are put in place to tackle
complex scenarios. Some of the groups are:
Data scientists and machine learning practitioners: One way of approaching the
problem is from the data scientists perspective. Statisticians and data scientists
are now facing data set size explosion; thus, coping with large-size data sets is a
must. These are users with strong mathematical background, proficient in statistics and mathematical software applications, such as R, Octave, MATLAB,
Mathematica, Python, SAS Studio, or IBMs SPSS, but less experienced in
coping with data sets of large dimensions, distributed computing, or software
development. Their expectation is to easily reuse the algorithms already available in their preferred language and be able to run them against large data sets on
distributed architectures (on-premise or cloud based). A later section in this
chapter entitled Distributed and Cloud-Based Execution Support in Popular
Machine Learning Tools overviews packages and tools available for this
purpose.
Software engineers and developers: Teams of software engineers often face
client requirements asking for the transition from available (large) data warehouse to actionable knowledge. These are users with a vast experience in
software development, skilled programmers in general-purpose programming
languages, and they speak parallel and distributed computing. Deep mathematics and statistics are not necessarily their preferred playground, as they
expect tools and libraries to enable them to integrate advanced ML algorithms
in their systems and thus quickly get actionable results. They need fast, easy-tocustomize (less number of parameters), and easy-to-integrate algorithms that run
on distributed architectures and are able to fetch data from large data repositories. Tools addressing these requirements are discussed in a later section in this
chapter entitled Distributed Machine Learning Platforms.
Domain experts: Domain experts (financial, telecommunications, physics,
astronomy, biotechnologies, etc.) know their data best, but they are less experienced in ML algorithms and software tools. Ideally, they need off-the-shelf
software applications, easy to install and use, or cloud-based Software-as-aService solutions allowing them to get insights on their data and produce reports
and executable models for further usage. A later section on Machine Learning
as a Service presents several machine learning services providers.
The dynamic of natural, social, and economic systems raises new challenges for
data scientists, such as:

144

D. Pop et al.

Massive data sets. Data sets are growing faster, being common now to reach
numbers of 100 TB or more. The Sloan Digital Sky Survey occupies 5 TB of
storage, the Common Crawl Web corpus is 81 TB in size, and the 1000 Genomes
Project requires 200 TB of space, just to name a few.
Large models. Massive data sets need large models to be learned. Some deep
neural networks are comprised of more than ten layers with more than a billion
parameters [24, 25], collaborative filtering for video recommendation on Netflix
comprises 110 billion parameters, and multitask regression model for simplest
whole-genome analysis may reach 1 billion parameters as well.
Inadequate ML tools and libraries. Traditional ML algorithms used for decades
(k-means, logistic regression, decision trees, Nave Bayes) were not designed for
handling large data sets and huge models; they were not developed for parallel/
distributed environments.
Operationalization of predictive models. Operationalize refers to integrate
predictive models into automated decision-making systems and processes on a
large scale in order to deliver predictions to end users, who will ultimately
benefit from them. Integrating these models into multiple platforms (Web,
stand-alone, mobile) across different business units requires a high degree of
customization, which slows deployment, drives up costs, and limits scalability.
Lack of clear contracts. More recently, terms such as Analytics as a Service
(AaaS) and Big Data as a Service (BDaaS) are becoming popular. They comprise services for data analysis similarly as IaaS offers computing resources.
Unfortunately, the analytics services still lack well-defined service-level agreements available for IaaS because it is difficult to measure quality and reliability
of results and input data, to provide promises on execution times and guarantees
on methods for analyzing the data. Therefore, there are fundamental gaps on
tools to assist service providers and clients to perform these tasks and facilitate
the definition of contracts for both parties [2].
Inadequate staffing. Market research shows that inadequate staffing and skills,
lack of business support, and problems with analytics software are some of the
barriers faced by corporations when performing analytics [36].
In the next three sections, we discuss various machine learning tools.

7.3

Distributed and Cloud-Based Execution Support

in Popular Machine Learning Tools

Annual Nuggets survey [23] shows that R, Python, SQL, and SAS have been rated
the preferred languages of choice for the past 3 years. One of the early trends
matching cloud computing and data analysis was, around 2010s, the provision of
virtual machine images (VMI) for these popular systems (R, Octave, or Maple)
integrated within public cloud service providers, such as Amazon Web Services or
Rackspace. After several proofs of concept were successfully built, such as

7 Distributed Platforms and Cloud Services: Enabling Machine Learning for Big Data

145

Table 7.1 Distributed processing and storage

Environment
R

Python

Package
RHadoop
RHIPE
Segue for R
RHive
Snow
H5
Pbd*
pyDoop
Anaconda
IPython.
parallel

Distributed processing
support
Hadoop
Hadoop
Amazon Elastic MapReduce
HIVE
Socket-based, MPI, PVM

MPI
Hadoop
Distributed and GPU
Distributed and parallel

Distributed file system

access
HDFS
HDFS

HIVE

HDF5
NetCDF
HDFS
HDFS, HDF5

Cloudnumbers,1 CloudStat,2 Opani,3 and Revolution R Enterprise,4 the practice

today is to provide VMI through the public cloud providers marketplaces, such as
Amazon Marketplace. One can find Amazon Machine Images (AMI), via the
marketplace, for all the popular mathematical and statistics environments. Examples include Predictive Analytics Framework and Data Science Toolbox5 that
support both Python and R, BF Accelerated Scientific Compute for R with accelerated math libraries for boosted performance, or SAS University Edition for SAS
Studio.
Much more effort has been invested in the development of plug-ins for the most
popular machine learning platforms to allow data scientists to easily create and run
time-consuming jobs over clusters of computers. This approach allows ML practitioners to reuse their existing code and adapt it for large data set processing, into the
same environment they used for prototyping. It also leverages existing infrastructure (grids, clusters) for large-scale distributed computation and data storage.
Table 7.1 synthesizes available plug-ins for distributed storage and processing for
the most popular languages of big data: R and Python.
Since R is the preferred option among machine learning practitioners, several
packages were developed in order to enable big data processing within R, most of
them being available under CRAN6 package Web page. These R extensions make
possible to distribute the computational workload on different types of clusters,
while accessing data from distributed file systems. First example is the RHadoop
[33], a collection of five R packages, that enables R users to run MapReduce jobs on
1

http://cloudnumbers.com
http://cs.croakun.com
3
http://opani.com
4
www.revolutionanalytics.com
5
http://datasciencetoolbox.org
6
http://cran.r-project.org/web/packages/available_packages_by_name.html
2

146

D. Pop et al.

Hadoop by writing R functions for mapping and reducing. Similarly, RHIPE7 is

another R package that brings MapReduce framework to R practitioners, providing
seamless access to Hadoop cluster from within R environment. Using specific R
functions, programmers are able to launch MapReduce jobs on the Hadoop cluster,
with results being easily retrieved from HDFS. Segue8 for R project makes it easier
to execute MapReduce jobs from within the R environment on elastic clusters at
Amazon Elastic MapReduce,9 but lacks support for handling large data sets. RHive
is an extension enabling distributed computing via HIVE in R, by a seamless
integration between HQL (Hive Query Language) and R objects and functions.
Snow (Simple Network of Workstations) [41] and its variants (snowfall, snowFT,
doSnow) implement a framework that is able to express an important class of
parallel computations and is easy to use within an interactive environment like
R. It supports three types of clusters: socket based, MPI, and PVM. Support for
manipulating large data sets in R is available in H5 plug-in, which provides an
interface to the HDF5 API through S4 objects, supporting fast storage and retrieval
of R objects to/from binary files in a language-independent format. The pbd*
(pbdBASE, pbbMPI, pbdNCDF4, pbdSLAP, etc.) series is a collection of R packages for programming with big data, enabling MPI distributed execution, NetCDF
file system access, or tools for scalable linear algebra.
As far as Python is concerned, we should start by mentioning pyDoop,10 a
Python MapReduce and HDFS API for Hadoop [26]. Anaconda11 is a free, scalable
Python distribution for large-scale data analytics and scientific computing. It is a
collection of Python packages (NumPy, SciPy, Pandas, IPython, Matplotlib,
Numba, Blaze, Bokeh) that enables fast large data set access, GPU computation,
access to distributed implementations of ML algorithms, and more. IPython.parallel12 provides a sophisticated and powerful architecture for parallel and distributed
computing [14] that enables IPython to support many different styles of parallelism
including single program multiple data (SPMD), multiple program multiple data
(MPMD), message passing using MPI, data parallel, and others. In a tutorial at
PyCon 2013, Grisel [15] presented how scikit-learn [32], a popular open-source
library for machine learning in Python, can be used to perform distributed machine
learning algorithms on a cheap Amazon EC2 cluster using IPython.parallel and
StarCluster.13 We should note as well that most of the libraries and frameworks
considered in the next sections offer Python language bindings, but we choose not
to include them in this section.

http://www.stat.purdue.edu/~sguha/rhipe/doc/html/index.html
http://code.google.com/p/segue
9
http://aws.amazon.com/elasticmapreduce
10
https://github.com/crs4/pydoop
11
https://store.continuum.io/cshop/anaconda
12
http://ipython.org/ipython-doc/dev/parallel/
13
http://star.mit.edu/cluster/
8

7 Distributed Platforms and Cloud Services: Enabling Machine Learning for Big Data

147

Other mathematical and statistics environments have seen similar interest in

embracing big data processing. For example, HadoopLink14 is a package that
allows MapReduce programs being implemented in Mathematica and to run them
on a Hadoop cluster. It looks more like a proof of concept (PoC), being stalled since
2013. MATLAB has its Parallel Computing Toolbox which extends the capabilities
of MATLAB MapReduce and Datastore15 in order to run big data application.
MATLAB Distributed Computing Server also supports running parallel
MapReduce programs on Hadoop clusters.16
There are extensions to traditional machine learning libraries that enable execution on top of Hadoop or Spark clusters. Weka [16], one of the most popular
libraries for data mining, supports both Hadoop and Spark execution through
Weka Hadoop integration [17]. There is also a commercial distribution, Pentaho
[34], that offers a complete solution for big data analytics, supporting all phases of
an analytics process from preprocessing to advanced data exploration and visualization, which uses distributed Weka execution for analytics. Another example is
the KNIMEs [4] big data extension,17 which enables the access to Hadoop via
Hive. RapidMiner [20] has Radoop18 that enables the deployment of workflows on
Hadoop.

7.4

Distributed Machine Learning Platforms

After distributed processing and storage environments (Hadoop, Dryad, MPI)

reached an acceptable level of maturity, they became an increasingly appealing
foundation for the design and implementation of new platforms for machine
learning algorithms. These provide users out-of-the-box algorithms, which are
run in parallel mode over a cluster of (commodity) computers. These solutions do
not use statistics, or mathematics software packages, rather they offer selfcontained, optimized implementations in general-purpose programming languages
(C/C++, Java) of state-of-the-art ML methods and algorithms. This section focuses
on ML platforms specifically designed for distributed and scalable computing.
Table 7.2 summarizes some recent platforms.
The IBM Research Lab has been one of the pioneers who invested in distributed
machine learning frameworks. Nimble [12] and SystemML [13] are two high-level
conceptual frameworks supporting the definition of ML algorithms and their execution on Hadoop clusters. Nimble, a sequel to IBMs Parallel Machine Learning
Toolbox [31], features a multilayered framework enabling developers to express

14

https://github.com/shadanan/HadoopLink
http://www.mathworks.com/help/matlab/large-files-and-big-data.html
16
http://www.mathworks.com/help/distcomp/big-data.html
17
https://www.knime.org/knime-big-data-extension
18
https://rapidminer.com/products/radoop/
15

148

D. Pop et al.

Table 7.2 Distributed ML frameworks

Name
Petuum

Jubatus

License
Open
source
(Sailing
Lab)
LGPL v2.1

ML Problem
DL, CLS, CLU,
RGR, MET, TOP

Distr. Env.
Clusters or Amazon EC2, Google
CE

Comm.
Medium

Lang.
C++

Zookeeper

Medium

C++

Spark

Large

Hadoop, Spark,
H2O

Medium

Scala,
Java
Java

Hadoop, Spark

Low

Java

Storm

Low

Java

Hadoop

Medium

Java

Hadoop, Spark,
MPI
Hadoop

High

C++

Ms-PL

CLS, RGR, ANO,

CLU, REC, Graph
RGR, CLS, REC,
CLU
Collaborative filtering, CLS, CLU, DR,
TOP
REC, CLS, RGR,
CLU
CLS, RGR, CLU,
DR
DL, RGR, CLS,
CLU, DR
CLU, CLS, RGR,
DL, REC
CLS, RGR, CLU

Medium

C++

Apache 2.0

DL

Medium

MIT
License
Apache 2.0

CLS

Hadoop, Spark,
AWS, Akka
Julia

Java,
Scala
Julia

Flink Hadoop

Low

Scala

Dryad, Hadoop
YARN
Hadoop
Hadoop

None

C#,
LINQ
Java
DML

MLlib
(MLBase)
Mahout

Apache 2.0

Oryx

Apache 2.0

Trident-ML

Apache 2.0

H2O

Apache 2.0

GraphLab
Create
Vowpal
Wabbit
Deeplearning4J

Apache 2.0

Julias MLBase
Flink-ML

Apache 2.0

CLS, RGR, CLU,

REC

DryadLINQ
Nimble
SystemML

NA
NA

CLU, FRQ, ANO,

RGR, PageRank

None
None

ANO anomaly detection, CLS classification, CLU clustering, DL deep learning, DR dimensionality
reduction, FRQ frequent pattern, MET metrics learning, REC recommendation, RGR regression,
TOP topic modeling

their ML algorithms as tasks, which are then passed to the next layer, an
architecture-independent layer, composed of one queue of DAGs of tasks, plus
worker thread pool that unfolds this queue. The bottom layer is an architecturedependent layer that translates the generic entities from the upper layer into various
runtimes, the only distributed environment supported within the proof of concept
being Hadoop alone. The layered architecture of the system hides the low-level
control and choreography details of most of the distributed and parallel programming paradigms (MR, MPI, etc.), it allows developers to compose parallel ML
algorithms using reusable (serial and parallel) building blocks, but also it enables
portability and scalability. SystemML proposes an R-like language (declarative

7 Distributed Platforms and Cloud Services: Enabling Machine Learning for Big Data

149

machine learning language) that includes linear algebra primitives and shows how
it can be optimized and compiled down to MapReduce. Authors report an extensive
performance evaluation on three ML algorithms (group nonnegative matrix factorization, linear regression, PageRank) on varying data and Hadoop cluster sizes.
These two systems are purely research endeavors, and they are not available to the
community.
Most of the frameworks rely on Hadoops MapReduce paradigm and the underlying distributed file storage system (HDFS) because it simplifies the design and
implementation of large-scale data processing systems. Only a few frameworks
(e.g., Jubatus, Petuum, GraphLab Create) have tried to propose novel distributed
paradigms, customized to machine learning for big data, in order to optimize the
complex, time-consuming ML algorithms.
Recognizing the limitations and difficulties of adapting general-purpose distributed frameworks (Hadoop, MPI, Dryad, etc.) to ML problems, a team at CMU
under E. P. Xing lead designed a new framework for distributed machine learning
able to handle massive data sets and cope with big models. Petuum19 (from
Perpetuum Mobile) [8, 43] takes advantage of data correlation, staleness, and
other statistical properties to maximize the performance for ML algorithms, realized through core features such as a distributed Parameter Server and a distributed
Scheduler (STRADS). It may run either on-premise clusters or on cloud computing
resources like Amazon EC2 or Google Compute Engine (GCE).
Jubatus20 is a distributed computing framework specifically designed for online
machine learning on big data. A loose model sharing architecture allows it to
efficiently train and share machine learning models by defining three fundamental
operations, viz., update, mix, and analyze [19]. Comparing to Apache Mahout,
Jubatus offers stream processing and online learning, which means that the model is
continuously updated with each data sample that is coming in, by fast, not memoryintensive algorithms. It requires no data storage nor sharing, only model mixing. In
order to efficiently support online learning, Jubatus operates updates on local
models and then each server transmits its model difference that are merged and
distributed back to all servers. The mixed model improves gradually thanks to all
servers work.
GraphLab Create,21 formerly GraphLab project [27], is a framework for
machine learning that expresses asynchronous, dynamic, graph-parallel computation while ensuring data consistency and achieving a high degree of parallel
performance, in both shared-memory and distributed settings. It is an end-to-end
platform enabling data scientists to easily create intelligent apps at scale, from
cleaning the data, developing features, training a model, and creating and
maintaining a predictive service. It runs on distributed Hadoop/YARN clusters, as

19

http://petuum.org
http://jubat.us
21
https://dato.com/products/create
20

150

D. Pop et al.

well on local machine or on EC2, and it exposes a Python interface for an easy
accessibility.
Apache Mahout [29] is a scalable machine learning framework built on top of
Hadoop that features a rich collection of distributed implementations of machine
learning and data mining algorithms. Although initially created on top of Hadoop,
starting with version 0.10, it supports additional execution engines such as Spark
and H20, while Flink22 is a project in progress. The same release introduces
Mahout-Samsara, a new math environment created to enable users to develop
their own extensions, using Scala language, based on general linear algebra and
statistical operations. Mahout-Samsara comes with an interactive shell that runs
distributed operations on a Spark cluster. This makes prototyping or task submission much easier and allows users to customize algorithms with a whole new degree
of freedom.
H2O23 and MLlib [10] are two of the most actively developed projects. Both
feature distributed, in-memory computations and are certified for Apache Spark
(MLlib being part of Spark), as well as for Hadoop platforms. This in-memory
capability means that in some instances these frameworks outperform Hadoopbased frameworks [44]. MLlib has been shown to be more scalable than Vowpal
Wabbit. One important distinction when comparing H2O with other MapReduce
applications is that each H2O node (which is a single JVM process) runs as a
mapper in Hadoop. There are no combiners nor reducers. Also, H2O has more builtin analytical features and a more mature REST API for R, Python, and JavaScript
than MLlib.
Vowpal Wabbit24 [38] is an open-source, fast, out-of-core learning system,
currently sponsored by Microsoft research. It has an efficient implementation of
online machine learning, using the so-called hash trick [35] as the core data
representation, which results in significant storage compression for parameter
vectors. VW reduces regression, multiclass, multi-label, or structured prediction
problems to a weighted binary classification problem. A Hadoop-compatible computational model called AllReduce [1] has been implemented in order to eliminate
MPI and MapReduce drawbacks, which relate to machine learning. Using this
model, a 1000-node cluster was able to learn a terafeature data set in one hour [1].
Julia25 is high-level, high-performance, and dynamic programming language. It
is designed for computing and provides a sophisticated compiler, distributed parallel execution, and numerical accuracy and has an extensive mathematical function
library. Comparing to traditional MPI, Julias implementation of message passing is
one sided, thus simplifying the process management. Furthermore, these operations typically do not look like message send and message receive but rather
resemble higher-level operations like calls to user functions. It also provides a

22

https://flink.apache.org/
http://0xdata.com/product/
24
https://github.com/JohnLangford/vowpal_wabbit/
25
http://julialang.org/
23

7 Distributed Platforms and Cloud Services: Enabling Machine Learning for Big Data

151

powerful browser-based notebook using IPython. It also possesses a built-in package manager and it is able to call C functions directly. It is specially designed for
parallelism and distributed computation. It also provides a variety of classification,
clustering, and regression analysis packages26 implemented in Julia.
Another framework focusing on real-time online machine learning is TridentML [22], built on top of Apache Storm, a distributed stream-processing framework.
It processes batches of tuples in a distributed way, which means that it can scale
horizontally. However, Storm does not allow state updates to append simultaneously, a shortage that hinders distributed model learning.
The Apache Oryx 227 framework is a realization of the lambda architecture built
on top of Spark and Apache Kafka. It is a specialized framework that provides realtime, large-scale machine learning. It consists of three tiers: lambda, machine
learning, and application. The lambda tier is further split up into batch, speed,
and serving tier, respectively. Currently, it has only three end-to-end
implementations for the batch, speed, and serving layers (collaborative filtering,
k-means clustering, classification, and regression based on random forest).
Although it has only these three complete implementations, its main design goal
is not that of a traditional machine learning library but more of a lambda
architecture-based platform for MLlib and Mahout. At this point, it is important
to note several key differences between Oryx 128 and Oryx 2. Firstly, Oryx 1 has a
monolithic tier for lambda architecture, while Oryx 2 has three as mentioned in the
previous paragraph. The streaming-based batch layer in Oryx 2 is based in Spark,
while in the first version, it was a custom MapReduce implementation in the
computational layer. Two of the most important differences relate to the deployment of these frameworks. Oryx 2 is faster yet more memory hungry than the
previous version because of its reliance on Spark. Second, the first version
supported local (non-Hadoop) deployment, while the second version does not.
DryadLINQ29 [5] is LINQ30 (Language Integrated Query) subsystem developed
at Microsoft Research on top of Dryad [21], a general-purpose architecture for
execution of data-parallel applications. A DryadLINQ program is a sequential
program composed of LINQ expressions performing arbitrary side effect-free
transformations on data sets and can be written and debugged using standard .
NET development tools. The system transparently translates the data-parallel
portions of the program into a distributed execution plan, which is passed to the
Dryad execution platform that ensures efficient and reliable execution of this plan.
Following Microsofts decision to focus on bringing Apache Hadoop to Windows
systems, this platform has been abandoned, and Daytona project took off, which has
recently became Windows Azure Machine Learning platform.

26

http://mlbasejl.readthedocs.org/en/latest/
http://oryxproject.github.io/oryx/
28
https://github.com/cloudera/oryx
29
http://research.microsoft.com/en-us/projects/dryad/
30
http://msdn.microsoft.com/netframework/future/linq/
27

152

D. Pop et al.

Deeplearning4J31 is an open-source distributed deep-learning library written in

Java and Scala. It is largely based on ND4J library for scientific computation that
enables GPU, as well as native code integration. It is also deployable on Hadoop,
Spark, and Mesos. The main difference between this library and the others mentioned above is that it is mainly focused on business use cases, not on research. This
means that some features, such as parallelism, is automatic, meaning that worker
nodes are set up automatically.
Apache SAMOA32 is a distributed streaming machine learning framework. It
contains abstractions for distributed streaming machine learning algorithms. This
means that users can focus on implementing distributed algorithms and not worry
about the underlying complexities of the stream processing engines it supports
(Storm, S4, Samza, etc.).

7.5

Machine Learning as a Service (MLaaS)

This section focuses on Software-as-a-Service providers provision of machine

learning services as MLaaS. These services are accessible via RESTful interfaces,
and in some cases, the solution may also be installed on-premise (e.g., ersatz). The
favorite class of machine learning problems addressed by these services is predictive modeling (BigML, Google Prediction API, EigenDog), while clustering and
anomaly detection receive far less attention. We did not include in this category the
fair number of SQL over Hadoop processing solutions (e.g., Cloudera Impala,
Hadapt, Hive), because their main target is not machine learning problems, rather
fast, elastic, and scalable SQL processing of relational data using the distributed
architecture of Hadoop.
Table 7.3 presents current MLaaS solutions as well as some of their key
characteristics. We have identified four characteristics: machine learning problem
support, data sources, model exporting, and model deployment. It is easily observable that most MLaaS are designed to deal with common problems such as
classification, regression, and clustering. When it comes to data acquisition facilities, all platforms support data upload (various formats csv, arff, etc.); some even
feature integration with different storage solutions (S3, HDFS, etc.). Predictive
model training, verification, and visualization are supported by all the solutions
listed in Table 7.3; however, not all support predictive model exporting via
PMML.33 Last but not least, some platforms support local Web services as well
as cloud deployment. In the next few paragraphs, we detail some of the more
important services from Table 7.3.

31

http://deeplearning4j.org
https://samoa.incubator.apache.org/
33
http://www.dmg.org/v4-1/GeneralStructure.html
32

7 Distributed Platforms and Cloud Services: Enabling Machine Learning for Big Data

153

Table 7.3 Machine learning as a service

Name
Azure ML

Ersatz Labs
ScienceOpsa
(ScienceBox)
Skymind
BigMLb

ML problems
CLS, RGR,
CLU, ANO
RGR, CLS,
REC, CLU
DL
RGR, CLS,
REC, CLU
DL
CLS, RGR, CLU

Amazon ML
BitYotac
Google Prediction
API
EigenDogd
Metamarketse

CLS, RGR, CLU

CLS, RGR, CLU
CLS, RGR,
CLU, ANO
CLU, RGR
CLU, ANO

Zementis ADAPAf

CLS, RGR, CLU

PredictionIO

Data source
Upload, Azure

Model
export
None

Deployment
Cloud

Upload, Hbase

None

Local, cloud

Upload
S3, Upload

None
PMML

Cloud, local
Cloud, local

Upload
Upload, S3, Azure,
OData
S3, Redshift Upload
S3, Azure
Upload, Google Cloud
Storage
Upload, S3
Upload, HDFS

None
PMML

Cloud, Local
Cloud

None
None
PMML

Cloud
Cloud
Cloud

None
None

S3, Azure, Upload,

SAP HANA

PMML

Local, cloud
Local (Druid),
cloud
Local, cloud

https://yhathq.com/products/scienceops
http://bigml.com
c
http://bityota.com
d
https://eigendog.com/\#home
e
http://metamarkets.com/
f
http://zementis.com/products/adapa/amazon-cloud/
b

Windows Azure Machine Learning, formerly project Daytona, was officially

launched in February 2015 as a cloud-based platform for big data processing. This
comes with a rich set of predefined templates for data mining workflows, as well
with a visual workflow designer that allows end users to compose complex machine
learning workflows. In addition, it supports the integration of R and Python scripts
within workflows and is able to run the jobs on Hadoop and Spark platforms. The
built models are deployed in a highly scalable cloud environment and can easily be
accessed via Web services [39].
PredictionIO34 is based on an open-source software such as Spark. This means
that the solution can be deployed and hosted on any infrastructure. This is in sharp
contrast with Azure, which requires the data to be uploaded into Azure. Also, it is
possible to write custom distributed data processing tasks in Scala while on Azure
custom scripts can only be run on a single node. There are no restrictions on the size
of the training data not on the number of concurrent request. It can be deployed on
Amazon WS, Vagrant, Docker, or even starting from source code.

34

https://prediction.io/

154

D. Pop et al.

The recent popularity of deep learning has resulted in the creation of various
services which bundle deep-learning libraries (Theano, pylearn2, Deeplearning4j,
etc.) into a MLaaS format. Some good examples are Ersatz Labs35 and Skymind.36
These provide similar services and support distributed as well as GPU deployment.
Amazon Machine Learning service37 allows users to train predictive models in
the cloud. It targets a similar use case as Azure Machine Learning from Microsoft
and Googles Predictive API. It has similar features to many large-scale learning
applications including visualization and basic data statistics. The exact learning
algorithm it uses is not known; however, it is similar to Vowpal Wabbit. There are
some limitations such as the inability to export the learned model or to access data
which is not stored inside Amazon (Amazon S3 or Redshift).
Google Prediction API38 is Googles cloud-based machine learning tools that
can help to analyze data. It is closely connected to Google Cloud Storage39 where
training data is stored and offers its services using a RESTful interface, client
libraries allowing programmers to connect from Java, JavaScript, .NET, Ruby,
Python, etc. In the first step, the model needs to be trained on data, supported
models being classification and regression for now. After the model is built, one can
query this model to obtain predictions on new instances. Adding new data to a
trained model is called Streaming Training and it is also nicely supported. Recently,
PMML preprocessing feature has been added, i.e., Prediction API supports
preprocessing your data against a PMML transform specified using PMML 4.0
syntax and does not support importing of a complete PMML model that includes
data. Created models can be shared as hosted models in the marketplace.

7.6

Related Studies

Since 1995, when Thearling [40] presented a massively parallel architecture and the
algorithms for analyzing time series data, allegedly, one of the first approaches to
parallelization of ML algorithms, many implementations were proposed for ML
algorithm parallelization for both shared and distributed systems. Consequently,
many studies tried to summarize, classify, and compare these approaches. We will
address in this section only the most recent ones.
Upadhyaya [42] presents an overview of machine learning efforts since 1995
onward grouping the approaches based on prominent underlying technologies:
those employed on GPUs (20002005 and beyond), those using MapReduce technique (2005 onward), the ones that did not consider neither MapReduce nor GPUs

35

http://www.ersatzlabs.com/
http://www.skymind.io/about/
37
http://docs.aws.amazon.com/machine-learning/latest/mlconcepts/mlconcepts.html
38
https://developers.google.com/prediction/
39
https://developers.google.com/storage/
36

7 Distributed Platforms and Cloud Services: Enabling Machine Learning for Big Data

155

(19992000 and beyond), and, finally, few efforts discussing the MapReduce
technique on GPU. Contrasting to this extensive overview, we focus on more recent
distributed and cloud-based solutions, regardless if they are coming from academia
or industry.
The book Scaling Up Machine Learning: Parallel and Distributed Approaches
by Bekkerman et al. [3] presents an integrated collection of representative
approaches, emerged in both academic (Berkeley, NYU, University of California,
etc.) and industrial (Google, HP, IBM, Microsoft) environments, for scaling up
machine learning and data mining methods on parallel and distributed computing
platforms. It covers general frameworks for highly scalable ML implementations,
such as DryadLINQ and IBM PMLT, as well as specific implementations of ML
techniques on these platforms, like ensemble decision trees, SVM, or k-means. The
book is a good starting point, but it does not aim at providing a structured view on
how to scale out machine learning for big data applications, which is central to our
study.
A broader study is conducted by Assunc~ao et al. [2], who discuss approaches and
environments for carrying out analytics on clouds for big data applications. Model
development and scoring, i.e., machine learning, is one of the areas they considered,
alongside other three: data management and supporting architectures, visualization
and user interaction, and business models. Through a detailed survey, they identify
possible gaps in technology and provide recommendations for the research community on future directions on cloud-supported big data computing and analytics
solutions. With respect to this study, our work goes into deeper details on the
specific topic of distributed machine learning approaches, synthesizing and classifying existing solutions to give data scientists a comprehensive view of the field.
Interesting analyses have been made available through online press and blogs
[7, 9, 18]; they have reviewed open-source or commercial players for big data
analytics and predictions.

7.7

Conclusion and Guidelines

Analyzing big data sets gives users the power to identify new revenue sources,
develop loyal and profitable customer relationships, and run the organization more
efficiently and cost-effectively overall, giving them competitive advantage over
other competitions. Big data analytics is still a challenging and time-demanding
task that requires important resources, in terms of large e-infrastructure, complex
software, skilled people, and significant effort, without any guarantee on ROI. After
reviewing more than 40 solutions, our key findings are summarized below.
Both, research and industry, have invested efforts in developing as-a-Service
solutions for big data problems (Analytics-as-a-Service, Data-as-a-Service,
Machine-Learning-as-a-Service) in order to benefit of the advantages cloud computing provides such as resources on demand (with costs proportional to the actual
usage), scalability, and reliability.

156

D. Pop et al.

Existing programming paradigms for expressing large-scale parallelism

(MapReduce, MPI) are the de facto choices for implementing distributed machine
learning algorithms. The initial enthusiastic interest devoted to MapReduce has
been balanced in recent years by novel distributed architectures specifically
designed for machine learning problems. Nevertheless, Hadoop remains the stateof-the-art platform for processing large data sets stored on HDFS, either in
MapReduce jobs or using higher-level languages and abstractions.
Although state-of-the-art tools and platforms provide intuitive graphical user
interfaces, current environments lack an interactive process, and techniques should
be developed to facilitate interactivity in order to include analysts in the loop, by
providing means to reduce time to insight. Systems and techniques that iteratively
refine answers to queries and give users more control of processing are desired.
From the perspective of resource management and data processing, new frameworks able to combine applications from multiple programming models (e.g., MPI,
MapReduce, workflows) on a single solution need to be further investigated.
Optimization of resource usage and energy consumption, while executing dataintensive applications, is another challenging research direction in the next decade.
Developing software that meets the high-quality standards expected for
business-critical applications remains a challenge, and quality-driven development
methodology and new tools need to be created. Building on the principles of modeldriven development and on popular standards, e.g., UML or MARTE, such an
approach will guide the simulation, verification, and quality evolution of big data
applications.
Further exploiting the scalability, availability, and elasticity of cloud computing
for model building and exposing of prediction and analytics as hosted services is
opening a competitive and challenging market. Tools and frameworks to support
the integration of mobile and sensor data into cloud platforms need to be further
developed.
At the end of this chapter, we review some of the best practices that the recently
published literature recommends, namely:
Understand the business problem. Having a well-defined problem, knowing
specific constraints available for the problem under investigation, can greatly
improve performance of the ML algorithms.
Understand the ML task. Is it supervised or unsupervised? What activities are
required to get the data labeled? The same features (attributes, domains, labels)
need to be available at both times, training and testing. Pick a machine learning
method appropriate to the problem and the data set. This is the most difficult
task, and here are some questions you should consider: Do human users need to
understand the model? Is the training time a constraint for your problem? What
is an acceptable trade-off between having an accurate answer and having the
answer quickly? Keep in mind that there is no single best algorithm; experiment
with several algorithms and see which one gives better results for your problem.
In case of predictive modeling, carefully select and partition the data at hand in
training and validation set, which will be used to build the model, versus the test

7 Distributed Platforms and Cloud Services: Enabling Machine Learning for Big Data

157

set that you will use to test the performance of your model. More data for
training the model results in better predictive performance. Better data always
beats a better algorithm, no matter how advanced it is. Visualize the data with at
least univariate histograms. Examine correlations between variables.
Well prepare your data. Deal with missing and invalid values (misspelled words,
values out of range, outliers). Take enough time, because no matter how robust a
model is, poor data will yield poor results.
Evaluate your model using confusion matrix, ROC (receiver operating characteristic) curve, precision, recall, or F1 score. Do not overfit your model, because
the power lies in good prediction of unseen examples.
Use proper tools for your problems. For low-level programming environments
you might find difficult to use, try first the machine learning services offered by
cloud service providers, which are easy to use and powered by state-of-the-art
algorithms.
Acknowledgments This work was supported by the European Commission H2020 co-funded
with project DICE (GA 644869).

References
1. Agarwal A, Chapelle O, Dudik M, Langford J (2014) A reliable effective terascale linear
learning system. J Mach Learn Res 15:11111133
2. Assunc~ao MD, Calheiros RN, Bianchi S, Netto MAS, Buyya R (2014) Big data computing and
clouds: trends and future directions. J. Parallel Distrib Comput. http://dx.doi.org/10.10.16/j.
jpdc.2014.08.003
3. Bekkerman R, Bilenko M, Langford J (eds) (2012) Scaling up machine learning: parallel and
distributed approaches. Cambridge University Press, Cambridge
4. Berthold MR, Cebron N, Dill F, Gabriel TR, K
otter T, Meinl T, Ohl P, Sieb C, Thiel K,
Wiswedel B (2008) KNIME: The Konstanz Information Miner. In: Preisach C, Burkhardt H,
Schmidt-Thieme L, Decker R (eds) Studies in classification, data analysis, and knowledge
organization. Springer, Berlin/Heidelberg
5. Budiu M, Fetterly D, Isard M, McSherry F, Yu Y (2012) Large-scale machine learning using
DryadLINQ. In: Bekkerman R, Bilenko M, Langford J (eds) Scaling up machine learning.
Cambridge University Press, Cambridge
6. Chandola V, Banerjee A, Kumar V (2009) Anomaly detection: a survey. ACM Comput Surv
41(3):15:115:58
7. Charrington S (2012) Three new tools bring machine learning insights to the masses, February,
Read Write Web. http://www.readwriteweb.com/hack/2012/02/three-new-tools-bringmachine.php
8. Dai W et al (2015) High-performance distributed ML at scale through parameter server
consistency models, AAAI
9. Eckerson W (2012) New technologies for big data. http://www.b-eye-network.com/blogs/
eckerson/archives/2012/11/new_technologie.php
10. Franklin M et al (2015) MLlib: Machine Learning in apache Spark
11. Gander M et al (2013) Anomaly detection in the cloud: detecting security incidents via
machine learning, trustworthy eternal systems via evolving software, data and knowledge,
vol 379. Springer, Berlin/Heidelberg, pp 103116

158

D. Pop et al.

12. Ghoting A, Kambadur P, Pednault E, Kannan R (2011) NIMBLE: a toolkit for the implementation of parallel data mining and machine learning algorithms on MapReduce. In: Proceedings
of the 17th ACM SIGKDD international conference on Knowledge Discovery and Data
mining KDD11, ACM, New York, NY, USA, pp 334342
13. Ghoting A, Krishnamurthy R, Pednault E, Reinwald B, Sindhwani V, Tatikonda S, Tian Y,
Vaithyanathan S (2011) SystemML: declarative machine learning on MapReduce. In: Proceedings of the 2011 I.E. 27th International Conference on Data Engineering (ICDE 11).
IEEE Computer Society, Washington, DC, USA, pp 231242
14. Granger B, Perez F, Ragan-Kelley M (2011) Using IPython for parallel computing. http://
minrk.github.com/scipy-tutorial-2011. Accessed 13 May 2015
15. Grisel O (2013) Advanced machine learning with scikit-learn, PYCON tutorial. https://us.
pycon.org/2013/schedule/presentation/23/
16. Hall M et al (2009) The WEKA data mining software: an update. ACM SIGKDD Explor
Newsl 11(1):1018
17. Hall M (2013) Weka and Spark http://markahall.blogspot.co.nz/. Accessed 13 May 2015
18. Harris D (2015) 5 low-profile startups that could change the face of big data. http://gigaom.
com/cloud/5-low-profile-startups-that-could-change-the-face-of-big-data/. Accessed 15 July
2015
19. Hido S, Tokui S, Oda S (2013) Jubauts: an open source platform for distributed online machine
learning, NIPS workshop on Big Learning, Lake Taho
20. Hofmann M, Klinkenberg R (2013) RapidMiner: data mining use cases and business analytics
applications. Chapman &Hall/CRC, Boca Raton
21. Isard M et al. (2007) Dryad: distributed data-parallel programs from sequential building
blocks. SIGOPS Oper Syst Rev 41:5972. doi:10.1145/1272998.1273005
22. Jain A, Nalya A (2014) Learning storm. Packt Publishing, Birmingham
23. Nuggets KD (2014) http://www.kdnuggets.com/polls/2014/languages-analytics-data-miningdata-science.html. Accessed 15 May 2015
24. Krizhevsky A, Sutskever I, Hinton GE ImageNet (2012) Classification with deep
convolutional neural networks. NIPS 2012: neural information processing systems, Lake
Tahoe, Nevada
25. Le Q, Ranzato MA, Monga R, Devin M, Chen K, Corrado G, Dean J, Ng A (2012) Building
high-level features using large scale unsupervised learning, international conference in
machine learning, Edinburgh, UK
26. Leo S, Zanetti G (2010) Pydoop: a Python MapReduce and HDFS API for Hadoop. In:
Proceedings of the 19th ACM international symposium on high performance distributed
computing, Chicago, IL, USA, pp 819825
27. Low Y et al. (2012) Distributed GraphLab: a framework for machine learning and data mining
in the cloud. In: Proceedings of the VLDB endowment, vol 5, no 8, August 2012, Istanbul,
Turkey
28. Mohri M, Rostamizadeh A, Talwalkar A (2012) A foundations of machine learning. The MIT
Press, Cambridge, MA
29. Owen S, Anil R, Dunning T, Friedman E (2011) Mahout in action. Manning Publications Co.,
Shelter Island
30. Patcha A, Park JM (2007) An overview of anomaly detection techniques: existing solutions
and latest technological trends. Comput Netw Elsevier, North-Holland, Inc., 51:34483470
31. Pednault E, Yom-Tov E, Ghoting A (2012) IBM parallel machine learning toolbox. In:
Bekkerman R, Bilenko M, Langford J (eds) Scaling up machine learning. Cambridge University Press, New York
32. Pedregosa F et al (2011) Scikit-learn: machine learning in Python. J Mach Learn Res
12:28252830
33. Piccolboni A (2015) RHadoop. https://github.com/RevolutionAnalytics/RHadoop/wiki.
Accessed 13 May 2015
34. Roldn MC (2013) Pentaho data integration beginners guide. Packt Publishing, Birmingham

7 Distributed Platforms and Cloud Services: Enabling Machine Learning for Big Data

159

35. Rosen J et al (2013) Iterative MapReduce for large scale machine learning, CoRR,
abs/1303.3517
36. Russom P Big data Analytics (2011) TDWI best practices report, The Data Warehousing
Institute (TDWI) Research
37. Sagha H, Bayati H, Millan JDR, Chavarriaga R (2013) On-line anomaly detection and
resilience in classifier ensembles. Pattern Recogn Lett, Elsevier Science Inc., 34:19161927
38. Shi Q et al (2009) Hash kernels for structured data. J Mach Learn Res JMLR.org,
10:26152637
39. Elston SF (2015) Data science in the cloud with Microsoft Azure Machine Learning and R,
OReilly
40. Thearling KK (1995) Massively parallel architectures and algorithms for time series analysis.
In: Nadel L, Stien D (eds) Lectures in complex systems. Addison-Wesley, Reading
41. Tierney L, Rossini AJ, Snow NL (2009) A parallel computing framework for the R system. Int
J Parallel Prog 37:7890. doi:10.1007/s10766-008-0077-2
42. Upadhyaya SR (2013) Parallel approaches to machine learning a comprehensive survey. J
Parallel Distrib Comput 73(3):284292. ISSN 07437315. http://dx.doi.org/10.1016/j.jpdc.
2012.11.001
43. Wei D, Wei J, Zheng X, Kim JK, Lee S, Yin J, Ho Q, Xing EP (2013) Petuum: a framework for
iterative-convergent distributed ML. arxiv.org/abs/1312:7651
44. Zaharia M et al. Spark: cluster computing with working sets. In: Proceedings of the 2nd
USENIX conference on hot topics in Cloud, Computing, USENIX Association, pp 1010

Chapter 8

An Analytics-Driven Approach to Identify

Duplicate Bug Records in Large Data
Repositories
Anjaneyulu Pasala, Sarbendu Guha, Gopichand Agnihotram,
Satya Prateek B, and Srinivas Padmanabhuni

Abstract Typically, the identification and analysis of duplicate bug records of a

software application are mundane activities, carried out by software maintenance
engineers. As the bug repository grows in size for a large software application, this
manual process becomes erroneous and a time-consuming activity. Automatic
detection of these duplicate bug records will reduce the manual effort spent by
the maintenance engineers. It also results in the reduction of costs of software
maintenance. There are two types of duplicate bug records: (1) the records that
describe the same problem using similar vocabulary, and (2) the records that
describe different problems using dissimilar vocabulary but share the same underlying root cause. Each of these types of records needs a different set of techniques to
identify the duplicate bug records. In this chapter, we explain the various machine
learning techniques that are used to detect both types of duplicate bug records.
Some of these duplicate bug records reappear, that is, they show up continuously
over a long period of time. Here, we present a framework that can be used to
automate the entire process of detection of both types of duplicates and recurring
bug records. Using the framework, we conducted empirical studies on the opensource Chrome bug data records that are accessible online and the results are
reported.
Keywords Software maintenance Mining software repositories Duplicate data
records Vector space model Clustering techniques Co-occurrence models Bug
fix

A. Pasala (*) S. Guha G. Agnihotram S. Prateek B S. Padmanabhuni

Infosys Labs, Infosys Ltd., Electronics City, Hosur Road, 560 100 Bangalore, India
e-mail: Anjaneyulu_Pasala@infosys.com; Sarbendu_Guha@infosys.com;
Gopichand_A@infosys.com; srinivas_p@infosys.com
Springer International Publishing Switzerland 2016
Z. Mahmood (ed.), Data Science and Big Data Computing,
DOI 10.1007/978-3-319-31861-5_8

161

162

8.1

A. Pasala et al.

Introduction

In recent years, the software industry has been growing in leaps and bounds. It was
estimated that in 2014, the global software market was valued at $299 billion
[1]. The maintenance activities alone account for 6080 % of the overall software
budget [2]. In the USA alone, the cost of maintenance was calculated to be more
than $70 billion for ten billion lines of existing code [3]. It has been reported that
more than 50 % of software personnel are engaged in various activities [4] of
software maintenance. That shows the enormity of software maintenance activities
in organizations.
Therefore, organizations are increasingly focusing on cost optimization techniques for software maintenance. One such technique is an automation of mundane
but highly important activities relating to software maintenance. One of the areas
where automation is used is in bug tracking and handling systems. Bug tracking and
handling systems are being extensively used by organizations to allow software
users to report on issues faced or bugs detected while using the software application. These are called bug reports or bug records. The set of bug records for a
particular software application is termed as a bug repository. For large software
applications, the number of bugs reported grows exponentially in size, resulting in a
large data repository. Sometimes, these repositories grow to the levels of millions
and millions of records in number. Hence, there arises the challenge of managing
and analyzing these large bug data repositories.
One major challenge of these systems is managing and analyzing duplicates in
such massive data repositories. As several records come in concurrently, a large
number of these records tend to be duplicates. Studies have shown that duplicate
bug records can amount to as much as 20 % of total bugs reported [5]. As a result of
this, maintenance engineers spend a lot of time and effort, dealing with issues that
are being handled elsewhere or have already been dealt with. Thus, automatic
detection of duplicate data records helps the maintenance engineers to dispose of
issues faster and in a timely fashion. Additionally, it enables them to gain the
confidence of software users.
Typically, issue management is carried at multiple levels. In software maintenance parlance, service level 1 is where the bugs or incidents are initially received
immediately after they are reported. The maintenance engineers are assigned bugs
which are subsequently resolved. However the amount of analytics that gets carried
out at level 1 is very minimal. The analytics are usually carried at the level 2 where
the bugs reported over a period of time are analyzed for bug patterns and
recurring bugs.
For instance, a well-known software organization develops and maintains
200 software applications. On average 900 bugs or incidents are reported every
month on each of these applications by the users. It results in 2.16 millions of bug
records in a year for the organization to analyze. Among 900 reported bugs, close to
250 of them are duplicates. It shows the enormity of duplicate bugs reported by the
users. Hence, there arises a need for a system that automatically identifies and

8 An Analytics-Driven Approach to Identify Duplicate Bug Records in Large. . .

163

Table 8.1 Examples of apple-to-apple pairs

1.
2.

Gmail app should allow associating notifications/

ringtones with individual labels
No master password option

Gmail should allow per-label

notification settings
Master password option is
missing

Table 8.2 Examples of apple-to-orange pairs

1.

Home if force closing

2.

Version check for browser type causes failure

emulation mode to mimic IE ver. response for
such queries

ClassCastException in the Launcher

with certain widgets on Donut/master
Cant watch Netflix

groups all these duplicate bugs present in a massive data repository together. Such a
system would enable the maintenance engineer to identify frequently occurring
bugs and forward them to level 3 for further analysis and closure. At level 3, the root
cause for these duplicates or recurring bugs is identified and resolved. In this way
the organization ensures that the bug is resolved and it does not recur in the future.
This is how the prevention of recurring bugs is handled manually in the organizations. This manual process becomes a highly error-prone and time-consuming
activity in the case of massive bug repositories. And the cost grows exponentially
with the size of the repository. Therefore, automatic detection of these duplicate
bug records will reduce the manual effort spent by maintenance engineers. Hence, it
will result in the reduction of costs spent on software maintenance.
Magnus et al. [6] describe two types of bug data records: (1) records that
describe the same problem using similar vocabulary are called apple-to-apple
pairs, and (2) records that describe different problems textually (using dissimilar
vocabulary) but share the same underlying cause are called apple-to-orange pairs.
Some examples of reported apple-to-apple and apple-to-orange pairs are shown in
Tables 8.1 and 8.2, respectively. In our analysis, we find that around 11 % of the
validated duplicates in the Chrome bug repository are apple-to-orange pairs.
Currently, the apple-to-apple pairs can be detected using natural language
processing (NLP) and information retrieval (IR) techniques as they are textually
similar. However, the apple-to-orange pairs cannot be detected using these techniques as they generally use different vocabulary to describe the problem they
narrate.
Given a dataset of validated duplicate data records, our first task is to identify all
the apple-to-apple pairs in it. A couple of well-known techniques are described to
detect such apple-to-apple pairs. The idea being that any duplicate pair that cannot
be detected using simple NLP techniques is likely to be an apple-to-orange pair.
Given the list of apple-to-orange pairs, we build a matrix of the words found in the
training data that keeps track of the number of times a word, in one bug of an apple-

164

A. Pasala et al.

to-orange pair, occurs along with another word in the second bug of that same pair.
After that, when the user searches for duplicate bug data records, the word matrix is
used to retrieve the keywords that are most common among the query and appended
with the keywords. This expanded query is then used to search for the duplicates.
This is the process we aim to use to detect the apple-to-orange pairs.
The organization of this chapter is as follows. Section 8.2 briefly presents the
different techniques proposed and used to identify the similar records in the
literature. Section 8.3 discusses our proposed framework to identify both types of
duplicate bug records. The next two sections, 8.4 and 8.5, describe in detail the
techniques presented to identify the duplicate records of apple-to-apple pairs and
apple-to-orange pairs, respectively. These are explained using example data repository available online. Section 8.6 presents the case study conducted on Chrome bug
data and its results. Section 8.7 briefly explains the proposed recurring duplicate
bugs prevention framework. Finally, Section 8.8 provides conclusions and future
challenges of bugs detection in big data repositories and the challenges of analytics
to be carried in the future.

8.2

Literature Survey

The most well-known information retrieval (IR) technique used in identifying the
duplicate records is vector space model (VSM). It was initially used by Magnus
et al. [6]. In their approach a defect report is represented as a weighted vector of
words, and the similarity between two records is calculated as the distance between
their vectors. Using this approach, they were able to filter out 40 % of the marked
duplicate bug records. Nicholas et al. [7] build on Magnus model by using a linear
model classifier using textual similarity and a clustering algorithm combined. The
clustering is done on a graph built with nodes as defect records and edges link
records with similar text. They analyzed bug records from the Mozilla data and
were able to filter out 8 % of the duplicate bug records.
Further machine learning approaches such as Latent Dirichlet Allocation (LDA),
a topic modeling method, along with clustering have also been successfully used.
These approaches use clustertopic distribution and topicword distribution for
extracting keywords from the corpus [8]. Support Vector Machine (SVM) technique has also been used with various features of keywords like frequency, part of
speech, distance from the beginning of document, etc. to detect textually similar
data records.
Statistical approaches such as the term frequency, which is the number of times a
word appears in the corpus, are also used extensively along with the above
approaches. These approaches improve the detection of duplicates in the text
documents. Further improvements in the results are achieved using a term
frequencyinverse document frequency (TF-IDF), mainly to obtain the frequency
of the word in the corpus [9].

8 An Analytics-Driven Approach to Identify Duplicate Bug Records in Large. . .

165

Xiaoyin et al. [10] present an approach that involves execution information

along with NLP techniques. In their approach, when a new bug record arrives, its
language information and execution information are compared with the existing
bug record information. Based on this comparison, a small number of existing bug
records are suggested to the triager as the similar bug records to the new bug.
Finally, the triager determines whether the new bug record is a duplicate of the
existing bug record.
Chengnian et al. [11] leveraged discriminative models for information retrieval
to detect duplicate bug records more accurately. They claim that their technique
could result in 1731 %, 2226 %, and 3543 % relative improvement over stateof-the-art techniques. Furthermore, retrieval function (REP) has been used to
measure the similarity between two bug records [12]. It fully utilizes the information available in a bug record including not only the similarity of textual
content in summary and description fields but also similarity of non-textual fields
such as product, component, version, etc. For more accurate measurement of
textual similarity, they extended BM25F weighting technique, especially for
duplicate record retrieval. They also used a two-round stochastic gradient descent
to automatically optimize REP for specific bug repositories in a supervised
learning manner.
Yuan et al. [13] extended Nicholas work by improving the accuracy of automated duplicate bug record identification. They used a support vector classifier to
identify a duplicate bug. They conducted empirical studies on bug records from
Mozilla bug tracking system and found that they could improve the accuracy of
finding duplicate records.
Anahita et al. [14] extended the state of the art by using contextual information,
relying on prior knowledge of software quality, software architecture, and system
development topics, to improve bug deduplication detection. However, Ashish
et al. [15] used a character N-gram-based model for duplicate bug detection.
They use a feature extraction module that extracts all the character N-grams from
the title and description of each bug record and then calculate their similarity score
as a function of shared N-gram characters. It is more promising in detecting
duplicate bug data records submitted by developers.
Techniques also have been proposed to detect certain types of bug records with
different vocabulary such as synonym replacement, semantic matching using
WordNet, etc. However, the existing techniques can only detect duplicate bug
records with similar text and cannot detect dissimilar duplicate bug records as
they do not share the common words. Also, synonym replacement techniques do
reasonably well only when two bug records describe the same problem using
different words but totally fail in case of dissimilar duplicate bug records. This is
because while the underlying cause for the two bugs may be the same, they are
describing separate problems so the vocabulary for the two will be completely
different. There is no system, currently, where both types of duplicates can be
detected simultaneously in a real-time scenario.

166

A. Pasala et al.

Hence, there is a need for a method and system for detection of duplicate bug
records by expanding the queries using a word matrix. The word matrix models the
underlying relationships between the words present in the two dissimilar bug
records and by extension models the relation between the bug records themselves.
Further, it can also be extended such that it can be used in online scenario for
detection of all types of duplicates.

8.3

The Proposed System to Identify Duplicate Records

In this section, we propose and present a system to detect both types of duplicate
bug data records in an issue management system. A high-level depiction of the
system is shown in Fig. 8.1.
The system concurrently searches for both types of duplicate records using
different techniques. The resulting bug records from these searches are aggregated
and presented as a single list of duplicates. When the user submits a query (or a bug
arrives at the system through online process), the search for duplicate bug records is
carried out both with the normal query and with the expanded query. The details of
the expanded query is discussed in the subsequent sections. The normal query is
used to get the duplicate records that are similar in vocabulary of the query itself.
Hence, these records are termed as apple-to-apple pairs of duplicates. The
expanded query is used to get the duplicates that are dissimilar in vocabulary of
this query but represent a common cause to the problem of this bug occurrence.
These duplicate records are termed as apple-to-orange pairs. The records thus
fetched in both searches are then aggregated and displayed as a list of duplicates
to the user. In the following sections, we discuss the different techniques that are
used to detect the apple-to-apple and apple-to-orange pairs.

Fig. 8.1 A system to detect

all duplicate bugs

Query

Search with
query

Apple-to-apple
pairs

Aggregate
the results

Search with
expanded
query

Apple-to-orange
pairs

List of
duplicates

8 An Analytics-Driven Approach to Identify Duplicate Bug Records in Large. . .

8.4

167

Detecting Apple-to-Apple Pairs

In this section, we describe two well-known machine learning techniques to detect

apple-to-apple records. The techniques are vector space model (VSM) and clustering approaches.

8.4.1

Vector Space Model

In VSM [16, 17], both the collection of records and the query are represented as
vectors of unique words present in these records. These vectors usually result in a
high-dimensional Cartesian space depending on the number of unique words
present in these records. Each vector component corresponds to a unique word in
the vocabulary of these records. Given a query vector and a set of record vectors, it
ranks the records by computing a similarity measure between the query vector and
each record vector. This ranking is based on comparing the angle between these
vectors. The smaller the angle, the more similar are the vectors representing the
respective record and the query. The similarity in vectors means the record and
query representing it are similar in terms of the words present in the vocabulary of
the respective record and the query. Hence, these two are duplicate in terms of the
vocabulary they represent. For example, consider that records d1, d2, and d3 and a
query q are having the following vocabulary:
d1 <Gossip, Gossip, Gossip, Gossip, Gossip, Jealous>
d2 <Gossip, Gossip, Gossip, Gossip, Gossip, Gossip, Gossip, Jealous, Jealous,
Jealous, Jealous, Jealous, Jealous, Jealous>
d3 <Jealous, Jealous, Jealous, Jealous, Jealous, Gossip>
q <Gossip, Gossip, Gossip, Jealous, Jealous, Jealous>
These records can then be represented in a two-dimensional vector as shown in
Fig. 8.2.
Fig. 8.2 Vector space
representation of records
and query

168

A. Pasala et al.

The records d 1 , d2 ; and d 3, and the query q are distributed in a vector space defined
by the terms Gossip and Jealous. Here the record d2 is most similar to query q
because the angle between d2 and q is the smallest. A mathematical notion to the
VSM technique has been defined as follows.
In general, two n -dimensional vectors, say, vectors x and y, are represented as
~
x < x1 , x2 , . . . , xn >
~
y < y1 , y2 , . . . , yn >
We have
~
x :~
y j~
x jj~
y j cos
where ~
x :~
y represents the dot product between the vectors and j~
x j and j~
yj
represent the magnitude of these vectors, respectively. The dot product is also
defined as
~
x :~
y

Xn

xy
i1 i i

The length of a vector is computed from the Euclidean distance formula as

xj
j~

q
Xn
x2
i1 i

Substituting and rearranging these three equations, we get

cos

Pn
~
~
xi yi
x
y
pP

*
p
P n i1
n
2
2
x j j~
yj
j~
x
i1 i
i1 yi

If 0, then cos 1 so the two vectors are collinear and as similar as

possible. If =2 , then cos 0 and vectors are orthogonal and as dissimilar
as possible.
In VSM, given a record vector ~
d and a query vector ~
q , the cosine similarity


sim ~
d, ~
q is given as

 ~
~
d
q
sim ~
d, ~
q   *
~
j
~
d  q j
It is the dot product of the record vector and query vector normalized to unit
length. Provided all components of the vectors are nonnegative, the value of this
cosine similarity ranges from 0 to 1, with its value increasing with increasing
similarity.

8 An Analytics-Driven Approach to Identify Duplicate Bug Records in Large. . .

169

Naturally, for a collection of even modest size of data records, this vector space
model produces vectors having millions of dimensions. Though this high dimensionality may appear inefficient, in many circumstances the query vector is sparse
with all but a few components being zero. For example, consider the entire
Shakespeares Romeo and Juliet novel represented as a set of records by considering each sentence as a record. Also consider the vector corresponding to a query
<William Shakespeare marriage>. The vector will have only three nonzero components. To compute the length of the vector or its dot product with these record
vectors, we need to consider the components corresponding to these three terms
only. On the other hand, a record vector typically has a nonzero component for each
unique term contained in the record. However, the length of a record vector is
independent of the query. It may be precomputed and stored in a frequency or
positional index along with other record-specific information. Otherwise, it may
also be applied to normalize the record vector in advance, with the components of
the normalized vector taking the place of term frequencies in the posting lists.
Based on these concepts, the weightage techniques are formulated.

8.4.1.1

Weightage Techniques

As discussed above, in VSM, the records are represented as a set of vectors. Such
vectors are represented by their projections on each axis, which represent the terms
in the entire corpus. These projections of the vector along each axis (term) have
magnitude values which can be treated as weight of that term. Therefore, we
propose to use the weights of the terms.
Term frequencyinverse document frequency (TF-IDF) is the most popular
weighting scheme. In TF-IDF method [17], weights are assigned to terms in a
vector that represents a query or a record. In representing a record or a query as a
vector, a weight is assigned to each term that represents the value of corresponding
component of the vector. The assignment of weights to text tokens in a TF-IDF
scheme potentially changes the unstructured text data into a set of numeric data, and
thus it becomes easier to perform mathematical and statistical operations on the
data. The weight of a term denotes the magnitude of the component of the record/
query vector along the dimension of the vector represented by that term. When
assigning a weight in a record vector, the TF-IDF weights are computed by taking
the product of a function of the term frequency (f t, d ) and a function of the inverse
document frequency (1=N t), where N t is the number of documents in which the term
t occurs. When assigning weight to a query vector, the within-query term frequency
qt may be substituted for f t, d in essence treating the query as a tiny document. A
TF-IDF weight is a product of functions of the term frequency and inverse document frequency. A common error is the use of the raw term frequency or f t, d . This
may lead to poor performance as very commonly occurring words will get assigned
higher weightage compared to distinctive words which may occur infrequently.

170

A. Pasala et al.


The IDF function typically relates the document frequency ( N t to the total
number of records in the collection (N). The basic intuition behind the IDF is that a
term appearing in many records should be assigned a lower weight or importance
than a term appearing in only a few records. Of the two functions, IDF comes
closer to having a standard form:
IDF log N=N t
The basic intuition behind the TF function is that a term appearing many times in
a record should be assigned a higher weight for that record than for a record in
which it appears less number of times. Another important consideration for the TF
function is that the value should not necessarily increase linearly with ft,d. Although
two occurrences of a term should be given more weight than single occurrence, they
should not necessarily be given twice the weight. The following definition of TF
meets these requirements:

TF

 
log f t, d 1
0

if
f t, d > 0
otherwise

When the above equation is used with a query vector, f t, d is replaced with the
term qt , the query term frequency of t in q.

8.4.1.2

Example

Consider the text fragments from Shakespeares Romeo and Juliet, Act I, Scene I,
as listed in Table 8.3.
In Table 8.3, there are five records in the collection and the sir word appears in
four of them. Therefore, the IDF value for sir is
log N = f sir log 5=4  0:32
Also, sir word appears twice in record 2. Therefore, the TF-IDF value for the
corresponding component of its vector is

Table 8.3 Text fragment collection

Record ID
1
2
3
4
5

Record content
Do you quarrel, sir?
Quarrel sir! No, sir!
If you do sir, I am for you. I serve as good a man as you
No better
Well sir!

8 An Analytics-Driven Approach to Identify Duplicate Bug Records in Large. . .

171

log f sir ; 2 1*log N = f sir log 2 1*log5=4  0:64

Computing the TF-IDF values for the remaining components and the remaining
records gives the following set of vectors:
d~1
d~2
d~3
d~4
d~5

< 0:00, 0:00, 0:00, 0:00, 1:32, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 1:32, 0:00, 0:32, 0:00, 1:32 >
< 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 1:32, 1:32, 0:00, 0:64, 0:00, 0:00 >
< 2:32, 2:32, 4:64, 0:00, 1:32, 2:32, 2:32, 4:64, 2:32, 2:32, 0:00, 0:00, 2:32, 0:32, 0:00, 3:42 >
< 0:00, 0:00, 0:00, 2:32, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 1:32, 0:00, 0:00, 0:00, 0:00, 0:00 >
< 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:32, 2:32, 0:00 >

where the components are sorted alphabetically according to their corresponding

terms. Normalizing these vectors by dividing with their lengths produces unit
vectors, as follows:
 
d~1 = d~1  < 0:00, 0:00, 0:00, 0:00, 0:57, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:57, 0:00, 0:14, 0:00, 0:57 >
 
d~2 = d~2  < 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:67, 0:67, 0:00, 0:33, 0:00, 0:00 >
 
d~3 = d~3  < 0:24, 0:24, 0:48, 0:00, 0:14, 0:24, 0:24, 0:48, 0:24, 0:24, 0:00, 0:00, 0:24, 0:03, 0:00, 0:35 >
 
d~4 = d~4  < 0:00, 0:00, 0:00, 0:87, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:49, 0:00, 0:00, 0:00, 0:00, 0:00 >
 
d~5 = d~5  < 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:14, 0:99, 0:00 >

To rank these five records with respect to the query < quarrel, sir >, we first
construct the unit query vector as stated below:
~
q
< 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:00, 0:97, 0:00, 0:24, 0:00, 0:00 >
qj
j~

The dot product between these unit query vector and each unit record vector
gives the cosine similarity values. The cosine similarity values for all these records
are computed and tabulated in Table 8.4. Based on these cosine similarity values,
the final record rankings are 2,1,5,3,4.

8.4.2

Clustering Approaches

Clustering approaches [18] are used in unsupervised methods to group the similar
text records together which will help the user in identifying the apple-to-apple
Table 8.4 Cosine similarity for text fragment collection
Record ID
1
2
3
4
5

Cosine similarity
0.59
0.73
0.01
0.00
0.03

172

A. Pasala et al.

pairs. These approaches group the similar records together by identifying the same
words in the records. There are different clustering approaches available, e.g.,
K-means clustering [8, 19] and supervised methods such as the nearest neighbor
(NN) method [20, 21].
We use tokenization technique and TF-IDF scores to identify similar words in
the records so that they can be grouped together in one group. Tokenization is the
process of chopping up a given stream of text or character sequence into words,
phrases, or some meaningful words. This information is used as input for further
processing such as obtaining the similar phrases, apple-to-apple and apple-toorange phrases. In the following sub-section, we describe the K-means clustering
approach to identify similar records.

8.4.2.1

K-Mean Clustering

It is one of the most efficient methods of clustering approaches. From the given set
of data records, words or phrases are obtained using tokenization and form the k
different clusters. Each of these clusters characterized with a unique centroid (e.g.,
mean of the words in that cluster) is partitioned using the K -means algorithm. The
elements belonging to one cluster are close to the centroid of that particular cluster
and dissimilar to the elements belonging to the other clusters.
The k in the K -means algorithm refers to the number of groups to be assigned
for a given dataset. If n objects have to be grouped into k clusters, cluster
centers have to be initialized. These objects are referred to words or phrases.
Further, each object is assigned to its closest cluster center. And the center of the
cluster is updated until there is no change in each cluster center. From these centers,
we can define a clustering by grouping objects according to the center the object is
assigned to. The data is partitioned into k clusters where each clusters center is
represented by the mean value of the objects in the cluster. Each of these clusters
forms the apple-to-apple pairs. The algorithm is explained as follows.

8.4.2.2

Algorithm: Cluster Algorithm to Group Similar Bug Data

Records

Input: Data records and k the number of clusters

Output: A set of k clusters:
Step 1: Choose k number of clusters to be determined.
Step 2: Choose Ck centroids randomly as the initial centers of the clusters.
Step 3: Repeat:
3.1: Assign each object to its closest cluster center using Euclidean distance.
3.2: Compute new cluster center by calculating mean points.

8 An Analytics-Driven Approach to Identify Duplicate Bug Records in Large. . .

173

Step 4: Until:
4.1: No change in cluster center OR.
4.2: No object changes its clusters.
The flow of different activities of the K -means algorithm is shown in Fig. 8.3.

8.4.2.3

Example

From the given documents or records, derive the keywords using TF-IDF approach.
After that, cluster the similar keywords together form the apple-to-apple pair in
each of the cluster.
Consider the given four clusters as shown in Table 8.5. In cluster 1 we are having
similar words (AbAb). In cluster 2, we are having (GhiGhi) similar words.
Similarly, in cluster 3, we are having similar keywords (DepDep). Likewise, in
cluster 4, we are having similar keywords (RffRff). This algorithm helps in
identifying the duplicates of apple-to-apple pairs.

Start

Input: Preprocessed
Documents/Incidents

Represent the documents as

words/Phrases using VSM or
Tokenization

Apply -means Clustering Algorithm

as given above

Output: Cluster containing similar Apple to Apple

pairs

Fig. 8.3 Flow chart of K -means algorithm to obtain the similar records

174

A. Pasala et al.

Table 8.5 Identifying the similar words together using K-means clustering
Cluster1
Ab
Ab
Ab
Ab
Ab
Ab
Ab
Ab
Ab
Ab
Ab
Ab
Ab
Ab

8.4.2.4

Cluster2
Ghi
Ghi
Ghi
Ghi
Ghi
Ghi
Ghi
Ghi
Ghi
Ghi
Ghi
Ghi
Ghi
Ghi

Cluster3
Dep
Dep
Dep
Dep
Dep
Dep
Dep
Dep
Dep
Dep
Dep
Dep
Dep
Dep

Cluster4
Rff
Rff
Rff
Rff
Rff
Rff
Rff
Rff
Rff
Rff
Rff
Rff
Rff
Rff

Nearest Neighbor Classifier

The nearest neighbor (NN) classifier is a supervised approach which is also called
as K-nearest neighbor (K -NN) classifier [15, 21]. This approach is a nonparametric
approach and used for both classification and regression. It is the simplest of all
machine learning algorithms, and it can be used to weight the contributions of the
neighbors, so that the nearer neighbors contribute more to the average than the
distant neighbor.
Consider the documents containing the bug data and extract the keywords using
TF-IDF scheme. Run the NN algorithm, and if the new keyword appears, it will be
classified as its similar keyword which is the nearest sample. In this way we will
group all similar keywords. In other words, the classes which are similar to appleto-apple pairs will be assigned in one class, as shown below:
Input
D - > Training data
K -> Number of the nearest neighbors
S -> New unknown sample which is a word or a phrase
Output
C -> Class label assigned to new sample S
The Algorithm Is
Let I1, I2, . . ., Ik denote the k instances from training set D that are nearest to new
unknown sample S.
C The class from k-nearest neighbor samples with maximum count.
k 1; then the output is simply assigned to the class of that single nearest
neighbor.

8 An Analytics-Driven Approach to Identify Duplicate Bug Records in Large. . .

8.5

175

An Approach to Detect Apple-to-Orange Pairs

In this section, we describe an approach to detect the apple-to-orange pairs of

duplicates. The approach is based on building word co-occurrence model. The idea
is to model the underlying relations between two dissimilar duplicates and use the
model thus built for future detection of duplicates. To accurately learn these
relations and build the model, we need to analyze the apple-to-orange pairs that
have been previously identified and stored during the maintenance of the software
manually by the maintenance engineers. Typically, this process is called learning or
training phase of the technique. After building the model, we use the model thus
built in our system to provide real-time detection of both apple-to-apple and appleto-orange pairs. This process is carried automatically by the system as and when a
new bug record arrives. The process is called an online phase of the technique.
Therefore, the process of identifying duplicates in dissimilar textual records consists of two phases as shown in Fig. 8.4. These phases are further described in detail
in the following subsections, respectively.

8.5.1

Training Phase

The process of building a model using the previously known data records is known
as the training phase. In this phase, we train the model that will be used to expand

Fig. 8.4 Approach to detect apple-to-orange pairs

176

A. Pasala et al.

the query that is submitted by the user of the duplicate bug record detection system.
For training the model, we need the previous history of duplicate bug records that
has been manually verified by the maintenance engineers and marked as duplicates.
The training phase can be divided into two steps, namely, identify the apple-toorange pairs and build a word co-occurrence model. These steps are
elaborated here.

8.5.1.1

Step 1

Identify the apple-to-orange pairs present in bug record history. Initially to train
our model, we have to identify the list of apple-to-orange pairs present in the
validated duplicate dataset. This is done by using vector space model. Typically
each record is represented by a vector with one component in the vector for every
term in the entire vocabulary present in the bug records processed till that time.
These components are usually calculated using the TF-IDF weighting scheme:
wi tf i  idf i
where
wi is the weight assigned to each term in the vector.
tf i represents the term frequency, i.e., the number of occurrences of a term in that
record.
idf i represents the inverse-document frequency and is a measure of whether a
term is common or rare across all records.
Inverse-document frequency is calculated as
idf i log

Dtot
D ti

where
Dtot is the total number of records processed.
Dti is the number of records containing that particular term.
The similarity between two records is calculated as the deviation of angles
between each record or the cosine of the angle between the vectors. We calculate
the similarity between all the duplicate pairs present in the bug repositories. The
ones which have no similarity are identified as apple-to-orange pairs.

8.5.1.2

Step 2

Build a word co-occurrence model by capturing the underlying relations between

known apple-to-orange pairs. The concept of co-occurrence has been slightly

8 An Analytics-Driven Approach to Identify Duplicate Bug Records in Large. . .

177

modified in our approach than in the normal use. We do not consider the frequency
of co-occurrence between two words belonging to the same pair. Rather we
consider only the number of times a word in one bug record occurs along with a
word in that records validated duplicate. The model is represented in a word matrix
which is of size NxN (N being the size of the vocabulary), and the value of the cell
aij will be the co-occurrence score between the word with index i and the word with
index j. This score is representative of the relationship between the two words; the
greater the score, the more related the two words are. As there will be huge number
of words in our vocabulary, representing the model as a simple two-dimensional
array will not be feasible. As the majority of the words do not co-occur, there is a
high level of sparsity in the matrix which allows us to use simpler sparse representations. In this way, we plan to capture the relationships between apple-to-orange
pairs. For example, if server failure and login issue are two validated dissimilar
duplicates, then we map server with login in the matrix. The next time any server
issue is reported, then we can use this model to predict that there might have been
some sort of login issue also. We build the co-occurrence model for all the known
apple-to-orange pairs in the available bug record history.

8.5.2

Online Phase

In a real-time scenario with new bug records constantly being reported by the users
of the software application, our system performs two searches for each of these bug
records (or queries) as shown in Fig. 8.1. These are (1) the standard text similarity
algorithm that is used to find all the similar duplicates for that bug record and (2) the
technique to find the apple-to-orange pairs present in the repository with extended
query. The procedure has the following three steps:
1. Compare and augment the query with the word co-occurrence model built
during the training phase. It has the following two subphases:
a. Extract the keywords relevant to the query from the co-occurrence model.
This is done by picking up those related keywords that are most common to
all the words present in the query. These are the words that have the highest
co-occurrence score in the co-occurrence model.
b. Expand the query with the words extracted from the model. The above word
matrix is then used to expand the query submitted by the user. To do this, the
query is first processed and all the stop words in the query are removed. Then
for each word in the query, we extract the top 5 most commonly co-occurring
words from the matrix. All these related words are added into a collection,
sorted according to their count. The word with the greatest count would be the
word that is most common among the words of the original query. Thus from
the collection, we choose the top 5 words based on their count. These words

178

A. Pasala et al.

are added to the original query which is used to search for duplicates. Thus
the resulted query is called expanded query.
2. Search the repository with the above expanded query. This will return a list of
records that include duplicates that are dissimilar in text to the original query.
More details on the procedure of search are given below.
3. Retrain the model with the new data if maintenance engineer validates it to be
part of an apple-to-orange pair. The process of learning wherein model parameters are changed or tweaked whenever new data points emerge is known as
incremental learning.
Finally, we merge the results obtained via the two searches to provide the user
with an aggregated list of duplicates.
The final search is performed on the indexed data using three separate retrieval
algorithms implemented in Lucene. These are Okapi BM25 and language modeling
using JelinekMercer and Dirichlet smoothing.

8.5.3

Okapi BM25

The BM25 weighting scheme also known as Okapi weighting scheme is one of the
state-of-the-art information retrieval (IR) techniques and has been in widespread
use since its introduction [22]. It is a probabilistic model, which uses three key
document attributes, namely, term frequency, document frequency, and average
document length. The scoring function is implemented as
Sq; d

IDFt

t2q

tf td


ld
k1 1  b bavl
tf td
d

where
tf td is the frequency of term t in document d.
ld is the length of document d.
avld is the average document length along the collection.
k1 is a parameter used to tune the saturation of the term frequency and is usually
between 1.2 and 2.
b is a length tuning parameter whose range is [0,1].

The IDF for term t is computed as

IDFt log
where

Dtot  Dt 0:5
Dt 0:5

8 An Analytics-Driven Approach to Identify Duplicate Bug Records in Large. . .

179

Dtot is the total number of documents.

Dt is the number of documents where that particular term occurs.

8.5.4

Language Modeling with Smoothing

The central idea behind language modeling is to estimate a model for each document and then rank the documents by the query likelihood according to the
estimated language model. A big problem while calculating the maximum likelihood is that if a particular term has not been encountered before, then the maximum
likelihood would become 0 as the probability of the individual term would be 0. To
avoid this, a technique known as smoothing is used which adjusts the likelihood
estimator to account for that type of data sparseness. Here, we use two types of
smoothing methods JelinekMercer and Dirichlet priors. The JelinekMercer
method involves a simple mixture of a document-specific distribution and a distribution estimated from the entire collection:
 
p tjd 1  pml tjd pml tc
 
where pml td represents the conditional probability of the term belonging to that
 
document and pml tc is the conditional probability of the term belonging to that
particular language.
In the Dirichlet smoothing method, the language model is built from the whole
collection as the prior distribution of a Bayesian process. The conjugate priors for
this form a Dirichlet distribution hence the name:
 
tf td p td
p tjd
ld
 
Here, p td is the conditional probability of the term belonging to the document.
According to Zhai et al. [9], the optimal value for in JelinekMercer is 0.7, and in
Dirichlet prior is 2000.

8.6

Implementation and Case Study

The prototype of duplicate bug identifier is built in Java using the open-source API
Lucene (Apache 2.0 licence). Lucene allows easy implementation of indexing and
searching using the VSM algorithm described above. It calculates the TF-IDF
values for each word and implements its custom scoring technique to find the
similarity between bug records. We modify this custom scoring technique to

180

A. Pasala et al.

include some of the more recent and powerful algorithms such as Okapi BM25 and
language modeling as discussed earlier.
As there are often huge number of words in bug record vocabulary, representing
the model as a standard matrix becomes highly memory intensive, and it will not be
possible to run any program in memory. However, as a majority of the words do not
co-occur, there is a high level of sparsity (the presence of zeros) in the matrix which
allows us to use simpler sparse representations such as compressed row storage
(CRS) or compressed column storage (CCS). These representations save memory
by storing subsequent nonzero values in contiguous memory locations and ignoring
the zeros. This sparse matrix multiplication was done using another open-source
API called l4aj which has functions for CRS as well as CCS matrix representations.
We used CRS for sparse matrix representation.

8.6.1

Datasets

Typically a bug record contains several fields like summary, description, platform,
priority, etc. which are in a natural language format. There are also other fields with
non-textual information as attachments. The status of the bug is set by the developer
or triager who checks if the bug record is valid or not. And if the bug record has
been raised before or caused by the same issue as another bug record, then it is
marked as a duplicate. A typical example of a bug record from Chrome dataset,
taken from Chrome issue tracker, is shown in Fig. 8.5. We have used the Chrome
bug record dataset available on the Google code website. The Chrome dataset
consists of 142,286 issues out of which 25,756 issues (18.1 %) were marked by
the maintenance engineers as duplicates. These numbers on this bug data are from
the date which we have accessed. Figure 8.6 shows a sample of the validated
duplicate bug record dataset.
We calculate the similarity between all the duplicate pairs, and the ones which
have no similarity are identified as apple-to-orange pairs. From Fig. 8.7, it is
noticed that 2769 out of 25,756 duplicate pairs in chrome were apple-to-orange
pairs.

Fig. 8.5 Bug records in Chrome issue tracker

8 An Analytics-Driven Approach to Identify Duplicate Bug Records in Large. . .

181

Fig. 8.6 Validated duplicate pairs in Chrome dataset

Fig. 8.7 No. of apple-to-orange pairs found in Chrome dataset

The apple-to-orange dataset was split into 60 % training and 40 % testing

datasets. This split was done randomly and repeated five times for the purpose of
generalization.
Among the 1108 apple-to-orange pairs used for testing, our tool detected around
205 pairs. However the number of bug records detected in the top 10 is quite low
with the best result being only 14 records coming among the top 10. The average
recall rates at 10 for all three techniques are shown in Table 8.6, where recall rate at
k is defined as the fraction of duplicate pairs that were detected in the top k results of
the search. A recall rate at 10 of 1 means that all the duplicate bug report pairs

182

A. Pasala et al.

Table 8.6 Recall rate of apple-to-orange pairs after using query expansion
Rank size
10
20
30
50
100

Okapi BM25
0.003
0.010
0.017
0.034
0.065

JelinekMercer
0.002
0.008
0.014
0.029
0.068

Dirichlet priors
0.015
0.029
0.039
0.059
0.081

Table 8.7 Performance statistics of our technique versus standard search

Fig. 8.8 Performance

statistics of the technique

Time taken for Search

with Query Expansion

Number of bug records

Number of apple-to-apple pairs
Number of apple-to-orange pairs
Time taken for standard search (milliseconds)
Time taken for search with query expansion
(seconds)

120
100
80
60
40
20
0

10,000
112
1181
63
1.773

20,000
348
2907
62
5.431

50,000
1116
8317
62
22.56

100,000
2147
18,685
63
65.874

150,000
2796
25,755
75
98.5

2796
2147

Se
112

348

1116

Values
above data
points
represent

Number of Incidents
were detected among the top 10 search results. We can observe that the best
performance was obtained using language modeling with Dirichlet smoothing.
Still it only amounts to 1.5 % of the apple-to-orange pairs being detected in the
top 10 ranks.
The performance of the technique is rather low because there is very little
training dataset. We can see that a significant number of duplicates that were not
being detected at all before are becoming visible now and with more data that is
spread over a longer period of time. It will be possible to increase the rank of those
results and improve the recall rates.
The technique as such does not add much time costs to the standard search
algorithms. Table 8.7 describes the time overhead in detecting the duplicates using
VSM and our technique. The program was run on a machine with a 32 bit single
core processor and 4GBs of RAM. We can observe that there is an increase in time
taken to display the results as the number of bugs increases, but at the maximum
for 150,000 bugs present, it can return the results for a new query in less than 2 min
on even a slower machine.
Figure 8.8 is a graphical representation of the time taken to get the results for a
bug given the total number of bugs already present in the repository. The values

8 An Analytics-Driven Approach to Identify Duplicate Bug Records in Large. . .

183

displayed over each data point correspond to the number of apple-to-orange pairs
that have been used to train the model.

8.7

Recurring Bug Prevention Framework

It is observed that some of the bugs reappear now and then. Such bugs are close to
20 % of the total bugs reported by the end users for that particular software
application. Preventing these bugs will save the effort and time of maintenance
engineers. These types of bugs need to be monitored over a period of time and a root
cause will be identified. Once the root cause is identified, a solution to the problem
can be designed and implemented. Based on this approach, we propose the following technical solution that automates the process of identifying the recurring bugs
and how well it prevents them to recur in the future. The framework consists of four
components as shown in Fig. 8.9. Brief detail of the components appears below.

8.7.1

Knowledge-Enriched Bug Repository

Typically, this component is a bug repository system that stores all the reported
bugs by the users for that particular software from the start. This repository is
maintained using any issue tracking system. The term knowledge enriched is used
because in addition to storing the bug details as reported by the user, it also stores
specific information retrieval-related information. This additional information consists of whether it is duplicate of some existing bug, whether it belongs to a
particular cluster, when that particular bug group had been fixed, what the solution
is, when it is applied, etc. Also if it has reappeared, it marks and links all of the
Fig. 8.9 Recurring bugs
prevention framework

Bug record
Knowledge
Enriched bug
records

Identify groups
of recurring
duplicate Bugs
records

Preventability
Analytics

Analyse the root

cause and apply
fix

184

A. Pasala et al.

duplicates. We can use this information to speed up the bug clustering process and
also to aid in the calculation of number of bugs prevented.

8.7.2

Identify Groups of Recurring Duplicate Bugs

At a given point of time, say once a month (or as decided by the organization based
on its policies toward bug analytics), using IR techniques the bugs in the
knowledge-enriched bug repository can be clustered into groups of similar bugs.
It is useful for identification of recurring bug patterns. IR technique is executed on
the above repository of reported bugs. One way is to iterate through all the bugs in
the repository and identify those duplicates with score greater than a threshold. Like
in ticket clustering tool, we can give the user the ability to vary the number of search
results and the threshold for the score. Two possible algorithms to identify this bug
patterns are (1) VSM and (2) clustering techniques.
Initial analysis on the Chrome data using VSM showed that out of 40,000 there
were around 2500 groups with cluster size greater than 10 and with a moderate
score (0.75) threshold. This grouping took 6 min which is not a huge overhead
considering the number of bugs reported every month is considerably lesser.
Further, through tooling we can also reduce the number of bugs to be grouped
each time increasing the performance.
Secondly, we used K-means algorithm to group the bugs into categories based
on similarity and provide the user the ability to choose number of clusters or give
keywords to aid and fine-tune the clustering. The problem is that the grouping
would be broader in nature and would not be feasible for large data. From the
previous example, we could see that there were more than 2500 clusters for the
Chrome incident data. Therefore, clustering technique could be helpful for small
datasets, which is a limitation.

8.7.3

Root Cause Analysis and Bug Fix

Once the patterns are identified, root cause analysis can be performed on the
clusters or patterns of duplicate/recurring bugs detected. The identification of root
cause for the repeatedly occurring bug is more of technical and manual process.
Based on the identified root cause for that particular bug, a solution to this root
cause problem is identified. The required solution is implemented. As the solutions
are stored for this recurring bug in the issue tracking system as and when these bugs
occurred, hence we can use machine learning algorithms to analyze all the applied
solutions, and a reasonable summary can be presented to the maintenance engineer
so that appropriate solution can be invented and implemented. Based on the
analysis, the maintenance engineer can devise a solution that is more suitable and
appropriate.

8 An Analytics-Driven Approach to Identify Duplicate Bug Records in Large. . .

8.7.4

185

Preventability Analytics

In this step, we store the information regarding the bug fixes in the knowledgeenriched bug repository to aid in the calculation of preventability. For example,
storing the time at which a particular pattern of reoccurring bugs are fixed can be
used to track further occurrences of the same. That is, if no further incidents of that
type are encountered, then we can use that to demonstrate the effectiveness of
preventive maintenance.
Let us assume that at some point of time t1, bug B1 has occurred the first time.
For the duration of n days, it was observed that there are m times the bug that
has been reported. That is, the Bm has occurred at time t2. At t2, the root cause
analysis is carried and the bug fixed with likely permanent solution at time t3. When
we carry the duplicate analysis at time t4, the duplicate bug doesnt appear at all.
Based on this analysis, we can conclude that the bug is not recurring. Therefore, we
will assume that all such bugs have been disappeared. Hence, using statistical
techniques the percentage of bugs detected and closed and prevented can be
derived.

8.8

Conclusions

In this chapter, the practical challenges of identifying duplicate bugs in large bug
data repositories are explained in detail with examples. To address these challenges,
a set of machine learning techniques are proposed and discussed. Based on the
proposed techniques, a tool has been built. Using the tool a case study has been
conducted on a practical open-source dataset which is available online. The results
have been presented in terms of the number of duplicate bugs detected in this
dataset.
The chapter proposed VSM and cluster algorithms to identify duplicate bugs
which are similar in vocabulary. It also proposed a method to detect the textually
dissimilar duplicate records by capturing the underlying root cause relations
between the two bug records. It used an expanded query by aggregating the
words present in it and also words captured through the co-occurrence relationships
between words present in the existing duplicate records of that repository. Both
types of duplicates are captured concurrently using two searches, one with the
original query and the other with the expanded query. This is a novel way of
identifying duplicate data records. It expands further by identifying patterns in
the history of previously validated duplicates. One of the limitations of the method
is that it requires a priory properly triaged bug data records which are used to build
the model. The process of triaging though is often neglected in organizations and
there can be a shortage of available data records.
Though we have conducted a case study on practical data, the results are not
quite encouraging on identification of apple-to-orange types of duplicates

186

A. Pasala et al.

especially on the top 10 lists of duplicates. Further, empirical analysis is required to

validate these proposed methods of duplicate identification and results. Then only
one can analyze the efficiency of the proposed methods. Further, there was a
limitation on the dataset size we have used. We have to experiment on a huge
datasets and conclude the results. Certainly the recurring bugs prevention framework takes a longer time to implement in an organization and also to validate the
framework. Therefore a lot of empirical analysis is warranted.
Further, the bug record not only contains textual information but also contains
the images of the screens. It is observed that few bugs have a simple screenshots
attached to the bug information for detailed explanation. However, these records
are out of our scope, as we perform analytics on text only. These images contain
more accurate information about the nature of the failure that occurred during the
access of the software application. To handle this additional information, given a
bug report with a screenshot of the error attached along with it, we can use optical
character recognition (OCR) techniques to extract any textual information that it
contains, and the text can be added to the description of the bug. This enhanced
description can now be used by the duplicate bug detection algorithms to provide
improved duplicate bug results. This could further increase the processing time for
duplicate search and building co-occurrence matrix. This shows the need for further
work in this direction to improve both the results fetched and as well as for
improving the processing time. Hence, there exists a wider scope for expanding
our work in applying analytics to large bug repositories.

References
1. Datamonitor (2011) Software: Global Industry Guide 2010. MarketLine
2. Canfora ACG (1995), Software maintenance. In: 7th international conference on software
engineering and knowledge engineering, pp 478486
3. Sutherland J (1995) Business objects in corporate information systems. ACM Comput Surv 27
(2):274276
4. Jones C (2006) http://www.compaid.com/caiinternet/ezine/capersjones-maintenance.pdf
5. Anvik J, Hiew L, Murphy GC (2005) Coping with an open bug repository. In: Proceedings of
the OOPSLA workshop on eclipse technology exchange, pp 3539
6. Alexandersson M, Runeson ONP (2007) Detection of duplicate defect reports using natural
language. In: Processing 29th international conference on software engineering, pp 499510
7. Jalbert N, Weimer W (2008) Automated duplicate detection for bug tracking systems. In: IEEE
international conference on dependable systems and networks with FTCS and DCC, pp 5261
8. Pudi V, Krishna PR (2009) Data mining, 1/e. Oxford University Press, India. 2009.x
9. Zhai C, Lafferty J (2004) A study of smoothing methods for language models applied to
information retrieval. ACM Trans Inf Syst 22(2):179214
10. Wang X, Zhang L, Xie T, Anvik J, Sun J (2008) An approach to detecting duplicate bug reports
using natural language and execution Information. In: Proceedings of ICSE 2008, Leipzig,
Germany
11. Sun C, Lo D, Jiang XWJ, Khoo S-C (2010) A discriminative model approach for accurate
duplicate bug report retrieval. In: The Proceedings of ICSE 2010, Cape Town, South Africa

8 An Analytics-Driven Approach to Identify Duplicate Bug Records in Large. . .

187

12. Sun C, Lo D, Khoo S-C, Jiang J (2011) Towards more accurate retrieval of duplicate bug
records. In: The Proceedings of IEEE/ACM international conference of automated software
engineering, IEEE Computer Society, Washington, DC, pp 253262
13. Tian Y, Sun C, Lo D (2012) Improved duplicate bug report identification. In: Proceedings of
16th European conference on software maintenance and reengineering
14. Alipour A, Hindle A, Stroulia E (2013) A contextual approach towards more accurate
duplicate bug detection. In: Mining Software Repositories (MSR), IEEE Press Piscataway,
pp 183192
15. Sureka A, Jalote P (2010) Detecting duplicate bug report using character N-Gram-based
features. In: Proceedings of the Asia Pacific software engineering conference, pp 366374
16. Manning C, Raghavan P, Sch
utze H (2008) Introduction to information retrieval. Cambridge
University Press, New York
17. Salton G, Wong A, Yang CS (1975) A vector space model for automatic indexing. Commun
ACM 18(11):613620
18. Han J, Kamber M, Pei J (2011) Data mining: concepts and techniques: concepts and techniques. Elsevier, Amsterdam
19. Forgy EW (1965) Cluster analysis of multivariate data: efficiency versus interpretability of
classifications. Biometrics 21:768769
20. Dasarathy BV (1990) Nearest neighbor (NN), norms: NN pattern classification techniques
ISBN: 0-8186-8930-7
21. Shakhnarovish G, Darrell T, Indyk P (2005) Nearest-neighbour methods in learning and
vision. MIT Press, Cambridge. ISBN 0-262-19547-X
22. Robertson SE, Walker S, Jones S (1995) Okapi at TREC-3. In: Proceedings of the third text
retrieval conference Gaithersburg, USA, pp 109126

Part III

Big Data Tools and Analytics

Chapter 9

Large-Scale Data Analytics Tools: Apache

Hive, Pig, and HBase
N. Maheswari and M. Sivagami

Abstract The Apache Hadoop is an open-source project which allows for the
distributed processing of huge data sets across clusters of computers using simple
programming models. It is designed to handle massive amounts of data and has the
ability to store, analyze, and access large amounts of data quickly, across clusters of
commodity hardware. Hadoop has several large-scale data processing tools and
each has its own purpose. The Hadoop ecosystem has emerged as a cost-effective
way of working with large data sets. It imposes a particular programming model,
called MapReduce, for breaking up computation tasks into units that can be
distributed around a cluster of commodity and server class hardware and thereby
providing cost-effective horizontal scalability. This chapter provides the introductory material about the various Hadoop ecosystem tools and describes their usage
with data analytics. Each tool has its own significance in its functions in data
analytics environment.
Keywords Hadoop Pig Hive HBase HDFS MapReduce Data analytics
Big data

9.1

Introduction

Big data refers to large or complex poly-structured data, including video, text,
sensor logs, and transactional records, that flows continuously through and around
organizations. Traditional data processing applications are not enough to handle
such big volumes of data. So, there are various challenges like data capture, data
sharing, data analysis, storage, security, and visualization. Organizations across
public and private sectors have made a strategic decision to turn big data into
competitive advantage. Extracting the required information from big data is similar
to refining the business intelligence from transactional data. The big data requires
extract, transform, and load (ETL) infrastructure [1] to handle it efficiently.

N. Maheswari (*) M. Sivagami

School of Computing Science and Engineering, VIT University, Vandalur-Kelambakkam
Road, 600 127 Chennai, Tamil Nadu, India
e-mail: maheswari.n@vit.ac.in; msivagami@vit.ac.in
Springer International Publishing Switzerland 2016
Z. Mahmood (ed.), Data Science and Big Data Computing,
DOI 10.1007/978-3-319-31861-5_9

191

192

N. Maheswari and M. Sivagami

When the source data sets are large, fast, and unstructured, the traditional ETL is
not suitable, because it is too complex to develop and too expensive to operate and
takes too long to execute.
Yahoo, Google, Facebook, and other companies have extended their services to
web scale; the amount of data they collect routinely from user interactions online
would have besieged the capabilities of traditional IT architecture. In the interest of
advancing the development of core infrastructure components rapidly, they
released relevant code for many of the components into open source. Apache
Hadoop has emerged from these components, as the standard for managing
big data.
Apache Hadoop is an open-source distributed software platform for storing and
processing data. Written in Java, it runs on a cluster of industry-standard servers
configured with direct-attached storage. Using Hadoop, petabytes of data can be
stored reliably on tens of thousands of servers while scaling performance costeffectively by merely adding inexpensive nodes to the cluster. Apache Hadoop has
the distributed processing framework known as MapReduce (Fig. 9.1) that helps
programmers solve data-parallel problems by subdividing large data sets into small
parts and processing independently. The system splits the input data set into
multiple chunks, each of which is assigned a map task that can process the data
in parallel. Each map task reads the input as a set of key-value pairs and produces a
transformed set of key-value pairs as the output. The framework shuffles and sorts

Interfaces

ETL
Tools

Business
Intelligence

RDMBS

Abstractions
PIG

Distributed
Data
Processing

Clustered
storage
system

Hive

Mapreduce

Hadoop Distributed File System

(HDFS)

Fig. 9.1 Hadoop architecture

HBASE
Database
real-time
access

9 Large-Scale Data Analytics Tools: Apache Hive, Pig, and HBase

193

outputs of the map tasks, sending the intermediate key-value pairs to the reduce
tasks, which group them into final results. MapReduce uses JobTracker and
TaskTracker mechanisms to schedule tasks, monitor them, and restart any that fail.
The Apache Hadoop platform also includes the Hadoop Distributed File System
(HDFS), which is designed for scalability and fault tolerance. HDFS stores large
files by dividing them into blocks (usually 64 or 128 MB each) and replicating the
blocks on three or more servers. HDFS provides APIs for MapReduce applications
to read and write data in parallel. Capacity and performance can be scaled by adding
Data Nodes, and a single NameNode mechanism manages data placement and
monitors server availability. HDFS clusters may have thousands of nodes and
each can hold petabytes of data.
Apache Hadoop includes many other components that are useful for ETL along
with MapReduce and HDFS.
In the following sections, more details about the large-scale data analytics tools
have been provided. Section 9.2 describes the Hive compilation stages, Hive
commands, partitioning and bucketing in Hive, and Hive performance and challenges. Section 9.3 describes the Pig compilation stages, Pig commands, Pig userdefined function, Pig scripts and issues, and challenges in Pig. Section 9.4 highlights the HBase architecture, HBase commands, and HBase performance. Conclusions are given in Sect. 9.5.

9.2

Apache Hive

The size of data sets being collected and analyzed in the industry for business
intelligence is growing rapidly, making traditional warehousing solutions prohibitively expensive. Apart from ad hoc analysis [2] and business intelligence applications used by analysts across the company, a number of Facebook products are also
based on analytics.
The entire data processing infrastructure in Facebook was developed around a
data warehouse built using a commercial RDBMS. The data generated was growing
very fast for example, it grew from a 15 TB data to a 700 TB data set. The
infrastructure was so inadequate to process the data and the situation was getting
worse. So there was an urgent need for infrastructure that could scale along with
large data. As a result, the Hadoop technology has been explored to address the
scaling needs. The same jobs that had taken more than a day to complete could now
be completed within a few hours using Hadoop. But Hadoop was not easy for end
users, especially for those users who were not familiar with MapReduce. End users
had to write MapReduce programs for simple tasks like getting row counts or
averages. Hadoop lacked the expressiveness of popular query languages like
SQL, and as a result, users ended up spending hours to write programs for even
simple analysis. To analyze the data more productively, there must be a need to
improve the query capabilities of Hadoop. As a result, Hive was introduced in

194

N. Maheswari and M. Sivagami

January 2007. It aimed to bring the familiar concepts of tables, columns, partitions,
and a subset of SQL to the unstructured world of Hadoop, along with maintaining
the extensibility and flexibility of the Hadoop. Hive and Hadoop are used extensively now for different kinds of data processing. Hive was open-source and since
then has been used and explored by a number of Hadoop users for their data
processing needs.
Apache Hive is a Hadoop ecosystem tool. It acts as a data warehouse infrastructure built on Hadoop for providing data aggregation, querying, and analysis. Data
warehouse applications maintain large data sets and can be mined for analytics.
Most of the data warehouse applications use SQL-based relational databases. In
order to do data analytics using Hadoop, data has to be transferred from warehouse
to Hadoop. Hive simplifies the process of moving the data from data warehouse
applications to Hadoop.
Hive QL is an SQL-like language provided by Hive to interact with it. It
facilitates a mechanism to project the structure on the data and query the data.
The traditional MapReduce pattern can be connected in Hive as a plug-in component. Hive does not create Java MapReduce programs. It uses the built-in generic
mapper and reducer modules which are driven by an XML file containing the job
plan. The generic modules serve like small language interpreters, and the language to force the computation is encoded using XML.

9.2.1

Hive Compilation and Execution Stages

The CLI (command line interface) is the most popular way to use Hive. The CLI
allows statements to be typed one at a time. It also facilitates to run Hive scripts
which are collection of Hive statements.
Hive components are shell, thrift server, metastore, driver, query compiler, and
execution engine. The shell allows the interactive queries to execute in CLI [3]. A
thrift service provides remote access from other processes. Access is provided by
using JDBC and ODBC. They are implemented on top of the thrift service. All Hive
installations require a metastore service, which Hive uses to store table schemas and
other metadata. It is typically implemented using tables in a relational database. The
driver manages the life cycle of Hive QL statement during compilation, optimization, and execution. Query compiler is invoked by the driver when Hive QL
statement is given. It translates the query statement into directed acyclic graph of
MapReduce jobs.
The driver submits the DAG-based MapReduce jobs to the execution engine of
the Hive and is available in Hadoop. By default, Hive uses a built-in Derby SQL
server, which provides limited single process storage. For example, when using
Derby, two simultaneous instances of the Hive CLI cannot run. A simple web
interface known as Hive Web Interface provides remote access to Hive.

9 Large-Scale Data Analytics Tools: Apache Hive, Pig, and HBase

9.2.2

195

Hive Commands

Apache Hive works in three modes [4]: local, distributed, and pseudo distributed
modes. Local mode refers to the files from local file system and the jobs run all the
tasks in a single JVM instance. Distributed mode accesses the files from HDFS and
the jobs are run as several instances in the cluster. By default, Hive stores the table
data in file:///user/hive/warehouse for local mode and in hdfs://
nnamenode_server/user/hive/ warehouse for the distributed mode.
Pseudo distributed mode is almost identical to distributed, but it is a one-node cluster.
This section explores the data definition and data manipulation language parts of
Hive QL which is used to create and store data into the Hive tables, extract data into
the file system, and manipulate data with queries.
The syntax and examples have been provided for the Hive commands. These
commands can be executed in the hive > command line interface (CLI).
9.2.2.1

Databases

The database in Hive is essentially a catalog or namespace of tables [4]. Databases

are very useful for larger clusters with multiple teams and users and help in
avoiding table name collisions. It is also common to use databases to organize
production tables into logical groups. In the absence of database name, the default
database is used. The simplest syntax for creating a database is shown in the
following example:
CREATE (DATABASE|SCHEMA) [IF NOT EXISTS]
database_name
[COMMENT database_comment]
[LOCATION hdfs_path]
[WITH DBPROPERTIES
(property_nameproperty_value, . . .)];
create database mydb;
The uses of SCHEMA and DATABASE are interchangeable they mean the
same thing. CREATE DATABASE was added in Hive 0.6. The WITH
DBPROPERTIES clause was added in Hive 0.7.
The USE database command informs the Hive of the database where the data is
to be stored or accessed from. It helps to group the logical related tables to avoid the
name collision:
use database_name;
use mydb;
The SHOW command displays the list of databases available in Hive:
show databases;

196

9.2.2.2

N. Maheswari and M. Sivagami

Tables

The CREATE TABLE command creates a table with the given name. An error is
thrown if a table or view with the same name already exists. One can use IF NOT
EXISTS statement to skip the error. It should be noted that:
Table names and column names are case insensitive but SerDe and property
names are case sensitive.
Table and column comments are string literals (single quoted).
The TBLPROPERTIES clause tags the table definition with the users own
metadata key-value pairs. Some predefined table properties also exist, such as
last_modified_user and last_modified_time which are automatically added and
managed by Hive. Other predefined table properties include: TBLPROPERTIES
("comment" "table_comment") and TBLPROPERTIES ("hbase.table.
name" "table_name") see HBase Integration.
To specify a database for the table, either issue the USE database_name statement prior to the CREATE TABLE statement (in Hive 0.6 and later) or qualify the
table name with a database name ("database_name.table.name" in Hive 0.7 and
later). The keyword default can be used for the default database:
CREATE [TEMPORARY] [EXTERNAL] TABLE [IF NOT EXISTS]
[db_name.]table_name -- (Note: TEMPORARY available in Hive
0.14.0and later)
[(col_namedata_type [COMMENT col_comment], . . .)]
[COMMENT table_comment]
[PARTITIONED BY (col_namedata_type [COMMENT
col_comment], . . .)]
[CLUSTERED BY (col_name, col_name, . . .) [SORTED BY
(col_name [ASC|DESC], . . .)] INTO num_buckets BUCKETS]
[ROW FORMAT row_format]
[STORED AS file_format]
[LOCATION hdfs_path]
[TBLPROPERTIES (property_nameproperty_value,
. . .)] -- (Note: Available in Hive 0.6.0 and
later)
[AS select_statement]; -- (Note: Available in
Hive 0.5.0and later; not supported for external
tables)
CREATE [TEMPORARY] [EXTERNAL] TABLE
[db_name.]table_name
LIKE existing_table_or_view_name
[LOCATION hdfs_path];
[ROW FORMAT row_format]
[STORED AS file_format]

[IF

NOT

EXISTS]

9 Large-Scale Data Analytics Tools: Apache Hive, Pig, and HBase

197

Use STORED AS TEXTFILE [4] if the data needs to be stored as plain text files.
TEXTFILE is the default file format. Use STORED AS SEQUENCEFILE if the
data needs to be compressed. Use STORED AS ORC if the data needs to be stored
in ORC file format. The Optimized Row Columnar (ORC) file format provides an
efficient way to store Hive data. Use ROW FORMAT SERDE for the serialization:
CREATE TABLE IF NOT EXISTS mydb.employees (
name STRING COMMENT Employee name,
salary FLOAT COMMENT Employee salary,
subordinates ARRAY<STRING>COMMENT Names of
subordinates,
deductionsMAP<STRING, FLOAT>
COMMENT Keys are deductions names, values are
percentages,
address STRUCT<street:STRING, city:STRING,
state:STRING, zip:INT>
COMMENT Home address)
COMMENT Description of the table
TBLPROPERTIES (creatorme, created_at201201-02 10:00:00, . . .)
LOCATION /user/hive/warehouse/mydb.db/employees;
The SHOW TABLES command lists all the tables and views in the current
working database:
show tables;
The DESCRIBE table command displays the structure of the table:
describe table table_name
describe table employees.

9.2.2.3

Loading Data into Table

The LOAD command loads the data either from local file system or from HDFS to
the Hive warehouse:
Load data local /path of a file/ overwrite into
table
table_ name.
load data local inpath /home/ponny/mar.txt
overwrite into table dt;

198

9.2.2.4

N. Maheswari and M. Sivagami

Retrieving Data from Table

The SELECT..CASE command helps to retrieve data based on the various conditions specified in each case:
select fieldnames from tablename where condition
select count(*), avg(salary) FROM employees;
select * FROM employees WHERE country US AND
state CA;
select fieldnames case condition then action . . .. end
from table name
select *, case when c>60 then first
when c<60 then second
else third
end from ma;

9.2.2.5

Drop Command for Database and Tables

The DROP command deletes the database/table from Hive:

DROP (DATABASE|SCHEMA) [IF EXISTS] database_name
[RESTRICT|CASCADE];
drop database mydp;
drop table tablename
drop table employee;

9.2.3

Partitioning and Bucketing

A simple query in Hive reads the entire data set even with the WHERE clause filter.
This becomes a bottleneck for running MapReduce jobs over a large table
[5, 6]. This issue can be overcome by implementing partitions in Hive. Hive
makes it very easy to implement partitions by using the automatic partition scheme
when the table is created. It organizes tables into partitions, a way of dividing a
table into related parts based on the values of partitioned columns such as date, city,
and department. Using partition, it is easy to query a portion of the data.
In Hives implementation of partitioning, data within a table is split across
multiple partitions. Each partition corresponds to a particular value(s) of partition
column(s) and is stored as a sub-directory within the tables directory on HDFS.
Though the selection of partition key is always a sensitive decision, it should always

9 Large-Scale Data Analytics Tools: Apache Hive, Pig, and HBase

199

be a low cardinal attribute, e.g., if the data is associated with time dimension, then
date could be a good partition key. Similarly, if data has an association with
location, e.g., a country or state, then it is a good idea to have hierarchical partitions
like country/state. When the table is queried and if it has partitions, only the
required partitions of the table are queried, thereby reducing the I/O and time
required by the query. Partitioning can carry out using static, dynamic, or hybrid
mode. Data type of the partitioning column has to be able to be converted to a string
in order to be saved as a directory name in HDFS.

9.2.3.1

Dynamic Partitioning

Dynamic partitioning has been done based on the attribute/attributes and not the
value. Table is dynamically partitioned based on the value of the corresponding
attribute in the insertion phase. The following steps are required for the partition.

Non-partitioned Table Creation

create table dt(a int,b string,c int,,d int,e int)
row format delimited fields terminated by ,
stored as textfile;

Load Data into Non-partitioned Table

load data local inpath /home/ponny/mar.txt
overwrite into table dt;

Dynamic Partition Table Creation

create table dtp(b string, c int, d int, e int) partitioned by
(a int) row format delimited fields terminated by , stored
as text file;

Flag Setting
set hive.exec.dynamic.partition true;
set hive.exec.dynamic.partition.mode nonstrict;

200

N. Maheswari and M. Sivagami

Load Data into Partitioned Table

insert overwrite table dtp partition(a) select
b,c,d,e,a from dt;
Here, the table dtp is partitioned by the attribute a. It will create n partitions
based on the total number of unique values of the attribute a.

9.2.3.2

Static Partitioning

Static partitioning is done based on the value/values of attribute/attributes. Table is

partitioned based on the value of the partitioned attribute in the insertion phase. The
following steps are required for the partition.

Non-partitioned Table Creation

create table dt(a int,b string,c int,,d int,e int)
row format delimited fields terminated by ,
stored as textfile;

Load Data into Non-partitioned Table

load data local inpath /home/ponny/mar.txt
overwrite into table dt;

Static Partition Table Creation

create table stp(a int,,b string,,d int,e int)
partitioned by (c int) row format delimited fields
terminated by , stored as textfile;

Flag Setting
set hive.exec.dynamic.partition.mode strict;

9 Large-Scale Data Analytics Tools: Apache Hive, Pig, and HBase

201

Load Data into Partitioned Table

insert overwrite table stp partition(c 90)
select a,b,d,e from dt where c 90;
Here the table stp is partitioned by the value of the attribute c. It will create a
partition based on the given value of the attribute c. It will create a directory with
the name c 90, and the partitioned values are stored in the file inside this
directory.

9.2.3.3

Hybrid Partitioning

Hybrid partitioning is a combination of dynamic and static partitioning. It should

have a minimum of one attribute pertaining to static partition and it must be
mentioned first in the order. If there are multiple partitioning columns, their order
is significant, since that translates to the directory structure in HDFS. The following
steps are required for the partition.

Non-partitioned Table Creation

create table dt(a int,b string,c int,,d int,e int)
row format delimited fields terminated by ,
stored as textfile;

Load Data into Non-partitioned Table

load data local inpath /home/ponny/mar.txt
overwrite into table dt;

Hybrid Partition Table Creation

create table stpy(a int,b string,e int) partitioned
by (c int,d int) row format delimited fields
terminated by , stored as textfile;

Flag Setting
set hive.exec.dynamic.partition.mode strict;

202

N. Maheswari and M. Sivagami

Load Data into Partitioned Table

insert overwrite table stpy partition(c 90,d)
select a,b,e,d from dt where c 90;
Here the table stpy is partitioned based on the unique values of d for the
records which have the value of c 90. For example, a table named Emp1 contains
employee data such as id, name, dept, and yoa (i.e., year of appointment). In order
to retrieve the details of all employees who joined in 2011, a query searches the
whole table for the required information. However, if the employee data is dynamically partitioned with the year and stored in a separate file, it reduces the query
processing time. The following example shows how to partition a file and its data:
The following file contains employee data table Emp1:
/tab1/employeedata/file1
id, name, dept, yoa
1, anbu, TP, 2011
2, jeeva, HR, 2011
3, kunal,SC, 2012
4, Praba, SC, 2012
The above data is partitioned into two files using year:
/tab1/employeedata/2011/file2
1, anbu, TP, 2011
2, jeeva, HR, 2011
/tab1/employeedata/2012/file3
3, kunal,SC, 2012
4, Praba, SC, 2012
In Hive, as data is written to disk, each partition of data will be automatically
split into different folders. During a read operation, Hive will use the folder
structure to quickly locate the write partitions and also return the partitioning
columns as columns in the result set. This can dramatically improve query performance, but only if the partitioning scheme reflects common filtering. Partitioning
feature is very useful in Hive; however, a design that creates too many partitions
may optimize some queries but damages other important queries. The other drawback is too many partitions lead to the large number of unnecessary Hadoop files
and directories. This creates overhead on the NameNode since it must keep all
metadata for the file system in memory. If the number of partitions rises above the
certain threshold, it can be run into out of memory errors when MapReduce jobs
are being generated. In this condition, even simple select statement can fail. In order
to solve this issue, Java heap size can be tuned/partitioning [7] and the scheme can
be modified to result as fewer partitions.

9 Large-Scale Data Analytics Tools: Apache Hive, Pig, and HBase

9.2.3.4

203

Bucketing

Bucketing is another technique [8] for decomposing data sets into more manageable
parts. For example, suppose a table using the date as the top-level partition and the
employee id as the second-level partition leads to too many small partitions. To
overcome this, Hive provides bucketing concept. Bucketing has several advantages.
The number of buckets is fixed so it does not fluctuate with data. If two tables are
bucketed by employee_id, Hive can create a logically correct sampling. Bucketing also
aids in doing efficient map-side joins, etc. Bucketing happens by using a hash algorithm
and then a modulo on the number of buckets. So, a row might get inserted into any of the
buckets. If the employee table had the buckets and employee_id had used as the
bucketing column, the value of this column will be hashed by a user-defined number
into buckets. Records with the same employee_id will always be stored in the same
bucket. Assuming the number of employee_id is much greater than the number of
buckets, each bucket will have huge employee_ids. Creating table, like CLUSTERED
BY (employee_id) INTO XX BUCKETS, can be specified, where XX is the number of
buckets. Bucketing can be used for sampling of data, as well as for joining two data sets
much more effectively. It reduces the I/O scans during the join process if the process is
happening on the same keys (columns). It also reduces the scan cycles to find a
particular key because bucketing ensures that the key is present in a certain bucket.
Example
Flag setting
Set hive.enforce.bucketing true;
Creation of buckets
create table dtb(a int,b string, c int, d int, e int)
clustered by (d) into 3 buckets row format delimited fields
terminated by , stored as textfile;
Data insertion
insert overwrite table dtb select a,b,c,d,e from dt;
Basically, both partitioning and bucketing slice the data for executing the query
much more efficiently than on the non-sliced data. The major difference is that the
number of slices will keep on changing in the case of partitioning as data is
modified, but with bucketing, the number of slices is fixed which are specified
during table creation.

9.2.4

External Table

In the managed table [3, 4], data is loaded using LOAD or INSERT command. Hive
will create appropriate directory and copy files into those directories. When
dropping those tables, Hive removes directory and data files both because Hive
owns data. So they are called Hive managed table. Here is an example:

204

N. Maheswari and M. Sivagami

create external table det(a int, c int, b string,d

int) row format delimited fields terminated by ,
location /user/ms/extq;
It is very common that data analytics applications need to use data within Hive
which is owned by some other. In that case, the data file is already there in HDFS
directory but it is owned and used by some other tool like Pig. In the case of
managed table, it is necessary to copy data into Hive warehouse. So it would be
better if there is a facility to map that data into Hive, perform some queries, prepare
output, remove mapping, and then leave data at the original place as it was. For this
task, Hive provides external table concept. In order to create external table, the
keyword external is to be mentioned in create table command. In external tables,
Hive assumes that it does not own data or data files. It means that when an external
table is dropped, Hive will remove metadata about external table but will leave
table data as it was.

9.2.5

Hive Performance

Hive is most suited for data warehouse applications, where relatively static data is
analyzed, fast response times are not required, and the data is not changing rapidly.
Besides, Hive has its own challenges: it does not support transactions since it is not
a full database. It does not provide record-level update, insert, or delete. New tables
can be created from queries or output query results to files.
Input formats play a critical role in Hive performance. For example, JSON, the
text type of input formats, is not a good choice for a large production system where
data volume is really high. These types of readable formats actually take a lot of
space and have some overhead of parsing (e.g., JSON parsing). To address these
problems, Hive comes with columnar input formats like RCFile, ORC, etc. Columnar formats reduce the read operations in analytics queries by allowing each column
to be accessed individually.
Hadoop can execute MapReduce jobs in parallel, and several queries executed
on Hive automatically use this parallelism. However, single, complex Hive queries
are normally translated to a number of MapReduce jobs that are executed by default
in a sequence. Often though, some of a querys MapReduce stages are not
interdependent and could be executed in parallel. They can take advantage of
spare capacity on a cluster and improve cluster utilization while at the same time
reducing the overall query execution time. Vectorization allows Hive to process a
batch of rows together instead of processing one row at a time [9]. Each batch
consists of a column vector which is usually an array of primitive types. Operations
are performed on the entire column vector, which improves the instruction pipelines
and cache usage.
Hive adds extensions to provide better performance in the context of Hadoop
and to integrate with external programs. It provides a familiar programming model

9 Large-Scale Data Analytics Tools: Apache Hive, Pig, and HBase

205

for people who know SQL and it also eliminates lots of hassle to code in Java. Hive
makes it easier for developers to port SQL-based applications to Hadoop, compared
with other Hadoop languages and tools. It focuses primarily on query part of SQL
and better reflects the underlying MapReduce process. Pig is a simple scripting
language and helps to write complex MapReduce transformations easily.

9.3

Apache Pig

Hadoop MapReduce has high coding complexity and it is highly favored with the
programmers with high Java programming skills. Here, when the input file size
increases, the execution time for MapReduce also increases proportionally. So there
is a requirement for a platform which supports ease of programming as well as
enhanced extensibility. In this context, Apache Pig reveals itself as a novel
approach in data analytics. Pig was initially developed at Yahoo [10] and later it
was open sourced. It allows people using Hadoop to focus more on analyzing large
data sets and spend less time to write mapper and reducer programs. Apache Pig is a
platform that supports substantial parallelization which enables the processing of
huge data sets.
Pig is an open-source high-level data processing language. This high-level data
processing language is known as Pig Latin which simplifies programming because
of ease of expressing the code in Pig Latin. Pig Latin can work on schema-less or
inconsistent environments and can operate on the available data as soon as it is
loaded into the HDFS. The Pig compiler compiles the script in a choice of
evaluation mechanisms and utilizes optimization opportunities in the script. This
avoids tuning the program manually. The progress in Pig compiler makes the Pig
Latin program get an automatic speedup. Pig Latin defines a set of transformations
on a data set such as aggregate, join, and sort. The statement written in Pig script is
translated into MapReduce so that it can be executed within Hadoop. Pig Latin
describes a directed acyclic graph (DAG) where the edges are data flows and nodes
are operators that process the data. Pig Latin stores the data after the execution in
HDFS. It can be extended using user-defined functions (UDFs), in which the user
can write the customized functions in Java, Python, etc., and those functions can be
invoked directly from the Pig Latin.

9.3.1

Modes of User Interaction with Pig

Pig has two execution types:

Local mode: To run Pig in local mode, users have access to a single machine. All
files are installed and run using the local host and file system. Local mode can be
specified using the -x flag (pig -x local). Local mode does not support parallel
mapper execution with Hadoop.

206

N. Maheswari and M. Sivagami

MapReduce mode: To run Pig in MapReduce mode, users need a setup of

Hadoop cluster with HDFS installation. MapReduce mode is the default mode
and users do not need to specify it using the -x flag (Pig or Pig -x MapReduce).
Pig allows three modes of user interaction:
Interactive mode: The user is presented with an interactive shell called Grunt,
which accepts Pig commands.
Batch mode: In this mode, a user submits a pre-written script called Pig script,
containing a series of Pig commands, typically ending with STORE. The
semantics are similar to interactive mode.
Embedded mode: Pig Latin commands can be submitted through method invocations from a Java program. This option permits dynamic construction of Pig
Latin programs, as well as dynamic control flow.

9.3.2

Pig Compilation and Execution Stages

A Pig script has different stages before being executed, as shown in Fig. 9.2. These
stages are described below.

9.3.2.1

Parsing

The role of the parser [11] is to verify that the program is syntactically correct and
whether all referenced variables are defined. Type checking and schema inference
can also be done by parser. Other checks, such as verifying the ability to instantiate
classes corresponding to user defined, also occur in this phase. A canonical logical
plan with one-to-one correspondence between Pig Latin statements and logical

PIG
Parse
Pig Latin Scripts

Mapreduce
Statements

Compile
&
Optimize
HDFS
Plan

Fig. 9.2 Pig compilation and execution stages

9 Large-Scale Data Analytics Tools: Apache Hive, Pig, and HBase

207

operators which are arranged in directed acyclic graph (DAG) is the output of the
parser. Logical optimizer optimizes the logical plan generated by the parser. In this
stage, logical optimizations such as projection and pushdown are carried out.

9.3.2.2

Compile and Optimize

The optimized logical plan is compiled into a series of MapReduce jobs, which then
pass through another optimization phase. An example of MapReduce-level optimization is utilizing the MapReduce combiner stage in order to perform early partial
aggregation [2], in the case of distributive or algebraic aggregation functions.

9.3.2.3

Plan

Topologically, the DAG of optimized MapReduce jobs is sorted and then submitted
to Hadoop HDFS for execution in that order. Pig (Hadoop job manager) usually
monitors the Hadoop execution status, and the user periodically gets the reports on
the progress of the overall program. Warnings or errors that arise during execution
are logged and reported to the user.

9.3.3

Pig Latin Commands

Pigs interactive shell Grunt allows to execute Pig commands manually. It is

typically used for ad hoc data analysis and the Grunt shell supports some basic
utility commands. Typing help will print out a help screen of such utility commands. The command quit can be used to exit the Grunt shell.
Pig Latin statements are the basic constructs to process data using Pig. A Pig
Latin statement is an operator that takes a relation as input and produces another
relation as output (except LOAD and STORE which read data from and write data
to the file system).
A relation can be defined as follows:

A relation is a bag (more specifically, an outer bag).

A bag is a collection of tuples.
A tuple is an ordered set of fields.
A field is a piece of data.

The names (aliases) of relations, Pig Latin functions, and fields are case sensitive.
The names of parameters and all other Pig Latin keywords are also case insensitive.
Fields are referred to by positional notation or by name (alias). Positional notation is
generated by the system. Positional notation is indicated with the dollar sign ($) and
begins with zero (0), for example, $0, $1, and $2. Pig Latin statements may include

208

N. Maheswari and M. Sivagami

expressions and schemas. It can be extended to multiple lines and must end with a
semicolon. Basically, Pig Latin statements are processed using multi-query execution.

9.3.3.1

LOAD Command

This command places the data from a file into a relation. It uses the PigStorage load
function as default with USING option unless specified. The data can be given a
schema name using the AS option:
alias LOAD file [USING function] [AS schema];

9.3.3.2

DUMP Command

This helps to execute the statements mentioned above DUMP [12]. It displays the
content of a relation. It helps in debugging. The relation should be small enough for
printing on screen. The LIMIT operation can be applied on an alias to make sure it
is small enough for display:
DUMP alias;

9.3.3.3

STORE Command

This helps to place the data from a relation into a directory. The directory must not
exist when this command is executed. Pig will create the directory and store the
relation in files named part-nnnnn in it. It uses the PigStorage store function as
default with the USING option unless specified:
STORE alias INTO directory [USING function];
In the following example, the names (aliases) of the relations are A and B:
A load student using PigStorage() AS
(name:chararray, age:int,gpa:float);
B limit A 4;
dump B;
(Ashu,18,5.0F)
(Maya,19,5.8F)
(Belly,20,7.9F)
(Joney,18,8.8F)
The limit command allows to specify how many tuples (rows) are to be
returned back.

9 Large-Scale Data Analytics Tools: Apache Hive, Pig, and HBase

9.3.3.4

209

DESCRIBE Command

It displays the schema of a relation:

describe alias;
describe A;

9.3.3.5

ILLUSTRATE Command

It displays step-by-step process of how data is transformed, starting with a load

command, to arrive at the resulting relation. To keep the display and processing
manageable, only a (not completely random) sample of the input data is used to
simulate the execution:
illustrate alias;
illustrate A;

9.3.3.6

Expressions

In Pig Latin, expressions are language constructs used with the FILTER,
FOREACH, UNION, GROUP, and SPLIT operators. The most salient characteristic about Pig Latin as a language is its relational operators. These operators, as
mentioned in the following subsections, define Pig Latin as a data processing
language.

9.3.3.7

UNION Command

This combines multiple relations together, whereas SPLIT partitions a relation into
multiple ones:
a load A using PigStorage(,) as (a1:int,
a2:int, a3:int);
b load B using PigStorage(,) as (b1:int,
b2:int, b3:int);
c UNION a, b;
dump c;

9.3.3.8

SPLIT Command

The SPLIT operation on relation s sends a tuple to f1 if its third field ($2) is greater
than 4000 and to f2 if it is not equal to 10,000. It is possible to write conditions such
that some rows will go to both f1 and f2 or to neither:

210

N. Maheswari and M. Sivagami

split s into f1 if $2>4000,f2 if $2 ! 10000;

dump f1;
dump f2;

9.3.3.9

FILTER Command

The FILTER operator trims a relation down to only tuples that pass a certain
condition:
p filter s by $2>5000;
dump p;
SPLIT can be simulated by multiple FILTER operators.

9.3.3.10

GROUP Command

When grouping a relation, the result will be a new relation with two columns,
group and the name of the original relation. The group column has the schema of
what it is grouped by. The new relations contain the tuples with the same group key.
Group key is the attribute on which the grouping has been performed:
l group s by $1;
dump l;
The group operator in Pig is a blocking operator and forces the Hadoop
MapReduce job. All the data is shuffled, so that rows in different partitions that
have the same grouping key group together. So grouping has overhead.

9.3.3.11

FOREACH Command

This can be used to project specific columns of a relation into the output. Arbitrary
expressions also can be applied on specific columns to produce the required output.
Nested form of FOREACH facilitates to do more complex processing of tuples:
h foreach s generate a*b;
dump h;

9.3.4

Pig Scripts

Pig scripts can run inside the Grunt shell and can be useful in debugging Pig scripts.
Pig Latin scripts are about combining the Pig Latin statements together. Pig scripts

9 Large-Scale Data Analytics Tools: Apache Hive, Pig, and HBase

211

can be executed in the Grunt shell using exec and run commands. The exec
command executes a Pig script in a separate space from the Grunt shell. The
command run executes a Pig script in the same space as Grunt (also known as
interactive mode [13]). It has the same effect as manually typing in each line of the
script into the Grunt shell. Pig Latin programs can be written in a sequence of steps
where each step is a single high-level data transformation. Pig Latin scripts are
generally structured as follows:

A LOAD statement to read data from the file system.

A series of transformation statements to process the data.
A DUMP statement to view results or a STORE statement to save the results.
DUMP or STORE statement is required to generate output.

9.3.4.1

ct.pig

u load /home/ponny/mar.txt as (f1:chararray);

y foreach u generate FLATTEN(TOKENIZE(f1));
r group y all;
c foreach r generate COUNT($1);
dump c;

9.3.4.2

comp.pig

u load /home/ponny/mar.txt as (f1:chararray);

y1 foreach u generate FLATTEN(TOKENIZE(f1)) as
word;
j filter y1 by wordabc;
k group j all;
l foreach k generate COUNT(j.$0);
dump l;
The left-hand side of equal sign in the Pig Latin commands refers to the relation
but not a single row; and on the right-hand side, the required operations can be done
for each row. The ct.pig script counts the total number of words in a file. The comp.
pig counts the occurrence of the particular word abc in a file. The script always
finds the count of the word abc. The script can be made to work for any userspecific word by parameter substitution. The third line of the above comp.pig script
can be replaced by the following statement for parameter substitution. For naming the
parameters in Pig Latin, the parameter name should start with the dollar ($) symbol:
j filter y1 by word$inp;
Here, $inp is the placeholder for the word entered by the user at the time of
execution. The variable name $inp in the above command is enclosed with single

212

N. Maheswari and M. Sivagami

quotes so that the word abc in the following execution command does not have the
quotes:
exec -param inp abc comp.pig

9.3.5

User-Defined Functions (UDFs) in Pig

Pig provides UDF as a piggy bank repository [14, 15]. These allow the users to
define their own user-defined functions for their requirement. UDFs can be
implemented using three languages, Java, Python, and Java script. Pig provides
extensive support for Java functions and limited support for other two languages. In
Java, UDFs can be developed using accumulator and algebraic interface.

9.3.5.1

Predefined UDF

Pig scripts can access the UDFs from the piggy bank repository. The sample script
using the same has been given below:
register
/home/ponny/hadoop/pig/pig/contrib/piggybank/java/
piggybank.jar;
define
Upper.org.apache.pig.piggybank.evaluation.string.
UPPER();
u load /home/ponny/mar.txt as (f1:chararray);
l foreach u generate Upper($0);

9.3.5.2

Customized Java UDFs

The steps below describe the user-defined uppercase conversion function using Java
UDF.
Step 1: Write a Java function, e.g.,
upper.java
package myudf;
import java.io.IOException;
import org.apache.pig.EvalFunc;
import org.apache.pig.data.Tuple;
public class upper extends EvalFunc<String>
{
public String exec(Tuple input) throws IOException
{
if (input null || input.size() 0)

9 Large-Scale Data Analytics Tools: Apache Hive, Pig, and HBase

213

return null;
try{
String str (String)input.get(0);
return str.toUpperCase();
}
catch(Exception e){
throw new IOException("Caught exception processing
input row", e);
}}}
Step 2: Compile the Java function, e.g.,
cd myudf
javac -cp /home/ponny/hadoop/pig/pig/pig-0.11.1.jar
upper.java
Step 3: Make a jar file, e.g.,
cd
jar -cf myudfs.jar myudf
Step 4: Develop a Pig script (s.pig) which uses the above UDF, e.g.,
s.pig
register myudfs.jar;
a load /home/ponny/myudf/b.csv using
PigStorage(,) as (l:chararray,c:int);
k foreach a generate myudf.upper(l);
dump k;
Pig is a higher-level data processing layer on top of Hadoop. Its Pig Latin
language provides programmers a more easy way to specify data flows. It supports
schemas in processing structured data, yet it is flexible enough to work with
unstructured text or semi-structured XML data. It is extensible with the use of
UDFs. It simplifies two aspects of MapReduce programming data joining and job
chaining.
Pig gains popularity in different areas of computational and statistical analysis. It
is widely used in processing weblogs and wiping off corrupted data from records.
Pig helps to build the behavior prediction models based on the user interactions
with a website. This feature can be used in displaying ads and news stories as per
the users choice. It can also be updated as the behavior prediction model with the
users data. This feature is largely used in social networking sites. When it comes to
small data or scanning multiple records in random order, Pig cannot be considered
as effective as MapReduce. It reduces the time consumed for working on the
complex programming constructs for implementing a typical MapReduce paradigm
using Java but still has the advantages of MapReduce task division method with the
help of the layered architecture and inbuilt compilers of Apache Pig. Hive and Pig

214

N. Maheswari and M. Sivagami

were designed for ad hoc batch processing of potentially large amount of data by
leveraging MapReduce. To store enormous amount of data, HBase can be used as a
key-value data store with low latency than HDFS.

9.4

Apache HBase

Hadoop works with various data formats such as arbitrary, semi-, or even unstructured and accesses them only in a sequential manner. Hadoop has to search the
entire data set even for the simplest jobs. A huge data set when processed results in
another huge data set, which should also be processed sequentially. So, a new
technique is required to access any point of data in a single unit of time in Hadoop
environment. HDFS in Hadoop is optimized for streaming access of large files. The
files stored in HDFS are accessed through MapReduce to process them in batch
mode. HDFS files are write-once files and read-many files. There is no concept of
random writes in HDFS and it also doesnt perform random reads very well.
HBase as an open-source project [16] provides the solution for it. It leverages the
fault tolerance provided by the Hadoop Distributed File System (HDFS). HBase is a
part of the Hadoop ecosystem and it provides random real-time read/write access to
data in the Hadoop File System.
It is a distributed column-oriented database, built on top of the Hadoop file
system, and horizontally scalable. Other random access databases are Cassandra,
CouchDB, Dynamo, MongoDB, etc. Column-oriented databases are those that store
data tables as sections of columns of data, rather than as rows of data. Facebook,
Twitter, Yahoo, Netflix, StumbleUpon, and Adobe use HBase internally to process
data. Data model of HBase is similar to Googles big table [17] and it stores data as
key-value pairs. HBase allows low latency access to small amounts of data from
within a large data set. Single rows can be accessed quickly from a billion row table.
HBase has flexible data model to work with and data is indexed by the row key. It
scans faster across the tables. Scaling is in terms of writes as well as total volume
of data.
The table schema defines only column families, which are the key-value pairs. A
table has multiple column families and each column family can have any number of
columns. Subsequent column values are stored contiguously on the disk. Each cell
value of the table has a time stamp. It provides data replication across clusters. In
Table 9.1, employee and occupation are shown as column families.

Table 9.1 Employee details

Row key
Eid
1
2

Employee
Name
Anu
Keen

Department
Sales
Marketing

Occupation
Designation
Manager
Associate manager

Salary
$12,000
$10,000

9 Large-Scale Data Analytics Tools: Apache Hive, Pig, and HBase

215

Master
Zookeeper

Region
Servers
memstore

HDFS

HFile

Fig. 9.3 HBase architecture

9.4.1

HBase Architecture

HBase architecture [18] as shown in Fig. 9.3 has master server and region servers as
major components. Region servers can be added or removed as per requirement.

9.4.1.1

Region Server

Regions are tables. They are split up and spread across the region servers. Region
servers have regions and stores. The regions in region servers handle data-oriented
operations by communicating with the client. The region server also performs read
and write requests for all the regions. It finalizes the size of the region, based on the
region size thresholds.
The store in the region server has memory store and HFiles. Memstore in store is
similar to a cache memory. Initially, the data is stored in memstore and transferred
to HFiles as blocks. Later, the memstore is flushed.

9.4.1.2

Master Server

The master server takes help from Apache Zookeeper to handle the load balancing
of region servers. It assigns regions to the region servers and unloads the busy
servers by shifting the regions to less occupied servers. It maintains the state of the
clusters. It takes the responsibility of schema changes and other metadata operations like creation of tables and column families.

216

9.4.1.3

N. Maheswari and M. Sivagami

Zookeeper

Zookeeper is an open-source project [19] and acts as a reliable centralized distributed coordinator that provides services like maintaining configuration information,
naming, providing distributed synchronization, etc. Zookeeper has transient nodes
representing different region servers. Master servers use these nodes to discover
available servers. The nodes are also used to track server failures or network
partitions. Clients communicate with region servers through Zookeeper. HBase
itself takes care of Zookeeper in pseudo and stand-alone modes.

9.4.2

HBase Commands

The command bin/start-hbase.sh script [20, 21] is used to start all HBase
daemons. HBase has a shell that helps to communicate with HBase. The instance of
HBase can be connected using the HBase shell command ($ ./bin/hbase
shell). The HBase shell prompt ends with a > character. The HBase shell
prompt will be hbase(main):001:0>.
Some of the commands and their output are discussed below.

9.4.2.1

Table Creation

Use the CREATE command to create a new table. Table and the column family
names must be specified:
hbase(main):001:0>create test1, cf1 (test table name
cf1- column family name)
0 row(s) in 0.4170 seconds
>Hbase::Table test1

9.4.2.2

List

Display the information about the created table test1 using the LIST command:
hbase(main):002:0>list test1
TABLE
test1
1 row(s) in 0.0180seconds
>["test1"]

9 Large-Scale Data Analytics Tools: Apache Hive, Pig, and HBase

9.4.2.3

217

Put

Write data into the table using the PUT command:

hbase(main):002:0>put test1, row1, cf1:x, xy
0 row(s) in 0.080seconds
hbase(main):003:0>put test1, row2, cf1:y, sb
0 row(s) in 0.01seconds
Here, two values have been inserted. The first insert is at row1, column cf1:x,
with a value of xy. Columns in HBase are comprised of a column family prefix,
cf1 in this example, followed by a colon and then a column qualifier suffix, x in this
case.

9.4.2.4

Get

This command helps to get a single row of data at a time:

hbase(main):006:0>get test1, row1
COLUMN CELL
cf:x timestamp1421762485768, valuexy
1 row(s) in 0.035seconds

9.4.2.5

Scan

Another way to get data from HBase is with the SCAN command. This command
can be used to scan all the rows from the table. The scan can be limited for a number
of records:
hbase(main):006:0>scan test1
ROW COLUMNCELL
row1 columncf1:x, timestamp1421762485768,
valuexy
row2 columncf1:y, timestamp1421762491785, valuesb
2 row(s) in 0.0230seconds

9.4.2.6

Disable/Enable a Table

Table can be disabled or enabled, using the DISABLE or ENABLE command:

hbase(main):008:0>disable test1
0 row(s) in 1.1820seconds
hbase(main):009:0>enable test1
0 row(s) in 0.1770seconds

218

N. Maheswari and M. Sivagami

9.4.2.7

Drop

To delete a table, this command can be used:

hbase(main):011:0>drop test1
0 row(s) in 0.137seconds

9.4.2.8

Exit/Quit

To quit the HBase shell and disconnect from the cluster, the QUIT command is
available. But HBase will be running in the background.

9.4.2.9

Stop

In the same way that the bin/start-hbase.sh script is used to start all HBase
daemons, the bin/stop-hbase.sh script stops them:
$ ./bin/stop-hbase.sh
stoppinghbase....................
$
After issuing this command, it can take several minutes for the processes to shut
down. Use the jps to be sure that the HMaster and HRegionServer processes are
shut down.
HBase is more suitable for the applications with large amounts of data which
require frequent random read/write operations. HBase does not use Hadoops
MapReduce capabilities directly, though it can integrate with Hadoop to serve as a
source or destination of MapReduce jobs. HBase is designed to support queries of
massive data sets and is optimized for read performance. For writes, HBase seeks to
maintain consistency. In contrast to eventually consistent [17], HBase does not
offer various consistency level settings. Due to HBases strong consistency, writes
can be slower. It prevents data loss from cluster node failure via replication. Hadoop
MapReduce is used together with HBase to process data in an efficient way. HBase is
used in the organizations with multi-petabyte databases, running them as missioncritical data stores. It also helps for click stream data storage and time series analysis.

9.5

Conclusion

The amount of existing data for processing is constantly increasing and becomes
more diverse. This chapter provides an insight into the various tools that can help in
analyzing big data. Apache Hadoop is one of the widely used software libraries to
perform large-scale data analytics tasks on clusters of computers in parallel.

9 Large-Scale Data Analytics Tools: Apache Hive, Pig, and HBase

219

Pig is a high-level data flow language and serves as an execution framework for
parallel computation. Hive acts as a data warehouse infrastructure that provides
data summarization and ad hoc querying. HBase is a NOSQL, scalable, distributed database that supports structured data storage for large tables. HBase is for
key-value data store and retrieve, but queries cannot be performed on rows
(key-value pairs). So Hive can be used as a query layer to an HBase data store
Pig also offers a way of loading and storing HBase data. These tools provide the
major functionality to store both structured and unstructured data and perform
refined processing and analysis. These tools also play pivotal role in designing
scalable data analytics system. This chapter describes how the acquisition of large
amounts of data is analyzed with the help of data analytics tools for further
process.

References
1. Intel (2013) White paper: extract, transform and load Big data with Hadoop. Available at:
hadoop.intel.com. Accessed 30 July 2015
2. Ashish et al (2010) Hive a Petabyte scale data warehouse using hadoop. IEEE International
Conference on Data Engineering, November 2010
3. Edward C et al (2012) Programming Hive. OReilly Media Inc, Sebastopol
4. Apache (2014) Language manual. Available at: https://cwiki.apache.org/confluence/display/
Hive/LanguageManualDDL#LanguageManualDDL-Overview. Accessed 10 May 2014
5. Tutorials point (2013) Hive partitioning. Available at: http://www.tutorialspoint.com/Hive/
Hive_partitioning.html. Accessed 15 June 2014
6. Rohit R (2014) Introduction to Hives partitioning. Available at: http://java.dzone.com/arti
cles/introduction-Hives. Accessed 25 Jan 2015
7. Peschka J (2013) Introduction to Hive partitioning. Available at: http://www.brentozar.com/
archive/2013/03/introduction-to-hive-partitioning/. Accessed 8 Aug 2015
8. Thrive school (2013) Available at: http://thriveschool.blogspot.in/2013/11/Hive-bucketedtables-and-sampling.html. Accessed 10 Jan 2015
9. Philip N (2014) 10 best practices for Apache Hive.Available at: www.qubole.com/blog/bigdata/hive-best-practices. Accessed 15 July 2015
10. Petit W (2014) Introduction to Pig. Available at: http://bigdatauniversity.com/bdu-wp/bducourse/introduction-to-pig/#sthash.HUcw7EZe.dpuf. Accessed 20 June 2014
11. Apache (2014) Hadoop online tutorial. Available at: http://hadooptutorial.info/tag/hadooppig-architecture-explanation. Accessed 13 Feb 2015
12. Hadoop (2010) Pig latin manual. Available at: https://pig.apache.org/docs/r0.7.0/piglatin_ref2.
html. Accessed 20 May 2014
13. Lam C (2010) Hadoop in action. Manning Publications, Greenwich
14. Gates A (2011) Programming Pig. OReilly Media Inc, Sebastopol
15. Apache (2007) Getting started-Pig, Apache Software Foundation
16. Apache (2015) When would I use Apache HBase. Available at: Hbase.apache.org. Accessed
10 Feb 2015
17. Grehan R (2014) Review: HBase is massively scalable and hugely complex. Available at:
http://www.infoworld.com/article/2610709/database/review--hbase-is-massively-scalable---and-hugely-complex.html. Accessed 10 July 2015

220

N. Maheswari and M. Sivagami

18. Servelets C (2012) HBase overview. Available at: www.coreservlets.com/hadoop-tutorial/

#HBase. Accessed 12 Jan. 2015
19. Apache (2010) Apache Zookeeper. Available at: zookeeper.apache.org. Accessed 15 March
2015
20. Tutorials Point (2014) HBase tutorial. Available at: Tutorialspoint.com/hbase. Accessed
18 Feb 2015
21. George L (2011) HBase definitive guide. OReilly Media Inc, Sebastopol

Chapter 10

Big Data Analytics: Enabling Technologies

and Tools
Mohanavadivu Periasamy and Pethuru Raj

Abstract The era of big data is at its full bloom with the data being generated,
captured, stocked, polished, and processed in astronomical proportions. This is
mainly due to the unprecedented levels of technology adoption and adaption
resulting in the connectivity technologies, network topologies, and tools that have
enabled seamless connectivity between billions of physical, mechanical, electrical,
electronic, and computer systems. This data explosion also challenges the way
corporate decisions are being taken. It is for this reason that data is being widely
recognized as one of the core strategic assets, and extraction of value from the
collected data, using data science approaches, is gaining importance. Apart from
the analytics, data management, mining, and visualization are also becoming
feverishly articulated tasks on such data. In the knowledge economy and digital
world of today, insight-driven decision-making is being tipped as the most important and influential factor. Precisely speaking, competent and cognitive data analytics is capable of bringing forth better and hitherto unforeseen approaches and
solutions for a variety of perpetual problems that are constantly hitting hard our
society and the global economy. This chapter looks into the big data domain and
discusses the relevant analytical methods, platforms, infrastructures, and
techniques.
Keywords AaaS BI Business intelligence big data CSP data science Data
virtualization Hadoop Information visualization MapReduce NoSQL
NoSQL databases SAP HANA

M. Periasamy (*)
TCL Canada, 1441 rue Carrie-Derick, Montreal H3C 4S9, QC, Canada
e-mail: mohanavadivu@gmail.com
P. Raj
IBM Global Cloud Center of Excellence, Bangalore 560045, India
e-mail: peterindia@gmail.com
Springer International Publishing Switzerland 2016
Z. Mahmood (ed.), Data Science and Big Data Computing,
DOI 10.1007/978-3-319-31861-5_10

221

222

10.1

M. Periasamy and P. Raj

Introduction

There is an infamous adage based on the Parkinsons Law: Work expands so as to

expand the time available for its completion. So is the case with the generation of
more data, which explodes to fill the space available for data storage. We now live
in an age where the number of electronic devices has already overtaken the human
population worldwide. There are more than seven billion mobile devices on this
planet. The data generated by billions of users over billions of devices for humanto-human interaction, human-to-machine interaction, and machine-to-machine
interaction is now measured in zettabytes.
Classifying data generated online, these days, is a definite challenge. As an
example, for Twitter messages, people try to classify or categorize the message
using hash tags (#). It is easy to categorize a message with a single # tag into a
category, e.g., I am #happy can be put into a category named happy. However,
how does one categorize the following message: When youre #happy, you enjoy
the #music. But, when youre #sad, you understand the #lyrics. We can fit this into
the category of music, happy, sad, or lyrics. In this case, how do we establish the
logic for selecting the category based on the # tag? A seemingly simple statement
categorization snowballs into a problem of huge proportion when we take into
account the fact that a billion tweets are generated every day, resulting in massive
amount of information being uploaded or revised on line and generated through
various instant-messaging services. This is where the need for systems for
processing the vast amounts of data kicks in.
Big data is one such data framework that aims to manage a vast amount of data
very efficiently. A commonly accepted definition of big data can be stated as:
information assets characterized by such a high volume, velocity, and variety to
require specific technology and analytical methods for its transformation into
value. Big data has been used successfully by various government departments
including the National Security Agency in the USA to process information collected over the Internet. In the UK, the National Health Service uses big data to
process information on drug intake; the Indian government uses big data to access
the effect of electorate response on various government policies. Corporate agencies use big data to improve the customer response, improve the efficiency in supply
chain, etc. Citibank uses big data to assess the credit card spending of customers to
recommend discounts. Interestingly, political parties have also taken the big data in
a big way. Big data seems to have played a stellar role in the analysis of electorate
opinion and tune the successful campaign strategies of President Barrack Obama in
the USA in the year 2012 and in the prime ministerial campaign of Narendra Modi
in India in the year 2014. Some of the political analysts attribute both these victories
to the smart campaigning methods, where big data has played a decisive role. From
these examples, we can easily infer that the application of big data has spread into
all walks of life.
Organizations face a major challenge in managing and processing big data
effectively to work out venerable solutions. The job of data scientists is to discover

10

Big Data Analytics: Enabling Technologies and Tools

223

patterns and relationships and synthesize them into resultant information that
addresses the needs of the organizations. Handling, processing, and managing big
data is a complex task. These complexities have limited the capabilities of traditional RDBMS and have paved the way for exploring a host of new technologies,
approaches, and platforms that are based on the cloud computing paradigm. This
has opened up a wide range of possibilities for organizations to use big data network
to understand the needs and expectations of customers so as to optimize the use of
resources. This requires a rich mixture of skills on the part of data scientists. They
need to know how to handle different types of big data and storage mechanisms
(e.g., RDBMS, Hadoop HDFS), develop advanced algorithms, develop codes (e.g.,
Java, Python, R), access data (SQL, Hive, etc.), analyze any inherent constraints,
perform regression analysis, communicate to social networks, and finally present
the findings to management in business terms through briefings and reports. Invariably, data scientists have established their impact in the advancement of all
analytical fields such as statistics, management science, operations research, computer science, logistics, and mathematics.
From the examples mentioned above, one can easily notice two types of requirements or usage categories when it comes to big data application, the first one being
real-time operational systems such as capital markets which require instant streaming, analysis, and feedback and the second one being analytical systems such as
market research, surveillance, government databases, etc. The concurrency of data
users and velocity are critical in real-time operations, whereas data scope and
variety are important in analytical systems. Big data handles the operational
systems and analytical systems differently as needs of the systems are different.
Database management systems for operational systems involve not only SQL
(NoSQL) database systems such as MongoDB, whereas analytical systems use
DBMS such as MPP-compliant MapReduce.
In the ensuing sections, we aim to focus on big data, its special characteristics,
how to capture and crunch big data to extract valuable insights, the technologies
and tools available for simplifying and streamlining the process of knowledge
discovery and dissemination across a variety of industry verticals, etc. There are
some challenges and concerns relating to big data. We also discuss the platforms
and products emerging and evolving in order to speed up big data analytics.
Finally, we discuss the role of the cloud paradigm in shaping up data analytics
and insight-driven enterprises. There are integrated platforms being deployed
in cloud environments so that the raging domain of analytics can be taken to
cost-effective, elastic, and elegant clouds.

10.2

Characterizing Big Data

Big data typically represents massive amounts of heterogeneous data that are not
necessarily stored in the relational form in traditional enterprise-scale databases.
Therefore, new-generation database systems are being unearthed in order to store,

224

M. Periasamy and P. Raj

retrieve, aggregate, filter, mine, and analyze large amount of multi-structured data
efficiently. The following are the general characteristics of big data:
Data storage is defined in the order of petabytes and exabytes in volume as
opposed to the current storage limits in terms of gigabytes and terabytes.
Big data contains structured, semi-structured, and totally unstructured data.
Different types of data sources (sensors, machines, mobiles, social sites, etc.)
and resources for big data.
Data is time sensitive (near real time as well as real time) which means that big
data consists of data collected with relevance to the time zones so that timely
insights can be extracted.
The recent developments in big data analytics have generated a lot of interest
among industry professionals as well as academicians. Big data has become an
inevitable trend, and it has to be solidly and compactly handled in order to derive
time-sensitive and actionable insights. There is a dazzling array of tools, techniques, and tips evolving in order to quickly capture data from different distributed
resources and processes; analyze and mine the relevant data to extract actionable
business-to-insight technology-sponsored business transformation and sustenance.
In short, analytics is a flourishing phenomenon in every sphere and segment of the
international world today. Especially, with the automated capture, persistence, and
processing of tremendous amount of multi-structured data getting generated by men
as well as machines, the analytical value, scope, and power of data are bound to
blossom further. Precisely, data is a strategic asset for organizations to insightfully
plan to sharply enhance their capabilities and competencies and to venture on the
appropriate activities that decisively and drastically power up their short- as well as
long-term offerings, outputs, and outlooks. Business innovations can happen in
plenty and be maintained when there is a seamless and spontaneous connectivity
between data-driven and analytics-enabled business insights and business
processes.
In the recent past, real-time analytics has brought in much excrescence, and
several product vendors have been flooding the market with a number of elastic and
state-of-the-art solutions (software as well as hardware) for alleviating on-demand,
ad hoc, real-time, and runtime analysis of batch, online transaction, social, machine,
operational, and streaming data. There are numerous advancements in this field due
to its high potential for worldwide companies in considerably reducing operational
expenditures while gaining operational insights. Hadoop-based analytical products
are capable of processing and analyzing any data type and quantity across huge
volume of commodity server clusters. Stream computing drives continuous and
cognitive analysis of massive volumes of streaming data with sub-millisecond
response times. There are enterprise data warehouses, analytical platforms,
in-memory appliances, etc. Data warehousing delivers deep operational insights
with advanced in-database analytics. Here are some examples:

10

Big Data Analytics: Enabling Technologies and Tools

225

The EMC Greenplum Data Computing Appliance (DCA) is an integrated analytics platform that accelerates the analysis of big data assets within a single
integrated appliance.
IBM PureData System for Analytics architecturally integrates database, server,
and storage into a single, purpose-built, easy-to-manage system.
SAP HANA is an exemplary platform for efficient big data analytics. Platform
vendors are conveniently tied up with infrastructure vendors especially Cloud
Service Providers (CSPs) to take analytics to cloud so that the goal of Analytics
as a Service (AaaS) sees a neat and nice reality, sooner than later.
There are multiple start-ups with innovative product offerings to speed up and
simplify the complex part of big data analysis.

10.3

The Inherent Challenges

There are some uncertainties, limitations, and potential roadblocks that could
probably unsettle the expected progress, since big data is still in the emerging
era. Let us consider a few of these which are more pertinent.
Technology Precedence Technologies and tools are very important for creating
business value from big data. There are multiple products and platforms from
different vendors. However, the technology choice is very important for firms to
plan and proceed without any hitch in their pursuit. The tools and technology
choices will vary depending on the types of data to be manipulated (e.g., XML
documents, social media, sensor data, etc.), business drivers (e.g., sentiment
analysis, customer trends, product development, etc.), and data usage (analytic
or product development focused).
Data Governance Any system has to be appropriately governed in order to be
strategically beneficial. Due to the sharp increase in data sources, types, channels, formats, and platforms, data governance is an important component in
efficiently regulating the data-driven tasks. Other important motivations include
data security while in transit and in persistence, data integrity, and data confidentiality. Furthermore, there are governmental regulations and standards from
world bodies, and all these have to be fully complied with, in order to avoid any
kind of ramifications at a later point in time.
Resource Deficit It is predicted by MGI that there will be a huge shortage of
human talent for organizations providing big data-based services and solutions.
There will be requirements for data modelers, scientists, and analysts in order to
get all the envisaged benefits of big data. This is a definite concern to be
sincerely attended to by companies and governments across the world.
Ramification in Framework Big data product vendors need to bring forth
solutions that simplify the complexities of the big data framework to enable
users to extract business value. The operating interfaces need to be intuitive and

226

M. Periasamy and P. Raj

informative so that the goal of ease of use can be ensured for people using big
data solutions.
Big datas reputation has taken a battering lately, e.g., in the case of the NSA for
collecting and storing peoples web and phone records, without proper consent. It
has led to a wider debate about the appropriateness of such extensive data-gathering
activities, but this negative publicity should not detract people from the reality of
big data. The fact is that big data is ultimately to benefit society as a whole.
In short, big data applications, platforms, appliances, and infrastructures need to
be contrived in a way to facilitate their usage and leverage for everyday purposes.
The awareness about the potential of big data needs to be propagated widely, and
professionals need to be trained in order to extract better business value out of big
data. Competing technologies, enabling methodologies, prescribing patterns, and
evaluating metrics, key guidelines, and best practices need to be unearthed and
made as reusable assets.

10.4

Big Data Infrastructures, Platforms, and Analytics

The future of business definitely belongs to those enterprises that embrace the big
data analytics movement and use it strategically to their own advantages. It is
pointed out that business leaders and other decision-makers, who are smart enough
to adopt a flexible and futuristic big data strategy, can take their businesses to
greater heights. Successful companies are already extending the value of classic and
conventional analytics by integrating cutting-edge big data technologies and
outsmarting their competitors [8]. There are several forecasts, exhortations, expositions, and trends on the discipline of big data analytics. Market research and
analyst groups have come out with positive reports and briefings, detailing its key
drivers and differentiators, the future of this brewing idea, its market value, the
revenue potentials and application domains, the fresh avenues and areas for
renewed focus, the needs for its sustainability, etc.
The cloud computing movement is expediently thriving and trendsetting a host
of delectable novelties. A number of tectonic transformations on the business front
are being activated and accentuated with faster and easier adaptability of the cloud
IT principles. The cloud concepts have opened up a deluge of fresh opportunities
for innovators, individuals, and institutions to conceive and formalize
new-generation business services and solutions. A dazzling array of pathbreaking
and mission-critical business augmentation models and mechanisms have already
emerged, and they are consistently evolving toward perfection as the cloud technology grows relentlessly and rewardingly in conjunction with other enterpriseclass technologies.

10

Big Data Analytics: Enabling Technologies and Tools

227

10.4.1 Unified Platforms for Big Data Analytics

Highly synchronized and integrated platforms are prerequisites to automate several
tasks enshrined in capturing of data, analysis, and knowledge discovery processes.
A converged platform brings out a reliable workbench to empower developers to
assist application development and other related tasks such as data security,
virtualization, integration, visualization, and dissemination. Special consoles are
being used with new-generation platforms for performing other important activities
such as management, governance, and enhancement. An example is Hadoop [7, 10]
that is a disruptive technology for data distribution among hundreds of commodity
compute machines for parallel data crunching, and any typical Big Data platform is
blessed with Hadoop software suite.
Furthermore, the big data platform enables entrepreneurs; investors; chief executive, information, operation, knowledge, and technology officers (CXOs); and
marketing and sales people to explore and perform their experiments on big data,
at scale at a fraction of time and cost to what was required previously. That is,
platforms are to bestow all kinds of stakeholders and end users with actionable
insights that in turn lead to consider and take informed decisions. Knowledge
workers such as business analysts and data scientists could be the other main
beneficiaries through these empowered platforms. Knowledge discovery is a vital
portion, and the platform has to be chipped in with real-time and real-world tips,
associations, patterns, trends, risks, alerts, and opportunities. In-memory and
in-database analytics are gaining momentum for high-performance and real-time
analytics. New advancements in the form of predictive and prescriptive analytics
are rapidly emerging with the maturity and stability of big data technologies,
platforms, infrastructures, tools and finally a cornucopia of sophisticated data
mining and analysis algorithms. Thus, platforms need to be fitted with new features,
functionalities, and facilities in order to provide new insights. Here is a simple
strategy:
Develop Infrastructure The infrastructure required to support the acquisition of
big data must deliver low and predictable latency both in capturing data and in
executing short and simple queries. It should be able to handle very high
transaction volumes often in a distributed environment that also support flexible
and dynamic data structures. NoSQL databases are the leading infrastructure to
acquire and store big data. NoSQL databases are well suited for dynamic data
structures and are highly scalable. The data stored in a NoSQL database is
typically of a high variety because the systems are intended to simply capture
all kinds of data without categorizing and parsing the data. For example, NoSQL
databases are often used to collect and store social media data. While customerfacing applications frequently change, underlying storage structures are kept
simple. Instead of designing a schema with relationships between entities, these
simple structures often just contain a major key to identify the data point and
then a content container holding the relevant data. This extremely simple and

228

M. Periasamy and P. Raj

nimble structure allows changes to take place without any costly reorganization
at the storage layer.
Integrate Big Data In classical data warehousing terms, organizing data is
called data integration. Because there is such a huge volume of data, there is a
tendency and trend to organize data at the original storage locations. This saves a
lot of time and money as there is no movement of data. The brewing need is to
have a robust infrastructure that is innately able to organize big data and process
and manipulate data in the original storage location. It has to support very high
throughput (often in batches) to deal with large data processing steps and handle
a large variety of data formats (unstructured, semi-structured, and fully
structured).
Analyze Big Data The data analysis can also happen in a distributed environment. That is, data stored in diverse locations can be accessed from a data
warehouse to accomplish the intended analysis. The appropriate infrastructure
required for analyzing big data must be able to support in-depth analytics such as
statistical analysis and data mining [2] on a wider range of data types stored in
those diverse systems to scale the extreme data volumes and to deliver faster
response times driven by changes in behavior and also to automate decisions
based on analytical models. Most importantly, the infrastructure must be able to
integrate analysis on the combination of big data and traditional enterprise data
to produce exemplary insights for fresh opportunities and possibilities. For
example, analyzing inventory data from a smart vending machine in combination with the event calendar for the venue in which the vending machine is
located will dictate the optimal product mix and replenishment schedule for the
vending machine.

10.4.2 Newer and Nimbler Applications

The success of any technology is to be squarely decided based on the number of
mission-critical applications it can create and sustain. That is, the applicability or
employability of the new paradigm to as many application domains as possible is
the major deciding factor for its successful journey. As far as the development is
concerned, big data applications could differ from other software applications to a
larger extent. Web and mobile enablement of big data applications are also important. As big insights are becoming mandatory for multiple industry segments, there
is a bigger scope for big data applications. Therefore, there is a big market for big
data application development platforms, patterns, metrics, methodology, reusable
components, etc.

10

Big Data Analytics: Enabling Technologies and Tools

229

10.4.3 Tending Toward a Unified Architecture

It is well recognized that an integrated IT environment is a minimum requirement
for attaining the expected success out of the big data paradigm. Refer to Fig. 10.1
for an integrated platform. Deploying big data platforms in an IT environment that
lacks a unified architecture and does not seamlessly and spontaneously integrate
distributed and diverse data sources, metadata, and other essential resources would
not produce the desired insights. Such deployments will quickly lead to a torrent of
failed big data projects, and in a fragmented setup, achieving the desired results
remains a dream forever. Hence, a unified and modular architecture is a need of the
hour for taking forward the ideals of the big data discipline. Deployment of big data
applications in a synchronized enterprise environment or cloud IT environment
makes analytics simpler, faster, cheaper, and accurate while reducing deployment
and operational costs.
In the ensuing era of big data, there could be multiple formats for data representation, transmission, and persistence. The related trend is that there are databases
without any formal schema. SQL is the standard query language for traditional
databases, whereas in the big data era, there are NoSQL databases that do not
support the SQL, which is the standard for structured querying. Special file systems
such as Hadoop Distributed File System (HDFS) are being produced in order to
facilitate big data storage and access. Thus analytics in the big data period is quite
different from the analytics on the SQL databases. However, there is a firm place for

Fig. 10.1 Big Data analytics platforms, appliances, products, and tools

230

M. Periasamy and P. Raj

SQL-based analytics, and hence there is an insistence on converging both to fulfill

the varying needs of business intelligence (BI). Tools and technologies that provide
a native coalesce of classic and new data analytics techniques will have an inherent
advantage.

10.4.4 Big Data Appliances and Converged Solutions

Appliances (hardware and virtual) are being prescribed as a viable and valueadding approach for scores of business-critical application infrastructure solutions
such as service integration middleware, messaging brokers, security gateways, load
balancing, etc. They are fully consolidated and prefabricated with the full software
stack so that their deployment and time to operation are quick and simple. There are
XML and SOA appliances plenty in the marketplace for eliminating all kinds of
performance bottlenecks in business IT solutions. In the recent past, EMC
Greenplum and SAP HANA appliances are stealing and securing the attention.
SAP HANA is being projected as a game-changing and real-time platform for
business analytics and applications. While simplifying the IT stack, it provides
powerful features like the significant processing speed, ability to handle big data,
predictive capabilities, and text mining capabilities. Thus, the emergence and
evolution of appliances represent a distinct trend as far as big data is concerned.
Besides converged architecture, infrastructures, application domains, and platforms and synchronized processes are very important in order to augment and
accelerate big data analytics. Already, analytics-focused processes are emerging
and evolving consistently. Analytics becomes an important activity to tightly
integrate with processes. Also, the AaaS paradigm is on the verge of massive
adaptation, and hence the analytics-oriented process integration, innovation, control, and management aspects will gain more prominence and dominance in very
near future.

10.4.5 Big Data Frameworks

Big data processing is generally of two types: real-time and batch processing. The
data is flowing endlessly from countless sources these days. Data sources are on the
climb. Innumerable sensors, varying in size, scope, structure, and smartness, are
creating data continuously. Stock markets are emitting a lot of data every second;
system logs are being received, stored, processed, analyzed, and acted upon ceaselessly. Monitoring agents are working tirelessly producing a lot of usable and useful
data, business events are being captured, and knowledge discovery is initiated. At
the same time, information visualization is being realized to empower enterprise
operations. Stream computing is the latest paradigm being aptly prescribed as the
best course of action for real-time receipt, processing, and analysis of online, live,

10

Big Data Analytics: Enabling Technologies and Tools

231

and continuous data. Real-time data analysis through in-memory and in-database
computing models is gaining a lot of ground these days with the sharp reduction in
computer memory costs. For the second category of batch processing, the Hadoop
technology is being recommended as there is a need for competent products,
platforms, and methods for efficiently and expectantly working with both realtime as well as batch data.
As elucidated before, big data analysis is not a simple affair, and there are
Hadoop-based software programming frameworks, platforms, and appliances
emerging to ease the innate complications. The Hadoop programming model
[7, 10] has turned out to be the central and core method to further extend the field
of big data analysis. The Hadoop ecosystem is continuously spreading its wings
wider, and the enabling modules are being incorporated to make Hadoop-based big
data analysis simpler, succinct, and quicker.

10.4.6 The Hadoop Software Family

Apache Hadoop [7, 10] is an open-source framework that allows distributed
processing of large-volume data sets across clusters of computers using a simple
programming model. Hadoop was originally designed to scale up from a single
server to thousands of machines, each offering local computation and storage.
Rather than relying on hardware to deliver high availability, the Hadoop software
library itself is designed to detect and handle failures at the application layer.
Therefore, it delivers a highly available service on top of a cluster of cheap
computers, each of which may be prone to failures. Hadoop [7, 10] is based on
the modular architecture, and thereby any of its components can be swapped with
competent alternatives if such a replacement brings noteworthy advantages.
Despite all the hubbub and hype around Hadoop, few IT professionals know its
key drivers, differentiators, and killer applications. Because of the newness and
complexity of Hadoop, there are several areas where confusion reigns and restrains
its full-fledged assimilation and adoption. The Apache Hadoop product family
includes the Hadoop Distributed File System (HDFS), MapReduce, Hive, HBase,
Pig, Zookeeper, Flume, Sqoop, Oozie, Hue, and so on. HDFS and MapReduce
together constitute the core of Hadoop, which is the foundation for all Hadoopbased applications. For applications in business intelligence (BI), data warehousing
(DW), and big data analytics, the core Hadoop is usually augmented with Hive and
HBase and sometimes Pig. The Hadoop file system excels with big data that is file
based, including files that contain nonstructured data. Hadoop is excellent for
storing and searching multi-structured big data, but advanced analytics is possible
only with certain combinations of Hadoop products, third-party products, or extensions of Hadoop technologies. The Hadoop family has its own query and database
technologies, and these are similar to standard SQL and relational databases. That
means BI/DW professionals can learn them quickly.

232

M. Periasamy and P. Raj

The HDFS is a distributed file system designed to run on clusters of commodity

hardware. HDFS is highly fault-tolerant because it automatically replicates file
blocks across multiple machine nodes and is designed to be deployed on low-cost
hardware. HDFS provides high-throughput access to application data and is suitable
for applications that have large data sets. As a file system, HDFS manages files that
contain data. HDFS itself does not offer random access to data since it is file based
and provides only limited metadata capabilities when compared to a DBMS.
Likewise, HDFS is strongly batch oriented and hence has limited real-time data
access functions. To overcome these challenges, it is possible to layer HBase over
HDFS to gain some of the mainstream DBMS capabilities. HBase is a product from
Apache Hadoop family and is modeled after Googles Bigtable, and hence HBase,
like Bigtable, excels with random and real-time access to very large tables
containing billions of rows and millions of columns. Currently, HBase is limited
to straightforward tables and records with little support for more complex data
structures. The Hive meta-store gives Hadoop some DBMS-like metadata
capabilities.
When HDFS and MapReduce are combined, Hadoop easily parses and indexes
the full range of data types. Furthermore, as a distributed system, HDFS scales well
and has a certain amount of fault tolerance based on data replication even when
deployed on top commodity hardware. For these reasons, HDFS and MapReduce
can complement existing BI/DW systems that focus on structured and relational
data. MapReduce is a general-purpose execution engine that works with a variety of
storage technologies including HDFS, other file systems, and some DBMSs.
As an execution engine, MapReduce and its underlying data platform handle the
complexities of network communication [1], parallel programming, and fault
tolerance. In addition, MapReduce controls hand-coded programs and automatically provides multi-threading processes, so they can be executed in parallel for
massive scalability. The controlled parallelization of MapReduce can apply to
multiple types of distributed applications, not to analytic ones. In a nutshell,
Hadoop MapReduce is a software programming framework for easily writing
massively parallel applications which process massive amounts of data in parallel
on large clusters (thousands of nodes) of commodity hardware in a reliable and
fault-tolerant manner. A MapReduce job usually splits the input data set into
independent chunks which are processed by the map tasks in a completely parallel
manner. The framework sorts the outputs of the maps as input to the reduced tasks
which, in turn, assemble one or more result sets.
Hadoop is not just for new analytic applications; it can revamp old ones too. For
example, analytics for risk and fraud that is based on statistical analysis or data
mining benefits from the much larger data samples that HDFS and MapReduce can
wring from diverse Big Data. Furthermore, most 360
customer views include
hundreds of customer attributes. Hadoop can provide insight and data to bump up
to thousands of attributes, which in turn provide greater detail and precision for
customer-based segmentation and other customer analytics. Hadoop is a promising
and potential technology that allows large data volumes to be organized and
processed while keeping the data on the original data storage cluster. For example,

10

Big Data Analytics: Enabling Technologies and Tools

233

Weblogs can be turned into browsing behavior (sessions) by running MapReduce

programs (Hadoop) on the cluster and generating aggregated results on the same
cluster. These aggregated results are then loaded into a relational DBMS system.
HBase is the mainstream Apache Hadoop database. It is an open-source,
non-relational (column oriented), scalable, and distributed database management
system that supports structured data storage. Apache HBase, which is modeled after
Google Bigtable, is the right approach when you need random and real-time read/
write access to big data. It is for hosting of very large tables (billions of rows X
millions of columns) on top of clusters of commodity hardware. Just like Google
Bigtable leverages the distributed data storage provided by the Google File System,
Apache HBase provides Bigtable-like capabilities on top of Hadoop and HDFS.
HBase does support writing applications in Avro, REST, and Thrift.

10.5

Databases for Big Data Management and Analytics

10.5.1 NoSQL Databases

Next-generation databases are mandated to be non-relational, distributed, open
source, and horizontally scalable. The original inspiration is the modern
web-scale databases. Additional characteristics such as schema free, easy replication support, simple API, eventually consistent/BASE (not Atomicity, Consistency,
Isolation, Durability (ACID)), etc., are also being demanded. The traditional Relational Database Management Systems (RDBMSs) use Structured Query Language
(SQL) for accessing and manipulating data that reside in structured columns of
relational tables. However, unstructured data is typically stored in key-value pairs
in a data store and therefore cannot be accessed through SQL. Such data are stored
in NoSQL [9] data stores and accessed via GET and PUT commands. There are
some big advantages of NoSQL databases compared to the relational databases as
illustrated in the following web page: http://www.couchbase.com/why-nosql/nosqldatabase.
Flexible Data Model Relational and NoSQL data models are extremely
different. The relational model takes data and separates them into many interrelated tables containing rows and columns. Tables reference each other through
foreign keys stored in columns. When looking up for data, the desired information needs to be collected from many tables and combined before it can be
provided to the application. Similarly, when writing data, the write process
needs to be coordinated and performed on many tables.
NoSQL databases [9] follow a different model. For example, a documentoriented NoSQL database takes the data that needs to be stored and aggregates it
into documents using the JSON format. Each JSON document can be thought of as
an object to be used by the application. A JSON document might, for example, take

234

M. Periasamy and P. Raj

all the data stored in a row that spans 20 tables of a relational database and
aggregate it into a single document/object. The resulting data model is flexible,
making it easy to distribute the resulting documents. Another major difference is
that relational technologies have rigid schemata, while NoSQL models are schemaless. Changing the schema once data is inserted is a big deal, extremely disruptive,
and frequently avoided. However, the exact opposite of the behavior is desired for
the big data processing. Application developers need to constantly and rapidly
incorporate new types of data to enrich their applications.
High Performance and Scalability To handle the large volumes of data (big
data) and their concurrent user accesses (big users), applications and underlying
databases need to be scaled using one of two choices: scale-up or scale-out.
Scaling-up implies a centralized approach that relies on bigger and even bigger
servers. Scaling-out implies a distributed approach that leverages many commodity physical or virtual servers. Prior to NoSQL databases, the default scaling
approach at the database tier was to scale up, dictated by the fundamentally
centralized, shared-everything architecture of relational database technology. To
support more concurrent users and/or store more data, here is a requirement for a
bigger server with more CPUs, memory, and disk storage to keep all the tables.
Big servers are usually highly complex, proprietary, and disproportionately
expensive.
NoSQL databases [9] were developed from the ground up to be distributed and
scale out databases. They use a cluster of standard, physical, or virtual servers to
store data and support database operations. To scale, additional servers are joined to
the cluster, and the data and database operations are spread across the larger cluster.
Since commodity servers are expected to fail from time to time, NoSQL databases
are built to tolerate and recover from such failures to make them highly resilient.
NoSQL databases provide a much easier and linear approach to database scaling. If
10,000 new users start using an application, one simply needs to add another
database server to a cluster, and there is no need to modify the application as per
the scale since the application always sees a single (distributed) database. NoSQL
databases share some characteristics, as presented below, with respect to scaling
and performance.
Auto-Sharding A NoSQL database automatically spreads data across different
servers without requiring applications to participate. Servers can be added or
removed from the data layer without any application downtime, with data (and
I/O) automatically spread across the servers. Most NoSQL databases also support data replication, storing multiple copies of data across the cluster and even
across data centers to ensure high availability (HA) and to support disaster
recovery (DR). A properly managed NoSQL database system should never be
taken offline.
Distributed Query Support Sharing a relational database can reduce or eliminate in certain cases the ability to perform complex data queries. NoSQL

10

Big Data Analytics: Enabling Technologies and Tools

235

database system retains their full query expressive power even when distributed
across hundreds of servers.
Integrated Caching To reduce latency and increase sustained data throughput,
advanced NoSQL database technologies transparently cache data in system
memory. This behavior is transparent to the application developer and the
operations team, compared to relational technology where a caching tier is
usually a separate infrastructure tier that must be developed to and deployed
on separate servers and explicitly managed by the operations team.
There are some serious flaws with respect to relational databases that come in the
way of meeting up the unique requirements in the modern-day social web applications, which gradually move to reside in cloud infrastructures. Another noteworthy
factor is that doing data analysis for business intelligence (BI) is increasingly
happening in the cloud era. There are some groups in academic and industrial
circles striving hard for bringing in the necessary advancements in order to prop up
the traditional databases to cope up with the evolving and enigmatic requirements
of social networking applications. However NoSQL and NewSQL databases are the
new breeds of versatile, vivacious, and venerable solutions capturing the imagination and attention.
The business requirement to leverage complex data can be driven through the
adoption of scalable and high-performance NoSQL databases. This new entrant is
to sharply enhance the data management capabilities of various businesses. Several
variants of NoSQL databases have emerged in the past decade in order to appropriately handle the terabytes, petabytes, and even exabytes of data generated by
enterprises and consumers. They are specifically capable of processing multiple
data types such as text, audio, video, social network feeds, and Weblogs that are
difficult to handle by traditional databases. These data are highly complex and
deeply interrelated, and therefore the demand is to unravel the truth hidden behind
these huge yet diverse data assets besides understanding the insights and acting on
them to enable businesses to plan and surge ahead.
Having understood the changing scenario, web-based businesses have been
crafting their own custom NoSQL databases to elegantly manage high volume of
data and its diversity. Amazon Dynamo and Googles Bigtable are the shining
examples of homegrown databases that can store huge amounts of data. These
NoSQL databases were designed for handling highly complex and heterogeneous
data. The key differentiation is that they are not built for high-end transactions but
for analytic purposes.

10.5.2 Why NoSQL Databases?

Business-to-consumer (B2C) e-commerce and business-to-business (B2B)
e-business applications are highly transactional, and the leading enterprise application frameworks and platforms such as Java Enterprise Edition (JEE) can directly

236

M. Periasamy and P. Raj

and distinctly support a number of transaction types (simple, distributed, nested,

etc.). For a trivial example, flight reservation application has to be rigidly transactional otherwise everything is bound to collapse. As enterprise systems are increasingly distributed, the need for transaction feature is being pronounced as a
mandatory one.
In the recent past, social applications have grown fast, and especially the
younger generation is totally fascinated by a stream of social computing sites
which has resulted in an astronomical growth of further data. It is no secret that
the popularity, ubiquity, and utility of Facebook, LinkedIn, Twitter, Google , and
other blogging sites are surging incessantly. There is a steady synchronization
between enterprise and social applications with the idea of adequately empowering
enterprise applications with additional power and value. For example, the online
sellers understand that businesses should be more interactive, open, and inclined
toward customers participation to garner and glean their views to reach out to even
more people across the globe and to pour in Richer Enterprise Applications (REAs).
There are specialized protocols and web 2.0 technologies (e.g., Atom, RSS, AJAX,
and mash-ups) to programmatically tag information about people and places and
proclivity to dynamically conceive, conceptualize, and concretize more and more
people-centric and premium services.
The dormant and dumb database technology has to evolve faster in order to
accomplish these new-generation IT abilities. With the modern data being more
complicated and connected, the NoSQL databases need to have the implicit and
innate strength to handle the multi-structured and massive amounts of data. A
NoSQL database should enable high-performance queries on such data. Users
should be able to ask questions such as the following: Who are all my contacts
in Europe? Which of my contacts ordered from this catalog? A white paper
entitled NoSQL for the Enterprise by Neo Technology [9] points to the uniqueness of NoSQL databases for enterprises. Some of the key points have been
reproduced below, from that paper:
A Simplified Data Representation A NoSQL database should be able to easily
represent complex and connected data that makes up todays enterprise applications [3]. Unlike traditional databases, a flexible schema that allows for
multiple data types also enables developers to easily change applications without
disrupting live systems. Databases must be extensible and adaptable. With the
massive adoption of clouds, NoSQL databases ought to be more suitable for
clouds.
End-to-End Transactions Traditional databases are famous for all or nothing
transactions, whereas NoSQL databases give a kind of leeway on this crucial
property. This is due to the fact that the prime reason for the emergence and
evolution of NoSQL databases was to process massive volumes of data quickly
and to come up with actionable inputs. In other words, traditional databases are
for enterprise applications, whereas NoSQL databases are for social applications. Specifically, the consistency aspect of ACID transactions is not rigidly
insisted in NoSQL databases. Sometimes an operation can fail in a social

10

Big Data Analytics: Enabling Technologies and Tools

237

application and it does not matter much. For instance, there are billions of short
messages being tweeted every day, and Twitter will probably survive if a few
Tweets do get lost. But online banking applications relying on traditional
databases have to ensure a very tight consistency in order to be meaningful.
That does not mean that NoSQL databases are off the ACID hook. Instead they
are supposed to support ACID transactions including XA-compliant distributed
two-phase commit protocol. The connections between data should be stored on a
disk in a structure designed for high-performance retrieval of connected data sets
all while enforcing strict transaction management. This design delivers significantly better performance for connecting data than the one offered by relational
databases.
Enterprise-Grade Durability Every NoSQL database for the enterprise needs to
have the enterprise-class quality of durability. That is, any transaction committed to the database will not be lost at any cost under any circumstances. If there is
a flight ticket reserved and the system crashes due to an internal or external
problem thereafter, then when the system comes back, the allotted seat still has
to be there. Predominantly the durability feature is ensured through the use of
database backups and transaction logs that facilitate the restoration of committed
transactions in spite of any software or hardware hitches. Relational databases
have employed the replication method for years successfully to guarantee the
enterprise-strength durability.

10.5.3 Classification of NoSQL Databases

There are four categories of NoSQL databases available today: key-value stores,
document databases, column-family databases, and graph databases. Each was
designed to accommodate the huge volumes of data as well as to have room for
future data types. The choice of NoSQL database depends on the type of data one
needs to store, its size, and complexity. Here is a brief mention of the four types:
Key-Value Stores A key-value data model is quite simple. It stores data in key
and value pairs where every key maps to a value. It can scale across many
machines but cannot support other data types. Key-value data stores use a data
model similar to the popular memcached distributed in-memory cache, with a
single key-value index for all the data. Unlike memcached, these systems
generally provide a persistence mechanism and additional functionality as well
including replication, versioning, locking, transactions, sorting, and several
other features. The client interface provides inserts, deletes, and index lookups.
Like memcached, none of these systems offer secondary indices or keys. A
key-value store is ideal for applications that require massive amounts of simple
data like sensor data or for rapidly changing data such as stock quotes.
Key-value stores support massive data sets of very primitive data. Amazon
Dynamo was built as a key-value store.

238

M. Periasamy and P. Raj

Document Databases (DocumentDB) A document database contains a collection of key-value pairs stored in documents. The documentDB support more
complex data than the key-value stores. While it is good at storing documents, it
was not designed with enterprise-strength transactions and durability in mind.
DocumentDB are the most flexible of the key-value style stores, perfect for
storing a large collection of unrelated and discrete documents. Unlike the
key-value stores, these systems generally support secondary indexes and multiple types of documents (objects) per database and nested documents or lists. A
good application would be a product catalog, which can display individual items,
but not related items. One can see what is available for purchase, but one cannot
connect it to what other products similar customers bought after they viewed
it. MongoDB and CouchDB are examples of documentDB systems.
Column-Family Databases A column-family database can handle semistructured data because in theory every row can have its own schema. It has
few mandatory attributes and few optional attributes. It is a powerful way to
capture semi-structured data but often sacrifices consistency for ensuring the
availability attribute. Column-family databases can accommodate huge amounts
of data, and the key differentiator is that it helps to sift through the data very fast.
Writes are really faster than reads so one natural niche is real-time data analysis.
Logging real-time events is a perfect use case, and another one is random and
real-time read/write access to the Big Data. Googles Bigtable was built on a
column-family database. Apache Cassandra [6], the Facebook database, is
another well-known example, which was developed to store billions of columns
per row. However, it is unable to support unstructured data types or end-to-end
query transactions.
Graph Databases A graph database uses nodes, relationships between nodes,
and key-value properties instead of tables to represent information. This model
is typically substantially faster for associative data sets and uses a schema-less
and bottom-up model that is ideal for capturing ad hoc and rapidly changing
data. Much of todays complex and connected data can be easily stored in a
graph database where there is great value in the relationships among data sets. A
graph database accesses data using traversals. A traversal is how a graph is
queried, navigating from starting nodes to related nodes according to an algorithm, finding answers to questions like what music do my friends like that I
dont yet own? or if this power supply goes down, what web services are
affected? Using traversals, one can easily conduct end-to-end transactions that
represent real user actions.

10.5.4 Cloud-Based Databases for Big Data

Relational database management systems are an integral and indispensable component in enterprise IT, and their importance is all set to grow. However, with the
advances of cloud-hosted and managed computing, network, and storage

10

Big Data Analytics: Enabling Technologies and Tools

239

infrastructures, the opportunity to offer a DBMS, as an offloaded and outsourced

service, is gradually gaining momentum. Carlo Curino and his team members have
introduced a new transactional database as a service (DBaaS) [5]. A DBaaS
promises to move much of the operational burden of provisioning, configuration,
scaling, performance tuning, backup, privacy, and access control from the database
users to the service operator, effectively offering lower overall costs to users.
However, the DBaaS being provided by leading CSPs do not address three important challenges: efficient multi-tenancy, elastic scalability, and database privacy.
The authors argue that before outsourcing database software and management into
cloud environments, these three challenges need to be surmounted. The key technical features of DBaaS are:
1. A workload-aware approach to multi-tenancy that identifies the workloads that
can be colocated on a database server achieving higher consolidation and better
performance over existing approaches
2. The use of a graph-based data partitioning algorithm to achieve near-linear
elastic scale-out even for complex transactional workloads
3. An adjustable security schema that enables SQL queries to run over encrypted
data including ordering operations, aggregates, and joins
An underlying theme in the design of the components of DBaaS is the notion of
workload awareness. By monitoring query patterns and data accesses, the system
obtains information useful for various optimization and security functions, reducing
the configuration effort for users and operators. By centralizing and automating
many database management tasks, a DBaaS can substantially reduce operational
costs and perform well. There are myriad of advantages of using cloud-based
databases, some of which are as follows:
Fast and automated recovery from failures to ensure business continuity
A built-in larger package with nothing to configure or comes with a straightforward GUI-based configuration
Cheap backups, archival, and restoration
Automated on-demand scaling with the ability to simply define the scaling rules
or facility to manually adjust
Potentially lower cost, device independence, and better performance
Scalability and automatic failover/high availability
Anytime, anywhere, any device, any media, any network discoverable, accessible, and usable
Less capital expenditure and usage-based payment
Automated provisioning of physical as well as virtual servers in the cloud
Some of the disadvantages include:

Security and privacy issues

Constant Internet connection (bandwidth costs!) requirement
Loss of controllability over resources
Loss of visibility on database transactions
Vendor lock in

240

M. Periasamy and P. Raj

Thus, newer realities such as NoSQL and NewSQL database solutions are fast
arriving and being adopted eagerly. On the other hand, the traditional database
management systems are being accordingly modernized and migrated to cloud
environments to substantiate the era of providing everything as a service. Data as
a service, insights as a service, etc., are bound to grow considerably in the days to
come as their realization technologies are fast maturing.

10.6

Data Science

Data science [4] is an independent discipline that emerged since the beginning of
the twenty-first century. It suggests models and frameworks to support the analysis
of complex data. Data analysis has a long history; it has been practiced by
librarians, statisticians, scientists, sales executives, financial experts, and marketing
technicians for years to induce attention of the customers to focus on their products.
In 1994, many companies introduced their marketing strategies based on databases, as marketing prediction to sell the products to overwhelm the competition.
The first conference on data science was held in 1996. The terminologies such as
data mining, data analysis, and data science were used by researchers and scientists
were introduced at that time. Data mining is defined as an application for the
extraction of patterns from data using specific algorithm. Statistics plays a major
role in processing of data, through unique statistical methods. Data mining refers to
extracting, importing, and analyzing data from huge databases to determine the
relationship among internal and external factors and the impacts on sales, profits,
and customer satisfaction. At the moment, analyzing the data sounds difficult as it
requires information from stochastic statistical model and unknown algorithmic
model. Statisticians observed uncertainty in their results while analyzing these data
models. This has resulted in research in computing in the field of data analysis,
enabling the statisticians to merge statistical methods and computing using
web/Internet to produce innovations in the field of data science. The view and the
idea are to find pace-based decision-making, instead of traditional calculations, and
to find predictive model using quantitative analysis as primary object.
Data science brings together expertise from data visualization, hacker mindsets,
and various domains to process data that are generated in a specific domain or
across domains. Rightly, the journal of data science defines or claims data science
as almost everything that has something to do with data: collecting, analyzing,
modeling, its application, and all sorts of application. For example, data mining
the processing of data to find interesting patterns in data that are not easily observed
using basic queries is widely used for decision-making in diverse domains such as
surveillance, market research, and asset management of municipal infrastructure
such as pipe lines. This pattern recognition could be performed on multiple types of
data such as audio, video, images, and texts. However there are limitations when we

10

Big Data Analytics: Enabling Technologies and Tools

241

apply these techniques to vast amounts of data, and the traditional data processing
techniques become inadequate.
This necessitates for appropriate data-handling frameworks and tools that could
handle the vast amount of data in a timely and efficient manner. It should be noted
that a lot of junk is also being generated online and that can creep into important
decision-making tasks. Though we boast of efficient decision-making techniques in
place, most of these systems are also fragile and feeble. For example, a fake tweet
on the hacked Associated Press Twitter account that President Barrack Obama was
injured in a bomb blast in White House caused a momentary decline in the value of
S&P 500 index about 0.9 %. This is just enough to wipe out $ 130 billion in a matter
of milliseconds. This is a crude reminder of how sensitive the systems are toward
the information generated online. A sensible recommendation based on a thorough
data analysis for such a highly impacting decision-making would require the
process of streaming large quantities of data from various sources; verifying the
authenticity of news from various news agency sites that are in audio (podcasts),
video (online live telecasts), and text formats (tweets, webpage contents); arriving
at a conclusion based on the analysis as and when the information is generated; and
instantly feeding it to the capital markets. All the aforementioned processes will
need to be performed at the lightning speed. Delays even by a couple of seconds
might render this information completely useless or might create negative and
irreparable consequences.

10.6.1 Basic Concepts

Data science is an activity to identify the problem (automated), process the requirement, and visualize the solution for the user. The activity involves collection of
data, preparation of the file system and storage, and analysis of data and the request,
focusing on result visualization and management to satisfy the customers requirements. It also involves predictive analysis of other aspects such as scalability,
storage, safety, and security.
Key factors involved with data science include classification, probability estimation to predict, value estimation by regression analysis, identification of similarities, determining unity between entities, describing attributes characterization,
link prediction from the basis, and reduction of data and technology to come up
with innovative models and modeling. Data science concepts and operations are
extracted from data engineering, programming analysis with statistics, social data
warehousing, and through other related processes. Data science [4] is of interest to
business by improving the decision-making concepts. Generally, Data science is the
function of quantitative and qualitative methods to solve related questions and
forecast conclusions. Businesses require real-time analysis using Data science.
The necessary skills relating to data science are as follows:
Technical communication with users, explaining the statistical aspect and problem requests to clear the technicality of the problem.

242

M. Periasamy and P. Raj

Expertise in adopting the application and getting the required data for the given
particular context.
For complex systems, data scientist has to create constraints according to the
problem request, to thoroughly understand the issues from the user side.
Skills in data description, defining the variables, and accordingly collecting
those data and making it available for analysis through network sharing.
For process and data transformation, the data scientist must have the ability to
resolve issues with feasible solutions and to interpret data in presentable form.
Visualization and keeping the presentable results with precision, detail, and
charts for a more effective communication with customers.
Expertise in handling of ethical issues such as security violation, privacy, and
communication limitation and keeping the data away from malpractice.

10.6.2 The Role of Data Scientist

A data scientist could be a single person or a group of people working in an
organization, whose main role is to perform statistical analysis, data mining, and
retrieval processes on a huge amount of data to determine trends and other
diagnose-related information. They normally work on data analysis on huge data
sets in data warehouse or common access data centers to resolve many issues with
respect to optimization of performance and factors relating to business intelligence
(BI). This will be useful for the business to act upon the inherent trends, so they can
plan and take correct corporate decisions.
In general, the data scientists role is to analyze big data or data repository that
are preserved by an organization or some other sources and to analyze the objectives using mathematics and statistical models to obtain instructions and recommendation to help with excellent business and marketing decision-making. The
responsibilities include handling the product development, technology transformation to fit with current trends, and building architecture in the back end for the
analysis and front end for demonstration of the results.

10.7

Conclusion

Big data analytics is now moving beyond the realm of intellectual curiosity and
propensity to make tangible and trendsetting impacts on business operations,
offerings, and outlooks. It is no longer hype or a buzzword and is rapidly becoming
a core requirement for every sort of business enterprise to be relevant and rightful to
their stakeholders and end users. Being an emerging and evolving technology, it
needs a careful and cognitive analysis before its adoption.
In this chapter, we have provided some detail on enabling big data analytics and
data science concepts. We have commenced the chapter with the key drivers for big

10

Big Data Analytics: Enabling Technologies and Tools

243

data. Then we proceeded by describing the significance of generating actionable

insights out of data sets in huge amounts in order to determine sustainable business
value. We have incorporated the details regarding the proven and potential tools
and provided tips for simplifying and streamlining big data analytics. We have also
discussed the prominent database systems such as NoSQL databases. This is an
introductory chapter to expose what is generally needed for readers to start with in
the fast-moving big data field.
Acknowledgment Sincere thanks are due to the following individuals for their contribution and
support in developing this chapter:
Ernestasia Siahaan, PhD candidate, Department of Intelligent Systems Multimedia Computing,
Faculty of Electrical Engineering, Mathematics, and Computer Science, Delft University of
Technology, the Netherlands
Mohanasundar Radhakrishnan, PhD fellow, UNESCO-IHE Institute for Water Education, Delft,
the Netherlands
Vadivazhagan Meiyazhagan, lead administrator, Wipro Infotech, India

References
1. Liang C, Yu FR (2014) Wireless network virtualization: a survey, some research issues and
challenges, communications surveys & tutorials. IEEE 17(1):358380. doi:10.1109/COMST.
2014.2352118
2. Han J, Kamber M, Pei J (2011) Data mining: concepts and techniques. Elsevier Publications,
Amsterdam/Boston. ISBN 978-0-12-381479-1
3. Grolinger K, Higashino WA, Tiwari A, Capretz MAM (2013) Data management in cloud
environments: NoSQL and NewSQL data stores. J Cloud Comput: Adv Syst Appl 2:22. doi:10.
1186/2192-113X-2-22
4. Lillian Pierson, Wiley, Carl Anderson, Ryan Swanstrom (2015) Data science for dummies.
John Wiley & Sons Inc, New York, United States. ISBN 9781118841556
5. Raj P (2012) Cloud enterprise architecture. CRC Press, Boca Raton. ISBN 9781466502321
6. Strickland R (2014) Cassandra high availability. Packt Publishing, Birmingham, Ebook
ISBN:978-1-78398-913-3 | ISBN 10:1-78398-913-0
7. White T, Cutting D (2009) Hadoop: the definitive guide. OReil-ly Media, Beijing, ISBN
0596521979; ISBN13: 9780596521974
8. McKinsey Global Institute (2011) Big data: the next frontier for innovation, competition, and
productivity. Available at: http://www.mckinsey.com/insights/business_technology/big_data_
the_next_frontier_for_innovation
9. Neo Technology (2011) NoSQL for the enterprise, white paper. Available at: http://www.
neotechnology.com/tag/nosql/
10. Russom P (2011) Hadoop: revealing its true value for business intelligence, white paper.
Available at: http://www.tdwi.org

Chapter 11

A Framework for Data Mining

and Knowledge Discovery in Cloud
Computing
Derya Birant and Pelin Yldrm

Abstract The massive amounts of data being generated in the current world of
information technology have increased from terabytes to petabytes in volume. The
fact that extracting knowledge from large-scale data is a challenging issue creates a
great demand for cloud computing because of its potential benefits such as scalable
storage and processing services. Considering this motivation, this chapter introduces a novel framework, data mining in cloud computing (DMCC), that allows
users to apply classification, clustering, and association rule mining methods on
huge amounts of data efficiently by combining data mining, cloud computing, and
parallel computing technologies. The chapter discusses the main architectural
components, interfaces, features, and advantages of the proposed DMCC framework. This study also compares the running times when data mining algorithms are
executed in serial and parallel in a cloud environment through DMCC framework.
Experimental results show that DMCC greatly decreases the execution times of
data mining algorithms.
Keywords DMCC Data mining Cloud computing Knowledge discovery
Classification Clustering Association rule mining

11.1

Introduction

As a result of technological innovations and developments, enormous amounts of

data are generated every day in a wide range of areas such as business, education,
healthcare, government, finance, social media, and many others. The increase in the
amount of data that is generated creates the potential to discover valuable

D. Birant (*)
Department of Computer Engineering, Dokuz Eylul University, Izmir, Turkey
e-mail: derya@cs.deu.edu.tr
P. Yldrm
Department of Software Engineering, Celal Bayar University, Manisa, Turkey
e-mail: pelin.yildirim@cbu.edu.tr
Springer International Publishing Switzerland 2016
Z. Mahmood (ed.), Data Science and Big Data Computing,
DOI 10.1007/978-3-319-31861-5_11

245

246

D. Birant and P. Yldrm

knowledge from it. However, it also generates a need to deal with data processing in
an efficient and low-cost way.
Data mining is the process of extracting interesting (previously unknown and
potentially useful) knowledge or patterns from large data repositories such as
databases, data warehouses, data marts, extensible markup language (XML) data,
files, and so on. Data mining is regarded as one of the main steps of Knowledge
Discovery in Databases (KDD).
The main data mining methods are classification, clustering, and association rule
mining (ARM). Classification is the process of developing an accurate model
according to classes that use the features in data. It enables the categorization of
new data into these predefined classes. Clustering is used to identify groups and
discover structure in unlabelled data. ARM is useful for finding the rules among
different items in large datasets.
The data mining in cloud computing allows to centralize the management of
software and data storage. Cloud computing is a new paradigm that relies on
sharing computing resources by providing parallel processing and data storage
over the Internet on a pay-for-use basis. Cloud systems can be effectively used to
handle the parallel mining of large datasets since they provide scalable storage and
processing services as well as software platforms for running data mining
applications.
This study presents a novel cloud computing-based framework, data mining in
cloud computing (DMCC), that has the ability to apply classification, clustering,
and ARM methods on huge amounts of data efficiently. We propose this framework
to enhance the existing data mining applications to take advantage of features
inherent to cloud characteristics that include scalability, efficiency, and ease of
use. This study compares the running times when data mining algorithms are
executed in serial and parallel in a cloud environment. A performance comparison
between the traditional and DMCC frameworks shows that cloud-based applications can run in a parallel manner and that DMCC greatly decreases the execution
time of data mining algorithms.
The focus of this study is to present the advantages of combining three core
technologies: data mining, cloud computing, and parallel computing. The novelty
and main contributions of the DMCC framework proposed in this study are as
follows:
It reduces the running time of data mining algorithms by simultaneously
exploiting both the scalable properties of cloud computing and parallel programming methods.
It enables on-demand pricing options of cloud platforms to utilize knowledge
discovery processes with affordable prices.
It allows data mining applications to be accessible from anywhere.
It makes data storage and management easier.
Experimental results in this study show that it is possible to decrease the
execution time of data mining techniques with the help of cloud computing
deployments and implementation of parallel programming structures.

11

A Framework for Data Mining and Knowledge Discovery in Cloud Computing

247

The rest of this chapter is structured as follows: Sect. 11.2 summarizes the
related literature and previous studies. While Sect. 11.3 explains the basic concepts
and methods of data mining, Sect. 11.4 presents the deployment models and service
models of cloud computing. Section 11.5 introduces the proposed DMCC framework and describes its characteristics, promises, and benefits. In Sect. 11.6, the
results obtained from experiments are presented, focusing on an analysis of the
various data mining techniques with different input parameters on several datasets
and also more briefly commenting on the comparison of parallel and serial executions. Section 11.7 presents concluding remarks and possible future studies.

11.2

Related Work

In this section, some related studies that apply data mining techniques on large
datasets using cloud computing technology are described.
Using the advantages of cloud computing technology to tackle the performance
problems in large-scale data mining applications is an approach that has gained
increasing attention [14]. The most common way to perform data mining algorithms effectively on high-volume data is to use MapReduce, a programming model
for parallel data processing that is widely used in cloud environments [5, 6]. An
open-source implementation of MapReduce, Hadoop, can be used for distributed
data processing [7]. However, many scientific applications, which have complex
communication patterns, still require low-latency communication mechanisms. For
this reason, in this study, we introduce a novel framework to help fill this
critical void.
The Sector/Sphere framework is another way to mine large-scale datasets using
computer clusters connected with wide-area high-performance networks [8
10]. While high-level data parallel frameworks like Sector/Sphere simplify the
design and implementation of large-scale data processing systems, they do not
naturally support many important data mining algorithms and can lead to inefficient
learning systems [11]. Thus, in this study, we introduce a novel framework that
supports many different data mining algorithms.
In addition to MapReduce, Hadoop, and Sector/Sphere frameworks, other largescale computation frameworks have been proposed: the GraphLab, Data Mining
Cloud App, and Define-Ingest-Preprocess-Analyze-Report (DIPAR) frameworks.
The GraphLab abstraction was proposed for machine learning and data mining in
the cloud, and it directly targets dynamic, graph-parallel computation in the shared
memory setting [11]. The Data Mining Cloud App framework performs parameter
sweeping operations on clustering and classification algorithms [12]. The DIPAR
framework, which consists of five stages, define, ingest, preprocess, analyse, and
report, was proposed as a way to implement Big Data Science (BDS) in
organizations [13].
Association rule mining (ARM) is a major data mining task in the application of
large-scale and parallel data mining in cloud computing environments

248

D. Birant and P. Yldrm

[14, 15]. Apiletti et al. [16] proposed a novel cloud-based service, SEARUM
(SErvice for Association RUle Mining), that applies ARM on the large volume of
data using MapReduce jobs that are run in the cloud. In another study [17], the
Apriori algorithm, which is a well-known ARM algorithm, was selected as a case
study and a pipelined MapReduce framework was utilized.
In addition to ARM, other well-known data mining tasks have also been investigated in cloud environments such as classification [5, 6], clustering [12, 18, 19],
sequential pattern mining [20], XML data mining [21], web mining [22], semantic
web mining [23], and text mining [24].
In contrast to these studies, the work presented in this contribution (the DMCC
framework) focuses on the implementation of different classification, clustering,
and ARM algorithms on large-scale data using the Microsoft Azure cloud platform
to reduce local workload. In contrast to existing systems, the DMCC is off the shelf,
and thus (1) it is ready to use from anywhere, (2) there is no need to move data,
(3) there is no cost for managing the underlying hardware/software layers, and
(4) the installation of software is not required.

11.3

Data Mining

This section presents most commonly used data mining tasks and the algorithms
which DMCC framework offers.

11.3.1 Classification
Classification is the most studied and commonly applied data mining task. It
develops a model to assign new patterns into one of several predefined classes.
Classification uses a training set D (R1, R2, . . ., Rn) that has some records R,
which consist of a number of attributes R (a1, a2, . . ., am) of which one (aj) is a
target outcome. Classification algorithms try to determine relationships between
attributes in a training set to classify new observations.
In this study, four classification algorithms are developed as Platform as a
Service (PaaS) for cloud end users: Naive Bayes, Decision Tree (C4.5), Random
Forest, and AdaBoost.

11.3.1.1

Naive Bayes Classifier

The Naive Bayes classifier is a well-known statistical classifier that uses Bayes
theorem to calculate unknown conditional probabilities to identify classes of input
samples [25]. The algorithm performs well and returns successful results in many

11

A Framework for Data Mining and Knowledge Discovery in Cloud Computing

249

different areas such as medical diagnosis, pattern recognition, document categorization, banking, and marketing.
The Naive Bayes classifier is generally applied to data that consists of discretevalued attributes. However, in this study, we used a Gaussian distribution to deal
with continuous-valued attributes. The Gaussian distribution, sometimes called the
normal distribution, uses mean and variance 2 values to calculate its probability
density function as follows:
x2
1
f x; ; p e 22
2

11.3.1.2

11:1

Decision Tree (C4.5)

A Decision Tree is one of the commonly used supervised classification methods

that aims to develop a tree to identify unknown target attribute values based on
input parameters. A Decision Tree consists of internal nodes for attributes, branches
to represent attributes values, and leaf nodes to assign classification labels. In this
study, we used the C4.5 algorithm, which computes information gain to construct
the tree based on entropy and the fraction ( fi) of items labelled with value i in the set
as follows:
m
X
I E f  f i log2 f i

11:2

i1

11.3.1.3

Random Forest

A Random Forest is an ensemble learning method that grows multivalued Decision

Trees in different training sets. To classify a new sample, input parameters are
given to each tree in the forest. The trees return predicted outcomes and the result is
selected by taking the majority vote over all the trees in the forest. The Random
Forest is an ideal learning model with the goal of reducing variance and increasing
ease of use.

11.3.1.4

AdaBoost

AdaBoost, an acronym of Adaptive Boosting, combines a number of weak learners

to form a weighted voting machine to achieve better separation between classes. It
is the most appropriate solution for binary classification problems in particular.
AdaBoost specifies a boost classifier FT for each weak learner ft that takes an object

250

D. Birant and P. Yldrm

x as input, and after T layers, it returns a real value indicating the class of the object
as follows:
FT x

T
X

f t x

11:3

t1

11.3.2 Clustering
Clustering, an unsupervised learning technique, is the process of grouping a set of
objects into meaningful clusters in such a way that similar objects are placed within
a cluster. Clustering analysis is currently used in many areas such as image
processing, pattern recognition, segmentation, machine learning, and information
retrieval. The main task of clustering is to compare, measure, and identify the
resemblance of objects based on their features by using a similarity measure such as
the Manhattan, Euclidean, or Minkowski distance for numerical attributes and
Jaccards distance for categorical values. Clustering algorithms are categorized
into five major types, as listed in Table 11.1. One of these, the K-Means, is briefly
described below.

11.3.2.1

K-Means

In this study, the K-Means clustering algorithm is developed as a PaaS for the cloud
end users because of its popularity, computational simplicity, efficiency, and
empirical success. K-Means is a well-known and simple unsupervised learning
algorithm that groups a given dataset into k clusters. Let X {x1, x2, . . ., xn} be a set
of observations. The K-Means algorithm divides n observations into k cluster sets
C {c1, c2, . . ., ck} by minimizing the Sum of Squared Errors (SSE) of the
k clusters. The algorithm involves the following steps:
1. The k initial centroids are chosen randomly.
2. Each object is assigned to the cluster associated with the nearest centroid.
Table 11.1 Categorization of well-known clustering algorithms
Cluster models
Partitioning methods
Hierarchical
methods
Density-based
methods
Grid-based methods
Model-based
methods

Clustering algorithms
K-Means, C-Means, K-Medoids, CLARANS, etc.
Single/complete/average link, BIRCH, ROCK, CAMELEON, etc.
DBSCAN, DENCLUE, OPTICS, etc.
STING, CLIQUE, WaveCluster, etc.
SOM (Self-Organizing Maps), COBWEB, EM (Expectation Maximization), etc.

11

A Framework for Data Mining and Knowledge Discovery in Cloud Computing

251

3. Centroids are updated to the mean of each segment.

4. Steps 2 and 3 are repeated until no object switches clusters.

11.3.3 Association Rule Mining

Association rule mining (ARM), one of the most important and well-researched
techniques of data mining, is the extraction of interesting correlations, relationships, frequent patterns or associations, or general structures among sets of items in
the transactions.
Let I {i1, i2, . . ., im} be a set of m distinct literals called items, T be a
transaction that contains a set of items such that T  I, and D be a dataset D {t1,
t2, . . ., tn} that has n transaction records T. An association rule is an implication of
the form X ) Y, where X  I and Y  I are sets of items called frequent itemsets, and
X \ Y . The rule X ) Y can be interpreted as if itemset X occurs in a transaction,
then itemset Y will also likely occur in the same transaction.
There are two important basic measures for association rules: support and
confidence. Usually thresholds of support and confidence are predefined by users
to drop those rules that are not particularly interesting or useful. In addition to these
measures, additional constraints can also be specified by the users such as time,
item, dimensional, or interestingness constraints.
Support of an association rule in the form of X ) Y is defined as the percentage
of records that contain both X and Y itemsets to the total number of transactions in
the dataset D. Support is calculated by the following formula:
SupportX ) Y

Number of transactions contain both X and Y

Total number of transactions in D

11:4

Suppose the support of rule X ) Y is 1 %. This means that 1% of the transactions

contain X and Y items together.
The Confidence of an association rule in the form of X ) Y is defined as the
percentage of the number of records that contain both X and Y itemsets with respect
to the total number of transactions that contain X, as follows:
ConfidenceX ) Y

SupportX [ Y
SupportX

11:5

Suppose the confidence of the association rule X ) Y is 85 %. This means that 85 %

of the transactions that contain X also contain Y.
In this study, the Apriori algorithm is developed as a PaaS for cloud end users.
Apriori is a well-known ARM algorithm that requires many passes over the
database, generating many candidate itemsets, pruning those itemsets whose supports are below the predefined threshold, and storing frequent itemsets. In each

252

D. Birant and P. Yldrm

pass, k itemsets are found. For example, the rule X, Y ) Z can be produced as a
frequent three itemset, where X, Y, and Z represent in any items in the dataset.

11.4

Cloud Computing

Cloud computing offers resources, software applications, and infrastructures over

the Internet to users [26]. This section explains the basic deployment models and
service models of cloud computing.

11.4.1 Deployment Models

Cloud computing broadly breaks down into three different deployment models:
public, private, and hybrid cloud [27].

11.4.1.1

Public Cloud

The public cloud is the common cloud computing model in which a service
provider hosts resources, such as applications and data storage, available to the
general public over the Internet.

11.4.1.2

Private Cloud

The private cloud (also called the internal cloud) consists of cloud infrastructure
that is operated solely for a single organization, whether managed internally or by a
third party.

11.4.1.3

Hybrid Cloud

A hybrid cloud comprises both private (internal) and public (external) cloud
services, offering the benefits of multiple deployment models.

11.4.2 Service Models

Cloud computing is divided into three broad service categories: Infrastructure as a
Service (IaaS), Platform as a Service (PaaS), and Software as a Service (SaaS) [28].

11

A Framework for Data Mining and Knowledge Discovery in Cloud Computing

11.4.2.1

253

Infrastructure as a Service (IaaS)

Infrastructure as a Service (IaaS) is a cloud computing model that provides

virtualized computing resources over the Internet, including some virtual machines
and hardware resource units. IaaS can also be classified into two categories:
Computation as a Service (CaaS) and Data as a Service (DaaS). CaaS refers to
the ability to buy or rent specific resourcesmainly processor, RAM, operating
system, or deployed softwarefor a period of time to perform difficult computation
calculations. DaaS refers to the ability to buy or rent storage space to store the users
data, regardless of its type. IaaS clouds often offer additional resources such as a
virtual-machine disc image library, block- and file-based storage, load balancers, IP
addresses, firewalls, and virtual local area networks. Typical examples are Amazon
EC2 (Elastic Cloud Computing), GoGrid, Rackspace Cloud, Amazon S3 (Simple
Storage Service), and Google Compute Engine.

11.4.2.2

Platform as a Service (PaaS)

Platform as a Service (PaaS) is a form of cloud computing in which cloud providers

offer a computing platform that typically includes an operating system, database,
web server, and environments for building, testing, and deploying custom applications. Application developers can develop and run their software solutions on a
cloud platform without the cost of managing the underlying hardware and software
layers. Some common examples of PaaS include Microsoft Azure, Heroku, Force.
com, Google App Engine, and Apache Stratos.

11.4.2.3

Software as a Service (SaaS)

Software as a Service (SaaS) is a cloud computing model in which users are

provided access to application software and databases. It allows users to run
existing applications remotely from the cloud to meet their requirements. Instead
of cloud users, cloud providers manage the infrastructure and platforms that run the
applications. Some common examples of SaaS include Google Calendar, Microsoft
Office 365, Yahoo Maps, Flickr, and Gmail.
This study focuses on developing and deploying a framework in the public cloud
using a PaaS approach.

11.5

The Proposed DMCC Framework

DMCC (data mining in cloud computing) is a framework that uses current algorithms to support data mining in a cloud environment. This section describes the
design and implementation of the DMCC framework that we developed to archive,

254

D. Birant and P. Yldrm

analyse, and mine large datasets in the cloud environment. The DMCC framework
offers the accessibility to increase computational power for users who need to
perform large-scale data/compute intensive data mining applications. However, to
perform such computations, two major conditions need to be satisfied: (i) the data
mining application should be deployed on the cloud, and (ii) an appropriate parallel
runtime support should be implemented.

11.5.1 Overview
In the DMCC framework, data mining techniques are implemented as a dynamic
library, and data mining applications are stored on the cloud. The aim is to develop
a framework suitable for executing data analysis and knowledge discovery tasks
over a wide range of different applications within different sectors such as medicine, education, entertainment, transportation, and accounting. It allows the user to
discover useful knowledge from virtually stored data that reduces the costs of
infrastructure and storage.
The DMCC framework provides a user interface that facilitates the utilization of
data mining applications in a cloud environment. Figure 11.1 shows a screenshot
from a data mining application developed in Visual Studio. Weka open-source data
mining library is used to perform classification, clustering, and ARM on large-scale
datasets. It is hosted on the Microsoft Azure cloud platform. Users can access this

Fig. 11.1 Screenshot from a data mining application deployed on the cloud

11

A Framework for Data Mining and Knowledge Discovery in Cloud Computing

255

data mining application using a web browser, regardless of their location. They
perform mining algorithms on selected massive amounts of data, independently
from the installed operating system.

11.5.2 DMCC Framework Architecture

When designing a data mining system based on cloud computing, Lin [29] divided
the platform, bottom-up, into three layers, the algorithm layer, task layer, and user
layer. In contrast to previous studies, the DMCC framework is typically built using
a four-layer architecture model consisting of an application layer, data mining
engine layer (DMEL), data access layer (DAL), and data layer. The application
layer provides the application-user interactions, the DMEL executes the data
mining algorithms, the DAL encapsulates the data access functionality, and the
data layer is responsible for data persistence and data manipulation. The DMCC
framework architecture is a multilayered, parallel processing-enabled, cloud-based
architecture, as shown in Fig. 11.2.
The framework is designed so that parallel computing can be done very simply.
In particular, if a user defines a function f on the parts of a dataset d, then he invokes
the command

Fig. 11.2 DMCC architecture

256

D. Birant and P. Yldrm

Fig. 11.3 Pseudocode for the classification method in DMCC

Parallel.For(0, d.parts.count, i >

{
f(d.part[i]);
});
This applies the user-defined function f to each part of the dataset d in parallel. In
other words, the Parallel. For command executes the same pieces of code several
times in parallel with different data parts.
Our framework was designed to significantly decrease the amount of time necessary to run n-fold cross-validation on a dataset using any given classifier. Figure 11.3
presents the pseudocode for this classification task in DMCC. It takes the number of
folds (n) as an input parameter. In n-fold cross-validation, the entire dataset is divided
into n nonoverlapping parts, and then the algorithm is trained on n-1 parts and tested
on the remaining one. The Parallel. For command executes both the training and
testing operations for a C4.5 Decision Tree algorithm in parallel. Other classification
algorithms such as Random Forest can also be called instead of the C4.5 algorithm.
The accuracy rate is calculated as the average of the n prediction accuracies.
Figure 11.4 presents the pseudocode for a clustering task in DMCC. It executes
the K-Means algorithm n times in parallel with different k values to determine the
optimal number of clusters. It does this because determining the optimal k value is a
frequent problem in data clustering that affects the amount of error in the resulting
clusters. Increasing k will always reduce the Sum of Squared Errors (SSE) in the
resulting clustering; hence the elbow method looks at the percentage of variance
explained as a function of the number of clusters and determines the point at which
the marginal gain drops, creating an angle in the graph. Finally, the clusters are
created using the optimal k value.
Figure 11.5 presents the pseudocode for ARM in DMCC. It takes the minimum
support and confidence values as input parameters and executes the Apriori algorithm with these parameters to output the rules.

11

A Framework for Data Mining and Knowledge Discovery in Cloud Computing

257

Fig. 11.4 Pseudocode for the clustering method in DMCC

Fig. 11.5 Pseudocode for ARM in DMCC

11.5.3 DMCC Framework Features

The DMCC framework has many advantages over traditional grid-based, distributed, or parallel computing techniques.
Grid computing is a collection of computer resources from multiple locations,
acting in concert to perform very large tasks using distributed and parallel computation. However, cloud computing provides the technologies to compute parallel
data mining applications at much more affordable prices than traditional parallel
computing techniques.
Distributed computing is an environment in which the application and data is
broken into pieces and executed on individual machines. However, cloud computing stores the data at a data centre and does not distribute it to individual machines
for processing.

258

D. Birant and P. Yldrm

The DMCC framework deploys data mining applications in cloud environments

for the following reasons:
Cloud resources are dynamically reallocated per demand; hence the cloud
approach maximizes the use of computing power and thus significantly reduces
the running time of data mining algorithms.
Cloud computing eliminates the investment needed for stand-alone software or
servers. It allows developers to avoid costs for items such as servers, air
conditioning, and rack space, allowing many data mining developers to reduce
their costs.
Cloud computing allows developers to get their data mining applications operational within a short time with improved manageability and less maintenance.
Developers focus on data mining projects instead of infrastructure.
Cloud computing offers on-demand pricing so that commercial data mining
software can be provided at affordable prices [30].
In cloud environments, providers are able to devote resources to solving security
issues, so security can be improved by data centralization, increased securityfocused resources, and so on.
Cloud computing enables users to access systems using a web browser, regardless of their location or which device they use (e.g. PC, laptop, or mobile phone)
so that data mining applications can be accessed via the Internet and users can
connect from anywhere.
Cloud computing offers cost-effective and policy-driven solutions to store big
data so that large amount of data can be processed in place without needing to
move it.
As a result, the major features of the DMCC framework include high performance, scalability, simplified access interfaces, quality of service, ease of use, costeffectiveness, and conceptual simplicity in a large-scale and data-massive environment. The DMCC framework is a practical alternative to traditional distributed or
grid-based data mining solutions. Data mining in cloud computing can be considered as the future of data mining because of the advantages listed above.

11.6

Experimental Results

Because of explosive data growth and the amount of computation involved in data
mining, a high-performance cloud computing is an excellent resource that is
necessary for efficient data mining. This study presents the importance of parallel
data analysis for a data mining application in a cloud environment. The aim of this
study is to show that the execution time of data mining algorithms can be significantly reduced by combining cloud computing and parallel computing
technologies.

11

A Framework for Data Mining and Knowledge Discovery in Cloud Computing

259

11.6.1 Dataset Description

We performed experimental studies on four different large datasets obtained from
the UCI Machine Learning Repository: the EEG (electroencephalography) Eye
State, Skin Segmentation, KDD Cup 1999, and Census Income datasets.

11.6.1.1

EEG Eye State Dataset

This dataset is used to classify an eye state as eye closed (1) or eye opened (0).
There are 15 attributes in the dataset that were detected via a camera during the
EEG measurement. The dataset consists of 14,980 samples. A brief description of
the dataset is presented in Table 11.2.

11.6.1.2

Skin Segmentation Dataset

The Skin Segmentation dataset consists of R (red), G (green), and B (blue) colour
values from face images of various age groups (young, middle, or old), race groups
(white, black, or Asian), and gender. RGB attributes have continuous values
ranging from 0 to 255. The dataset contains 245,057 labelled samples, where
50,859 are skin/face samples (class label 1) and 194,198 are non-skin/not-face
samples (class label 2). Table 11.3 shows statistical information about the data.
Table 11.2 Description of the EEG Eye State dataset
Attributes
AF3
F7
F3
FC5
T7
P7
O1
O2
P8
T8
FC6
F4
F8
AF4
Class (eye detection)

Min value
Max value
1030.77
309,231
2830.77
7804.62
1040
6880.51
2453.33
642,564
2089.74
6474.36
2768.21
362,564
2086.15
567,179
4567.18
7264.10
1357.95
265,641
1816.41
6674.36
3273.33
6823.08
2257.95
7002.56
86.67
152,308
1366.15
715,897
0 (eye open) 1 (eye closed)

Mean
4321.92
4009.77
4264.02
4164.95
4341.74
4644.02
4110.4
4616.06
4218.83
4231.32
4202.46
4279.23
4615.20
4416.44

Std. dev.
2492.07
45.94
44.29
5216.40
34.74
2924.79
4600.93
29.29
2136.41
38.05
37.79
41.54
1208.37
5891.29

260

D. Birant and P. Yldrm

Table 11.3 Description of the Skin Segmentation dataset

Attributes
B (blue)
G (green)
R (red)
Class (skin detection)

11.6.1.3

Min value
Max value
0
255
0
255
0
255
1 (skin) 2 (non-skin)

Mean
125.07
132.51
123.18

Std. dev.
62.26
59.94
72.56

KDD Cup 1999 Dataset

The purpose of this dataset is to build a network intrusion detector capable of

distinguishing between attacks and normal connections. In this study, we used the
10 % KDD Cup 1999 dataset, which contains 494,021 connection records, each
with 38 numeric features (i.e. duration, number of failed login attempts, number of
data bytes, number of urgent packets, and so on), except three categorical attributes
(protocol type, service, and status of the connection flag). It contains one normal
and 22 different attack types (class labels): normal, back, buffer_overflow,
ftp_write, guess_passwd, imap, ipsweep, land, loadmodule, multihop, neptune,
nmap, perl, phf, pod, portsweep, rootkit, satan, smurf, spy, teardrop, warezclient,
and warezmaster.

11.6.1.4

Census Income Dataset

The Census Income dataset has 48,842 individual records and contains demographic information about people. Each record has 15 attributes for each person:
age, working class, final weight, education, years of education, marital status,
occupation, type of relationship, race, sex, capital gain, capital loss, weekly working hours, native country, and salary.

11.6.2 Classification Results

We compared the performances of different classification algorithms in the cloud
environment in terms of both their accuracy rates and execution times when running
in parallel and serial. The average times shown in the figures and tables were
calculated using the results of five repeated runs.
The accuracy rate of a classification algorithm is generally defined as the
closeness of the predicted values to the actual values and shows the success of
the algorithm on the selected dataset. To evaluate the accuracy rates, we used a
fivefold cross-validation technique. Thus, the dataset is divided into five parts, 4/5
for training and 1/5 for testing, and the accuracy rates of the algorithms were
evaluated according to five experiments that were acquired independently on
different test sets.

11

A Framework for Data Mining and Knowledge Discovery in Cloud Computing

261

Tables 11.4, 11.5, and 11.6 show the execution times and accuracy rates of three
classification algorithms, C4.5 Decision Tree, AdaBoost, and Random Forest, on
three datasets, EEG Eye State, Skin Segmentation, and KDD Cup 1999. According
to the results, the Random Forest algorithm has the highest accuracy rate, but
performs poorly on the datasets in terms of execution times. The AdaBoost algorithm shows the best performance for the Eye State dataset in the case of parallel
execution, but the accuracy rate of this classification algorithm is the worst.
Classification algorithms show similar accuracy rates for the KDD Cup 1999
dataset, highlighting the benefit of using a large training set to obtain a better
learning mechanism.
Comparing the execution times, we find that the running times of different data
mining algorithms on the same dataset are different. However, from these results, it
is clearly evident that the parallel run times in the cloud environment perform
competitively well for all classification algorithms. Although we used different
datasets with different sizes, the Random Forest algorithm shows considerably high
accuracy rates compared to the other algorithms; however we still need to compare
it with other algorithms or use alternative approaches for different types of problems and datasets.
We measured the performance of the Naive Bayes classification algorithm in a
cloud environment. The classifier created from the training set used a Gaussian
distribution to deal with continuous data. Figure 11.6 compares the parallel and

Table 11.4 Comparison of classification algorithms on EEG Eye State data

EEG Eye State data
Classification algorithms
C4.5 Decision Tree
AdaBoost
Random Forest

Serial (sec.)
4.184
2.020
6.078

Parallel (sec.)
0.696
0.345
0.610

Accuracy rate (%)

84.52
67.16
90.24

Table 11.5 Comparison of classification algorithms on Skin Segmentation data

Skin Segmentation data
Classification algorithms
C4.5 Decision Tree
AdaBoost
Random Forest

Serial (sec.)
13.298
16.778
46.070

Parallel (sec.)
3.250
3.423
5.890

Accuracy rate (%)

99.92
89.62
99.94

Table 11.6 Comparison of classification algorithms on KDD Cup 1999 data

KDD Cup 1999 data
Classification algorithms
C4.5 Decision Tree
AdaBoost
Random Forest

Serial (sec.)
232.822
285.734
262.862

Parallel (sec.)
73.876
64.085
48.297

Accuracy rate (%)

99.94
97.78
99.97

262

D. Birant and P. Yldrm

serial execution of the Naive Bayes algorithm on different datasets in a cloud

environment. From these results, it is clearly evident that it is possible to decrease
long-term consumption fairly well using many computing resources concurrently.
Hence, this study confirms the importance of parallel data analysis in data mining
applications to provide good performance in a cloud environment.
Figures 11.7 and 11.8 show the average running times of the Naive Bayes
algorithm for varying input sizes, 5000, 10,000, and 14,980 records for the EEG
Eye State dataset and 50,000, 100,000, 150,000, 200,000, and 245,057 records for
the Skin Segmentation dataset. The results show that the running time increases
more sharply when the data size increases. When the amount of data is increased,
the gap between serial and parallel executions increases significantly.

Fig. 11.6 Parallel and

serial executions of the
same data mining algorithm
in a cloud environment
Execution time (sec.)

Naive Bayes with Gaussian Distribution

Eyestate Dataset

Skin Dataset

Serial

1.15

3.43

Parallel

0.26

1.23

Naive Bayes - EEG Eye State Dataset

Execution time (sec.)

Fig. 11.7 Average running

times for varying input sizes
of the EEG Eye State
dataset

4
3.5
3
2.5
2
1.5
1
0.5
0

1.2
1
0.8
0.6
0.4
0.2
0

5000
records

10000
records

14980
records

Serial

0.39

0.84

1.14

Parallel

0.17

0.29

0.28

A Framework for Data Mining and Knowledge Discovery in Cloud Computing

Fig. 11.8 Average running

times for varying input sizes
of the Skin Segmentation
dataset

263

Naive Bayes - Skin Segmentation Dataset

Execution time (sec.)

11

4.5
4
3.5
3
2.5
2
1.5
1
0.5
0

50000
records

100000
records

150000
records

200000
records

245057
records

Serial

0.38

1.61

2.38

3.38

3.99

Parallel

0.24

0.56

0.81

1.05

1.12

11.6.3 Clustering Results

We specifically analysed the performance of clustering applications deployed in a
cloud environment to determine the overhead of virtualized resources and understand how clustering algorithms with different numbers of clusters perform on
cloud resources. We also evaluated the K-Means clustering algorithm for different
numbers of clusters in order to understand clustering errors.
Tables 11.7, 11.8, and 11.9 show the execution times and SSE for the K-Means
clustering algorithm on three different datasets. The K-Means algorithm was
executed with varying k input values from two to ten in increments of two. These
results allow us to predict how the execution time changes as the number of clusters
increases. Results show that even as the amount of data grows and the number of
clusters increases, the execution time increases within an acceptable range
(a clearly normal way), and the application completes in reasonable time. In other
words, even when the data size and number of clusters increases rapidly, the
execution time increases in a much slower manner than the predicted time.

11.6.4 A Study of Association Rule Mining

We also specifically analysed the number of rules generated by the ARM application deployed in a cloud environment to determine the overhead of virtualized
resources and understand how ARM applications with different numbers of support
values perform on cloud resources. We also evaluated the Apriori algorithm in a
cloud environment for different numbers of confidence values.
The graphs in Figs. 11.9 and 11.10 show the number of rules generated by the
Apriori algorithm for varying minimum support and confidence values, respectively. When the minimum support value is kept constant and confidence values are
changed, the number of rules decreases almost linearly. However, in the other case,

264

D. Birant and P. Yldrm

Table 11.7 Analyses of the K-Means algorithm in a cloud environment for varying numbers of
clusters on the Eye State dataset
Eye State data
K value
2
4
6
8
10

Execution time (sec.)

1.267
1.330
3.020
5.710
5.675

Sum of Squared Error (SSE)

14.45
13.17
10.29
6.94
6.79

Table 11.8 Analyses of K-Means algorithm in a cloud environment for varying numbers of
clusters on the Skin Segmentation dataset
Skin Segmentation data
K value
2
4
6
8
10

Execution time (sec.)

3.120
4.700
6.990
6.630
6.285

Sum of Squared Error (SSE)

20336.13
9842.09
6568.03
5063.96
4055.97

Table 11.9 Analyses of the K-Means algorithm in a cloud environment for varying numbers of
clusters on the KDD Cup 1999 dataset
KDD Cup 1999 data
K value
2
4
6
8
10

Execution time (sec.)

51.130
84.530
89.500
102.530
147.620

Sum of Squared Error (SSE)

551115.68
260735.65
83524.36
57807.90
61061.86

Apriori Algorithm
3000

Number of Rules

2500

2486

2000
1500
1000
652

500

188

106

0
20

30

40

50

50
60

24
70

80
Support (%)

Fig. 11.9 Evaluation of the Apriori algorithm for varying minimum support values and a constant
confidence value of 90

11

A Framework for Data Mining and Knowledge Discovery in Cloud Computing

265

Apriori Algorithm
1200

1118
1031

Number of Rules

1000

859

800

622

600

380

400

188

200

0
40

50

60

70

80

90
99
Confidence (%)

Fig. 11.10 Evaluation of the Apriori algorithm for varying minimum confidence values and a
constant support value of 40

the number of rules decreases exponentially. Results show that deploying an ARM
application on a cloud environment is more critical when the application is executed with small support values.
Cloud infrastructures allow users to access the hardware nodes in which the
virtual machines are deployed. With a cloud infrastructure, we have complete
access to both virtual-machine instances and the underlying nodes; as a result, we
can deploy different virtual-machine configurations, allocating different CPU cores
to each virtual machine. In shared scale mode, six instances are served to an
application, while in standard mode, ten instances that have four cores and 7 GB
of memory are offered for usage. We increased the core count and memory capacity
of each instance by changing the settings to provide better throughput and
performance.
As a result, it is clear that the proposed DMCC framework can effectively
process very large data in a cloud environment. Based on our observations, we
conclude that the DMCC framework can improve the efficiency of data mining
applications.

11.7

Conclusion and Suggestions for Future Work

This study proposes DMCC, a cloud-based framework, developed to perform the

main data mining tasks (classification, clustering, and ARM) over large datasets.
We discussed the main architectural components, interfaces, and features of the
DMCC framework. The DMCC framework has many advantages over traditional
grid-based, distributed, or parallel computing techniques. The important advantages provided by DMCC framework are that a large amount of data can be
processed from anywhere, without moving it, and with on-demand pricing options,
increased security, improved manageability, less maintenance, and increased

266

D. Birant and P. Yldrm

computing power. Furthermore, there is no investment in stand-alone software or

servers.
The general aim of the DMCC framework is to reduce the execution times of
data mining algorithms using the advantages of both cloud computing and parallel
programming methods. We performed experimental studies on the DMCC framework using three real-world datasets. The performance of data mining algorithms
within the DMCC framework was measured by executing them in parallel and
serial modes. In the experiments, (i) classification algorithms (C4.5 Decision Tree,
Random Forest, and AdaBoost) were compared with each other, (ii) the K-Means
clustering algorithm was evaluated in terms of cluster numbers, and (iii) the number
of rules generated by the Apriori algorithm was measured for different support and
confidence values. The experimental results show that cloud systems can be
effectively used to handle the parallel mining of large datasets because they provide
virtualized, dynamically-scalable computing power as well as storage, platforms,
and services.
As future work, this approach could be specialized for the private cloud computing deployment model to enable data mining applications for a specific organization. In this way, corporate companies could benefit from the proposed
framework on their private network, providing only authorized access to the
system. In addition, it would also be possible to expand this framework by integrating different data mining algorithms.

References
1. Geng X, Yang Z (2013) Data mining in cloud computing. International conference on
information science and computer applications, Atlantis Press, 17
2. Petre R (2012) Data mining in cloud computing. Database Syst J 3(3):6771
3. Kamala B (2013) A study on integrated approach of data mining and cloud mining. Int J Adv
Comput Sci Cloud Comput 1(2):3538
4. Hu T, Chen H, Huang L, Zhu X (2012) A survey of mass data mining based on cloudcomputing. IEEE conference on anti-counterfeiting, security and identification, 14
5. Zhou L, Wang H, Wang W (2012) Parallel implementation of classification algorithms based
on cloud computing environment. Indonesian J Electr Eng 10(5):10871092
6. Tan A X, Liu VL, Kantarcioglu M, Thuraisingham B (2010) A comparison of approaches for
large-scale data mining utilizing MapReduce in large-scale data mining, Technical Report
7. Nappinna V, Revathi N (2013) Data mining over large datasets using hadoop in cloud
environment. Int J Comput Sci Commun Netw 3(2):7378
8. Grossman RL, Gu Y (2008) Data mining using high performance data clouds: experimental
studies using sector and sphere. In: Proceedings of the 14th ACM SIGKDD international
conference on knowledge discovery and data mining, 920927
9. Mishra N, Sharma S, Pandey A (2013) High performance cloud data mining algorithm and
data mining in clouds. IOSR J Comput Eng 8(4):5461
10. Mahendra TV, Deepika N, Rao NK (2012) Data mining for high performance data cloud using
association rule mining. Int J Adv Res Comput Sci Softw Eng 2(1)

11

A Framework for Data Mining and Knowledge Discovery in Cloud Computing

267

11. Low Y, Gonzalez J, Kyrola A, Bickson D, Guestrin C, Hellerstein JM (2012) Distributed

graphLab: a framework for machine learning and data mining in the cloud. Proc Very Large
Data Bases (VLDB) Endowment 5(8):716727
12. Marozzo F, Talia D, Trunfio P (2011) A cloud framework for parameter sweeping data mining
applications. In: Proceedings of the IEEE 3th international conference on cloud computing
technology and science, 367374
13. Villalpando LEV, April A, Abran A (2014) DIPAR: a framework for implementing big data
science in organizations. In: Mahmood Z (ed) Continued rise of the cloud: advances and trends
in cloud computing. Springer, London
14. Qureshi Z, Bansal J, Bansal S (2013) A survey on association rule mining in cloud computing.
Int J Emerg Tech Adv Eng 3(4):318321
15. Kamalraj R, Kannan AR, Vaishnavi S, Suganya V (2012) A data mining based approach for
introducing products in SaaS (Software as a service). Int J Eng Innov Res 1(2):210214
16. Apiletti D, Baralis E, Cerquitelli T, Chiusano S, Grimaudo L (2013) SeARuM: a cloud-based
service for association rule mining. In: Proceedings of the 12th IEEE international conference
on trust, security and privacy in computing and communications, 2013, 12831290
17. Wu Z, Cao J, Fang C (2012) Data cloud for distributed data mining via pipelined mapreduce,
vol 7103, Lecture notes in computer science. Springer, Berlin/Heidelberg, pp 316330
18. Ismail L, Masud MM, Khan L (2014) FSBD: a framework for scheduling of big data mining in
cloud computing. In: Proceedings of the IEEE international congress on big data, 514521
19. Masih S, Tanwani S (2014) Distributed framework for data mining as a service on private
cloud. Int J Eng Res Appl 4(11):6570
20. Huang JW, Lin SC, Chen MS (2010) DPSP: Distributed progressive sequential pattern mining
on the cloud, vol 6119, Lecture notes in computer science. Springer, Berlin/Heidelberg, pp
2734
21. Li Z (2014) Massive XML data mining in cloud computing environment. J Multimed 9
(8):10111016
22. Ruan S (2012) Based on cloud-computings web data mining, vol 289, Communications in
computer and information science. Springer, Berlin/Heidelberg, pp 241248
23. Lal K, Mahanti NC (2010) A novel data mining algorithm for semantic web based data cloud.
Int J Comput Sci Secur 4(2):160175
24. Ioannou ZM, Nodarakis N, Sioutas S, Tsakalidis A, Tzimas G (2014) Mining biological data
on the cloud a MapReduce approach, IFIP advances in information and communication
technology, vol. 437. Springer, 96105
25. Yldrm P, Birant D (2014) Naive bayes classifier for continuous variables using novel method
(NBC4D) and distributions. In: Proceedings of the IEEE international symposium on innovations in intelligent systems and applications, 110115
26. Erl T, Puttini R, Mahmood Z (2013) Cloud computing: concepts, technology, & architecture.
Prentice Hall, Upper Saddle River
27. Mahmood Z (2011) Cloud computing for enterprise architectures: concepts, principles and
approaches. In: Mahmood Z, Hill R (eds) Cloud computing for enterprise architectures.
Springer, London/New York
28. Fernandez A, Rio S, Herrera F, Benitez JM (2013) An overview on the structure and
applications for business intelligence and data mining in cloud computing, vol 172, Advances
in intelligent systems and computing. Springer, Berlin/Heidelberg, pp 559570
29. Lin Y (2012) Study of layers construct for data mining platform based on cloud computing, vol
345, Communications in computer and information science. Springer, Berlin/Heidelberg
30. Wu X, Hou J, Zhuo S, Zhang W (2013) Dynamic pricing strategy for cloud computing with
data mining method, vol 207, Communications in computer and information science. Springer,
Berlin/Heidelberg, pp 4054

Chapter 12

Feature Selection for Adaptive Decision

Making in Big Data Analytics
Jaya Sil and Asit Kumar Das

Abstract Rapid growth in technology and its accessibility by general public

produce voluminous, heterogeneous and unstructured data resulted in the emergence of new concepts, viz. Big Data and Big Data Analytics. High dimensionality,
variability, uncertainty and speed of generating such data pose new challenges in
data analysis using standard statistical methods, especially when Big Data consists
of redundant as well as important information. Devising intelligent methods is the
need of the hour to extract meaningful information from Big Data. Different
computational tools such as rough-set theory, fuzzy-set theory, fuzzy-rough-set
and genetic algorithm that are often applied to analyse such kind of data are the
focus of this chapter. But sometimes local optimal solution is achieved due to
premature convergence, so hybridization of genetic algorithm with local search
methods has been discussed here. Genetic algorithm, a well-proven global optimization algorithm, has been extended to search the fitness space more efficiently in
order to select global optimum feature subset. Real-life data is often vague, so fuzzy
logic and rough-set theory are applied to handle uncertainty and maintain consistency in the data sets. The aim of the fuzzy-rough-based method is to generate
optimum variation in the range of membership functions of linguistic variables. As
a next step, dimensionality reduction is performed to search the selected features for
discovering knowledge from the given data set. The searching of most informative
features may terminate at local optimum, whereas the global optimum may lie
elsewhere in the search space. To remove local minima, an algorithm is proposed
using fuzzy-rough-set concept and genetic algorithm. The proposed algorithm
searches the most informative attribute set by utilising the optimal range of membership values used to design the objective function. Finally, a case study is given
where the dimension reduction techniques are applied in the field of agricultural
science, a real-life application domain. Rice plants diseases infect leaves, stems,
roots and other parts, which cause degradation of production. Disease identification
and taking precaution is very important data analytic task in the field of agriculture.
Here, it is demonstrated in the case study to show how the images are collected from

J. Sil (*) A.K. Das

Department of Computer Science and Technology, Indian Institute of Engineering Science and
Technology, Shibpur, Howrah, West Bengal, India
e-mail: jayasil@hotmail.com
Springer International Publishing Switzerland 2016
Z. Mahmood (ed.), Data Science and Big Data Computing,
DOI 10.1007/978-3-319-31861-5_12

269

270

J. Sil and A.K. Das

the fields, diseased features are extracted and preprocessed and finally important
features are selected using genetic algorithm-based local searching technique and
fuzzy-rough-set theory. These features are important to develop a decision support
system to predict the diseases and accordingly devise methods to protect the most
important crops.
Keywords High-dimensional data Adaptive decision making Optimization
Local search Uncertainty Big Data Analytics

12.1

Introduction

Due to the advancement of technology and its accessibility to general public, data is
generated by leaps and bounds demanding the need for data analysis. The concept
of Big Data [1] opens up different areas of research including social networking,
bioinformatics, e-health and becomes a big challenge in devising new methods for
extracting knowledge. Big Data is characterised by massive sample size, high
dimensionality and intrinsic heterogeneity. Moreover, noise accumulation, spurious
correlation and incidental endogeneity are the common features in highdimensional data set. Due to the accumulation of noise, true signals are often
dominated and handled by sparsity assumption in high-dimensional data.
Big Data processing is difficult using standard statistical methods due to its
volume, variability, rate of growth and unstructured form [2]. In traditional statistical data analysis, we think of observations of instances of a particular phenomenon. We assumed many observations and a few selected variables are well chosen
to explain the phenomenon. Today the observations could be curves, images or
movies resulting dimension in the range of billions for only hundreds of observations considered for study. The problems include collection, storage, searching,
sharing, transfer, visualisation and analysis. A hybrid model may be appropriate to
model Big Data. Analysing the highly dynamic Big Data requires simultaneous
estimation and testing of several parameters. Errors associated with estimation of
these parameters help to design a decision rule or prediction rule depending on the
parameters. Big Data can make important contributions in real-life decision making
by developing cost-effective methods with minimum risk factor.
The basic approach of processing Big Data is to divide and analyse the subproblems, and finally results from subproblems are combined to obtain the final
solution. Dimensionality reduction plays an increasingly important role in the
analytics of Big Data and is applied before partitioning of the problem. Dimensionality reduction techniques aim at finding and exploiting low-dimensional structures in high-dimensional data. The methods overcome the curse of dimensionality
and reduce computation and storage burden of the decision-making systems.
Designing very large-scale, highly adaptive and fault-tolerant computing systems
is the solution to manage and extract information from Big Data.

12

Feature Selection for Adaptive Decision Making in Big Data Analytics

271

The objective of the chapter is to integrate different fields of machine learning

for devising novel dimensionality reduction techniques. In this chapter, two
different concepts on dimensionality reduction techniques have been proposed to
improve decision making. Finally, we discuss how dimension reduction is important for Big data analytics and present a case study in the field of agricultural
science.
In the rest of this chapter, we first discuss dimension reduction techniques in
Sect. 12.2. This section explores how the important features are selected using
genetic algorithm-based optimization technique. The vagueness and uncertainty is
also handled using fuzzy-rough-set theory in this section. A case study is provided,
based on agricultural data, in Sect. 12.3. This section gives a precise idea about how
a decision support system is designed in case of a Big data environment. The
section also includes acquisition of images, filtering and extraction of features,
selection of important diseased features, etc. The methodology explained in
Sect. 12.2 is used for important diseased feature selection. Finally, Sect. 12.4
presents our conclusions.

12.2

Dimension Reductions

Importance of features may vary from one application to another, and often it is
difficult to know exactly which features are relevant for a particular task. The
irrelevant features not only increase complexity of the systems but affect performance of the system. Dimension reduction plays a vital role in discovering knowledge from high-dimensional data set by removing redundant and irrelevant
attributes or features.
The goal of dimensionality reduction is to avoid selecting too many or too few
features than is necessary. Too few features result loss of information, while too
many irrelevant features dominate important features that overshadow the information present in the data set. The aim of dimension reduction is to select a
minimum set of relevant features by preserving essential information content of
the data set. This minimum set of features is used instead of the entire attribute set
for decision making.
Feature selection [3] is basically searching of a feature space to select the
optimal subset of features that can fully characterise the whole data set. There are
various optimised searching techniques such as ant colony-based search, Tabu
search and simulated annealing which are used frequently for feature subset
selection. Memetic algorithms [4] are quite popular for successful searching in a
large search space where traditional exhaustive search methods often fail.
Genetic algorithms [5, 6] have great potential for finding the best approximate
solutions in the evolutionary optimization community. However, genetic algorithms (GA) in some specific applications may provide local optimal solution due
to premature convergence. To overcome this problem, several hybridization of
genetic algorithms with local search methods are proposed [68]. In [9], the GA

272

J. Sil and A.K. Das

incorporates two new operations, namely, cleaning operator and proposition

improvement operator to maintain sustained diversity among the population. The
work in [10] proposes a multi-objective genetic algorithm integrating local search
for finding a set of solutions of a flow shop scheduling problem. In [11], a multiobjective time-dependent route planning problem is solved by integrating NSGA-II
[12] with local search.
In this section, two different feature selection methods are proposed. Firstly, a
novel neighbourhood-based local search strategy in genetic algorithm for obtaining
optimal set of important features is described. Kullback-Leibler (KL) divergence
method is used as fitness function with binary-coded chromosomes as population.
Steady-state selection, single-point crossover and jumping gene mutation [13]
scheme are applied stepwise to implement the proposed algorithm. Secondly,
dimension reduction is performed utilising fuzzy-rough concept and genetic algorithm. Optimised range of value associated with each linguistic label of the attributes is evaluated using rough-fuzzy concept. We build an efficient classifier using
genetic algorithm (GA) to obtain optimal subset of attributes considering optimal
membership values of the linguistic variables. The proposed algorithm reduces
dimensionality to a great extent without degrading the accuracy of classifier and
avoids being trapped at the local minima.

12.2.1 Hybrid Genetic Search Model (HGSM)

This method deals with selecting relevant features by hybridising genetic algorithm
with local neighbourhood-based searching method. Each chromosome is encoded
as a binary string where 1/0 in i-th position of the string depicts the presence/
absence of the corresponding feature. Population size is initialized randomly to
explore the search space. Genetic operators are described below.
Selection
A steady-state selection mechanism is implemented where each member of population (i.e. chromosome) is equal probable to get into mating pool.
Crossover
Every pair of parents is made sure to take part in the crossover operation to produce
offspring using single-point crossover scheme. Offspring are kept if only if they fit
better than the least good individual of the population.
Mutation
Mutation plays a vital role in preserving diversity among the population over
generations. Single-bit mutation is most popular in genetic algorithm, but it lacks
diversity in population since the most significant bit of a binary string generally
does not change. To overcome the demerit, jumping gene or mutation methodology
is used for mutating the genes. Jumping genes are a set of genes which can jump
dynamically within the chromosome. These genes have a great potential for

12

Feature Selection for Adaptive Decision Making in Big Data Analytics

273

maintaining diversity throughout the entire population, crucial for evolutionary

searching algorithm.
Let a chromosome in population be (a1, a2, . . ..., an). Jumping genes of length
q (q < <n), say (b1, b2, . . ., bq), are selected randomly and replace a portion of the
chromosome, decided randomly. Let k be the starting position, so after mutation the
muted chromosome is (a1, a2, . . ., ak-1, b1,b2, . . ., bq, ak+q, . . ., an).
New Population Generation
At each generation, new population is obtained using local search strategy. In a
particular generation, proximity of an individual is computed with respect to other
members of the population. Let D be the snapshot of current population with four
members {1 1 0 1}, {1 0 0 1}, {1 1 1 1} and {1 0 0 0}. Proximity of first member {1
1 0 1} is evaluated using Hamming distance considering the rest of the members in
the population. Integer mean of the Hamming distances is taken for selecting the
neighbouring region of the first member. In this example, the mean is 1, so
assuming the member {1 1 0 1} as schema, specified number of neighbours is
generated each with Hamming distance 1, as illustrated in Fig. 12.1. If any binary
string in the neighbours is already present in the current population, then it is
discarded and another neighbour is created in its place. Now search is performed
within the neighbours, and original member {1 1 0 1} is replaced by a neighbour of
better fitness value. Similarly, neighbours of the entire population are built and the
population is modified by the local search strategy. In this way, searching is finetuned resulting better performance in the output.
Fitness Function
Fitness function determines the quality of a chromosome in the population. So, a
strong fitness function is imperative for giving good result. The method uses
Kullback-Leibler (KL) divergence [14] for computing fitness of the chromosomes.
The KL divergence is a primary equation of information theory that quantifies the
proximity of two probability distributions. It is a measure in statistics that quantifies
in bits how close a probability distribution p {pi} is to a model (or candidate)
distribution q {qi}. The KL divergence is measured using Eq. (12.1).
 
no:of
feature
X


p
DKL pjjq
pi log2 i
qi
i1

12:1

DKL is non-negative (0) and not symmetric in p and q. Its value is zero if the
distributions match exactly and can potentially equal perpetuity. In the binary string
Fig. 12.1 Neighbourhood
generation method

274

J. Sil and A.K. Das

population, candidate feature subset is created and associated probability distribution (qi) is computed. Then the probability distribution ( pi) based on decision
attribute over all samples is calculated. We evaluate pairwise DKL value and
mean of these DKL value is used as fitness function of the genetic algorithm. The
workflow diagram of the methodology is described in Fig. 12.2.

12.2.2 Fuzzy-Rough-Set Approach

Real-world data sets are often vague and redundant, creating problem to take
decision accurately. Very recently, rough-set theory [15, 16] has been used successfully with the goal of removing vagueness and extracting knowledge after
dimension reduction of the data set. However, rough-set theory is applied on
discrete data set, but in real life most often the attributes are continuous.
Discretisation of data leads to information loss and may add inconsistency in the
data sets. Fuzzy-rough concept [17, 18] has been applied to overcome the above
limitations. After removing inconsistency, rough-set theory [19] is applied to obtain
the reduced attribute set called reduct. However, handling of non-discretised values
increases computational complexity of the system. Therefore, to build an efficient
classifier, genetic algorithm (GA) has been applied to obtain optimal subset of
attributes, sufficient to classify the objects. The proposed algorithm reduces dimensionality to a great extent without degrading the accuracy of classifier and avoids
being trapped at local minima. As a next step genetic algorithm has been applied to
obtain global optima in the search space for most informative attributes.
The following subsection describes how to generate fuzzy rule set with optimum
variation in the range of linguistic labels representing antecedent of the rules. Fuzzy
inference system is invoked to obtain membership value of linguistic variables. In
the next subsection, we present the dimension reduction algorithm for obtaining
reduced attribute set or reduct.

12.2.2.1

Data Preparation

To design the fitness function of GA, two parameters are first evaluated, i.e. the
membership values of each instances or objects in different clusters and the
membership value of the objects in different classes. The former parameters are
obtained by applying fuzzy c-means clustering algorithm [20] and the latter by
invoking a Mamdani-type fuzzy inference system [19]. Different steps of the
procedure are described below:
Step 1: Input data is fuzzified based on the range of minimum and maximum
value of each attribute, which determines the spread of membership value of
respective attribute.

12

Feature Selection for Adaptive Decision Making in Big Data Analytics

275

Start

Initialize GA parameters

Randomly generate
initial population

Evaluate fitness function

Satisfying
terminate
criteria?

Yes

No
Steady state selection of parent
chromosomes for next generation

Apply single point crossover of parent chromosomes

Apply jumping gene

mutation

New population

Deploy local search on

chromosome

Fig. 12.2 Workflow of hybrid genetic search model

Best chromosome gives

the optimal feature subset

Stop

276

J. Sil and A.K. Das

Step 2: Objects are grouped using fuzzy c-means clustering algorithm, where c is
equal to the number of class labels.
Step 3: For each attribute and its respective class label, linguistic variables are
assigned following semantic of the dataset. The value of the linguistic variable is
set based on the range of values to which an attribute value is spread for a
particular class label.
Step 4: For each linguistic variable, corresponding membership curves are
drawn.
Step 5: The rule set is designed by randomly choosing data elements from the
training set and evaluating their membership values. Using the attribute values
we judge in which linguistic label corresponding attribute value belongs, and
finally the rule is framed with the linguistic label along with the class label of
the rule.
Step 6: After generating the total rule base, a fuzzy inference system (FIS) has
been built using Mamdani model.
Finally, the FIS is used to produce the membership values of each object
belonging to different classes (second parameter) for evaluation of fuzzy-rough
dependency factor, which is treated as fitness function of GA. A data set with two
attributes and three output class label is considered in Table 12.1 to illustrate the
above procedure. The required steps are:
Step 1: The minimum and maximum ranges of the attributes for spread of
corresponding membership curve, as shown in Table 12.1, are Attr#1 215,
Attr#2 525.
Step 2: Now reconstruct the decision table by grouping the class label as shown
in Table 12.2.
Step 3: For each attribute, membership values are assigned by utilising the range
of attribute values corresponding to individual class labels:
Attr#1: LOW (28), MED (612), HIGH (1015)
Attr#2: VERY LITTLE (516), LITTLE (1020), MORE (2125)

Table 12.1 A decision table

Objects
O1
O2
O3
O4
O5
O6
O7
O8
O9
O10

Attr#1
2
7
5
6
12
8
10
15
1
13

Attr#2
10
5
15
8
16
20
25
22
17
21

Class label
1
2
1
2
2
1
3
3
1
3

12

Feature Selection for Adaptive Decision Making in Big Data Analytics

277

Table 12.2 Output of step 2

Objects
O1
O3
O6
O9
O2
O4
O5
O7
O8
O10

Attr#1
2
5
8
7
6
12
10
10
15
13

Attr#2
10
15
20
5
8
16
25
25
22
21

Class label
1
1
1
2
2
2
3
3
3
3

Fig. 12.3 Membership curve of two attributes

Step 4: The membership curves for the attributes are shown in Fig. 12.3.
Step 5: The rule set corresponding to the decision system is given below:
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.

IF Attr#1 is LOW and Attr#2 is VERY LITTLE THEN class is 1.

IF Attr#1 is LOW and Attr#2 is LITTLE THEN class is 1.
IF Attr#1 is LOW and Attr#2 is MORE THEN class is 1.
IF Attr#1 is MED and Attr#2 is LITTLE THEN class is 1.
IF Attr#1 is MED and Attr#2 is MORE THEN class is 1.
IF Attr#1 is LOW and Attr#2 is VERY LITTLE THEN class is 2.
IF Attr#1 is MED and Attr#2 is VERY LITTLE THEN class is 2.
IF Attr#1 is MED and Attr#2 is LITTLE THEN class is 2.
IF Attr#1 is HIGH and Attr#2 is VERY LITTLE THEN class is 2.
IF Attr#1 is HIGH and Attr#2 is LITTLE THEN class is 2.
IF Attr#1 is MED and Attr#2 is MORE THEN class is 3.
IF Attr#1 is HIGH and Attr#2 is MORE THEN class is 3.

278

J. Sil and A.K. Das

Finally, Mamdani model has been applied for evaluating degree of membership
value of each attribute in different classes.

12.2.2.2

DIM-RED-GA Algorithm

In the proposed method, the dependency of the decision attribute or class label on
different set of condition attribute is calculated, and attributes with highest dependency factor are selected as optimum set of reduct by applying genetic algorithm.
In the proposed DIM-RED-GA, population size is considered as the size of the
total data set, and fuzzy-rough dependency factor is designed as the fitness function.
The length of each chromosome is equal to the number of attributes, and each
component represents corresponding attribute value of different objects,
i.e. population. Two parents are chosen randomly and crossing over between
them is performed by selecting a crossover point with a probability of 0.10.
Mutation is performed with a probability of 0.02. In each generation, a combination
of different attributes is formed and from them few are selected based on the fitness
value. Termination condition is kept as combination of two conditions: (1) number
of generation is greater than a maximum value set previously or (2) no change in
fitness value.
It has been observed that for the UCI database, the proposed algorithm DIMRED-GA shows appreciable performance both in terms of dimensionality reduction
and classification accuracy.

12.3

Case Study

Rice is cropped worldwide and over half of the world population depends on rice as
the main food. To cope with the increasing demand for rice, a key element is the
development and implementation of effective rice insect management strategies.
The use of high-yielding varieties and greater nitrogen fertiliser increased rice
yields. Still disease and pests cause 25 % loss of rice in India, which need to be
controlled for more production of rice. Losses may be reduced using pesticides.
However, the use of pesticides increases the cost of production, and in addition it
reduces food quality creating adverse effect on the environment. So instead of using
pesticides, researchers try to develop sustainable farming practice that depends on
many factors including effective and timely detection of diseases and pest management to protect the crops. Timely diagnosis of crop diseases in the fields is
critical for precision on farm disease management.
Rapid development of digital devices for acquiring and storing of data attracts
researchers to apply analytical methods for automatic identification of rice diseases
using computational frameworks. However, due to the diversity of crops and their
associated diseases, application of technology in agriculture is still in research
stage. Therefore, the challenge is elaborate investigation and analysis of the big
data set for developing automated rice disease classification system. To design an

12

Feature Selection for Adaptive Decision Making in Big Data Analytics

279

efficient classification system, different types of features are acquired from the rice
plant images generating huge volume of data set. Using the proposed feature
selection methods, most important features are obtained based on which the rice
diseases are classified.

12.3.1 Rice Diseases

Three different types of attacks, namely, weeds, insect/pest and diseases, cause
substantial loss in crops. In the present study, five rice plant diseases are considered
that are grouped as follows:
Diseases caused by fungus
Diseases caused by bacteria
Diseases caused by virus
To describe the diseases, locations of infection in different parts of the rice plant
are to be detected accurately. The rice plant consists of roots, stem, leaves and
panicle, as illustrated in Fig. 12.4.

12.3.1.1

Leaf Brown Spot

Brown spot symptoms are observed at the tillering stage and beyond. Small foliar
lesions are found and the shapes of the lesions vary from circular to oval with light
brown to grey at the centre and reddish brown margin, as shown in Fig. 12.5.

12.3.1.2

Rice Blast

Rice blast is observed in both lowland and upland. Lesions on the leaves are,
generally, brown to dark brown colour and may enlarge and coalesce to kill the
Fig. 12.4 Mature rice plant

280

J. Sil and A.K. Das

Fig. 12.5 Leaf infected by

brown spot disease

Fig. 12.6 Rice leaf

infected by blast disease

Fig. 12.7 Rice stem

infected by sheath rot

entire leaves. Initially white to greyish green circular lesions or spots with dark
green borders are found on the leaves. Older lesions on the leaves are elliptical or
spindle shaped with more or less pointed ends. A leaf infected by blast is shown in
Fig. 12.6.

12.3.1.3

Sheath Rot

Sheath rot occurs on the leaf sheath that encloses the young panicles. The lesions
start as oblong or irregular spots with variation in colour from grey to light brown at
centres, surrounded by distinct dark reddish brown margins, shown in Fig. 12.7. As
the disease progresses, the lesions enlarge and coalesce and may cover most of the
leaf sheath. Lesions may also consist of diffuse reddish brown discolorations in the

12

Feature Selection for Adaptive Decision Making in Big Data Analytics

281

sheath. An abundant whitish powdery growth may be found inside the affected
sheaths, although the leaf sheath looks normal from outside.

12.3.1.4

Bacterial Blight

Bacterial blight disease appears on old plants, usually at or after maximum tillage or
at rooting stage. Water-soaked lesions usually start at leaf margins, a few cm from
the tip and spreading towards the leaf base. Length and width of the affected areas
increase, while colour changes from yellowish to light brown due to drying.
Yellowish border between dead and green areas are found on the leaf, as shown
in Fig. 12.8.

12.3.1.5

Rice Hispa

Rice hispa is a defoliator during the vegetative stage of the rice plant. Scraping of
the upper surface of the leaf blade leaving only the lower epidermis as white streaks
parallel to the midrib is observed, while tunnelling of larvae through leaf tissue
causes irregular translucent white patches that are parallel to the leaf veins, as
shown in Fig. 12.9.
Fig. 12.8 Rice leaf
infected by bacterial blight
disease

Fig. 12.9 Rice leaf

infected by pest rice hispa
disease

282

J. Sil and A.K. Das

12.3.2 Methodology
Image processing has wide range of applications in biometric identification, defect
detection, quality measurement, medical imaging and many more fields. However,
in agriculture domain, it has limited applications. One potential area of application
in agriculture is rice disease classification where image processing and computational methodologies can be applied to develop an autonomous system. Many
researchers have attempted to develop such systems with the capability to understand and reason from observations using intelligent techniques [21, 22]. Real-life
implementation has very limited scope due to variety of diseases and variation of
symptoms that increase complexity of the system. Presently, rice diseases are
detected manually depending on the visual symptoms and opinion of field experts.
Visual symptom-based common features are change of colour of the plant, shape of
the infected region, colour of different subregions with respect to the background
region, texture of infected regions and location of infection. However, evaluation of
manually identified visual symptoms are subjective and observer dependent, and
lack in timely diagnose often fail to classify diseases accurately. Different types of
features are identified and extracted from the infected regions using geometric
property of the affected region, statistical characteristics, intensity and texture
values. All such features are not significant, and the presence of redundant features
affect classification accuracy and increases complexity of the systems. Therefore,
features are selected using evolutionary algorithms to reduce complexity of the
decision-making system.

12.3.2.1

Image Filtering

The basic task of image processing is to extract pixel information from the image
and provide suitable interpretation of such information for understanding of human
observers. However, when images are captured optically or electronically, the
original image is degraded by the environment and image acquisition system. So
the very first step is to filter noise content from the acquired image with an objective
to recover the image from degraded observations. Filtering is a low-level imageprocessing task, which removes high-frequency noise component from the
degraded image. Noise removal using different types of low-pass filtering techniques is an important research area in image and signal processing. Linear filter
[23] is sufficient in removing additive noise like Gaussian noise [24] but delivers
poor performance due to the presence of impulse noise. Non-linear filter like
median filter [25] is used to de-noise the salt and pepper noise from the degraded
images. In linear filter, the value of the output pixel is obtained using the weighted
sum of the input pixels, while non-linear filtering method assigns a value to the
output pixel, based on the neighbouring pixel information of the input pixel.
Mean filter, also called average filter, is one kind of low-pass linear filter that
produces output as average intensity of the neighbouring pixels. The average filter

12

Feature Selection for Adaptive Decision Making in Big Data Analytics

283

smoothes local variation in an image, and thus noise is reduced due to blurring
effect in the image. In average filtering, if the template size is larger, the blurring
effect is more in the output image. Median filtering is a non-linear low-pass filter
that is adequate for removing the outliers present in an image. It uses
neighbourhood operations to process an image around the input pixel. The pixel
values in a neighbourhood region are ordered, and the median value is used to
substitute the pixel at the same location of the output image.

12.3.2.2

Feature Extraction

Feature extraction is an important step towards disease identification on which the

success of the system depends. A visual feature is defined to capture certain visual
properties of an image, either globally for the entire image or locally for a small
group of pixels. Most of the common features observed by the farmers in the
diseased plants are the visual symptoms in the spatial domain such as change of
colour, shape of the infected portion, location of infection and texture. Spots created
by the diseases have colour difference between its boundary region, centre or core
region and uninfected region. Change of colour of the spot at boundary and core
regions with respect to the uninfected portion have also been used as colour features
to address the age and nutrient level of the plant. Another set of colour features
based on distribution of hue are obtained since hue represents the true colour
recognised by the human visual system. Distribution of hue value from the centre
towards the boundary of the spot along the radius in eight different directions has
been considered as colour features to distinguish the diseases.
Different types of features are extracted from the infected plant images to
classify the rice diseases. However, it has been observed that classification accuracy
to detect diseases is not proportional to the number of features, rather less number
of features but more significant ones. Therefore, feature selection is an important
step for accurately classifying rice diseases based on the extracted features. The
irrelevant and redundant features should be removed prior to building the classifier
in order to achieve better performance and reducing complexity of the systems.
Most common features used for building image-based recognition/classification
system are colour features, shape-based features and texture features.

12.3.2.3

Colour Features

Colour is a major distinctive attribute used to distinguish diseases. Colour features

are geometric transformation invariance and generally global so is considered to
extract colour features. Consumers initially accept or reject a food based on its
colour and other visual attributes, which is simulated using machine vision method.
To identify different types of diseased plant images in remote-sensing techniques,
different types of vegetative indices are used to represent quantitative measurements indicating the vigour of vegetation. Bannari et al. [26] suggest that over forty

284

J. Sil and A.K. Das

vegetation indices have been developed during the last two decades and used for
classification of different diseases. Vegetative index based on the visual images are
also proposed in literature where normalised difference greenness index (NDGI)
and redness index (RI) are defined as
gr
gr
rg
RI
rg

NDGI

where r and g represent the red and green value of a pixel, used to identify stress of
the images.
Colour is used not only for disease identification but also as a major feature for
quality, composition and standards of identity of food. Colour features are widely
used for grain quality measurement. Wang et al. [27] developed a real-time grain
inspection instrument to classify durum wheat kernels based on their virtuousness.
To develop the system, acquired images are normalised into 70  200 size and
converted to HSI colour plane. Here it has been assumed that an artificial neural
network (ANN) would prefer the HSI model over the RGB model because ANNs
are designed based on an analogy to the human neural system. Using various
artificial neural network models, they got accuracy in the range of 5890 %.

12.3.2.4

Shape-Based Features

Shape of the spots created by infection plays an important role for disease identification. Some simple geometric features are used to describe the shapes of the
infected regions. Usually, the simple geometric features can only discriminate the
shapes with large differences, therefore usually used as filters to eliminate false hits
or combined with other shape descriptors to discriminate shapes. A shape can be
described by different aspects. Shape parameters available in literature are centre of
gravity, axis of least inertia, digital bending energy, eccentricity, circularity ratio,
elliptic variance, rectangularity, convexity, solidity, Euler number, profiles and
hole-area ratio. Different shape parameters are described below:
The position of centre of gravity should be fixed in relation to the shape [28] and
so used to recognise the shape.
The axis of least inertia is unique to the shape and serves as a unique reference
line to preserve the orientation of the shape. The axis of least inertia (ALI) of a
shape is defined as the line for which the integral of the square of the distances to
points on the shape boundary is a minimum [29].
Average bending energy is defined by the average of squared curvature of the
shape and is a useful property of shape [30].
Eccentricity is the measure of aspect ratio, defined as the ratio of the length of
major axis to the length of minor axis [28].

12

Feature Selection for Adaptive Decision Making in Big Data Analytics

285

Minimum bounding rectangle is also called minimum bounding box, the

smallest rectangle that contains every point in the shape.
Solidity describes the extent to which the shape is convex or concave [31] and
defined by the ratio of area of the shape region where H is the convex hull area of
the shape.
Euler number describes the relation between the number of contiguous parts and
the number of holes on a shape [32].
Circularity ratio represents how a shape is similar to a circle [33]. Three different
definitions are available in literature to measure the circularity. Circularity ratio
is (i) ratio of the area of a shape to the area of a circle having the same
perimeters, (ii) ratio of the area of a shape to the shapes perimeter square and
(iii) circle variance, ratio of standard deviation and mean of the radial distance
from the centroid of the shape to the boundary points.
Ellipse variance is a mapping error of a shape to fit an ellipse that has an equal
covariance matrix as the shape [28].
Rectangularity represents how rectangular a shape is, i.e. how much it fills its
minimum bounding rectangle and defined as a ratio of area of the shape and the
area of the minimum bounding rectangle [28].
Convexity is defined as the ratio of perimeters of the convex hull over that of the
original contour [34].
Hole-area ratio (HAR) is defined as the ratio of area of a shape and the total area
of all holes present in the shape. Hole-area is the most effective in discriminating
between symbols that have big holes and symbols with small holes [32].
Shape-based features for rice disease classification include area, perimeter
length and minimum enclosing rectangle of infected spot as reported in
[35]. The shape features used for grain classification and grading system include
perimeter, area and shape descriptors (kernel circularity and shape compactness), selected to classify group of kernels which are placed in random direction.
Area and perimeter of the spot are calculated using chain code.

12.3.2.5

Texture Features

Texture is one of the most important characteristics of images in identifying defects

or flaws that provides important information for recognition and interpretation of
images. The term texture may be used to characterise the surface of a given object
or region. Success of texture classification depends on the appropriate method of
texture analysis for feature extraction, and selection of such method is a critical one.
Texture features based on the extraction methods are broadly divided into four
categories:

Statistical features
Model-based features
Structural features
Filter-based features

286

J. Sil and A.K. Das

Statistical Features
Statistical features are determined by applying statistical procedures. Most common
statistical texture features are based on the co-occurrence matrix. A variety of
concurrence matrix is found in [36, 37]. Most common is the grey label occurrence
matrix (GLCM) [36]. It is a two-dimensional histograms of the occurrence of pairs
of grey levels for a given displacement vector. Based on which 14 features are
calculated such as angular second moment, contrast, correlation, variance, inverse
difference moment, sum average, sum variance, sum entropy, entropy, difference
variance, difference entropy, information measures of co-relation and maximal
correlation coefficient using grey-level co-occurrence matrices.
Model-Based Features
A number of random field models (i.e. models of two-dimensional random processes) have been used for modelling and synthesis of texture. If a model is shown
to be capable of representing and synthesising a range of textures, then its parameters may provide a suitable feature set for classification and/or segmentation of the
textures. For a model-based approach to be successful, there must exist a reasonably
efficient and appropriate parameter estimation scheme, and the model itself should
be parsimonious, i.e. use the minimum number of parameters. Popular random field
models used for texture analysis include fractals, autoregressive models, fractional
differencing models and Markov random fields.
Autoregressive models have been used to model images as random fields
(two-dimensional random processes) by a number of researchers [3840]. In the
two-dimensional spatial case, the previous values of the time series process are
replaced by the grey values of local neighbourhood pixels. Unlike the temporal
case, there is normally no preferred direction in a lattice, and neighbourhoods are
therefore normally defined to consist noncausal (two-sided) variables. Again the
parameters may be estimated either by using least square error or maximum
likelihood techniques.
Markov random fields (MRFs) have been popular for modelling images. They
are able to capture the local (spatial) contextual information in an image. These
models assume that the intensity at each pixel in the image depends on the
intensities of only the neighbouring pixels. MRF models have been applied to
various image-processing applications such as texture synthesis, texture classification, image segmentation, image restoration and image compression.
Structural Features
Structural approaches generally model a texture as the deterministic or stochastic
placement of texture primitives called textons with the emphasis on texton characterisation such as size and shape or both local properties. Structural methods are
based on regular or semi-regular placements of textural primitives. In the case of
observable or visual textures, it is usually quite difficult to extract the primitives and
their placements. Therefore, these approaches are more appropriate only for highly
regular deterministic textures.

12

Feature Selection for Adaptive Decision Making in Big Data Analytics

287

Filtered-Based Features
Filter-based texture features are gaining popularity for different classification
purposes due to its ability to represent texture precisely and overcome the drawbacks of statistical and structural texture features. Filter-based methods can be
grouped into three categories: spatial domain filtering, frequency domain filtering
and spatial-frequency domain filtering.
Texture-based approach can fail where texton primitives are not readily identifiable. Statistical approaches focus on the global spatial relationships between
intensity variations and often fail to capture local properties of the texture. Texture
characterisation requires both local texton primitive and global spatial organisation
description. Neither structural nor statistical methods satisfy this requirement fully.
Statistical methods have been widely accepted over the past two decades.

12.3.3 Position Detection

Another important feature to classify diseases is the position of infection (PSI) with
respect to the top boundary of the leaf. A tree structure method has been described
here to identify the PSI [41]. The background of a segmented image is represented
by grey value, and noninfected region is marked by black, while the infected region
is represented by white as shown in Fig. 12.10. A rectangle is drawn across the leaf
so that it covers the infected region and boundary of the leaf, called image block, as
shown in Fig. 12.10. Image blocks of different infected leaves are standardised to
600  900 image size. A vertical line from the middle of the length and a horizontal
line from the middle of the width of the rectangle are drawn, which split the image
block into four equal subblocks. Thus, the main image block, say B, is divided into
four subblocks, named as B00, B01, B10 and B11, and a tree is constructed as follows:
A vertex is considered for each block.
An edge is drawn from parent block B to subblock Bij with edge label {ij}, where
ij is marked as combination of 0 and 1.

Fig. 12.10 Splitting of regions

288

J. Sil and A.K. Das

The subblocks are either non-uniform or uniform. The non-uniform block

contains the mixture of black and white pixels, while uniform block contains either
black or white pixels.
Next, all subblocks are checked for uniformity. The subblocks of a parent block
are considered as leaf nodes of the tree if at most one subblock is the non-uniform
block. Otherwise, each non-uniform subblock of the parent block is divided in the
same way, and the tree is expanded until all leaf nodes are generated. Figure 12.11
shows the tree structure of the block, corresponding to the image shown in
Fig. 12.10. The block B representing the whole image is the root (R) of the tree,
while the non-uniform subblocks are the leaf nodes represented by grey colour. The
leaf nodes with grey colour carry boundary information of the infected region,
while white-coloured leaf nodes, i.e. uniform subblocks, have no importance. The
intermediate nodes including the root node are used only to measure the distance of
the infected region from the top boundary of the leaf. To measure the distance, a
binary string is associated to each node of the graph according to Eq. (12.2).
BSR f0g
BSC BSpfijg; if P is parent node other than root node of child node C
fijg
; other wise
12:2
where {ij} represent the edge label from P to C.
The binary string associated with the non-uniform leaf node (NL) ie BS (NL) is
converted to its decimal equivalent and provides distance in terms of blocks from
the top of the image. The width of the subblock corresponding to non-uniform node
is considered as unit block distance.
Let, for a node T, binary string contain n-components, i.e. BS (T ) {i1j1}{i2j2}
{i3j3}. . ...{injn}. The binary code BC(T ) is given as {i1}{i2}{i3}. . ..{in}

Fig. 12.11 Tree structure of the image in Fig. 12.10

12

Feature Selection for Adaptive Decision Making in Big Data Analytics

289

{i1i2i3. . .in}, whose decimal equivalent is DC(T ). The left most bit of each component of the string is considered for binary code generation because most significant bit carries maximum weight.
For example, in Fig. 12.11, binary string for non-uniform leaf node T is {01}
{11}{11}{01}, and its corresponding binary code is {0}{1}{1}{0} {0110}. It is
6 in decimal code, which implies that subblock T is six block distance apart from
the top of the main block (i.e. the image).
In reality, infected region created by different diseases vary in size, and as a
result a global block with minimum width is used for distance measurement. The
width W(T ) of a subblock associated with non-uniform leaf T is given by Eq. (12.3).
W T

w
2l

12:3

where w is equal to 900 representing standard width of the image block and l is the
label of the node T in the quad tree.
The procedure is applied on large number of infected leaves; corresponding trees
and width of non-uniform leaf nodes are calculated using Eq. (12.3). Therefore, all
the non-uniform nodes in the same label of the tree have associated blocks with
same width. Similarly, a node at highest label of the tree has associated block with
minimum width. Say, lmax is the highest label among all trees, so the minimum
0
width of the block associated to non-uniform node T is obtained using Eq. (12.4).
 0
wmin T

w
2lmax

12:4

Thus, the actual block distance D(T ) of the subblock T from the top of the image
block is given by Eq. (12.5) considering minimum width subblock as the unit block
distance.
DT DCT *

W T

0
W min T

2lmax
2l
lmax  l
DCT  2

DCT 

12:5

So, if a tree contains b non-uniform leaf nodes {T1, T2, . . .,Tb}, then the position of
the spot in the image from the top is given by Eq. (12.6).
pos

b


1X
D Ti
b i1

where D(Ti) represents block distance for ith non-uniform leaf node.

12:6

290

J. Sil and A.K. Das

For rice disease identification, a wide variation of features are obtained and
values are acquired over a period of time. The data set represent characteristics of
the images when attacked by different types of diseases. However, the presence of
large number of feature may increase redundancy, and dominance of less important
features reduces classification accuracy. Therefore, feature selection is an important
step for accurately classifying rice diseases; here we have considered five different
diseases. The irrelevant and redundant features are removed using feature selection
methods prior to classify the diseases.

12.4

Conclusions

Systematic and unbiased approach to classification is of great importance, and

therefore, automated discovery of this small and good feature subset is highly
desirable. In this chapter, a novel local search based genetic algorithm approach
has been proposed to select informative feature subset, which classify the dataset
effectively and efficiently. In the second feature selection method, vagueness and
continuous domain dataset is managed using fuzzy-rough approach, and optimum
features are selected using conventional genetic algorithm.
Finally a real-life case study in agricultural domain is illustrated with sufficient
features and large data set. The features are extracted from the objects which are
robust, accurate and independent of transformation variance. Colour-based features
are robust, transform invariant and represent most important visual changes
recognised by human eye. Shape features are used for disease classification, since
depending on the type and severity of infection shape of the infected region varies
considerably. The presence of strips on the plant parts provides texture information,
which are affected by diseases, and so different statistical texture features are also
used for classifications. For disease detection, variability of features plays vital role,
and at the same time complexity of the decision-making system is important for
real-life problem. Genetic algorithm-based optimal feature selection and cleaning
the data set using fuzzy-rough approaches are applied for minimising
misclassification rate.
Feature dependency and context-based feature selection methods can be taken as
future work for dimensionality reduction. More number of classes or diseases can
be considered with an aim to enhance its performance and ability with respect to all
categories of crops at an early stage. Time-series data can be analysed to understand
progression of the disease, and preventive measures could be taken providing
appropriate suggestions to control them.

12

Feature Selection for Adaptive Decision Making in Big Data Analytics

291

References
1. Hu H, Wen Y, Chua TS, Li X (2014) Toward scalable systems for big data analytics: a
technology tutorial. IEEE Access 2:652686
2. Jagadish HV, Gehrke J, Labrinidis A, Papakonstantinou Y, Patel JM, Ramakrishnan R,
Shahabi C (2014) Big data and its technical challenges. Commun ACM 57:8694
3. Blum A, Langley P (1997) Selection of relevant features and examples in machine learning.
Artif Intell 97(1.2):245271
4. Knowles JD, David WC (2000) M-PAES: a memetic algorithm for multiobjective optimization. In: Proceedings of the 2000 congress on evolutionary computation, 2000, vol. 1. IEEE
5. Goldberg DE, Holland JH (1988) Genetic algorithms and machine learning. Mach Learn
3.2:9599
6. Deb K et al (2002) A fast and elitist multiobjective genetic algorithm: NSGA-II. IEEE Trans
Evol Comput 6.2:182197
7. Coello CA et al (2007) Evolutionary algorithms for solving multi-objective problems.
Springer, New York
8. Kim KW, Yun YS, Yoon JM, Gen M, Yamazaki G (2005) Hybrid genetic algorithm with
adaptive abilities for resource constrained multiple project scheduling. Comput Ind 56
(2):143160
9. Diaz CA, Muro AG, Perez RB, Morales EV (2014) A hybrid model of genetic algorithm with
local search to discover linguistic data summaries from creep data. In: Proceedings of the
expert system with appllications, 2014, pp 20352042
10. Ishibuchi H, Murata T (1998) A multi-objective genetic local search algorithm and its
application to flowshop scheduling. IEEE Trans Syst Man Cybern 28(3):392403
11. Sharma S, Mathew TV (2011) Multiobjective network design for emission and travel-time
trade-off for a sustainable large urban transportation network. Environ Plan B: Plan Des
38.3:520538
12. Kalyanmoy D et al (2002) A fast and elitist multiobjective genetic algorithm: NSGA-II. IEEE
Trans Evol Comput 6.2:182197
13. Pati SK et al (2013) Gene selection using multiobjective genetic algorithm integrating cellular
automata and rough set theory, swarm, evolutionary, and memetic computing. Springer,
Cham, pp 144155
14. Kullback S, Leibler RA (1951) On information and sufficiency. Ann Math Stat 22(1):7986
15. Pawlak Z (1998) Rough set theory and its applications to data analysis. Cybern Syst
29:661688
16. Pawlak Z (1991) Rough sets theoretical aspects of reasoning about data. Kluwer Academic
Publishers, Boston/London/Dordrecht, p 229
17. Jensen R, Shen Q (2002) Fuzzy-rough sets for descriptive dimensionality reduction. In:
Procceding of the 11th international conference on fuzzy systems, pp 2934
18. Moumita S, Sil J (2011) Dimensionality reduction using genetic algorithm and fuzzy-rough
concepts. In: 2011 world congress on information and communication technologies, IEEE, pp
379384
19. Qinghua H, Daren Y, Zongxia X (2006) Information-preserving hybrid data reduction based
on fuzzy-rough techniques. Pattern Recogn Lett 27(5):414423
20. Jain A, Murty M, Flynn P (1999) Data clustering: a review. ACM Comput Surv 31(3):264323
21. Zhigang L, Zetian F, Yan S, Tiehua X (2003) Prototype system of automatic identification of
cotton insect pest and intelligent decision based on machine vision. American Society of
Agricultural and Biological Engineers
22. Qin Z, Zhang M, Christensen T, Li W, Tang H (2003) Remote sensing analysis of rice disease
stresses for farm pest management using wide band airborne data. IEEE 4:22152217
23. Gonzalez RC, Woods RE (2007) Digital image processing. Pearson Education, New Delhi
24. Pratt WK (2010) Digital image processing. Wiley, New York

292

J. Sil and A.K. Das

25. Chan RH, Ho CW, Nikolova M (2005) Salt-and-pepper noise removal by median-type noise
detectors and detail-preserving regularization. IEEE Trans Image Process 14(10):14791485
26. Bannari A, Morin D, Bonn F, Huete AR (1995) A review of vegetation indices. Remote Sens
Rev 13(1):95120
27. Wang N, Dowell FE, Zhang N (2003) Determining wheat vitreousness using image processing
and a neural network. Trans Am Soc Agric Eng 46(4):11431150
28. Mingqiang Y, Kidiyo K, Joseph R (2008) A survey of shape feature extraction techniques. In:
Pattern recognition techniques, technology and applications. I-Tech, Vienna
29. Horn B (1986) Robot vision. MIT Press, Cambridge
30. Loncaric S (1998) A survey of shape analysis techniques. Pattern Recogn 31(8):9831001
31. Cheng C, Liu W, Zhang H (2001) Image retrieval based on region shape similarity. In:
Proceedings of the 13th SPIE symposium on electronic imaging, storage and retrieval for
image and video databases
32. Soffer A (1997) Negative shape features for image databases consisting of geographic symbols. In: Proceedings of the 3rd international workshop on visual form
33. Zhang D, Lu G (2002) A comparative study of fourier descriptors for shape representation and
retrieval. In: Proceedings of the 5th asian conference on computer vision
34. Mukundan R (2004) A new class of rotational invariants using discrete orthogonal moments.
Sixth IASTED international conference on signal and image processing, pp 8084
35. Yao Q, Guan Z, Zhou Y, Tang J, Hu Y, Yang B (2009) Application of support vector machine
for detecting rice diseases using shape and colour texture features. IEEE International Conference on Engineering Computation, pp 7983
36. Robert M, Haralick KS, Itshak D (1973) Texture features for image classification. IEEE Trans
Syst Man Cybern SMC-3(6):610621
37. Weszka JS, Dyer CR, Rosenfeld A (1976) A comparative study of texture measures for terrain
classification. IEEE SMC-6:269285
38. Kashyap RL, Chellappa R (1983) Estimation and choice of neighbors in spatial interaction
models of images. IT V29:6072
39. Kartikeyan B, Sarkar A (1991) An identification approach for 2-D autoregressive models in
describing textures. CVGIP Graph Model Image Process 53:121131
40. Mao J, Jain AK (1992) Texture classification and segmentation using multi resolution simultaneous autoregressive models. Pattern Recogn 25(2):173188
41. Phadikar S, Sil J, Das AK (2012) Feature selection and rule generation to classify rice diseases
by extracting features using empirical methods. J Comput Electron Agric 75:304312

Chapter 13

Social Impact and Social Media Analysis

Relating to Big Data
Nirmala Dorasamy and Natasa Pomazalova

Abstract Social media is a component of a larger dynamic and complex media and
information domain. As the connection with Big Data grows, its impact in the social
media domain cannot be avoided. It is vital that while the positive impact needs to
be recognized, the negative impact emerging from Big Data analysis as a social
computational tool needs to be recognized and responded to by various agencies.
There have been major investments in the development of more powerful digital
infrastructure and tools to tackle new and more complex and interdisciplinary
research challenges. While there is a need to size the opportunities offered by
continuing advances in computational techniques for analyzing social media, the
effective use of human expertise cannot be ignored. Using the right data, in the right
way and for the right reasons, can change lives for the better, especially if Big Data
is used discriminately and transparently. This chapter analyzes the impact of Big
Data from social media platforms in the social, political, and economic spheres.
Further, the discriminate use of Big Data analysis from social media platforms is
explored, within the context of ethical conduct by potential users and proposes
important imperatives to minimize, if not control, the negative impact of Big Data
analysis from a social perspective.
Keywords Social media Digital technology Social networking sites Database
tools Big Data Big Data analysis Social media services

N. Dorasamy (*)
Department of Public management and Economics, Durban University of Technology,
Durban, South Africa
e-mail: nirmala@dut.ac.za
N. Pomazalova
Department of Regional Development, Mendel University, Brno, Czech Republic
Durban University of Technology, Durban, South Africa
e-mail: natta@atlas.cz
Springer International Publishing Switzerland 2016
Z. Mahmood (ed.), Data Science and Big Data Computing,
DOI 10.1007/978-3-319-31861-5_13

293

N. Dorasamy and N. Pomazalova

294

13.1

Introduction

Data, or segments of information, have been collected and used throughout history.
However, the potential to collect, store, and analyze data has significantly increased
with the advancement in digital technology. The emergence of Big Data has
escalated its usefulness for decision-making at various levels of analysis, including
individual, group, organizational, and national systems. Organizations have moved
from using data stored in relational databases to using data from data mining in
general ledger packages, weblogs, social media, e-mail, sensors, photographs,
corporate enterprise resource planning (ERP) systems, custom relationship management (CRM) programs, and social networks. While the growth of Big Data has
accelerated in the last few years, the ability to find useful information within the Big
Data is of crucial importance and requires careful consideration. Managing such
data has to be underpinned by quality, protecting privacy and ethical use.
The rapid growth of the web as a publishing tool and the recent explosion of
social media and social networking sites have generated opportunities and challenges to social researchers. Currently, there are many types of social media
services (SMS). The Personal SMS like Facebook allows users to create online
profiles and connect with other users, focusing on social relationships and information sharing such as ones gender, age, interests, and job profile. The Status SMS
like Twitter allows users to post short status updates to broadcast information
quickly and publicly with other users. The Location SMS like Foursquare and
Google Latitude, using GPS-based networks, broadcasts ones real-time location.
The Content-sharing SMS like YouTube and Flickr is designed as platforms for
sharing content, such as music, photographs, and videos [4]. The Shared-interest
SMS like LinkedIn is more a network for a subset of professional users to share
information interests like politics and education. These social media services
provide datasets that have expanded in size and complexity to the extent that
computer-based methods are now required to analyze mass volumes of information.
Data, with datasets whose size can range from a few dozen terabytes (TB) to
multiple petabytes (PB), is beyond the ability of typical database software tools
to capture, process, and store, manage, and analyze. Big Data technologies are
required to economically extract value and meaning from very large volumes of a
wide variety of data, by enabling high-velocity capture, discovery, and
analysis [21].
Today, social media is a key source for Big Data analysis from platforms such as
Twitter, Facebook, and Flickr, especially for businesses that depend on data-driven
intelligence [3, 12]. IT companies such as Google, Amazon, Facebook, and IBM are
fiercely competing in the Big Data analytics market. Social media systems provide
valuable information in terms of its detail, personal nature, and accuracy. Since the
data is not totally private, it is exposed to scrutiny within a users network, which
can increase the chances of accuracy when compared to data from other sources.
Big Data integration and predictive analytics can help overcome the challenges of
managing in an environment with increasing rates of change and innovation
[12]. Studies have shown that Big Data analytics has resulted in improvements in

13

Social Impact and Social Media Analysis Relating to Big Data

295

retail operating margins, reduction in national healthcare expenditures, and savings

in operational efficiencies. It has great potential, in that it can generate significant
value across sectors, such as healthcare, retail, manufacturing, and the public sector
[12]. However, seeing that data is dynamic, there has to be continuous integration of
existing data with living data, if companies want to reap the optimal benefits of
data analytics.
Big Data volume is expanding due to the increase of social media, online data,
and location data, often resulting from the accelerated usage of sensor-enabled
devices. This has resulted in mobile cloud computing being made possible by
focusing on Internet technologies that are built on web-based standards and protocols. The key drivers for cloud computing are bandwidth increase in networks,
cost reduction in storage systems, and advances in database [12]. It has created a
form of virtualization over the Internet, which involves data outsourcing that can be
provisioned with minimal management effort or service provider interaction, with
no up-front cost and provides just-in-time services. It is a model for enabling
convenient and on-demand network access to a shared pool of computing resources
and eliminates the cost for in house infrastructure [5]. An area of distinctive growth
in the IT ecosystem is social media with smart devices, such as smartphones and
tablets, providing a new communication platform with real-time access for customers. Various agencies have recognized such data as an important constituent of
innovation and have developed means to use Big Data to develop solutions to
business, technical, and social problems in innovative and collaborative ways.
While the use of Big Data is invaluable, businesses are challenged by Big Data
because it grows so large that they become awkward to work with using on-hand
database management tools [12].
While the value of Big Data is clear for tackling complex technical and business
problems, the question is on how well Big Data can solve complex social problems?
While, business and science have shown the value of Big Data, the social sector
needs to show how they can adopt this type of decision-making potential into their
operations. The issues that are being addressed in the social sector are more
complex than they are in business or science, making the use of Big Data more
challenging. In addition, greater focus must be given to the rights, privacy, and
dignity of different stakeholders. The large-scale collection, aggregation, analysis,
and disclosure of detailed and triangulated information offer the possibility of
powerful computational social science tools, but carries with it the potential for
abuse by various entities, especially if datasets are not reliable and representative.
In spite of these obstacles, progress continues.

13.2

Using Big Data and Social Media Innovatively

Big Data can encompass information such as transactions, social media, enterprise
content, sensors, and mobile devices. Since Big Data refers to datasets that extend
beyond single data repositories (databases or data warehouses), they are too large

296

N. Dorasamy and N. Pomazalova

and complex to be processed by traditional database management and processing

tools. New Big Data technologies have helped to capture analyze and store data to
solve problems. Organizations are employing computing power through hardware
and software advances to manipulate huge amounts of data. Some new approaches
include Hadoop, a software that takes a different approach to data management, and
HANA, a hardware approach that handles data manipulation in raw memory to
make real-time analytics of the Big Data a reality [7]. Such advances have enabled
the cost reduction and volume capacity increase of digital storage mediums and
unique ways of manipulating data from social media.
Big Data offers much potential for innovative use, thereby creating value.
Fanning and Grant [7] identified the following ways in which Big Data is valuable:
Big Data can reveal significant value by making information transparent and
usable at a much higher frequency. More people looking at the data will bring
different perspectives.
As the data proliferates, there is more accurate and detailed performance information that may expose variability and issues that need attention.
Controlled experiments using Big Data analysis can assist to make better
management decisions regarding tailored products and services.
Big Data can be used to improve the development of the next generation of
products and services through proactive maintenance and preventive measures
to minimize failures.

13.2.1 Social Media as Advertising and Marketing Medium

Social media platform providers provide numerous metrics for data analysis to
profile users for advertisers by showing the right advertisements to the right people,
using masses of hidden information to model which users are likely to respond to a
particular advertisement. Facebook can optimize advertising revenue by targeting
advertisements to achieve the greatest possible number of clicks. If used scrupulously, valuable information can reach the right users. An additional example is
Amazon, which allows users to exercise some control over their data by allowing
them to flag purchases not to be used to form recommendations [1]. This is useful
both in cases of gift purchases and in cases where the product is of a sensitive
nature. Further, by allowing users to instruct a website not to use a specific piece of
personal data for a particular purpose represents a significant improvement on the
current personal data free-for-all model used by both social networking companies
and their corporate customers [17]. Such platforms can also play a beneficial role in
exposing, predicting, and helping to eliminate disruptive behavior. Computational
social science can serve a positive role in promoting the interests of the community
in a social media platform, provided the ethical considerations are evaluated for
various purposes of use.

13

Social Impact and Social Media Analysis Relating to Big Data

297

The impact of Big Data can also be seen in marketing. There is a noticeable
move away from large-scale mailings of catalogues and offers to various individuals, based primarily on purchased mailing lists or phone directories. This is due to
businesses mining Big Data to target individuals based on knowing the preferences
of individuals in the population. For example, Amazon can almost instantaneously
offer additional purchase opportunities to individuals based on what others have
also purchased given a similar purchase showing up in the cart. It can also target
those who in the past purchased a product or those who are searching a particular
topic [2]. Further, Google offers marketers the opportunity to provide relevant
advertisements to individuals based on their search habits.
It is through the improvement in the tools to analyze and collect Big Data that
larger markets can be targeted. According to Fanning and Grant [7], digital marketing uses a combination of push-and-pull Internet technologies to execute marketing campaigns. The use of software vendors such as Adobe allows customers to
make each digital transaction layered, thus allowing the organization to see in real
time how a particular advertising campaign is performing, in terms of what is being
viewed, how often, how long, as well as other actions such as responses rates and
purchases made [8]. This information provides key information for marketers to
make real-time decisions. In addition, the proliferation of mobile phones, tablets,
and other means of accessing the Internet has facilitated multichannel marketing by
companies like Usablenet.
Applications such as Facebook, LinkedIn, and Google are driven by information
sharing that can be used by business clients, government, other users within the
social media platform, and the platform provider itself [7]. For example, business
clients draw on this computational social science to target users based on constructs
that range from age, gender, and geographical location, sexual preferences, education level, and employer.

13.2.2 Adding Value for Social Well-Being

The rapid use of mobile phones and Internet usage, especially in developing
economies, gives people the opportunity to improve the quality of their lives. A
mobile phone acts as an individual sensor, collecting pertinent information from its
environment, which when aggregated and analyzed with information from thousands of other mobile phones can provide vital information, which can then be
disseminated back to people on the ground via the same mobile phones. For
example, Cell Life, a South African organization, created a mass messaging mobile
service called Communicate, which reminds patients to take their medications,
links patients to clinics, and offers peer-to-peer support services such as counseling
and monitoring [6].
In addition, most modern mobile phones contain global positioning systems
technology, which identifies the geographic location of the phone, and other
information relating to social media postings. This is important, for example,

298

N. Dorasamy and N. Pomazalova

when researching migration patterns to understand the spread of infectious diseases

like Ebola and to help stop the disease from spreading. Information on the patterns
of human travel collected from mobile phone usage can be used to develop
predictive models to combat diseases in specific regions [13].
A further consideration is census data collected in different countries, which is
an important source of information for governments. In the 1800s, information from
the national census was logged by hand, microfilmed, and sent to be stored in state
archives, libraries, and universities. It took many years to properly tabulate census
data after the initial collection. In recent years, countries have streamlined their data
collection methods by adopting emerging technologies like geographic information
systems, social media, videos, intelligent character recognition systems, and sophisticated data-processing software and processing tools to survey the populace [6].
Thorough data management during crisis communication can provide major
benefits. The study by Proctera et al. [18] on the 2011 riots in England found that
mainstream media lagged behind crowdsourced (citizen journalism) reports
appearing in social media. Further, it was evidenced that collaborative efforts by
large numbers of producers of data can provide competing and, at times, better
coverage of events than mainstream media. While evidence cannot always be taken
at face value, social media provides a platform for robust mechanisms for authentication, so that false rumors are identified more quickly. This is supported by
Mendoza, Poblete, and Castillo [16] who noted that users deal with true and
false rumors differently: the former are affirmed more than 90% of the time,
whereas the latter are challenged (i.e., questioned or denied) 50% of the time.
There are several examples of social media being a valuable tool for information
gathering, for keeping the public informed, and for providing advice during crisis
situations. One such case was the August 2011 riots in England, which began as an
isolated incident in Tottenham. A study by Proctera, Visb, and Vossc [18] revealed
that Twitter was used overwhelmingly for more positive means, especially for the
organization of the riot cleanup. Even the police supported keeping social media
sites open during the crisis. However, the study confirmed the conclusions of other
studies that the police and government agencies in general still need to use social
media platforms like Twitter effectively [18].

13.2.3 Gauging Business Performance

The business community has also been a heavy user of Big Data. For example,
Netflix collects billions of hours of user data to analyze the titles, genres, time spent
viewing, and video color schemes to gauge customer preferences and to give the
customer the best possible experience. Following in the path of e-commerce, the
rise of social media, Big Data, and cloud computing has impacted businesses in the
following ways [11]:

13

Social Impact and Social Media Analysis Relating to Big Data

299

The ability to identify at the earliest opportunity those who are in danger of
leaving organizations and to action the retention of best talent
The ability to identify the high performers based on new live data highlighting
performance and ratings and profitability rather than old assumptive data such
as university attended and the grade of their degree
The ability to measure the real drivers of performance within the business,
thereby identifying hidden gems that make a real difference
The ability to fine-tune businesses based on fact and evidence rather than
fiction and emotion
The ability to understand that datasets can be combined to form a more intricate
and accurate picture from several data sources

13.2.4 Adding Value in Financial Markets

Financial markets have also benefitted from Big Data. Social media has played a
significant role in growing financial markets with respect to the following [11]:
Significant means by which crowdfunding helps the underserved small- and
mid-sized entities (SMEs) and start-ups to access capital in a cost-effective
manner.
New technologies fund opportunities for growing segments of the economy not
reached by traditional outlets and also create new jobs in a dynamic, technologybased business model.
Algorithms and soft-/hardware technology related to high-frequency trading
has exploded over the last 20 years, and the primary beneficiary has been the
market for existing shares and other financial instruments (secondary market).
With new cost-effective mechanisms to raise funds, underserved primary markets can now enjoy lower funding costs.
Provides investors with protection from fraudsters through both proactive education and appropriate regulation by governments through the creation of a new
and dynamic mechanism for allocating capital from traditional, institutional
stakeholders (such as banks) to individual-driven operations using current and
future technologies to touch millions of individuals looking for investment
opportunities.

13.3

Discriminate Use of Social Media Analysis

There are multiple dimensions to Big Data, which include volume (considers the
amount of data generated and collected), velocity ( refers to the speed at which data
are analyzed), variety (indicates the diversity of the types of data that are collected),
viscosity (measures the resistance to flow of data), variability (measures the

300

N. Dorasamy and N. Pomazalova

unpredictable rate of flow and types), veracity (measures the biases, noise, abnormality, and reliability in datasets), and volatility (indicates how long data are valid
and should be stored) [15]. The multiple dimensions make data searches and
retrieval more complex, as organizations have to find economical ways of integrating heterogeneous datasets while allowing for newer sources of data (in origin and
type) to be integrated within existing systems [6]. The proliferation of social
networks and social media requires much of the data being collected to be thoroughly analyzed before decision-making, as the data can be easily manipulated. In
view of the high cost of data collection and management, organizations need to
analyze the trade-off between accuracy and the cost of inaccurate data. If the quality
of data is poor and users cannot make sense of it, then the data has limited value
and use.

13.3.1 Unstructured Data

Unstructured data from multimedia networks cannot be categorized or analyzed
numerically, as it uses natural language. The explosive growth of social media
implies that the variety and quantity of Big Data is growing. A great deal of the
growth can be traced to unstructured data. For example, analyzing words and
pictures and then collating everything into meaningful and accurately interpreted
information requires diverse methods and can be time-consuming. The challenge is
exacerbated when time-sensitive issues like monitoring civil riots require data to be
aggregated and analyzed in the shortest time possible [19].
Some authors argue that there is no Big Data in the context of social problems, as
data is highly unstructured and generally not limited to numbers. In the case of child
pornography, the global industry lures thousands of children annually. Increasingly,
the producers of child pornography make use of various Internet platforms like
mobile phones, social media, and online classifieds. Although, many initiatives
exist to curb the problem, few initiatives have attempted to use Big Data. Data from
these technologies could be collected and used to identify, track, and prosecute
offenders. The problem is that the illicit nature of child pornography makes it
difficult to collect reliable primary data for some of the following reasons: there
are no valid indicators to measure antipornographic success and information collected often meets organizational needs and not global needs [6]. In addition,
because of data privacy and security issues, data held by various organizations
are seldom shared in raw form, thereby limiting the creation of global datasets. This
is accentuated by agencies combating social evils often competing with each other
for scarce resources, therefore not being eager to share data.
Different service providers have their own network management systems. Organizations cannot connect their datasets across other organizations, if such data is
immersed in their administrative systems for operational purposes. One such case is
the US healthcare industry, which is characterized by large volumes of health plans
offered by different service providers with their own network management. This

13

Social Impact and Social Media Analysis Relating to Big Data

301

invariably results in data being stored in multiple formats in multiple places. If this
data was more efficiently managed, then massive savings can accrue.

13.3.2 Gaps in Governance Standards

The lack of adequate data governance standards has failed to define how data is
captured, stored, and used for accountability in the social arena. As a result some of
the emergent challenges include integrating different datasets which lack good
metadata (data that describe data), poor quality data, and difficult-to-manipulate
forms such as PDFs or older file formats. As a result, large inconsistencies in the
captured data exist, further complicated by the need to transform data for analysis,
which is costly. The number of publicly available government datasets has accelerated, but only limited datasets where there is good metadata, ease of accessibility,
and manipulability are ever used. Further, integrating information from multiple
data sources requires skilled workers. In a report quoted by Fanning and Grant [7],
by 2018 there would be a shortage of 140,000190,000 people with deep analytical
skills in the USA. This could be a serious problem for the analysis of Big Data to
make effective decisions, if there are no initiatives to develop data analytics skills.
Without clarifying ethical issues on data storage, access, and use by different
stakeholders, advancements in computational social science may put the public at
increased risk. If society is to be protected, then there should be legal and ethical
limitations on how social media as a computational social science tool can be used.
Citing the sentiments of the CEO of Nasdaq, Mark Zuckerberg, to take risks is to
move fast and break things. Oboler et al. [17] argue for external constraints to
protect society from the cost of mistakes by social media innovators.
The risk posed by the capacity of computational social science tools and the
explosion in the corpus of data, free of the ethical constraints placed on researchers,
raises serious questions about the impact that those who control the data and the
tools can have on society. Many social media companies are driven by online
advertising revenue which places the individuals interest in privacy in conflict
with the interest of advertisers in extensive customer profiling. This is facilitated by
Web 2.0 sites which have higher advertising rates for advertisers to target selected
users. The increase in advertisers targeting specific users highlights unexpected
consequences. For example, Target analyzed purchasing patterns to identify potential customers of baby paraphernalia. The analysis, based on purchasing history of
unrelated items, highlighted potential pregnancies with a high degree of accuracy.
Target sent advertising material to its target market, creating an angry response
from a father whose teenage daughter received the advertising, but not knowing at
the time that she was really pregnant. The daughter was forced to confirm her
pregnancy because the retailer targeted her in the marketing as a result of the data
analysis. This does raise ethical concerns regarding privacy of information [10].
Further, social media sites can scrape a great deal of data from users age and
sexual preference to target advertisements for adult products, which can cause

302

N. Dorasamy and N. Pomazalova

distress or unauthorized disclosure of sensitive personal information. This poses

both technical and ethical questions like: Is any technically possible use of
personal data ethically acceptable? [17]. Social networking companies and advertisers need to consider such critical questions. By limiting data acquisition, sharing,
and use and by raising public awareness of the implications of its availability
through ethical considerations, the risk of abuse can be controlled.

13.3.3 Serving Self-Interest

Large volumes of data are not necessarily representative and reliable to solve
problems relating to public interest. Big Data users can exploit Big Data with no
regard for data quality, legality, data meaning, and process quality. For example, in
2011, the Rainforest Action Network in the USA discovered that the American
Petroleum Institute and its oil lobby allies were able to manipulate social media
opinion to show support for a pipeline project to carry oil from Canada to Texas by
using fake Twitter accounts to send large numbers of tweets to show support for the
project, which falsely represented public opinion. The Rainforest Action Network
(RAN) discovered that 14 of 15 accounts were faked and the tweets were generated
by an automated process [6].
Desouza and Smith [6] cite the case when public agencies and a newspaper in
New York released information about gun owners after the Connecticut school
mass shooting. Published information on the names and addresses of licensed gun
owners living in the neighborhood can be used by the wrong people like criminals
to target vulnerable homeowners who do not own guns or to target homeowners
who have guns in order to steal them. According to Oboler, Welsh, and Cruz [17],
methodology and ethics, drawn from the underlying fields of computational and
social sciences, need to be considered. The authors argue that considerations apply
not only to the research context but also to the worlds of government and commerce
where philosophical concerns are less likely to counter immediate practical benefits. Most significantly, these concerns need to be considered in the context of social
media platforms which have become computational social science tools that are
easily accessible to businesses, governments, private citizens, and the platform
operators themselves. Governments can exercise their power when they see social
media acting against their interests. One such example was when the US government asked Twitter through a court order for data on WikiLeaks founder Julian
Assange and those connected to him.
Social media can also promote the agenda of governments. Another example
was when the Egyptian government cut off the Internet during the 2011 riots, after it
realized that the US government provided training through the Internet to influence
social change among Egyptian dissidents. Computational social science tools
together with social media data can be used to reconstruct the movements of
activists, to locate dissidents, and to map their networks. Governments and their
security services have a strong interest in this activity [17].

13

Social Impact and Social Media Analysis Relating to Big Data

303

13.3.4 Moving Away from Comfort Metrics

The analysis of general trends and the profiling of individuals can be investigated
through social sciences. In this regard, Kettleborough [11] argues against comfort
metrics, whereby data that is not relevant and focuses merely on the process is
collected. There is a need to look at data differently and to be prepared to throw
away many old beliefs like forced ranking, which looks at normal distribution to
employee performance, as a good employee performance measurement tool.
According to Kettleborough [11], there is evidence that these approaches have
actually damaged organizations like Microsofts lost decade, which was as a direct
result of misplaced or misunderstood data techniques.
Research studies have also shown that many businesses were preselecting and
filtering candidates for employment based on social media. If job applicants do not
protect their reputation online, then this can compromise their applications, if the
interviewer has access to negative attributes posted online. Currently, business and
government control large volumes of data used for computational social science
analysis. The capacity to collect and analyze datasets on a vast scale provides the
magnitude to disclose patterns of individual and group behavior. The potential
damage from inappropriate disclosure of information is sometimes obvious. A lack
of transparency in the way data is analyzed and aggregated, combined with a
difficulty in predicting which pieces of information may later prove damaging,
means that many individuals have little knowledge of potential adverse effects of
the expansion in computational social science [17].
If data is not correctly understood, then massive mistakes can cause harm. Many
employment recruiters are already looking at the social media lives of job
applicants Although it is seen as justifiable, especially using work-related sites
such as LinkedIn, in cases where employers seek data from nonwork-related social
media life, there could be some potentially negative consequences. According
to the CIPD, using social media in recruitment or as part of career progression
carries the risk of a number of different claims if a candidate is not appointed as a
result of information gleaned. These include the following [11]:
A breach of the Human Rights Act 1998 (incorporating Article 8 of the European
Convention on Human Rights) to respect private and family life.
A breach of the Data Protection Act 1998, which states that data controllers such
as prospective employers should not hold excessive information and should
process information in a fair way.
It has been suggested that the over 50s age group will be more cautious with their
social media presence than the under 30s, resulting in more potential for
negative recruitment decisions for younger people.
Information about marital status, number of children, and sexual orientation may
incorrectly influence a selection decision.
Information about physical or mental state, such as revealing depression to
friends on social media, may be a disadvantage.

304

N. Dorasamy and N. Pomazalova

If life-changing judgments are to be made about people, then quality and

accuracy must be beyond reproach. If employees are hired, promoted or dismissed
based on Big Data discrimination, then there can be legal implications [11].

13.3.5 Using Power to Leverage Outcomes

Governments and powerful data-rich companies have the financial support and
powerful resources to access data. Such organizations, by their nature, tend at
times to assume that the risk of unjustified impacts on individuals is of little
consequence when compared with the potential to avert perceived calamities
[20]. It is easy to manipulate people, like using computational social science to
guide political or product advertising, selling messages that people will favor or
withhold information that may compromise support. Google, for example, can sway
an election by predicting messages that would engage an individual voter (positively or negatively) and then disseminate content to influence that users vote. The
predictions could be highly accurate by making use of a users e-mail in their
Google-provided Gmail account, their search history, and social network connections. The dissemination of information could include recommended videos on
YouTube to highlight where one political party agrees with the users views also
articles in Google News could be given higher visibility to help sway voters into
making the right choice [17]. Further, this can be complemented with negative
messages to appear to create a balance, but in reality may have little or no impact.
Such manipulation may not appear obvious, yet powerful to achieve the outcomes
of the manipulator.

13.3.6 Risks Relating to Social Media Platforms

Social media platforms have added to their data either by acquiring other technology companies as Google did when acquiring YouTube or by moving into new
fields as Facebook did when it created Facebook Places providing a geolocation
service which generates high value information [14]. The value of information can
be maximized by using a primary key that connects this data with existing information like a Facebook user ID or a Google account name, where a user is treated as
a single user across all products of the company [14]. One account can connect to
various types of online interactions, exposing greater breadth of a users profile. In
such a case all the data is immediately related and available to any query companies
like Facebook and Google may have. This can be alarming as there is little privacy,
since any information can be collected across platforms about users.
Accounts that are identity-verified, frequently updated, and used across multiple
aspects of a persons life present the richest data and pose the greatest risk. For
example, Facebooks Timeline feature allows users to mine social interactions that

13

Social Impact and Social Media Analysis Relating to Big Data

305

had long been buried. Further, since Timeline is not an option in Facebook, masses
of personal data can be held. Another challenge was the Beacon software which,
developed by Facebook, connected peoples purchases to their Facebook account. It
indicated what users had purchased, where they got it, and whether they got a
discount. It was eventually closed in view of legal, privacy, and ethical considerations. Further, the emergence of massive open online courses under MOOCs is
now causing a stir in the world of Big Data with evidence that student details,
including performance data, is being sold online. Recruiters looking for highly
motivated candidates with wisdom can hunt potential candidates on this
platform [11].

13.3.7 Research Methodology Challenges

The use of social media as a source of social research data can present various
methodological challenges. There can be sampling bias which can distort findings,
as any particular social medium is not representative of the population as a whole.
Avoiding sampling bias in social media sources is a great challenge for researchers
in the social sciences. If computational tools are to be appropriately used in social
research, then it is important that users are aware of the strengths and weaknesses of
such tools. Therefore, it is vital that the capacity of social researchers in developing
skills relating to computational methods and tools is developed, so that they can
decide when and how to apply them responsibly.
Anonymity among employees during surveys is also causing concern. While it is
believed that employee responses will be more truthful if they remain anonymous,
their identity can be traced from the demographic details in their social media
profiles. If used incorrectly, this honest data could be turned against the
employees [11]. Employees who are aware of this may not necessarily give the
correct picture.
In terms of reliability and validity, decisions cannot be made with incomplete or
incomprehensive data. Rational and fair decisions have to be based on representation. For example, if 20 people are happy with the service at a state hospital, this
does not exhibit behavior that is statistically significant for the whole population.
As the old adage states, one swallow does not a summer make. Data focusing on a
few does not paint a correct picture. Further, Kettleborough [11] contends that
correlation and causation should not be analyzed at face value, since if two items
correlate, it does not mean that one causes the other.
Data mining or scrapping of social media sites can result in personal data being
used against individuals, even if it has been cleaned to remove personal references.
One such example is a study by researchers from the Universite Catholique de
Louvain in Belgium who identified 95% of the unique users by analyzing only four
GPS time and location stamps per person. In addition, researchers at Carnegie
Mellon University were able to create a system to uncover Social Security numbers

306

N. Dorasamy and N. Pomazalova

from birthday and hometown information listed on social networking sites like
Facebook [11]. Large amounts of data can become the target of the unscrupulous.

13.3.8 Securing Big Data

Big Data, while sourcing data from multiple sources, relies on data that is available.
Further, such data must be secured. The challenge is how the data is collected and
stored. This raises security issues like internal employees adhering to confidentiality policies. Cases of storage abuses have occurred at Facebook sites. Data can also
be lost due to hackers and employees. One example is the two Aviva employees
who sold details of people who had accidents to claims companies. The fraud flag
was raised when the claimants received calls from firms persuading them to take
personal injury claims [11]. Information that is not secured can be used for
blackmailing or espionage.

13.3.9 Limitations of Addressing Social Problems

In the social arena, a major gap exists between the potential of data-driven information and its actual use in addressing social problems. Certain social problems can
be easily solved using Big Data, such as weather forecasting and areas with high
disease rates. However, pandemic problems like drug trafficking and unemployment cannot be easily resolved in a sustainable way with Big Data. According to
Desouza and Smith [6], these evil problems are more dynamic and complex than
their technical counterparts, because of the diversity of stakeholders involved and
the numerous feedback loops among the interrelated constituents. Government
agencies and nonprofits are involved in tackling these problems but face the
following challenges: limited cooperation and data sharing among them; inadequate information technology resources; their counterparts in the hard sciences
work on technical problems or in business who have ready access to financial,
product, and customer information; missing and incomplete data; and data stored in
silos or in forms that are inaccessible to automated processing. In addition, there are
regulatory constraints like policy relating to data sharing agreements, privacy and
confidentiality of data, and collaboration protocols among various stakeholders
tackling the same type of problem. While various agencies may invest in data
technologies, the return on investment for solving social problems is yet to be
convincing. This impact on the need is to be provided with information and advice
via sources that they can trust in a more timely way.

13

Social Impact and Social Media Analysis Relating to Big Data

13.4

307

Imperatives for Big Data Use from Social Media

In decision-making, context is the key; therefore knowledge of the domain such as

social media is crucial. Analysis of Big Data is directly linked to decision-making,
which has to be supported by very intricate techniques using wide and deep
extensive data sources as shown in Table 13.1. Big Data is about massive amounts
of different types of observational data, supporting different types of decisions and
decision time frames. According to Goes [9], analytics moves from data to information to knowledge and finally to intelligence. The generation of knowledge and
intelligence to support decision-making is critical as the Big Data world is moving
toward real-time or close to real-time decision-making. Therefore, the need for
context-dependent methodologies that strengthen prediction is pivotal for effective
data analysis.
Big Data analytics from social media has to consider the tools, software, and the
data to ensure quality results. While the technical ability may exist to gather data,
the analytical capacity to draw meaning from such data needs to be developed. For
example, visualization can be produced from real-time information as datasets
emerge from user activity. However, such visualizations can only be considered
powerful representations if visualization specialists are aware of which relationships benefit users. This requires an understanding of how meaning can be created
through and across various datasets in social media platforms.

13.4.1 Responsibility of Analytic Role Players

Big Data has enormous potential to inform decision-making to help solve the
worlds toughest social problems. But for this to happen, issues relating to data
collection, organization, and analysis must first be resolved. Much of this responsibility lies with the major analytic players, as shown in Table 13.2, who offer
valuable services that help users cope with using Big Data effectively.
The aforementioned major analytic players have to ensure effective use of Big
Data from social media platforms. This requires prudent use of analytic tools,
incorporating the following guidelines [7]:
Table 13.1 Big data
analytics

Decisions
Real time
Close to RT
Hourly
Weekly
Monthly
Yearly
Adapted from [9]

Analytics
Visualization
Exploration
Explanatory
Predictive

Techniques
Statistics
Econometrics
Machine learning
Computational
Linguistics
Optimization
Simulation

N. Dorasamy and N. Pomazalova

308
Table 13.2 Major analytic players
Firm
SAP

Products
HANA; Applied
analytics

Splunk

Tableau
Cloudera

Splunk Hadoop Connect,

Splunk enterprise
Spotfire
Big query
Pure data system
High-performance
analytics
Data pipes; Druid
Big data appliance;
Exadata; Exalytics
Tableau
Hbase

Qliktech
GoodData

QlikView
GoodData bashes

Tibco
Google
IBM
SAS
Metamarkets
Oracle

Website
http://www.sap.com/solutions/technology/in-mem
ory-computingplatform/hana/overview/index.epx;
http://www54.sap.com/solutions/analytics/applica
tions/software/overview.html
http://www.splunk.com/
http://spotfire.tibco.com/
https://cloud.google.com/products/big-query
http://www-01.ibm.com/software/data/puredata
http://www.sas.com/software/high-performance-ana
lytics/index.html
http://metamarkets.com/platform
http://www.oracle.com/us/technologies/big-data/
index.html
http://www.tableausoftware.com/
http://www.cloudera.com/content/cloudera/en/prod
ucts/clouderaenterprise-core.html
http://www.qlikview.com/
http://www.gooddata.com/what-is-gooddata/

Adapted from [7]

Use of in-memory database technology that avoids resources swapping databases between the storage medium and the memory, but rather operating within
memory with only limited accessing of alternative storage mediums.
The sheer size and complexity of data cannot be handled by traditional technologies built on relational or multidimensional databases, as there is a need to have
flexibility to have questions answered in real time.
Use caves of data from unstructured data to improve service levels, reduce
operations costs, mitigate security risks, and enable compliance.
Use tools that break down traditional data silos and attain operational intelligence that benefits both IT and the business, which is valuable for capturing
machine-generated data into a system that would provide operational realtime data.
The need to use technology to efficiently store, manage, and analyze unlimited
amounts of data that can process any type of data differently than relational
databases.
Since information in the various social media platforms is not static, if it is not
updated and cleaned, then dirty data will arise. Considering the garbage in,
garbage out syndrome, poor quality data will produce poor results.

13

Social Impact and Social Media Analysis Relating to Big Data

309

13.4.2 Evidence-Based Decision-Making

In addition, the following four recommendations have the potential to create
datasets useful for evidence-based decision-making [6]:
The global community needs to create large data banks on critical issues like
homelessness and malnutrition, which must have the capacity to hold multiple
different data types along with metadata that describes the datasets. This requires
multi-sector alliances that promote and create data sharing on sectoral issues. At
the 2012 G-8 Summit, leaders committed to the New Alliance for Food and
Nutrition Security to help 50 million people out of poverty over the next 10 years
through sustained agricultural growth. This is supported by a number of databases like Agrilinks.org, Feed the Future Initiative website, and Womens
Empowerment in Agriculture Index.
Citizens and professionals can help create and analyze these datasets. With the
growth of data through open data platforms, citizens are creating new ideas and
products through what has become known as citizen science. A bike map and
map of the London tube were created by citizens, using the raw data from the
London Datastore which is managed by the Greater London Authority.
Big Data cannot be left to the pure sciences and business, but needs analysts in
the social sciences to be statistically equipped to collect data for large-scale
datasets. Skills in data organization, preservation, visualization, search, and
retrieval, identifying networked relationships among datasets, and how to
uncover latent patterns in datasets need to be developed. These are valuable
skills that go beyond simply searching the web for information.
Virtual experimentation platforms which allow individuals to interact with
different ideas and work collaboratively to find solutions to problems can create
large datasets, develop innovative algorithms to analyze and visualize the data,
and develop new knowledge for tackling social challenges. The use of open
forums such as wikis and discussion groups can help the community share
lessons learned, collaborate, and advance new solutions.
In addition, Oboler et al. [17] argue that social networking has provided a diverse
range of datasets covering large sections of the population, granting researchers,
governments, and businesses the powerful ability to identify trends in behavior
among a large population and to find vast quantities of information on an individual
user. As the industry develops, social media computational tools will increase the
scope, accuracy, and usefulness of such datasets. In view of the ethical and privacy
implications, regulatory barriers restricting the collection, retention, and use of
personal information require consideration. While laws protect human rights, there
is a need for greater protection of the customer.

13.4.3 Protection of Rights

The rights of users need to be protected, as social media platform providers and
various agencies provide innovative services to targeted users. The debate is

310

N. Dorasamy and N. Pomazalova

whether consumers are protected from preventable harm only after proving damage
or are rules set by law. In the first approach, advertisers have more freedom to mine
data from various social media platforms, data over which the user has no control
especially if it is outdated or hacked by third parties. The safeguarding of personal
rights and freedoms is more favored through the setting of regulations and laws.
This would place the burden on social media to restrict the storage, accessibility,
and manipulation of data in ways that limit its usefulness. This can prevent
unscrupulous use of data. However, this will require legislators to use multilateral
legislating, since websites can freely choose the physical location of their hosting
infrastructure where there is least regulation.
The ethical barriers for data use in the social sciences are much higher than pure
science research as the data collection of personal information is higher in the social
sciences. Oboler et al. [17] illustrate some suggestions to manage ethical use of
social media data, as given below:
In keeping with the code of ethics developed by professional bodies, example for
engineers, these should be applied to social media as well. Such guidelines
commit members to act in public interest, by not causing harm or violating the
privacy of others. Social media platforms are a form of computational social
science which requires recognition of the ethical concerns in the social sciences.
This can reduce the opportunity for the abuse of a very powerful tool. Users of
social media have an ethical responsibility to one another.
A code of conduct for producers and consumers of online data which can
highlight the issues to be considered when publishing information. For example,
when a Twitter user uploads photographs, their action may reveal information
about others in their network; the impact on those other people should be
considered under a producers code of ethics. A consumer code of ethics is
also needed; such a code would cover users viewing information posted by
others through a social media platform. A consumer code could raise questions
of when it is appropriate to further share information, for example, by
retweeting it.
Guidelines for principles of engagement can help users determine what they are
publishing and to create awareness of the potential impact of publishing information. The power of social media can be used to warn the owner when the
content may pose a risk, especially when accesses open.
A cultural mind shift is needed to become more forgiving of information
exposed through social media, an acceptance that social media profiles are
private and must be locked down with more intricate filters and used only in
certain settings.
The aforementioned suggestions would change the nature of social media as a
computational social science tool, by filtering what should be included out of the
tools field of observation. As an instrument-based discipline, the way the field is
understood can be changed either by changing the nature of the tool or by changing
the way we allow it to be used.

13

Social Impact and Social Media Analysis Relating to Big Data

311

13.4.4 Knowing the Context

Understanding and knowledge of the context is fundamental. For example, marketing depends, to a large extent, on information technology accruing from social
networks. Researchers have to master the collection and analysis of web data and
user-generated content, using advanced techniques. This is necessitated by the
massive amounts of observational data, of different types, supporting different
types of decisions and decision time frames [9]. In this regard, if researchers want
to explain the growth in online shopping among teenagers in developing economies
using social media networks, then it is imperative to use models like longitudinal
models, latent models, and structured models to explain the causes within the
context. Since the Big Data environment is targeting real-time decision-making,
it is imperative that tools employed to analyze social media networks use contextdependent methodologies that enhance prediction in a valid and reliable way. The
reason being that not only are the networks intricate but also require knowledge of
the complex models. Consideration of these dynamics can produce valuable information from Big Data, allowing modeling of individuals at a very detailed level
with a rich proliferation of the environment surrounding them [9].
Big Data analytics has the ability to yield deeper insights and predictions about
the individuals. According to Waterman and Bruening [22], even though data may
be processed accurately, the results may have profound effect on personal life
choices. The authors argue that understanding the sources and limitations of data
is critical to mitigate harm to individuals. This necessitates understanding and
responding to the implications of choices about data and data analytic tools,
integrity of analytic processes, and the consequences of applying the outcomes of
analytic models to information about individuals [22].

13.5

Conclusion

The proliferation of Big Data has emerged as the new frontier in the wide arena of
IT-enabled innovations and opportunities. Attention has focused on how to harness
and analyze Big Data. Social media is one component of a larger dynamic and
complex information domain, and their interrelationships need to be recognized. As
the connection with Big Data grows, we cannot avoid its impact. Without being
familiar with the data, the benefits of Big Data cannot be reaped. Large volumes of
data cannot be analyzed using conventional media research methods and tools. The
current Big Data analytics trend has seen the tools used to analyze and visualize
data getting continuously better. There has been a major investment in the development of more powerful digital infrastructure and tools to tackle new and more
complex and interdisciplinary research challenges.
Current programs have seen companies like Splunk, GoodData, and Tibco
providing services to allow their users to benefit from Big Data. Users with the

312

N. Dorasamy and N. Pomazalova

ability to query and manipulate Big Data can achieve actionable information from
Big Data to derive growth by making informed decisions. Access to data is critical.
However, several issues require attention in order to benefit from the full potential
of Big Data. Policies dealing with security, intellectual property, privacy, and even
liability will need to be addressed in the Big Data environment. Organizations need
to institutionalize the relevant talent, technology, structure workflows, and incentives to maximize the use of Big Data.
It is imperative that apart from the power users in marketing, financial,
healthcare, science, and technical fields, those involved in daily decision-making
must be empowered to use analytics. As more and more analytical power reaches
decision-makers, enhanced and more accurate decision-making will emerge in the
future. While there is a need to size the opportunities offered by continuing
advances in computational techniques for analyzing social media, the effective
use of human expertise cannot be ignored. Using the right data in the right way and
for the right reasons to innovate, compete, and capture value from deep and realtime Big Data information can change lives for the better. Big Data has to be used
discriminately and transparently.

References
1. Boyd D, Crawford K (2012) Critical questions for big data: provocations for a cultural,
technological, and scholarly phenomenon. Inf Commun Soc 15(5):662679
2. Brown B, Chui M, Manyika J (2011) Are you ready for the era of big data? McKinsey Q
4:2435
3. Bughin J, Chui M, Manyika J (2010) Clouds, big data, and smart assets: ten tech-enabled
business trends to watch. McKinsey Q 56(1):7586
4. Chen H, Chiang RH, Storey VC (2012) Business intelligence and analytics: from big data to
big impact. MIS Q 36(4):11651188
5. Colgren D (2014) The rise of crowdfunding: social media, big data, cloud technologies.
Strategic Financ 2014:5557
6. Desouza KC, Smith KL (2014) Big data for social innovation. Stanf Soc Innov Rev
2014:3943
7. Fanning K, Grant R (2013) Big data: implications for financial managers. J Corporate Account
Financ 2013:2330
8. Gandomi A, Haider M (2014) Beyond the hype: big data concepts, methods, and analytics. Int
J Inf Manag 35:137144
9. Goes PB (2014) Big data and IS research. MIS Q 38(3):iiiviii
10. Hill K (2012) How target figured out a teen girl was pregnant before her father did. Forbes
(16 February). http://www.forbes.com/sites/kashmirhill/2012/02/16/how-target-figured-out-ateen-girl-was-pregnant-before-her-father-did/. Accessed 9 Dec. 2014
11. Kettleborough J (2014) Big data. Train J 2014:1419
12. Kim HJ, Pelaez A, Winston ER (2013) Experiencing big data analytics: analyzing social media
data in financial sector as a case study. Northeast Decision Sciences Institute Annual Meeting
Proceedings. Northeast Region Decision Sciences Institute (NEDSI), April 2013, 6269
13. LaValle S, Lesser E, Shockley R, Hopkins MS, Kruschwitz N (2013) Big data, analytics and
the path from insights to value. MIT Sloan Manag Rev 21:4050

13

Social Impact and Social Media Analysis Relating to Big Data

313

14. McCarthy J (2010) Blended learning environments: using social networking sites to enhance
the first year experience. Australas J Educ Technol 26(6):729740
15. McKelvey K, Rudnick A, Conover MD, Menczer F (2012) Visualizing communication on
social media: making big data accessible. arXiv (3):1014
16. Mendoza M, Poblete B, Castillo C (2010) Twitter under crisis: can we trust what we RT. In: 1st
workshop on Social Media Analytics (SOMA 10). ACM Press, Washington, DC
17. Oboler A, Welsh K, Cruz L (2012) The danger of big data: social media as computational
social science. First Monday 17(7):6065
18. Proctera R, Visb F, Vossc A (2013) Reading the riots on Twitter: methodological innovation
for the analysis of big data. Int J Soc Res Methodol 16(3):197214
19. Qualman E (2012) Socialnomics: how social media transforms the way we live and do
business. Wiley, New York
20. Wigan MR, Clarke R (2013) Big datas big unintended consequences. IEEE Comput Soc
2013:4653
21. Young SD (2014) Behavioral insights on big data: using social media for predicting biomedical outcomes. Trends Microbiol 22(11):601602
22. Waterman KK, Bruening PJ (2014) Big data analytics: risks and responsibilities. Int Data
Privacy law 4(2):8995

Index

A
Abnormal patterns, 47, 53
Actuation, 132
Adaptive Boosting, 249
Agglomerative approaches, 58, 59, 64, 65, 67,
71, 7476
AJAX, 236
AllReduce, 150
Amazon Dynamo, 235
Amazon Machine Images (AMI), 145
Analytics, 225, 242
Analytics-as-a-Service, 126, 127
Anomaly detection, 141, 142, 148
Apache, 140, 148152
Hadoop, 192, 193, 218
Kafka, 106
Pig, 205, 213
Apple-to-apple records, 167
Apple-to-orange pairs, 163, 164, 166, 175177,
180183
Apriori algorithm, 248, 251, 256, 263266
Artificial neural network, 284
Association rule mining, 246

B
Bacterial blight, 281
Big, 224234, 238
Big Data, 95106, 108, 113, 119128, 131,
132, 134
Big Data analysis, 294, 296
Big Data as a Service (BDaaS), 144
Big Data Interoperability Framework, 6
Big Data scenarios, 59, 65, 70, 71, 7476
Big Data Science (BDS), 247

Bigtable, 232, 233

Bio-informatics, 120
Blaze, 146
Brown spot symptoms, 279
Bucketing, 193, 198203
Bug records, 162, 164166, 176, 177, 179182,
185
Bulk-synchronous parallel (BSP), 58, 6061
Business intelligence (BI), 230232, 235, 242
Business Process Execution Language (BPEL),
13, 21, 25
Business-to-business (B2B), 235
Business-to-consumer (B2C), 235

C
C/C++, 147
Cartesian space, 167
Case study, 271, 290
Characteristics of big data, 120
Chrome bug data, 164
CLARANS, 250
Classification, 141, 142, 148, 246249, 254,
256, 260, 261, 265, 266
Classification algorithm, 248, 261
Cloud computing, 101, 103, 125, 144, 155,
246, 252253, 295, 298
Cloudera, 308
Cloudera Impala, 152
Cluster(ing), 58, 6476, 141, 142, 148,
246248, 250, 254, 256, 257, 263, 265,
266
Clustering algorithm, 274, 276
Column family databases, 238
Complex event processing, 4248, 54, 55

Springer International Publishing Switzerland 2016

Z. Mahmood (ed.), Data Science and Big Data Computing,
DOI 10.1007/978-3-319-31861-5

315

316
Compliance, 23, 25, 30, 34
Constrained Application Protocol (CoAP), 17
Content based image retrieval (CBIR), 60
Convexity, 285
Cosine similarity, 168, 171
Cost optimization, 162
Crop diseases, 278
Custom relationship management (CRM), 294
Cyber-infrastructures, 95, 96
Cyber physical cloud computing, 127
Cyber physical systems (CPSs), 96, 118, 119,
127, 132, 134
CyberWater, 97, 108113
Cyber world, 118, 119, 124, 127, 134

D
Data acquisition, 128
Data analysis, 131132
Data analytics, 119, 125, 223226, 230, 242
Data-as-a-service, 126
Database, 194198, 204, 214, 219
Data cleaning, 129
Data compression, 129
Data fusion, 129
Data gathering, 101102
Data management, 107108, 119, 125, 126,
128, 130, 131, 134
Data mining, 246253, 255
Data mining in cloud computing (DMCC)
framework, 246248, 253255, 257,
258, 265, 266
Data models, 100
Data Protection Act, 303
Data science, 240242
Data scientist, 242
Data storage, 130
Data warehousing, 228, 231, 241
Decision-making, 9799, 103, 108, 113
Decision support systems, 118, 119, 124,
126, 271
Decision Tree, 248, 249, 256, 261, 266
Description Logic, 21
Digital marketing, 297
Digital technology, 294
Dimensionality reduction, 141, 148, 270
Dimension reduction, 271
DIM-RED-GA, 278
Duplicate data records, 162, 163, 185
Duplicate image detection (DID), 60

E
Edge clustering value (ECV), 70, 72
E-health, 95

Index
Ellipse variance, 285
Enterprise architectures, 8589, 274278
Enterprise resource planning (ERP), 294
Environmental data, 96, 104
Evidence-based decision-making, 309
Extensible Markup Language (XML), 4, 5, 8,
9, 13, 1619, 21, 2426, 2931, 36, 37,
225, 230, 246, 248
External memory breadth first search
(EM BFS), 63
External memory techniques, 58

F
Facebook, 59, 60, 121, 122, 294, 296, 297,
304, 306
Face recognition, 60
Fast data, 4, 730, 34
Fast unfolding algorithm, 64, 68, 72
Feature extraction, 283
FIFO, 49
Fuzzy inference, 274, 276
Fuzzy propagation, 81, 83, 84, 91

G
Gaussian distribution, 249, 261
Genetic algorithm (GA), 272, 274
Geographical information system (GIS), 97,
103, 104, 107
Giraph, 58
Google, 192, 214
GPS, 50, 51
Graph databases, 238
GraphLab abstraction, 247
Graphics processing unit (GPU) framework,
8183
Greedy algorithm, 8083, 8991
Grid computing, 257

H
Hadoop, 4, 8, 36, 121, 140, 145153, 156, 223,
224, 227, 229, 231233, 247, 296, 308
Hadoop Distributed File System (HDFS), 193,
197, 198, 201, 204207, 214, 223, 229,
231233
HANA, 296, 308
HBase, 193, 196, 214219, 231233
HC-PIN algorithm, 71
HDFS. See Hadoop Distributed File System
(HDFS)
Healthcare, 96
Healthcare monitoring, 52, 54
Heterogeneity, 9597, 102, 108, 113

Index
Hive, 145, 146, 193205, 213, 219, 223,
231, 232
Hive Query Language (HQL), 146
HPC, 103
Hybrid cloud, 252
Hybridizing genetic algorithm, 272

I
Iaa. See Infrastructure as a Service (IaaS)
Image filtering, 282283
Image processing, 282
IMGPU, 82
Independent cascade model, 91
Information-as-a-service (IaaS), 126
Information retrieval, 164, 165,
178, 183
Infrastructure as a Service (IaaS), 125, 126,
252, 253
INSPIRE, 103
Instagram, 121
Internet of Things (IoT), 4, 6, 25, 26, 32, 35, 36,
42, 54, 96
Interoperability, 4, 5, 714, 1619, 2125,
2836
Intrusion detection, 45
IPython, 146

J
Java, 52, 54, 147, 148, 152, 154
Java Enterprise, 235
JavaScript Object Notation (JSON), 4, 5, 8,
13, 16, 18, 19, 21, 2427, 30, 31, 36,
37, 233
JDBC, 194
Jubatus, 148, 149

K
Kalman filter, 50
K-means, 142, 151, 155, 250251, 256, 263,
264, 266
algorithm, 250, 263
clustering, 172
K-Medoids, 250
Knowledge discovery, 246, 254
Knowledge discovery in databases, 246

L
Latent Dirichlet Allocation (LDA), 164
Leader follower algorithm, 69, 70

317
Leaders identification for community detection
(LICOD) algorithm, 65, 72
Linguistics, 307
Linkedin, 294
Local minima, 272, 274

M
Machine learning, 141, 154, 259
Machine-learning-as-a-service (MLaaS),
152, 154
Mahout, 148151
Mamdani model, 276, 278
Maple, 144
MapReduce, 6, 35, 60, 81, 83, 8591, 106,
108, 121, 140, 145147, 149151,
154, 156, 192, 193, 198, 223, 231,
232, 247, 248
Markov random field (MRF), 82
Massive data sets, 144
Matlab, 143, 147
Memetic algorithms, 271
Message passing interface (MPI), 61
Microsoft Azure, 248, 253, 254
Minimum-sized influential vertices (MIV), 79,
81, 83, 84, 89
MLlib, 148, 150, 151
Mutation, 272, 278

N
Naive Bayes, 248
National Oceanic and Atmospheric
Administration (NOAA), 104
Natural language processing, 141, 163
Nearest neighbor (NN)
classifier, 174
method, 172
Neighbourhood generation method, 273
N-gram based model, 165
NIMBLE, 147, 148
NoSQL, 140, 223, 227, 229, 233238, 240, 243
databases, 227, 233236

O
Octave, 143, 144
ODBC, 194
Okapi weighting scheme, 178
Open data, 97, 107
Oracle, 308
Oryx, 148, 151
OWL. See Web Ontology Language (OWL)

318
P
PaaS. See Platform as a Service (PaaS)
Parallel processing, 5860, 64, 71, 7376,
81, 82
Parsing, 206207
Partitioning, 58, 62, 64, 65, 71, 75, 76, 193,
198203
Partitioning methods, 250
Physical world, 118, 119, 124, 127, 128, 131
Pig, 193, 205214, 219, 231
Pig Latin, 205207, 209211, 213
PIR, 50, 51
Platform as a Service (PaaS), 125, 248,
250253
Pragmatic, 12, 13, 32
Pregel, 58, 60, 62
Preventability analytics, 185
Private cloud, 252
Probabilistic model, 178
Probability distribution, 273, 274
Public cloud, 252
Python, 140, 143146, 150, 153, 154

Q
Query processing, 130131

R
Radio-frequency identification (RFID), 6, 41,
44, 46, 48, 50, 51, 53, 54, 120
Random Forest, 248, 249, 256, 261, 266
Random walks algorithm, 64, 71, 72
RapidMiner, 147
Redness index, 284
Regression, 141, 148
Relational Database Management Systems
(RDBMS), 193, 223
Representational state transfer (REST), 233
RESTful applications, 8, 9, 15, 16, 18
RFID. See Radio-frequency identification
(RFID)
RHadoop, 145
RHive, 145, 146
Rice blast, 279
Rice hispa, 281
RSS feed, 44
R system, 140, 143145, 148, 150, 153

S
SaaS. See Software as a Service (SaaS)
SAP HANA, 225, 230

Index
Scalability, 119, 133
Scalable algorithm, 81
SCAN algorithm, 69, 72
SCHEMA, 195, 198
SEARUM. See SErvice for Association RUle
Mining (SEARUM)
Sector/sphere framework, 247
Self Organizing Maps (SOM), 250
Semantic intrusion detection system, 42,
48, 49
Sensor Alert Service (SAS), 103
Sensor Event Service (SES), 103
Sensor Observation Service (SOS), 101, 103
Sensors, 118120, 124, 127131, 133
networks, 47
web, 97, 98, 101, 105
Seraph, 58, 60, 6263
Service, 143, 144, 152155
SErvice for Association RUle Mining
(SEARUM), 248
Service-oriented architecture (SOA), 19, 31,
230
Sheath rot, 280
SIL Public Interface Descriptor (SPID), 26
Simple Object Access Protocol (SOAP), 4, 5,
13, 18, 19, 31
Smart cities, 95, 96
Smart devices, 295
Social media, 95, 96, 245, 294, 296, 299, 302,
304, 310, 311
Social media services (SMS), 294
Social network(ing), 59, 121, 294, 296, 302,
306, 309
Software as a Service (SaaS), 125,
252, 253
Spark, 4, 8, 106, 140, 147, 148, 150153
SPSS, 143
SQL server, 194
Sqoop, 231
Storm, 106, 148, 151, 152
STRADS, 149
Strong Connected Component (SCC), 82
Structural similarity, 68
Sum of Squared Errors (SSE), 256
SystemML, 147, 148

T
Term frequency-inverse document frequency
(TF-IDF), 164, 169
Textons, 286
Tokenization technique, 172
Twitter, 59, 121, 122, 294, 298, 302, 310

Index
U
Unstructured data, 300
User Datagram Protocol (UDP), 17

V
Variety, 6, 59, 98
Vector space model, 164, 167171, 176
Vehicular ad hoc networks (VANETs),
122124
Velocity, 6, 59, 99
Veracity, 99
Vicissitude, 99
Visualization, 307
Volatility, 99
Volume, 6, 59, 98
Voter model, 91

W
Water resource, 96
Web 3.0, 95, 96, 104
Web of services, 18

319
Web Ontology Language (OWL), 13, 21
Web servers, 120
Weighted graph, 79, 81, 8385
Windows Azure, 151, 153
Wireless body area networks (WBANs),
123124
Wireless sensor network (WSNs), 105,
120121

X
XML. See Extensible Markup Language
(XML)

Y
Yahoo, 192, 205, 214

Z
Zephyr BioHarness device, 54
Zookeeper, 231