Social Media Mining

An Introduction

Reza Zafarani
Mohammad Ali Abbasi
Huan Liu

Draft Version: April 20, 2014

By permission of Cambridge University Press, this preprint is free. Users can make
one hardcopy for personal use, but not for further copying or distribution (either print or
electronically). Users may link freely to the books website: http://dmml.asu.edu/smm, but
may not post this preprint on other web sites.

Social Media Mining

Contents

Introduction
1.1 What is Social Media Mining . . . .
1.2 New Challenges for Mining . . . . .
1.3 Book Overview and Readers Guide
1.4 Summary . . . . . . . . . . . . . . . .
1.5 Bibliographic Notes . . . . . . . . . .
1.6 Exercises . . . . . . . . . . . . . . . .

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

15
16
16
18
21
22
24

Essentials

27

Graph Essentials
2.1 Graph Basics . . . . . . . . . . . . . . . .
2.1.1 Nodes . . . . . . . . . . . . . . .
2.1.2 Edges . . . . . . . . . . . . . . . .
2.1.3 Degree and Degree Distribution
2.2 Graph Representation . . . . . . . . . .
2.3 Types of Graphs . . . . . . . . . . . . . .
2.4 Connectivity in Graphs . . . . . . . . . .
2.5 Special Graphs . . . . . . . . . . . . . . .
2.5.1 Trees and Forests . . . . . . . . .
2.5.2 Special Subgraphs . . . . . . . .
2.5.3 Complete Graphs . . . . . . . . .
2.5.4 Planar Graphs . . . . . . . . . . .
2.5.5 Bipartite Graphs . . . . . . . . .

29
30
30
31
32
35
37
39
44
44
45
47
47
48

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

2.6

2.7
2.8
2.9
3

2.5.6 Regular Graphs . . . . . . . .

2.5.7 Bridges . . . . . . . . . . . . .
Graph Algorithms . . . . . . . . . . .
2.6.1 Graph/Tree Traversal . . . . .
2.6.2 Shortest Path Algorithms . .
2.6.3 Minimum Spanning Trees . .
2.6.4 Network Flow Algorithms .
2.6.5 Maximum Bipartite Matching
2.6.6 Bridge Detection . . . . . . .
Summary . . . . . . . . . . . . . . . .
Bibliographic Notes . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . .

Network Measures
3.1 Centrality . . . . . . . . . . . .
3.1.1 Degree Centrality . . . .
3.1.2 Eigenvector Centrality .
3.1.3 Katz Centrality . . . . .
3.1.4 PageRank . . . . . . . .
3.1.5 Betweenness Centrality
3.1.6 Closeness Centrality . .
3.1.7 Group Centrality . . . .
3.2 Transitivity and Reciprocity . .
3.2.1 Transitivity . . . . . . .
3.2.2 Reciprocity . . . . . . .
3.3 Balance and Status . . . . . . .
3.4 Similarity . . . . . . . . . . . . .
3.4.1 Structural Equivalence .
3.4.2 Regular Equivalence . .
3.5 Summary . . . . . . . . . . . . .
3.6 Bibliographic Notes . . . . . . .
3.7 Exercises . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

Network Models
4.1 Properties of Real-World Networks
4.1.1 Degree Distribution . . . .
4.1.2 Clustering Coefficient . . .
4.1.3 Average Path Length . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

48
49
49
49
53
56
57
61
64
66
68
69

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

73
74
74
75
78
80
82
84
85
87
87
91
92
95
96
98
101
102
103

.
.
.
.

105
106
106
109
109

4.2

4.3

4.4

4.5
4.6
4.7
5

Random Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . 110

4.2.1 Evolution of Random Graphs . . . . . . . . . . . . . . 112
4.2.2 Properties of Random Graphs . . . . . . . . . . . . . . 115
4.2.3 Modeling Real-World Networks with Random Graphs118
Small-World Model . . . . . . . . . . . . . . . . . . . . . . . . 119
4.3.1 Properties of the Small-World Model . . . . . . . . . . 121
4.3.2 Modeling Real-World Networks with the Small-World
Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
Preferential Attachment Model . . . . . . . . . . . . . . . . . 125
4.4.1 Properties of the Preferential Attachment Model . . . 126
4.4.2 Modeling Real-World Networks with the Preferential Attachment Model . . . . . . . . . . . . . . . . . . 128
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
Bibliographic Notes . . . . . . . . . . . . . . . . . . . . . . . . 131
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132

Data Mining Essentials

5.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . .
5.1.1 Data Quality . . . . . . . . . . . . . . . . .
5.2 Data Preprocessing . . . . . . . . . . . . . . . . .
5.3 Data Mining Algorithms . . . . . . . . . . . . . .
5.4 Supervised Learning . . . . . . . . . . . . . . . .
5.4.1 Decision Tree Learning . . . . . . . . . . .
5.4.2 Naive Bayes Classifier . . . . . . . . . . .
5.4.3 Nearest Neighbor Classifier . . . . . . . .
5.4.4 Classification with Network Information
5.4.5 Regression . . . . . . . . . . . . . . . . . .
5.4.6 Supervised Learning Evaluation . . . . .
5.5 Unsupervised Learning . . . . . . . . . . . . . .
5.5.1 Clustering Algorithms . . . . . . . . . . .
5.5.2 Unsupervised Learning Evaluation . . .
5.6 Summary . . . . . . . . . . . . . . . . . . . . . . .
5.7 Bibliographic Notes . . . . . . . . . . . . . . . . .
5.8 Exercises . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

135
137
141
142
144
144
145
148
150
151
154
157
159
160
162
166
167
169

II
6

Communities and Interactions

173

Community Analysis
6.1 Community Detection . . . . . . . . . . . . . . . . .
6.1.1 Community Detection Algorithms . . . . . .
6.1.2 Member-Based Community Detection . . . .
6.1.3 Group-Based Community Detection . . . . .
6.2 Community Evolution . . . . . . . . . . . . . . . . .
6.2.1 How Networks Evolve . . . . . . . . . . . . .
6.2.2 Community Detection in Evolving Networks
6.3 Community Evaluation . . . . . . . . . . . . . . . . .
6.3.1 Evaluation with Ground Truth . . . . . . . .
6.3.2 Evaluation without Ground Truth . . . . . .
6.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . .
6.5 Bibliographic Notes . . . . . . . . . . . . . . . . . . .
6.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . .
Information Diffusion in Social Media
7.1 Herd Behavior . . . . . . . . . . . . . . . . . .
7.1.1 Bayesian Modeling of Herd Behavior
7.1.2 Intervention . . . . . . . . . . . . . . .
7.2 Information Cascades . . . . . . . . . . . . . .
7.2.1 Independent Cascade Model (ICM) .
7.2.2 Maximizing the Spread of Cascades .
7.2.3 Intervention . . . . . . . . . . . . . . .
7.3 Diffusion of Innovations . . . . . . . . . . . .
7.3.1 Innovation Characteristics . . . . . . .
7.3.2 Diffusion of Innovations Models . . .
7.3.3 Modeling Diffusion of Innovations . .
7.3.4 Intervention . . . . . . . . . . . . . . .
7.4 Epidemics . . . . . . . . . . . . . . . . . . . .
7.4.1 Definitions . . . . . . . . . . . . . . . .
7.4.2 SI Model . . . . . . . . . . . . . . . . .
7.4.3 SIR Model . . . . . . . . . . . . . . . .
7.4.4 SIS Model . . . . . . . . . . . . . . . .
7.4.5 SIRS Model . . . . . . . . . . . . . . .
7.4.6 Intervention . . . . . . . . . . . . . . .
7.5 Summary . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

175
179
179
181
188
197
198
201
204
204
209
211
212
214

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

217
220
222
224
225
226
229
231
232
232
233
236
239
240
242
243
245
246
248
249
251

7.6
7.7

III
8

Bibliographic Notes . . . . . . . . . . . . . . . . . . . . . . . . 252

Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254

Applications
Influence and Homophily
8.1 Measuring Assortativity . . . . . . . . . . . . . . . . . .
8.1.1 Measuring Assortativity for Nominal Attributes
8.1.2 Measuring Assortativity for Ordinal Attributes .
8.2 Influence . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.2.1 Measuring Influence . . . . . . . . . . . . . . . .
8.2.2 Modeling Influence . . . . . . . . . . . . . . . . .
8.3 Homophily . . . . . . . . . . . . . . . . . . . . . . . . . .
8.3.1 Measuring Homophily . . . . . . . . . . . . . . .
8.3.2 Modeling Homophily . . . . . . . . . . . . . . .
8.4 Distinguishing Influence and Homophily . . . . . . . .
8.4.1 Shuffle Test . . . . . . . . . . . . . . . . . . . . .
8.4.2 Edge-Reversal Test . . . . . . . . . . . . . . . . .
8.4.3 Randomization Test . . . . . . . . . . . . . . . .
8.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.6 Bibliographic Notes . . . . . . . . . . . . . . . . . . . . .
8.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . .

257
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

Recommendation in Social Media

9.1 Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.2 Classical Recommendation Algorithms . . . . . . . . . . . .
9.2.1 Content-Based Methods . . . . . . . . . . . . . . . . .
9.2.2 Collaborative Filtering (CF) . . . . . . . . . . . . . . .
9.2.3 Extending Individual Recommendation to Groups of
Individuals . . . . . . . . . . . . . . . . . . . . . . . .
9.3 Recommendation Using Social Context . . . . . . . . . . . .
9.3.1 Using Social Context Alone . . . . . . . . . . . . . . .
9.3.2 Extending Classical Methods with Social Context . .
9.3.3 Recommendation Constrained by Social Context . . .
9.4 Evaluating Recommendations . . . . . . . . . . . . . . . . . .
9.4.1 Evaluating Accuracy of Predictions . . . . . . . . . .
9.4.2 Evaluating Relevancy of Recommendations . . . . .

259
261
262
265
268
268
273
278
278
278
280
280
281
282
285
286
287
289
290
292
292
293
300
303
303
304
306
309
309
310

9.5
9.6
9.7

9.4.3 Evaluating Ranking of Recommendations

Summary . . . . . . . . . . . . . . . . . . . . . . .
Bibliographic Notes . . . . . . . . . . . . . . . . .
Exercises . . . . . . . . . . . . . . . . . . . . . . .

10 Behavior Analytics
10.1 Individual Behavior . . . . . . . . . . .
10.1.1 Individual Behavior Analysis .
10.1.2 Individual Behavior Modeling
10.1.3 Individual Behavior Prediction
10.2 Collective Behavior . . . . . . . . . . .
10.2.1 Collective Behavior Analysis .
10.2.2 Collective Behavior Modeling .
10.2.3 Collective Behavior Prediction
10.3 Summary . . . . . . . . . . . . . . . . .
10.4 Bibliographic Notes . . . . . . . . . . .
10.5 Exercises . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

311
314
315
316

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

319
319
321
326
327
332
332
337
338
340
341
343

Bibliography

345

Index

373

This chapter is from Social Media Mining: An Introduction.

By Reza Zafarani, Mohammad Ali Abbasi, and Huan Liu.
Cambridge University Press, 2014. Draft version: April 20, 2014.
Complete Draft and Slides Available at: http://dmml.asu.edu/smm

Preface

We live in an age of big data. With hundreds of millions of people spending

countless hours on social media to share, communicate, connect, interact,
and create user-generated data at an unprecedented rate, social media has
become one unique source of big data. This novel source of rich data
provides unparalleled opportunities and great potential for research and
development. Unfortunately, more data does not necessarily beget more
good, only more of the right (or relevant) data that enables us to glean
gems. Social media data differs from traditional data we are familiar with
in data mining. Thus, new computational methods are needed to mine
the data. Social media data is noisy, free-format, of varying length, and
multimedia. Furthermore, social relations among the entities, or social networks, form an inseparable part of social media data; hence, it is important
that social theories and research methods be employed with statistical and
data mining methods. It is therefore a propitious time for social media
mining.
Social media mining is a rapidly growing new field. It is an interdisciplinary field at the crossroad of disparate disciplines deeply rooted in
computer science and social sciences. There are an active community and
a large body of literature about social media. The fast-growing interests
and intensifying need to harness social media data require research and
the development of tools for finding insights from big social media data.
This book is one of the intellectual efforts to answer the novel challenges
of social media. It is designed to enable students, researchers, and practitioners to acquire fundamental concepts and algorithms for social media
mining.
9

Researchers in this emerging field are expected to have knowledge in

different areas, such as data mining, machine learning, text mining, social
network analysis, and information retrieval, and are often required to
consult research papers to learn the state of the art of social media mining.
To mitigate such a strenuous effort and help researchers get up to speed
in a convenient way, we take advantage of our teaching and research of
many years to survey, summarize, filter, categorize, and connect disparate
research findings and fundamental concepts of social media mining. This
book is our diligent attempt to provide an easy reference or entry point to
help researchers quickly acquire a wide range of essentials of social media
mining. Social media not only produces big user-generated data; it also
has a huge potential for social science research, business development, and
understanding human and group behavior. If you want to share a piece
of information or a site on social media, you would like to grab precious
attention from other equally eager users of social media; if you are curious
to know what is hidden or who is influential in the complex world of social
media, you might wonder how one can find this information in big and
messy social media; if you hope to serve your customers better in social
media, you certainly want to employ effective means to understand them
better. These are just some scenarios in which social media mining can
help. If one of these scenarios fits your need or you simply wish to learn
something interesting in this emerging field of social media mining, this
book is for you. We hope this book can be of benefit to you in accomplishing
your goals of dealing with big data of social media.

Book Website and Resources

The books website and further resources can be found at
http://dmml.asu.edu/smm
The website provides lecture slides, homework and exam problems, and
sample projects, as well as pointers to useful material and resources that
are publicly available and relevant to social media mining.

To the Instructors
The book is designed for a one-semester course for senior undergraduate
or graduate students. Though it is mainly written for students with a
10

background in computer science, readers with a basic understanding of

probability, statistics, and linear algebra will find it easily accessible. Some
chapters can be skipped or assigned as a homework assignment for reviewing purposes if students have knowledge of a chapter. For example,
if students have taken a data mining or machine learning course, they can
skip Chapter 5. When time is limited, Chapters 68 should be discussed in
depth, and Chapters 9 and 10 can be either discussed briefly or assigned
as part of reading material for course projects.

Reza Zafarani
Mohammad Ali Abbasi
Huan Liu
Tempe, AZ
August, 2013

11

Acknowledgments
In the past several years, enormous pioneering research has been performed by numerous researchers in the interdisciplinary fields of data
mining, social computing, social network analysis, network science, computer science, and the social sciences. We are truly dwarfed by the depth,
breadth, and extent of the literature, which not only made it possible for
us to complete a text on this emerging topic social media mining but also
made it a seemingly endless task. In the process, we have been fortunate
in drawing inspiration and obtaining great support and help from many
people to whom we are indebted.
We would like to express our tremendous gratitude to the current and
former members of the Data Mining and Machine Learning laboratory
at Arizona State University (ASU); in particular, Nitin Agrawal, Salem
Alelyani, Geoffrey Barbier, William Cole, Zhuo Feng, Magdiel GalanOliveras, Huiji Gao, Pritam Gundecha, Xia (Ben) Hu, Isaac Jones, Shamanth
Kumar, Fred Morstatter, Sai Thejasvee Moturu, Ashwin Rajadesingan,
Suhas Ranganath, Jiliang Tang, Lei Tang, Xufei Wang, and Zheng Zhao.
Without their impressive accomplishments and continuing strides in advancing research in data mining, machine learning, and social computing, this book would have not been possible. Their stimulating thoughts,
creative ideas, friendly aggressiveness, willingness to extend the research
frontier, and cool company during our struggling moments (Arizona could
be scorchingly hot in some months), directly and indirectly, offered us encouragement, drive, passion, and ideas, as well as critiques in the process
toward the completion of the book.
This book project stemmed from a course on social computing offered
in 2008 at ASU. It was a seminar course that enjoyed active participation by graduate students and bright undergraduates with intelligent and
provocative minds. Lively discussion and heated arguments were fixtures
of the seminar course. Since then, it has become a regular course, evolving
into a focused theme on social media mining. Teaching assistants, students,
and guest speakers in these annual courses were of significant help to us in
choosing topics to include, determining the depth and extent of each topic,
and offering feedback on lecture materials such as homework problems,
slides, course projects, and reading materials.
We would like to especially thank Denny Abraham Cheriyan, Nitin
Ahuja, Amy Baldwin, Sai Prasanna Baskaran, Gaurav Pandey, Prerna
12

Satija, Nitesh Kedia, Bernard Ma, Dhanyatha Manjunath, Girish Kumar

Reddy Marthala, Apurv Patki, Greggory Scherer, Nikhil Sunkesula Bhaskar,
Yajin Wang, and Ruozhou Yu for their detailed comments on the drafts of
this book. In addition, we owe our gratitude to Daria Bazzi and Michael
Meeder for their help in proofreading earlier versions of the book, Farzad
Zafarani for preparing the books website, Subbarao Kambhampati for
reading an earlier draft and offering encouragement, and Rebecca Goolsby
for continually challenging us in understanding social media and developing social computing tools for real-world applications of humanitarian
assistance and disaster relief.
The idea of having this book published by Cambridge University Press
began with a casual conversation with K. Selcuk Candan, a well-received
Cambridge author. It was an enjoyable and pleasant process working with
Cambridge University Press. We would like to thank Lauren Cowles, a
senior editor of Mathematics and Computer Sciences at Cambridge, for her
patience and kind support during the process, and the Cambridge team,
David Jou and Joshua Penney, as well as Adrian Pereira and his colleagues
at Aptara for their work on the production of the book.
Our research on social computing, data mining, and social network
analysis has been, in part, supported by the Office of Naval Research,
National Science Foundation, and Army Research Office.
We have truly enjoyed our collaboration in this arduous journey. We
will certainly miss our weekly meetings and many missed deadlines.

13

14

This chapter is from Social Media Mining: An Introduction.

By Reza Zafarani, Mohammad Ali Abbasi, and Huan Liu.
Cambridge University Press, 2014. Draft version: April 20, 2014.
Complete Draft and Slides Available at: http://dmml.asu.edu/smm

Chapter

Introduction
With the rise of social media, the web has become a vibrant and lively Social Media
realm in which billions of individuals all around the globe interact, share,
post, and conduct numerous daily activities. Information is collected,
curated, and published by citizen journalists and simultaneously shared Citizen
or consumed by thousands of individuals, who give spontaneous feed- Journalism
back. Social media enables us to be connected and interact with each other
anywhere and anytime allowing us to observe human behavior in an
unprecedented scale with a new lens. This social media lens provides us
with golden opportunities to understand individuals at scale and to mine
human behavioral patterns otherwise impossible. As a byproduct, by understanding individuals better, we can design better computing systems
tailored to individuals needs that will serve them and society better. This
new social media world has no geographical boundaries and incessantly
churns out oceans of data. As a result, we are facing an exacerbated problem of big data drowning in data, but thirsty for knowledge. Can data
mining come to the rescue?
Unfortunately, social media data is significantly different from the traditional data that we are familiar with in data mining. Apart from enormous
size, the mainly user-generated data is noisy and unstructured, with abundant social relations such as friendships and followers-followees. This
new type of data mandates new computational data analysis approaches
that can combine social theories with statistical and data mining methods. The pressing demand for new techniques ushers in and entails a new
interdisciplinary field social media mining.
15

1.1
Social Atom
Social Molecule

Social Media Mining

Data Scientist

What is Social Media Mining

Social media shatters the boundaries between the real world and the virtual
world. We can now integrate social theories with computational methods
to study how individuals (also known as social atoms) interact and how
communities (i.e., social molecules) form. The uniqueness of social media
data calls for novel data mining techniques that can effectively handle usergenerated content with rich social relations. The study and development
of these new techniques are under the purview of social media mining,
an emerging discipline under the umbrella of data mining. Social Media Mining is the process of representing, analyzing, and extracting actionable
patterns from social media data.
Social Media Mining, introduces basic concepts and principal algorithms
suitable for investigating massive social media data; it discusses theories and methodologies from different disciplines such as computer science, data mining, machine learning, social network analysis, network
science, sociology, ethnography, statistics, optimization, and mathematics.
It encompasses the tools to formally represent, measure, model, and mine
meaningful patterns from large-scale social media data.
Social media mining cultivates a new kind of data scientist who is well
versed in social and computational theories, specialized to analyze recalcitrant social media data, and skilled to help bridge the gap from what we
know (social and computational theories) to what we want to know about
the vast social media world with computational tools.

1.2

New Challenges for Mining

Social media mining is an emerging field where there are more problems
than ready solutions. Equipped with interdisciplinary concepts and theories, fundamental principles, and state-of-the-art algorithms, we can stand
on the shoulders of the giants and embark on solving challenging problems
and developing novel data mining techniques and scalable computational
algorithms. In general, social media can be considered a world of social
atoms (i.e., individuals), entities (e.g., content, sites, networks, etc.), and
interactions between individuals and entities. Social theories and social
norms govern the interactions between individuals and entities. For effective social media mining, we collect information about individuals and
16

entities, measure their interactions, and discover patterns to understand

human behavior.
Mining social media data is the task of mining user-generated content
with social relations. This data1 presents novel challenges encountered in
social media mining.
Big Data Paradox. Social media data is undoubtedly big. However, when Big Data
we zoom into individuals for whom, for example, we would like to make Paradox
relevant recommendations, we often have little data for each specific individual. We have to exploit the characteristics of social media and use its
multidimensional, multisource, and multisite data to aggregate information with sufficient statistics for effective mining.
Obtaining Sufficient Samples. One of the commonly used methods to Obtaining
collect data is via application programming interfaces (APIs) from social Sufficient
media sites. Only a limited amount of data can be obtained daily. With- Samples
out knowing the populations distribution, how can we know that our
samples are reliable representatives of the full data? Consequently, how
can we ensure that our findings obtained from social media mining are
any indication of true patterns that can benefit our research or business
development?
Noise Removal Fallacy. In classic data mining literature, a successful data Noise
mining exercise entails extensive data preprocessing and noise removal as Removal
garbage in and garbage out. By its nature, social media data can contain Fallacy
a large portion of noisy data. For this data, we notice two important observations: (1) blindly removing noise can worsen the problem stated in the
big data paradox because the removal can also eliminate valuable information, and (2) the definition of noise becomes complicated and relative
because it is dependent on our task at hand.
Evaluation Dilemma. A standard procedure of evaluating patterns in data Evaluation
mining is to have some kind of ground truth. For example, a dataset can be Dilemma
divided into training and test sets. Only the training data is used in learning, and the test data serves as ground truth for testing. However, ground
truth is often not available in social media mining. Evaluating patterns
from social media mining poses a seemingly insurmountable challenge.
On the other hand, without credible evaluation, how can we guarantee the
1

The data has a power-law distribution and more often than not, data is not independent and identically distributed (i.i.d.) as generally assumed in data mining.

17

validity of the patterns?

This book contains basic concepts and fundamental principles that will
help readers contemplate and design solutions to address these challenges
intrinsic to social media mining.

1.3

Book Overview and Readers Guide

This book consists of three parts. Part I, Essentials, outlines ways to represent social media data and provides an understanding of fundamental
elements of social media mining. Part II, Communities and Interactions,
discusses how communities can be found in social media and how interactions occur and information propagates in social media. Part III, Applications, offers some novel illustrative applications of social media mining.
Throughout the book, we use examples to explain how things work and to
deepen the understanding of abstract concepts and profound algorithms.
These examples show in a tangible way how theories are applied or ideas
are materialized in discovering meaningful patterns in social media data.
Consider an online social networking site with millions of members
in which members have the opportunity to befriend one another, send
messages to each other, and post content on the site. Facebook, LinkedIn,
and Twitter are exemplars of such sites. To make sense of data from these
sites, we resort to social media mining to answer corresponding questions.
In Part I: Essentials (Chapters 25), we learn to answer questions such as
the following:
1. Who are the most important people in a social network?
2. How do people befriend others?
3. How can we find interesting patterns in user-generated content?
These essentials come into play in Part II: Communities and Interactions
(Chapters 6 and 7) where we attempt to analyze how communities are
formed, how they evolve, and how the qualities of detected communities
are evaluated. We show ways in which information diffusion in social
media can be studied. We aim to answer general questions such as the
following:
1. How can we identify communities in a social network?
18

2. When someone posts an interesting article on a social network, how

far can the article be transmitted in that network?
In Part III: Application (Chapters 810), we exemplify social media
mining using real-world problems in dealing with social media: measuring influence, recommending in a social environment, and analyzing user
behavior. We aim to answer these questions:
1. How can we measure the influence of individuals in a social network?
2. How can we recommend content or friends to individuals online?
3. How can we analyze the behavior of individuals online?
To provide an overall picture of the book content, we created a dependency graph among chapters (Fig. 1.1) in which arrows suggest dependencies between chapters. Based on the dependency graph, therefore, a
reader can start with Chapter 2 (graph essentials), and it is recommended
that he or she read Chapters 5 (data mining essentials) and 8 (influence and
homophily) before Chapter 9 (recommendation in social media). We have
also color-coded chapter boxes that are of the same level of importance and
abstraction. The darkest chapters are the essentials of this book, and the
lightest boxes are those chapters that are more applied and have materials
that are built on the foundation of other chapters.

Who Should Read This Book?

A reader with a basic computer science background and knowledge of
data structures, search, and graph algorithms will find this book easily
accessible. Limited knowledge of linear algebra, calculus, probability, and
statistics will help readers understand technical details with ease. Having
a data mining or machine learning background is a plus, but not necessary.
The book is designed for senior undergraduate and graduate students.
It is organized in such a way that it can be taught in one semester to students
with a basic prior knowledge of statistics and linear algebra. It can also
be used for a graduate seminar course by focusing on more advanced
chapters with the supplement of detailed bibliographical notes. Moreover,
the book can be used as a reference book for researchers, practitioners, and
project managers of related fields who are interested in both learning the
19

Figure 1.1: Dependency between Book Chapters. Arrows show dependencies and colors represent book parts.
basics and tangible examples of this emerging field and understanding the
potentials and opportunities that social media mining can offer.

20

1.4

Summary

As defined by Kaplan and Haenlein [141], social media is the group of

internet-based applications that build on the ideological and technological foundations of Web 2.0, and that allow the creation and exchange
of user-generated content. There are many categories of social media
including, but not limited to, social networking (Facebook or LinkedIn),
microblogging (Twitter), photo sharing (Flickr, Photobucket, or Picasa),
news aggregation (Google reader, StumbleUpon, or Feedburner), video
sharing (YouTube, MetaCafe), livecasting (Ustream or Justin.TV), virtual
worlds (Kaneva), social gaming (World of Warcraft), social search (Google,
Bing, or Ask.com), and instant messaging (Google Talk, Skype, or Yahoo!
messenger).
The first social media site was introduced by Geocities in 1994, which
allowed users to create their own homepages. The first social networking
site, SixDegree.com, was introduced in 1997. Since then, many other social media sites have been introduced, each providing service to millions
of people. These individuals form a virtual world in which individuals
(social atoms), entities (content, sites, etc.) and interactions (between individuals, between entities, between individuals and entities) coexist. Social
norms and human behavior govern this virtual world. By understanding
these social norms and models of human behavior and combining them
with the observations and measurements of this virtual world, one can
systematically analyze and mine social media.
Social media mining is the process of representing, analyzing, and
extracting meaningful patterns from data in social media, resulting from
social interactions. It is an interdisciplinary field encompassing techniques
from computer science, data mining, machine learning, social network
analysis, network science, sociology, ethnography, statistics, optimization,
and mathematics. Social media mining faces grand challenges such as the
big data paradox, obtaining sufficient samples, the noise removal fallacy,
and evaluation dilemma.
Social media mining represents the virtual world of social media in a
computable way, measures it, and designs models that can help us understand its interactions. In addition, social media mining provides necessary
tools to mine this world for interesting patterns, analyze information diffusion, study influence and homophily, provide effective recommendations,
and analyze novel social behavior in social media.
21

1.5

Bibliographic Notes

For historical notes on social media sites and challenges in social media,
refer to [81, 173, 141, 150, 115]. Kaplan and Haenlein [141] provide a categorization of social media sites into collaborative projects, blogs, content
communities, social networking sites, virtual game worlds, and virtual
social worlds. Our definition of social media is a rather abstract one whose
elements are social atoms (individuals), entities, and interactions. A more
detailed abstraction can be found in the work of [149]. They consider the
seven building blocks of social media to be identity, conversation, sharing, presence, relationships, reputation, and groups. They argue that the
amount of attention that sites give to these building blocks makes them
different in nature. For instance, YouTube provides more functionality in
terms of groups than LinkedIn.
Social media mining brings together techniques from many disciplines.
General references that can accompany this book and help readers better
understand the material in this book can be found in data mining and web
mining [120, 280, 92, 174, 51], machine learning [40], and pattern recognition [75] texts, as well as network science and social network analysis
[78, 253, 212, 140, 28] textbooks. For relevant references on optimization
refer to [44, 219, 228, 207] and for algorithms to [61, 151]. For general
references on social research methods consult [36, 47]. Note that these are
generic references and more specific references are provided at the end of
each chapter. This book discusses nonmultimedia data in social media.
For multimedia data analysis refer to [49].
Recent developments in social media mining can be found in journal articles in IEEE Transactions on Knowledge and Data Engineering
(TKDE), ACM Transactions on Knowledge Discovery from Data (TKDD),
ACM Transactions on Intelligent Systems and Technology (TIST), Social
Network Analysis and Mining (SNAM), Knowledge and Information Systems (KAIS), ACM Transactions on the Web (TWEB), Data Mining and
Knowledge Discovery (DMKD), World Wide Web Journal, Social Networks, Internet Mathematics, IEEE Intelligent Systems, and SIGKDD Exploration. Conference papers can be found in proceedings of Knowledge
Discovery and Data Mining (KDD), World Wide Web (WWW), Association for Computational Linguistics (ACL), Conference on Information
and Knowledge Management (CIKM), International Conference on Data
Mining (ICDM), Internet Measuring Conference (IMC), International Con22

ference on Weblogs and Social Media (ICWSM), International Conference

on Web Engineering (ICWE), Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD), The European Conference on Machine
Learning and Principles and Practice of Knowledge Discovery in Data Basis (ECML/PKDD), Web Search and Data Mining (WSDM), International
Joint Conferences on Artificial Intelligence (IJCAI), Association for the
Advancement of Artificial Intelligence (AAAI), Recommender Systems
(RecSys), Computer-Human Interaction (CHI), SIAM International Conference on Data Mining (SDM), Hypertext (HT), and Social Computing
Behavioral-Cultural Modeling and Prediction (SBP) conferences.

23

Table 1.1: List of Websites

Amazon
Flickr
Facebook
BlogCatalog
MySpace Last.fm
LinkedIn
Reddit
Vimeo
StumbleUpon Yelp
YouTube

1.6

Twitter
Pandora
Del.icio.us
Meetup

Exercises

1. Discuss some methodologies that can address the grand challenges

of social media.
2. What are the key characteristics of social media that differentiate it
from other media? Please list at least two with a brief explanation.
3. What are the different types of social media? Name two, and provide
a definition and an example for each type.
4. (a) Visit the websites in Table 1.1 (or find similar ones) and identify
the types of activities that individuals can perform on each one.
(b) Similar to questions posed in Section 1.3, design two questions
that you find interesting to ask with respect to each site.
5. What marketing opportunities do you think exist in social media?
Can you outline an example of such an opportunity in Twitter?
6. How does behavior of individuals change across sites? What behaviors remain consistent and what behaviors likely change? What are
possible reasons behind these differences?
7. How does social media influence real-world behaviors of individuals? Identify a behavior that is due to the usage of, say, Twitter.
8. Outline how social media can help NGOs fulfill their missions better
in performing tasks such as humanitarian assistance and disaster
relief.
9. Identify at least three major side effects of information sharing on
social media.
24

10. Rumors spread rapidly on social media. Can you think of some
method to block the spread of rumors on social media?

25

26

Part I
Essentials

27

This chapter is from Social Media Mining: An Introduction.

By Reza Zafarani, Mohammad Ali Abbasi, and Huan Liu.
Cambridge University Press, 2014. Draft version: April 20, 2014.
Complete Draft and Slides Available at: http://dmml.asu.edu/smm

Chapter

Graph Essentials
We live in a connected world in which networks are intertwined with
our daily life. Networks of air and land transportation help us reach our
destinations; critical infrastructure networks that distribute water and electricity are essential for our society and economy to function; and networks
of communication help disseminate information at an unprecedented rate.
Finally, our social interactions form social networks of friends, family, and
colleagues. Social media attests to the growing body of these social networks in which individuals interact with one another through friendships,
email, blogposts, buying similar products, and many other mechanisms.
Social media mining aims to make sense of these individuals embedded
in networks. These connected networks can be conveniently represented
using graphs. As an example, consider a set of individuals on a social
networking site where we want to find the most influential individual.
Each individual can be represented using a node (circle) and two individuals
who know each other can be connected with an edge (line). In Figure 2.1,
we show a set of seven individuals and their friendships. Consider a
hypothetical social theory that states that the more individuals you know,
the more influential you are. This theory in our graph translates to the
individual with the maximum degree (the number of edges connected to
its corresponding node) being the most influential person. Therefore, in
this network Juan is the most influential individual because he knows
four others, which is more than anyone else. This simple scenario is
an instance of many problems that arise in social media, which can be
solved by modeling the problem as a graph. This chapter formally details
29

Figure 2.1: A Sample Graph. In this graph, individuals are represented

with nodes (circles), and individuals who know each other are connected
with edges (lines).
the essentials required for using graphs and the fundamental algorithms
required to explore these graphs in social media mining.

2.1

Graph Basics

In this section, we review some of the common notation used in graphs.

Any graph contains both a set of objects, called nodes, and the connections
between these nodes, called edges. Mathematically, a graph G is denoted
as pair G(V, E), where V represents the set of nodes and E represents the
set of edges. We formally define nodes and edges next.

2.1.1

Vertices and Actors

Nodes

All graphs have fundamental building blocks. One major part of any
graph is the set of nodes. In a graph representing friendship, these nodes
represent people, and any pair of connected people denotes the friendship
between them. Depending on the context, these nodes are called vertices
or actors. For example, in a web graph, nodes represent websites, and the
connections between nodes indicate web-links between them. In a social
setting, these nodes are called actors. The mathematical representation for
30

Figure 2.2: A Directed Graph and an Undirected Graph. Circles represent

nodes, and lines or arrows connecting nodes are edges.
a set of nodes is

V = {v1 , v2 , . . . , vn },

(2.1)

where V is the set of nodes and vi , 1 i n, is a single node. |V| = n is

called the size of the graph. In Figure 2.1, n = 7.

2.1.2

Edges

Another important element of any graph is the set of edges. Edges connect
nodes. In a social setting, where nodes represent social entities such as
people, edges indicate inter-node relationships and are therefore known
as relationships or (social) ties . The edge set is usually represented using E, Relationships
E = {e1 , e2 , . . . , em },

and Ties

(2.2)

where ei , 1 i m, is an edge and the size of the set is commonly shown

as m = |E|. In Figure 2.1, lines connecting nodes represent the edges, so in
this case, m = 8. Edges are also represented by their endpoints (nodes),
so e(v1 , v2 ) (or (v1 , v2 )) defines an edge e between nodes v1 and v2 . Edges
can have directions, meaning one node is connected to another, but not
vice versa. When edges are undirected, nodes are connected both ways.
Note that in Figure 2.2(b), edges e(v1 , v2 ) and e(v2 , v1 ) are the same edges,
because there is no direction in how nodes get connected. We call edges
in this graph undirected edges and this kind of graph an undirected graph.
Conversely, when edges have directions, e(v1 , v2 ) is not the same as e(v2 , v1 ).
31

Neighborhood

Graph 2.2(a) is a graph with directed edges; it is an example of a directed

graph. Directed edges are represented using arrows. In a directed graph,
an edge e(vi , v j ) is represented using an arrow that starts at vi and ends
at v j . Edges can start and end at the same node; these edges are called
loops or self-links and are represented as e(vi , vi ). For any node vi , in an
undirected graph, the set of nodes it is connected to via an edge is called
its neighborhood and is represented as N(vi ). In Figure 2.1, N(Jade) = {Jeff,
Juan}. In directed graphs, node vi has incoming neighbors Nin (vi ) (nodes
that connect to vi ) and outgoing neighbors Nout (vi ) (nodes that vi connects
to). In Figure 2.2(a), Nin (v2 ) = {v3 } and Nout (v2 ) = {v1 , v3 }.

2.1.3

Degree and Degree Distribution

The number of edges connected to one node is the degree of that node.
Degree of a node vi is often denoted using di . In the case of directed edges,
nodes have in-degrees (edges pointing toward the node) and out-degrees
(edges pointing away from the node). These values are presented using
din
and dout
, respectively. In social media, degree represents the number
i
i
of friends a given user has. For example, on Facebook, degree represents
the users number of friends, and on Twitter in-degree and out-degree
represent the number of followers and followees, respectively. In any
undirected graph, the summation of all node degrees is equal to twice the
number of edges.
Theorem 2.1. The summation of degrees in an undirected graph is twice the
number of edges,
X
di = 2|E|.
(2.3)
i

Proof. Any edge has two endpoints; therefore, when calculating the degrees di and d j for any connected nodes vi and v j , the edge between them
contributes 1 to both di and d j ; hence, if thePedge is removed,
P di and d j become di 1 and d j 1, and the summation k dk becomes k dk 2. Hence,
by removal of all m edges, the degree summation becomes smaller by 2m.
However, we know that when all edges are removed the degree summation
becomes zero; therefore, the degree summation is 2 m = 2|E|.

Lemma 2.1. In an undirected graph, there are an even number of nodes having
odd degree.
32

Proof. The result can be derived from the previous theorem directly because
the summation of degrees is even: 2|E|. Therefore, when nodes with even
degree are removed from this summation, the summation of nodes with
odd degree should also be even; hence, there must exist an even number
of nodes with odd degree.

Lemma 2.2. In any directed graph, the summation of in-degrees is equal to the
summation of out-degrees,
X
X
dout
=
din
(2.4)
j .
i
i

Proof. The proof is left as an exercise.

Degree Distribution
In very large graphs, distribution of node degrees (degree distribution) is an
important attribute to consider. The degree distribution plays an important
role in describing the network being studied. Any distribution can be
described by its members. In our case, these are the degrees of all nodes
in the graph. The degree distribution pd (or P(d), or P(dv = d)) gives the
probability that a randomly
P selected node v has degree d. Because pd is a
probability distribution d=0 pd = 1. In a graph with n nodes, pd is defined
as
nd
(2.5)
pd = ,
n
where nd is the number of nodes with degree d. An important, commonly
performed procedure is to plot a histogram of the degree distribution, in
which the x-axis represents the degree (d) and the y-axis represents either
(1) the number of nodes having that degree (nd ) or (2) the fraction of nodes
having that degree (pd ).
Example 2.1. For the graph provided in Figure 2.1, the degree distribution pd for
d = {1, 2, 3, 4} is
p1 = 17 , p2 = 74 , p3 = 17 , p4 = 17 .
(2.6)
Because we have four nodes have degree 2, and degrees 1, 3, and 4 are observed
once.
33

Figure 2.3: Facebook Degree Distribution for the US and Global Users.
There exist many users with few friends and a few users with many friends.
This is due to a power-law degree distribution.

Power-law
Distribution

Example 2.2. On social networking sites, friendship relationships can be represented by a large graph. In this graph, nodes represent individuals and edges
represent friendship relationships. We can compute the degrees and plot the degree
distribution using a graph where the x-axis is the degree and the y-axis is the fraction of nodes with that degree.1 The degree distribution plot for Facebook in May
2012 is shown in Figure 2.3. A general trend observable on social networking sites
is that there exist many users with few connections and there exist a handful of
users with very large numbers of friends. This is commonly called the power-law
degree distribution.
As previously discussed, any graph G can be represented as a pair
G(V, E), where V is the node set and E is the edge set. Since edges are
between nodes, we have
E V V.
(2.7)
Graphs can also have subgraphs. For any graph G(V, E), a graph G0 (V 0 , E0 )
1

This is similar to plotting the probability mass function for degrees.

34

Figure 2.4: A Graph and Its Corresponding Adjacency Matrix.

is a subgraph of G(V, E), if the following properties hold:
V 0 V,
E0 (V 0 V 0 ) E.

2.2

(2.8)
(2.9)

Graph Representation

We have demonstrated the visual representation of graphs. This representation, although clear to humans, cannot be used effectively by computers
or manipulated using mathematical tools. We therefore seek representations that can store the node and edge sets in a way that (1) does not lose
information, (2) can be manipulated easily by computers, and (3) can have
mathematical methods applied easily.
Adjacency Matrix
A simple way of representing graphs is to use an adjacency matrix (also
known as a sociomatrix). Figure 2.4 depicts an example of a graph and its Sociomatrix
corresponding adjacency matrix. A value of 1 in the adjacency matrix indicates a connection between nodes vi and v j , and a 0 denotes no connection
between the two nodes. When generalized, any real number can be used
to show the strength of connections between two nodes. The adjacency
matrix gives a natural mathematical representation for graphs. Note that
35

Table 2.1: Adjacency List

Node Connected To
v1
v2
v2
v1 ,v3 ,v4
v2 ,v4
v3
v4
v2 ,v3 ,v5 ,v6
v4
v5
v4
v6
in social networks, because of the relatively small number of interactions,
many cells remain zero. This creates a large sparse matrix. In numerical
analysis, a sparse matrix is one that is populated primarily with zeros.
In the adjacency matrix, diagonal entries represent self-links or loops.
Adjacency matrices can be commonly formalized as
(
1 if vi is connected to v j ,
Ai, j =
(2.10)
0 otherwise.
Adjacency List
In an adjacency list, every node is linked with a list of all the nodes that
are connected to it. The list is often sorted based on node order or some
other preference. For the graph shown in Figure 2.4, the corresponding
adjacency list is shown in Table 2.1.
Edge List
Another simple and common approach to storing large graphs is to save
all edges in the graph. This is known as the edge list representation. For the
graph shown in Figure 2.4, we have the following edge list representation:
(v1 , v2 )
(v2 , v3 )
(v2 , v4 )
(v3 , v4 )
(v4 , v5 )
(v4 , v6 )
36

In this representation, each element is an edge and is represented as

(vi , v j ), denoting that node vi is connected to node v j . Since social media
networks are sparse, both the adjacency list and edge list representations
save significant space. This is because many zeros need to be stored when
using adjacency matrices, but do not need to be stored in an adjacency or
an edge list.

2.3

Types of Graphs

In general, there are many basic types of graphs. In this section we discuss
several basic types of graphs.
Null Graph. A null graph is a graph where the node set is empty (there
are no nodes). Obviously, since there are no nodes, there are also no edges.
Formally,
G(V, E), V = E = .
(2.11)
Empty Graph. An empty or edgeless graph is one where the edge set is
empty:
G(V, E), E = .
(2.12)
Note that the node set can be non-empty. A null graph is an empty
graph but not vice versa.
Directed/Undirected/Mixed Graphs. Graphs that we have discussed thus
far rarely had directed edges. As mentioned, graphs that only have directed edges are called directed graphs and ones that only have undirected
ones are called undirected graphs. Mixed graphs have both directed and
undirected edges. In directed graphs, we can have two edges between i
and j (one from i to j and one from j to i), whereas in undirected graphs
only one edge can exist. As a result, the adjacency matrix for directed
graphs is not in general symmetric (i connected to j does not mean j is
connected to i, i.e., Ai, j , A j,i ), whereas the adjacency matrix for undirected
graphs is symmetric (A = AT ).
In social media, there are many directed and undirected networks. For
instance, Facebook is an undirected network in which if Jon is a friend
of Mary, then Mary is also a friend of Jon. Twitter is a directed network,
where follower relationships are not bidirectional. One direction is called
followers, and the other is denoted as following.
37

Simple Graphs and Multigraphs. In the example graphs that we have

provided thus far, only one edge could exist between any pair of nodes.
These graphs are denoted as simple graphs. Multigraphs are graphs where
multiple edges between two nodes can exist. The adjacency matrix for
multigraphs can include numbers larger than one, indicating multiple
edges between nodes. Multigraphs are frequently observed in social media
where two individuals can have different interactions with one another.
They can be friends and, at the same time, colleagues, group members,
or other relation. For each one of these relationships, a new edge can be
added between the individuals, creating a multigraph.
Weighted Graphs. A weighted graph is one in which edges are associated
with weights. For example, a graph could represent a map, where nodes
are cities and edges are routes between them. The weight associated
with each edge represents the distance between these cities. Formally, a
weighted graph can be represented as G(V, E, W), where W represents the
weights associated with each edge, |W| = |E|. For an adjacency matrix
representation, instead of 1/0 values, we can use the weight associated
with the edge. This saves space by combining E and W into one adjacency
matrix A, assuming that an edge exists between vi and v j if and only if
Wi j , 0. Depending on the context, this weight can also be represented by
wi j or w(i, j).

Signed Edges and

Signed Graphs

An example of a weighted graph is the web graph. A web graph is a way

of representing how internet sites are connected on the web. In general,
a web graph is a directed graph. Nodes represent sites and edge weights
represent number of links between sites. Two sites can have multiple links
pointing to each other, and individual sites can have loops (links pointing
to themselves).
A special case of a weighted graph is when we have binary weights
(0/1 or +/) on edges. These edges are commonly called signed edges,
and the weighted graph is called a signed graph. Signed edges can be
employed to represent contradictory behavior. For instance, one can use
signed edges to represent friends and foes. A positive (+) edge between
two nodes represents friendship, and a negative () edge denotes that the
endpoint nodes (individuals) are considered enemies. When edges are
directed, one endpoint considers the other endpoint a friend or a foe, but
not vice versa. When edges are undirected, endpoints are mutually friends
or foes. In another setting, a + edge can denote a higher social status, and a
38

Figure 2.5: A Signed Graph Example.

edge can represent lower social status. Social status is the rank assigned to
ones position in society. For instance, a school principal can be connected
via a directed + edge to a student of the school because, in the school
environment, the principal is considered to be of higher status. Figure 2.5
shows a signed graph consisting of nodes and the signed edges between
them.

2.4

Connectivity in Graphs

Connectivity defines how nodes are connected via a sequence of edges in a

graph. Before we define connectivity, some preliminary concepts need to
be detailed.
Adjacent Nodes and Incident Edges. Two nodes v1 and v2 in graph G(V, E)
are adjacent when v1 and v2 are connected via an edge:
v1 is adjacent to v2

e(v1 , v2 ) E.

(2.13)

Two edges e1 (a, b) and e2 (c, d) are incident when they share one endpoint
(i.e., are connected via a node):
e1 (a, b) is incident to e2 (c, d)
(a = c) (a = d) (b = c) (b = d).

(2.14)

Figure 2.6 depicts adjacent nodes and incident edges in a sample graph.
In a directed graph, two edges are incident if the ending of one is the
beginning of the other; that is, the edge directions must match for edges to
be incident.
39

Figure 2.6: Adjacent Nodes and Incident Edges. In this graph u and v, as
well as v and w, are adjacent nodes, and edges (u, v) and (v, w) are incident
edges.
Traversing an Edge. An edge in a graph can be traversed when one starts
at one of its end-nodes, moves along the edge, and stops at its other endnode. So, if an edge e(a, b) connects nodes a and b, then visiting e can start
at a and end at b. Alternatively, in an undirected graph we can start at b
and end the visit at a.

Open Walk and

Closed Walk

Random Walk

Walk, Path, Trail, Tour, and Cycle. A walk is a sequence of incident edges
traversed one after another. In other words, if in a walk one traverses
edges e1 (v1 , v2 ), e2 (v2 , v3 ), e3 (v3 , v4 ), . . . , en (vn , vn+1 ), we have v1 as the walks
starting node and vn+1 as the walks ending node. When a walk does
not end where it started (v1 , vn+1 ) then it is called an open walk. When
a walk returns to where it was started (v1 = vn+1 ), it is called a closed
walk. Similarly, a walk can be denoted as a sequence of nodes, v1 , v2 ,
v3 , . . . , vn . In this representation, the edges that are traversed are e1 (v1 , v2 ),
e2 (v2 , v3 ), . . . , en1 (vn1 , vn ). The length of a walk is the number of edges
traversed during the walk and in our case is n 1.
A trail is a walk where no edge is traversed more than once; therefore,
all walk edges are distinct. A closed trail (one that ends where it started)
is called a tour or circuit.
A walk where nodes and edges are distinct is called a path, and a closed
path is called a cycle. The length of a path or cycle is the number of edges
traversed in the path or cycle. In a directed graph, we have directed paths
because traversal of edges is only allowed in the direction of the edges. In
Figure 2.7, v4 , v3 , v6 , v4 , v2 is a walk; v4 , v3 is a path; v4 , v3 , v6 , v4 , v2 is a trail;
and v4 , v3 , v6 , v4 is both a tour and a cycle.
A graph has a Hamiltonian cycle if it has a cycle such that all the nodes
in the graph are visited. It has an Eulerian tour if all the edges are traversed
only once. Examples of a Hamiltonian cycle and an Eulerian tour are
provided in Figure 2.8.
One can perform a random walk on a weighted graph, where nodes
are visited randomly. The weight of an edge, in this case, defines the
40

Figure 2.7: Walk, Path, Trail, Tour, and Cycle. In this figure, v4 , v3 , v6 , v4 , v2
is a walk; v4 , v3 is a path; v4 , v3 , v6 , v4 , v2 is a trail; and v4 , v3 , v6 , v4 is both a
tour and a cycle.

Figure 2.8: Hamiltonian Cycle and Eulerian Tour. In a Hamiltonian cycle

we start at one node, visit all other nodes only once, and return to our start
node. In an Eulerian tour, we traverse all edges only once and return to
our start point. In an Eulerian tour, we can visit a single node multiple
times. In this figure, v1 , v5 , v3 , v1 , v2 , v4 , v6 , v2 , v3 , v4 , v5 , v6 , v1 is an Eulerian
tour.

41

Algorithm 2.1 Random Walk

Require: Initial Node v0 , Weighted Graph G(V, E, W), Steps t
1: return Random Walk P
2: state = 0;
3: vt = v0 ;
4: P = {v0 };
5: while state < t do
6:
state = state + 1;
7:
8:
9:
10:
11:
12:

select a random node v j adjacent to vt with probability wt,j ;

vt = v j ;
P = P {v j };
end while
Return P

probability of traversing it. For this to work correctly, we must make sure
that for all edges that start at vi we have
X

wi,x = 1, i, j wi,j 0.

(2.15)

The random walk procedure is outlined in Algorithm 2.1. The algorithm starts at a node v0 and visits its adjacent nodes based on the transition
probability (weight) assigned to edges connecting them. This procedure
is performed for t steps (provided to the algorithm); therefore, a walk of
length t is generated by the random walk.
Connectivity. A node vi is connected to node v j (or v j is reachable from
vi ) if it is adjacent to it or there exists a path from vi to v j . A graph is
connected if there exists a path between any pair of nodes in it. In a directed
graph, the graph is weakly connected if there exists a path between any
pair of nodes, without following the edge directions (i.e., directed edges
are replaced with undirected edges). The graph is strongly connected if
there exists a directed path (following edge directions) between any pair
of nodes. Figure 2.9 shows examples of connected, disconnected, weakly
connected, and strongly connected graphs.
Components. A component in an undirected graph is a subgraph such
that, there exists a path between every pair of nodes inside the component.
42

Figure 2.9: Connectivity in Graphs.

(a) A Graph with 3 Components

(b) A Graph with 3 Strongly Connected Components

Figure 2.10: Components in Undirected and Directed Graphs.

Figure 2.10(a) depicts an undirected graph with three components. A
component in a directed graph is strongly connected if, for every pair of
nodes v and u, there exists a directed path from v to u and one from u to
v. The component is weakly connected if replacing directed edges with
undirected edges results in a connected component. A graph with three
strongly connected components is shown in Figure 2.10(b).
Shortest Path. When a graph is connected, multiple paths can exist between any pair of nodes. Often, we are interested in the path that has the
shortest length. This path is called the shortest path. Applications for shortest paths include GPS routing, where users are interested in the shortest
path to their destination. In this chapter, we denote the length of the short43

est path between nodes vi and v j as li,j . The concept of the neighborhood
of a node vi can be generalized using shortest paths. An n-hop neighborhood
of node vi is the set of nodes that are within n hops distance from node vi .
That is, their shortest path to vi has length less than or equal to n.
Diameter. The diameter of a graph is defined as the length of the longest
shortest path between any pair of nodes in the graph. It is defined only for
connected graphs because the shortest path might not exist in disconnected
graphs. Formally, for a graph G, the diameter is defined as
diameterG =

2.5

max li,j .

(vi ,v j )VV

(2.16)

Special Graphs

Using general concepts defined thus far, many special graphs can be defined. These special graphs can be used to model different problems. We
review some well-known special graphs and their properties in this section.

2.5.1

Trees and Forests

Trees are special cases of undirected graphs. A tree is a graph structure that
has no cycle in it. In a tree, there is exactly one path between any pair of
nodes. A graph consisting of set of disconnected trees is called a forest. A
forest is shown in Figure 2.11.

Figure 2.11: A Forest Containing Three Trees.

44

Figure 2.12: Minimum Spanning Tree. Nodes represent cities and values
assigned to edges represent geographical distance between cities. Highlighted edges are roads that are built in a way that minimizes their total
length.
In a tree with |V| nodes, we have |E| = |V| 1 edges. This can be proved
by contradiction (see Exercises).

2.5.2

Special Subgraphs

Some subgraphs are frequently used because of their properties. Two such
subgraphs are discussed here.
Spanning Tree. For any connected graph, the spanning tree is a subgraph and a tree that includes all the nodes of the graph. Obviously,
when the original graph is not a tree, then its spanning tree includes
all the nodes, but not all the edges. There may exist multiple spanning trees for a graph. For a weighted graph and one of its spanning
trees, the weight of that spanning tree is the summation of the edge
weights in the tree. Among the many spanning trees found for a
weighted graph, the one with the minimum weight is called the minimum spanning tree (MST) .
For example, consider a set of cities, where roads need to be built to
connect them. We know the distance between each pair of cities. We
45

Figure 2.13: Steiner tree for V 0 = {v2 , v4 , v7 }.

can represent each city with a node and the distance between these
nodes using an edge between them labeled with the distance. This
graph-based view is shown in Figure 2.12. In this graph, nodes v1 ,
v2 , . . . , v9 represent cities, and the values attached to edges represent
the distance between them. Note that edges only represent distances
(potential roads!), and roads may not exist between these cities. Due
to construction costs, the government needs to minimize the total
mileage of roads built and, at the same time, needs to guarantee that
there is a path (i.e., a set of roads) that connects every two cities.
The minimum spanning tree is a solution to this problem. The edges
in the MST represent roads that need to be built to connect all of
the cities at the minimum length possible. Figure 2.2 highlights the
minimum spanning tree.

Terminal Nodes

Steiner Tree. The Steiner Tree problem is similar to the minimum

spanning tree problem. Given a weighted graph G(V, E, W) and a
subset of nodes V 0 V (terminal nodes) , the Steiner tree problem
aims to find a tree such that it spans all the V 0 nodes and the weight
of the tree is minimized. Note that the problem is different from the
MST problem because we do not need to span all nodes of the graph
V, but only a subset of the nodes V 0 . A Steiner tree example is shown
in Figure 2.13. In this example, V 0 = {v2 , v4 , v7 }.
46

Figure 2.14: Complete Graphs Ki for 1 i 4.

(a) Planar Graph

(b) Non-planar Graph

Figure 2.15: Planar and Nonplanar Graphs.

2.5.3

Complete Graphs

A complete graph is a graph where for a set of nodes V, all possible edges
exist in the graph. In other words, all pairs of nodes are connected with an
edge. Hence,
!
|V|
|E| =
.
(2.17)
2
Complete graphs with n nodes are often denoted as Kn . K1 , K2 , K3 , and
K4 are shown in Figure 2.14.

2.5.4

Planar Graphs

A graph that can be drawn in such a way that no two edges cross each
other (other than the endpoints) is called planar. A graph that is not planar
is denoted as nonplanar. Figure 2.15 shows an example of a planar graph
and a nonplanar graph.
47

(a) Bipartite Graph

(b) Affiliation Network

Figure 2.16: Bipartite Graphs and Affiliation Networks.

2.5.5

Bipartite Graphs

A bipartite graph G(V, E) is a graph where the node set can be partitioned
into two sets such that, for all edges, one endpoint is in one set and the
other endpoint is in the other set. In other words, edges connect nodes in
these two sets, but there exist no edges between nodes that belong to the
same set. Formally,
V = VL VR ,
VL VR = ,
E VL VR .

(2.18)
(2.19)
(2.20)

Figure 2.16(a) shows a sample bipartite graph. In this figure, VL =

{v1 , v2 } and VR = {v3 , v4 , v5 }.
In social media, affiliation networks are well-known examples of bipartite
graphs. In these networks, nodes in one part (VL or VR ) represent individuals, and nodes in the other part represent affiliations. If an individual is
associated with an affiliation, an edge connects the corresponding nodes.
A sample affiliation network is shown in Figure 2.16(b).

2.5.6

Regular Graphs

A regular graph is one in which all nodes have the same degree. A regular
graph where all nodes have degree 2 is called a 2-regular graph. More
generally, a graph where all nodes have degree k is called a k-regular graph.
48

Figure 2.17: Regular Graph with k = 3.

Regular graphs can be connected or disconnected. Complete graphs are
examples of regular graphs, where all n nodes have degree n 1 (i.e.,
k = n 1). Cycles are also regular graphs, where k = 2. Another example
for k = 3 is shown in Figure 2.17.

2.5.7

Bridges

Consider a graph with several connected components. Edges in this graph

whose removal will increase the number of connected components are
called bridges. As the name suggests, these edges act as bridges between
connected components. The removal of these edges results in the disconnection of formerly connected components and hence an increase in the
number of components. An example graph and all its bridges are depicted
in Figure 2.18.

2.6

Graph Algorithms

In this section, we review some well-known algorithms for graphs, although they are only a small fraction of the plethora of algorithms related
to graphs.

2.6.1

Graph/Tree Traversal

Among the most useful algorithms for graphs are the traversal algorithms
for graphs, and special subgraphs, such as trees. Consider a social media
49

Figure 2.18: Bridges. Highlighted edges represent bridges; their removal

will increase the number of components in the graph.
site that has many users, and we are interested in surveying it and computing the average age of its users. The usual technique is to start from
one user and employ some traversal technique to browse his friends and
then these friends friends and so on. The traversal technique guarantees
that (1) all users are visited and (2) no user is visited more than once.
In this section, we discuss two traversal algorithms: depth-first search
(DFS) and breadth-first search (BFS).
Depth-First Search (DFS)
Depth-first search (DFS) starts from a node vi , selects one of its neighbors
v j N(vi ), and performs DFS on v j before visiting other neighbors in
N(vi ). In other words, DFS explores as deep as possible in the graph using
one neighbor before backtracking to other neighbors. Consider a node vi
that has neighbors v j and vk ; that is, v j , vk N(vi ). Let v j(1) N(v j ) and
v j(2) N(v j ) denote neighbors of v j such that vi , v j(1) , v j(2) . Then for a
depth-first search starting at vi , that visits v j next, nodes v j(1) and v j(2) are
visited before visiting vk . In other words, a deeper node v j(1) is preferred
to a neighbor vk that is closer to vi . Depth-first search can be used both for
trees and graphs, but is better visualized using trees. The DFS execution
on a tree is shown in Figure 2.19(a).
The DFS algorithm is provided in Algorithm 2.2. The algorithm uses a
50

Algorithm 2.2 Depth-First Search (DFS)

Require: Initial node v, graph/tree G(V, E), stack S
1: return An ordering on how nodes in G are visited
2: Push v into S;
3: visitOrder = 0;
4: while S not empty do
5:
node = pop from S;
6:
if node not visited then
7:
visitOrder = visitOrder +1;
8:
Mark node as visited with order visitOrder; //or print node
9:
Push all neighbors/children of node into S;
10:
end if
11: end while
12: Return all nodes with their visit order.

stack structure to visit nonvisited nodes in a depth-first fashion.

Breadth-First Search (BFS)

Breadth-first search (BFS) starts from a node, visits all its immediate neighbors first, and then moves to the second level by traversing their neighbors.
Like DFS, the algorithm can be used both for trees and graphs and is provided in Algorithm 2.3.
The algorithm uses a queue data structure to achieve its goal of breadth
traversal. Its execution on a tree is shown in Figure 2.19(b).
In social media, we can use BFS or DFS to traverse a social network:
the algorithm choice depends on which nodes we are interested in visiting
first. In social media, immediate neighbors (i.e., friends) are often more
important to visit first; therefore, it is more common to use breadth-first
search.
51

(a) Depth-First Search (DFS)

(b) Breadth-First Search (BFS)

Figure 2.19: Graph Traversal Example.

52

Algorithm 2.3 Breadth-First Search (BFS)

Require: Initial node v, graph/tree G(V, E), queue Q
1: return An ordering on how nodes are visited
2: Enqueue v into queue Q;
3: visitOrder = 0;
4: while Q not empty do
5:
node = dequeue from Q;
6:
if node not visited then
7:
visitOrder = visitOrder +1;
8:
Mark node as visited with order visitOrder; //or print node
9:
Enqueue all neighbors/children of node into Q;
10:
end if
11: end while

2.6.2

Shortest Path Algorithms

In many scenarios, we require algorithms that can compute the shortest

path between two nodes in a graph. For instance, in the case of navigation,
we have a weighted network of cities connected via roads, and we are
interested in computing the shortest path from a source city to a destination
city. In social media mining, we might be interested in determining how
tightly connected a social network is by measuring its diameter. The
diameter can be calculated by finding the longest shortest path between
any two nodes in the graph.
Dijkstras Algorithm
A well-established shortest path algorithm was developed in 1959 by Edsger Dijkstra. The algorithm is designed for weighted graphs with nonnegative edges. The algorithm finds the shortest paths that start from a
starting node s to all other nodes and the lengths of those paths.
The Dijkstras algorithm is provided in Algorithm 2.4. As mentioned,
the goal is to find the shortest paths and their lengths from a source node
s to all other nodes in the graph. The distance array (Line 3) keeps track of
the shortest path distance from s to other nodes. The algorithm starts by
assuming that there is a shortest path of infinite length to any node, except
s, and will update these distances as soon as a better distance (shorter path)
is observed. The steps of the algorithm are as follows:
53

Algorithm 2.4 Dijkstras Shortest Path Algorithm

Require: Start node s, weighted graph/tree G(V, E, W)
1: return Shortest paths and distances from s to all other nodes.
2: for v V do
3:
distance[v] = ;
4:
predecessor[v] = 1;
5: end for
6: distance[s] = 0;
7: unvisited = V;
8: while unvisited , do
9:
smallest = arg minvunvisited distance(v);
10:
if distance(smallest)== then
11:
break;
12:
end if
13:
unvisited = unvisited \ {smallest};
14:
currentDistance = distance(smallest);
15:
for adjacent node to smallest: neighbor unvisited do
16:
newPath = currentDistance+w(smallest, neighbor);
17:
if newPath < distance(neighbor) then
18:
distance(neighbor)=newPath;
19:
predecessor(neighbor)=smallest;
20:
end if
21:
end for
22: end while
23: Return distance[] and predecessor[] arrays

1. All nodes are initially unvisited. From the unvisited set of nodes,
the one that has the minimum shortest path length is selected. We
denote this node as smallest in the algorithm.
2. For this node, we check all its neighbors that are still unvisited.
For each unvisited neighbor, we check if its current distance can be
improved by considering the shortest path that goes through smallest. This can be performed by comparing its current shortest path
length (distance(neighbor)) to the path length that goes through smallest (distance(smallest)+w(smallest, neighbor)). This condition is checked
in Line 17.
54

Figure 2.20: Dijkstras Algorithm Execution Example. The shortest path

between node s and t is calculated. The values inside nodes at each step
show the best shortest path distance computed up to that step. An arrow
denotes the node being analyzed.
3. If the current shortest path can be improved, the path and its length
are updated. The paths are saved based on predecessors in the path
sequence. Since, for every node, we only need the predecessor to
reconstruct a path recursively, the predecessor array keeps track of
this.
4. A node is marked as visited after all its neighbors are processed and
is no longer changed in terms of (1) the shortest path that ends with
it and (2) its shortest path length.
To further clarify the process, an example of the Dijkstras algorithm is
provided.
Example 2.3. Figure 2.20 provides an example of the Dijkstras shortest path
algorithm. We are interested in finding the shortest path between s and t. The
shortest path is highlighted using dashed lines. In practice, shortest paths are
saved using the predecessor array.
55

Algorithm 2.5 Prims Algorithm

Require: Connected weighted graph G(V, E, W)
1: return Spanning tree T(Vs , Es )
2: Vs = {a random node from V};
3: Es = {};
4: while V , Vs do
5:
e(u, v) = argmin(u,v),uVs ,vVVs w(u, v)
6:
Vs = Vs {v};
7:
Es = Es e(u, v);
8: end while
9: Return tree T(Vs , Es ) as the minimum spanning tree;

2.6.3

Minimum Spanning Trees

A spanning tree of a connected undirected graph is a tree that includes

all the nodes in the original graph and a subset of its edges. Spanning
trees play important roles in many real-life applications. A cable company
that wants to lay wires wishes not only to cover all areas (nodes) but
also minimize the cost of wiring (summation of edges). In social media
mining, consider a network of individuals who need to be provided with
a piece of information. The information spreads via friends, and there is a
cost associated with spreading information among every two nodes. The
minimum spanning tree of this network will provide the minimum cost
required to inform all individuals in this network.
There exist a variety of algorithms for finding minimum spanning trees.
A famous algorithm for finding MSTs in a weighted graph is Prims algorithm [233]. Interested readers can refer to the bibliographic notes for
further references.
Prims Algorithm
Prims algorithm is provided in Algorithm 2.5. It starts by selecting a
random node and adding it to the spanning tree. It then grows the spanning
tree by selecting edges that have one endpoint in the existing spanning tree
and one endpoint among the nodes that are not selected yet. Among the
possible edges, the one with the minimum weight is added to the set (along
56

Figure 2.21: Prims Algorithm Execution Example.

with its endpoint). This process is iterated until the graph is fully spanned.
An example of Prims algorithm is provided in Figure 2.21.

2.6.4

Network Flow Algorithms

Consider a network of pipes that connect an infinite water source to a water

sink. In these networks, given the capacity of these pipes, an interesting
question is, What is the maximum flow that can be sent from the source to
the sink?
Network flow algorithms aim to answer this question. This type of
question arises in many different fields. At first glance, these problems do
not seem to be related to network flow algorithms, but there are strong
parallels. For instance, in social media sites where users have daily limits
(the capacity, here) of sending messages (the flow) to others, what is the
maximum number of messages the network should be prepared to handle
at any time?
Before we delve into the details, let us formally define a flow network.
57

Flow Network
A flow network G(V, E, C)2 is a directed weighted graph, where we have
the following:
e(u, v) E, c(u, v) 0 defines the edge capacity.
When (u, v) E, (v, u) < E (opposite flow is impossible).
Source and Sink

s defines the source node and t defines the sink node. An infinite
supply of flow is connected to the source.
A sample flow network, along with its capacities, is shown in Figure 2.22.

Figure 2.22: A Sample Flow Network.

Flow
Given edges with certain capacities, we can fill these edges with the flow
up to their capacities. This is known as the capacity constraint. Furthermore,
we should guarantee that the flow that enters any node other than source
s and sink t is equal to the flow that exits it so that no flow is lost (flow
conservation constraint). Formally,
(u, v) E, f (u, v) 0 defines the flow passing through that edge.
(u, v) E, 0 f (u, v) c(u, v) (capacity constraint).
P
P
v V, v < {s, t}, k:(k,v)E f (k, v) = l:(v,l)E f (v, l) (flow conservation
constraint).
2

Instead of W in weighted networks, C is used to clearly represent capacities.

58

Commonly, to visualize an edge with capacity c and flow f , we use

the notation f /c. A sample flow network with its flows and capacities is
shown in Figure 2.23.

Figure 2.23: A Sample Flow Network with Flows and Capacities.

Flow Quantity
The flow quantity (or value of the flow) in any network is the amount
of outgoing flow from the source minus the incoming flow to the source.
Alternatively, one can compute this value by subtracting the outgoing flow
from the sink from its incoming value:
X
X
X
X
flow =
f (s, v)
f (v, s) =
f (v, t)
f (t, v).
(2.21)
v

Example 2.4. The flow quantity for the example in Figure 2.23 is 19:
X
X
flow =
f (s, v)
f (v, s) = (11 + 8) 0 = 19.
v

(2.22)

Our goal is to find the flow assignments to each edge with the maximum
flow quantity. This can be achieved by a maximum flow algorithm. A wellestablished one is the Ford-Fulkerson algorithm [90].
Ford-Fulkerson Algorithm
The intuition behind this algorithm is as follows: Find a path from source
to sink such that there is unused capacity for all edges in the path. Use that
capacity (the minimum capacity unused among all edges on the path) to
increase the flow. Iterate until no other path is available.
59

Before we formalize this, let us define some concepts.

Given a flow in network G(V, E, C), we define another network GR (V, ER , CR ),
called the residual network. This network defines how much capacity remains in the original network. The residual network has an edge between
nodes u and v if and only if either (u, v) or (v, u) exists in the original graph.
If one of these two exists in the original network, we would have two edges
in the residual network: one from (u, v) and one from (v, u). The intuition
is that when there is no flow going through an edge in the original network, a flow of as much as the capacity of the edge remains in the residual.
However, in the residual network, one has the ability to send flow in the
opposite direction to cancel some amount of flow in the original network.
The residual capacity cR (u, v) for any edge (u, v) in the residual graph is
(
c(u, v) f (u, v) if (u, v) E
cR (u, v) =
(2.23)
f (v, u)
if (u, v) < E
A flow network example and its resulted residual network are shown in
Figure 2.24. In the residual network, edges that have zero residual capacity
are not shown.
Augmentation and Augmenting Paths

Weak link

In the residual graph, when edges are in the same direction as the original
graph, their capacity shows how much more flow can be pushed along
that edge in the original graph. When edges are in the opposite direction,
their capacities show how much flow can be pushed back on the original
graph edge. So, by finding a flow in the residual, we can augment the flow
in the original graph. Any simple path from s to t in the residual graph
is an augmenting path. Since all capacities in the residual are positive,
these paths can augment flows in the original, thus increasing the flow.
The amount of flow that can be pushed along this path is equal to the
minimum capacity along the path, since the edge with that capacity limits
the amount of flow being pushed.3 Given flow f (u, v) in the original graph
and flow fR (u, v) and fR (v, u) in the residual graph, we can augment the
flow as follows:
faugmented (u, v) = f (u, v) + fR (u, v) fR (v, u).
3

This edge is often called the weak link.

60

(2.24)

(a) Flow Network

(b) Residual Network

Figure 2.24: A Flow Network and Its Residual.

Example 2.5. Consider the graph in Figure 2.24(b) and the augmenting path
s, v1 , v3 , v4 , v2 , t. It has a minimum capacity of 1 along the path, so the flow
quantity will be 1. We can augment the original graph with this path. The new
flow graph and its corresponding residual graph are shown in Figure 2.25. In the
new residual, no more augmenting paths can be found.
The Ford-Fulkerson algorithm will find the maximum flow in a network, but we skip the proof of optimality. Interested readers can refer to
the bibliographic notes for proof of optimality and further information.
The algorithm is provided in Algorithm 2.6.
The algorithm searches for augmenting paths, if possible, in the residual
and augments flows in the original flow network. Path finding can be
achieved by any graph traversal algorithm, such as BFS.

2.6.5

Maximum Bipartite Matching

Suppose we are trying to solve the following problem in social media:

61

(a) Flow Graph

(b) Residual Graph

Figure 2.25: Augmenting Flow Graph.

Given n products and m users such that some users are only interested in certain
products, find the maximum number of products that can be bought by users.

Matching

The problem is graphically depicted in Figure 2.26. The nodes on the

left represent products and the nodes on the right represent users. Edges
represent the interest of users in products. Highlighted edges demonstrate
a matching, where products are matched with users. The figure on the left
depicts a matching and the figure on the right depicts a maximum matching,
where no more edges can be added to increase the size of the matching.
This problem can be reformulated as a bipartite graph problem. Given
a bipartite graph, where VL and VR represent the left and right node sets
(V = VL VR ), and E represents the edges, we define a matching M, M E,
such that each node in V appears in at most one edge in M. In other words,
either the node is matched (appears in an edge e M) or not. A maximum
bipartite matching MMax is a matching such that, for any other matching M0
in the graph, |MMax | |M0 |.
Here we solve the maximum bipartite matching problem using the previously discussed Ford-Fulkerson maximum flow technique. The problem
62

Algorithm 2.6 Ford-Fulkerson Algorithm

Require: Connected weighted graph G(V, E, W), Source s, Sink t
1: return A Maximum flow graph
2: (u, v) E, f (u, v) = 0
3: while there exists an augmenting path p in the residual graph GR do
4:
Augment flows by p
5: end while
6: Return flow value and flow graph;

Figure 2.26: Maximum Bipartite Matching.

can be easily solved by creating a flow graph G(V 0 , E0 , C) from our bipartite
graph G(V, E), as follows:
Set V 0 = V {s} {t}.
Connect all nodes in VL to s and all nodes in VR to t,
E0 = E {(s, v)|v VL } {(v, t)|v VR }.

(2.25)

Set c(u, v) = 1, (u, v) E0 .

This procedure is graphically shown in Figure 2.27. By solving the max
flow for this flow graph, the maximum matching is obtained, since the
63

Figure 2.27: Maximum Bipartite Matching Using Max Flow.

maximum number of edges need to be used between VL and VR for the
flow to become maximum.4

2.6.6
Cut-Edges

Bridge Detection

As discussed in Section 2.5.7, bridges or cut edges are edges whose removal
makes formerly connected components disconnected. Here we list a simple algorithm for detecting bridges. This algorithm is computationally
expensive, but quite intuitive. More efficient algorithms have been described for the same task.
Since we know that, by removing bridges, formerly connected components become disconnected, one simple algorithm is to remove edges one
by one and test if the connected components become disconnected. This
algorithm is outlined in Algorithm 2.7.
The disconnectedness of a component whose edge e(u, v) is removed
can be analyzed by means of any graph traversal algorithm (e.g., BFS or
4

The proof is omitted here and is a direct result from the minimum-cut/maximum flow
theorem not discussed in this chapter.

64

Algorithm 2.7 Bridge Detection Algorithm

Require: Connected graph G(V, E)
1: return Bridge Edges
2: bridgeSet = {}
3: for e(u, v) E do
4:
G0 = Remove e from G
5:
Disconnected = False;
6:
if BFS in G0 starting at u does not visit v then
7:
Disconnected = True;
8:
end if
9:
if Disconnected then
10:
bridgeSet = bridgeSet {e}
11:
end if
12: end for
13: Return bridgeSet
DFS). Starting at node u, we traverse the graph using BFS and, if node v
cannot be visited (Line 6), the component has been disconnected and edge
e is a bridge (Line 10).

65

2.7

Summary

This chapter covered the fundamentals of graphs, starting with a presentation of the fundamental building blocks required for graphs: first nodes
and edges, and then properties of graphs such as degree and degree distribution. Any graph must be represented using some data structure for
computability. This chapter covered three well-established techniques: adjacency matrix, adjacency list, and edge list. Due to the sparsity of social
networks, both adjacency list and edge list are more efficient and save significant space when compared to adjacency matrix. We then described various types of graphs: null and empty graphs, directed/undirected/mixed
graphs, simple/multigraphs, and weighted graphs. Signed graphs are examples of weighted graphs that can be used to represent contradictory
behavior.
We discussed connectivity in graphs and concepts such as paths, walks,
trails, tours, and cycles. Components are connected subgraphs. We discussed strongly and weakly connected components. Given the connectivity of a graph, one is able to compute the shortest paths between different
nodes. The longest shortest path in the graph is known as the diameter.
Special graphs can be formed based on the way nodes are connected and
the degree distributions. In complete graphs, all nodes are connected to
all other nodes, and in regular graphs, all nodes have an equal degree. A
tree is a graph with no cycle. We discussed two special trees: the spanning tree and the Steiner tree. Bipartite graphs can be partitioned into two
sets of nodes, with edges between these sets and no edges inside these
sets. Affiliation networks are examples of bipartite graphs. Bridges are
single-point-of-failure edges that can make previously connected graphs
disconnected.
In the section on graph algorithms, we covered a variety of useful techniques. Traversal algorithms provide an ordering of the nodes of a graph.
These algorithms are particularly useful in checking whether a graph is
connected or in generating paths. Shortest path algorithms find paths
with the shortest length between a pair of nodes; Dijkstras algorithm is
an example. Spanning tree algorithms provide subgraphs that span all the
nodes and select edges that sum up to a minimum value; Prims algorithm
is an example. The Ford-Fulkerson algorithm, is one of the maximum flow
algorithms. It finds the maximum flow in a weighted capacity graph. Maximum bipartite matching is an application of maximum flow that solves
66

a bipartite matching problem. Finally, we provided a simple solution for

bridge detection.

67

2.8

Bibliographic Notes

The algorithms detailed in this chapter are from three well-known fields:
graph theory, network science, and social network analysis. Interested
readers can get better insight regarding the topics in this chapter by referring to general references in graph theory [43, 301, 71], algorithms and
algorithm design [151, 61], network science [212], and social network analysis [294].
Other algorithms not discussed in this chapter include graph coloring
[139], (quasi) clique detection [2], graph isomorphism [191], topological
sort algorithms [61], and the traveling salesman problem (TSP) [61], among
others. In graph coloring, one aims to color elements of the graph such as
nodes and edges such that certain constraints are satisfied. For instance,
in node coloring the goal is to color nodes such that adjacent nodes have
different colors. Cliques are complete subgraphs. Unfortunately, solving
many problems related to cliques, such as finding a clique that has more
that a given number of nodes, is NP-complete. In clique detection, the
goal is to solve similar clique problems efficiently or provide approximate
solutions. In graph isomorphism, given two graphs G and G0 , our goal is
to find a mapping f from nodes of G to G0 such that for any two nodes
of G that are connected, their mapped nodes in G0 are connected as well.
In topological sort algorithms, a linear ordering of nodes is found in a
directed graph such that for any directed edge (u, v) in the graph, node u
comes before node v in the ordering. In the traveling salesman problem
(TSP), we are provided cities and pairwise distances between them. In
graph theory terms, we are given a weighted graph where nodes represent
cities and edge weights represent distances between cities. The problem is
to find the shortest walk that visits all cities and returns to the origin city.
Other noteworthy shortest path algorithms such as the A [122], the
Bellman-Ford [32], and all-pair shortest path algorithms such as FloydWarshalls [89] are employed extensively in other literature.
In spanning tree computation, the Kruskal Algorithm [156] or Boruvka
[204] are also well-established algorithms.
General references for flow algorithms, other algorithms not discussed
in this chapter such as the Push-Relabel algorithm, and their optimality
can be found in [61, 7].

68

2.9

Exercises

Graph Basics
1. Given a directed graph G(V, E) and its adjacency matrix A, we propose
two methods to make G undirected,
A0i j = min(1, Ai j + A ji ),
A0i j

(2.26)

= Ai j A ji ,

(2.27)

where A0i, j is the (i, j) entry of the undirected adjacency matrix. Discuss the advantages and disadvantages of each method.

Graph Representation
2. Is it possible to have the following degrees in a graph with 7 nodes?
{4, 4, 4, 3, 5, 7, 2}.

(2.28)

3. Given the following adjacency matrix, compute the adjacency list

and the edge list.

A =

0
1
1
0
0
0
0
0
0

1
0
1
0
0
0
0
0
0

1
1
0
1
1
0
0
0
0

Special Graphs
4. Prove that |E| = |V| 1 in trees.
69

0
0
1
0
1
1
1
0
0

0
0
1
1
0
1
1
0
0

0
0
0
1
1
0
1
1
0

0
0
0
1
1
1
0
1
0

0
0
0
0
0
1
1
0
1

0
0
0
0
0
0
0
1
0

(2.29)

Figure 2.28: A Sample (Binary) Tree

Graph/Network Algorithms
5. Consider the tree shown in Figure 2.28. Traverse the graph using
both BFS and DFS and list the order in which nodes are visited in
each algorithm.
6. For a tree and a node v, under what condition is v visited sooner by
BFS than DFS? Provide details.
7. For a real-world social network, is BFS or DFS more desirable? Provide details.
8. Compute the shortest path between any pair of nodes using Dijkstras
algorithm for the graph in Figure 2.29.

Figure 2.29: Weighted Graph.

9. Detail why edges with negative weights are not desirable for computing shortest paths using Dijkstras algorithm.
70

Figure 2.30: Weighted Graph

10. Compute the minimal spanning tree using Prims algorithm in the
graph provided in Figure 2.30.
11. Compute the maximum flow in Figure 2.31.

Figure 2.31: Flow Graph

12. Given a flow network, you are allowed to change one edges capacity.
Can this increase the flow? How can we find the correct edge to
change?
13. How many bridges are in a bipartite graph? Provide details.

71

72

This chapter is from Social Media Mining: An Introduction.

By Reza Zafarani, Mohammad Ali Abbasi, and Huan Liu.
Cambridge University Press, 2014. Draft version: April 20, 2014.
Complete Draft and Slides Available at: http://dmml.asu.edu/smm

Chapter

Network Measures
In February 2012, Kobe Bryant, the American basketball star, joined Chinese microblogging site Sina Weibo. Within a few hours, more than 100,000
followers joined his page, anxiously waiting for his first microblogging post
on the site. The media considered the tremendous number of followers
Kobe Bryant received as an indication of his popularity in China. In this
case, the number of followers measured Bryants popularity among Chinese social media users. In social media, we often face similar tasks in
which measuring different structural properties of a social media network
can help us better understand individuals embedded in it. Corresponding measures need to be designed for these tasks. This chapter discusses
measures for social media networks.
When mining social media, a graph representation is often used. This
graph shows friendships or user interactions in a social media network.
Given this graph, some of the questions we aim to answer are as follows:
Who are the central figures (influential individuals) in the network?
What interaction patterns are common in friends?
Who are the like-minded users and how can we find these similar
individuals?
To answer these and similar questions, one first needs to define measures for quantifying centrality, level of interactions, and similarity, among
other qualities. These measures take as input a graph representation of a
73

social interaction, such as friendships (adjacency matrix), from which the

measure value is computed.
To answer our first question about finding central figures, we define
measures for centrality. By using these measures, we can identify various
types of central nodes in a network. To answer the other two questions,
we define corresponding measures that can quantify interaction patterns
and help find like-minded users. We discuss centrality next.

3.1

Centrality

Centrality defines how important a node is within a network.

3.1.1

Degree Centrality

In real-world interactions, we often consider people with many connections to be important. Degree centrality transfers the same idea into a
measure. The degree centrality measure ranks nodes with more connections higher in terms of centrality. The degree centrality Cd for node vi in
an undirected graph is
Cd (vi ) = di ,
(3.1)
where di is the degree (number of adjacent edges) of node vi . In directed
graphs, we can either use the in-degree, the out-degree, or the combination
as the degree centrality value:
Cd (vi ) = din
(prestige),
i
out
Cd (vi ) = di
(gregariousness),
in
out
Cd (vi ) = di + di .
Prominence or
Prestige

(3.2)
(3.3)
(3.4)

When using in-degrees, degree centrality measures how popular a node

is and its value shows prominence or prestige. When using out-degrees, it
measures the gregariousness of a node. When we combine in-degrees and
out-degrees, we are basically ignoring edge directions. In fact, when edge
directions are removed, Equation 3.4 is equivalent to Equation 3.1, which
measures degree centrality for undirected graphs.
The degree centrality measure does not allow for centrality values to
be compared across networks (e.g., Facebook and Twitter). To overcome
this problem, we can normalize the degree centrality values.
74

Figure 3.1: Degree Centrality Example.

Normalizing Degree Centrality
Simple normalization methods include normalizing by the maximum possible degree,
di
(vi ) =
,
(3.5)
Cnorm
d
n1
where n is the number of nodes. We can also normalize by the maximum
degree,
di
Cmax
.
(3.6)
d (vi ) =
max j d j
Finally, we can normalize by the degree sum,
d
d
d
Pi = i = i .
Csum
d (vi ) =
2|E| 2m
j dj

(3.7)

Example 3.1. Figure 3.1 shows a sample graph. In this graph, degree centrality
for node v1 is Cd (v1 ) = d1 = 8, and for all others, it is Cd (v j ) = d j = 1, j , 1.

3.1.2

Eigenvector Centrality

In degree centrality, we consider nodes with more connections to be more

important. However, in real-world scenarios, having more friends does
not by itself guarantee that someone is important: having more important
friends provides a stronger signal.
75

Eigenvector centrality tries to generalize degree centrality by incorporating the importance of the neighbors (or incoming neighbors in directed
graphs). It is defined for both directed and undirected graphs. To keep
track of neighbors, we can use the adjacency matrix A of a graph. Let ce (vi )
denote the eigenvector centrality of node vi . We want the centrality of vi to
be a function of its neighbors centralities. We posit that it is proportional
to the summation of their centralities,
n

1X
A j,i ce (v j ),
ce (vi ) =
j=1

(3.8)

where is some fixed constant. Assuming Ce = (Ce (v1 ), Ce (v2 ), . . . , Ce (vn ))T
is the centrality vectors for all nodes, we can rewrite Equation 3.8 as
Ce = AT Ce .

Perron-Frobenius
Theorem

(3.9)

This basically means that Ce is an eigenvector of adjacency matrix AT

(or A in undirected networks, since A = AT ) and is the corresponding
eigenvalue. A matrix can have many eigenvalues and, in turn, many corresponding eigenvectors. Hence, this raises the question: which eigenvalue
eigenvector pair should we select? We often prefer centrality values to
be positive for convenient comparison of centrality values across nodes.
Thus, we can choose an eigenvalue such that the eigenvector components
are positive.1 This brings us to the Perron-Frobenius theorem.
Theorem 3.1 (Perron-Frobenius Theorem). Let A Rnn represent the adjacency matrix for a [strongly] connected graph or A : Ai,j > 0 (i.e. a positive n
by n matrix). There exists a positive real number (Perron-Frobenius eigenvalue)
max , such that max is an eigenvalue of A and any other eigenvalue of A is
strictly smaller than max . Furthermore, there exists a corresponding eigenvector
v = (v1 , v2 , . . . , vn ) of A with eigenvalue max such that vi > 0.
Therefore, to have positive centrality values, we can compute the eigenvalues of A and then select the largest eigenvalue. The corresponding
eigenvector is Ce . Based on the Perron-Frobenius theorem, all the components of Ce will be positive, and this vector corresponds to eigenvector
centralities for the graph.
1

This constraint is optional and can be lifted based on the context.

76

Figure 3.2: Eigenvector Centrality Example.

Example 3.2. For the graph shown in Figure 3.2(a), the adjacency matrix is

0 1 0

A = 1 0 1 .
(3.10)

0 1 0
Based on Equation 3.9, we need to solve Ce = ACe , or
(A I)Ce = 0.
Assuming Ce = [u1 u2 u3 ]T ,

1
0
1
0

0
1

0
1
0

(3.11)

u1 0
u2 = 0 .


u3
0

Since Ce , [0 0 0]T , the characteristic equation is


1
0
 0
0
1
det(A I) =  1

0
1
0



 = 0,


(3.12)

(3.13)

or equivalently,
()(2 1) 1() = 2 3 = (2 2 ) = 0.
(3.14)

So the eigenvalues are ( 2, 0, + 2). We select the largest eigenvalue: 2.

We compute the corresponding eigenvector:

0 2
u1 0
1
0

(3.15)

1
0 2
1 u2 = 0 .

u3
0
0
1
0 2
77

Assuming Ce vector has norm 1, its solution is

u1 1/2

Ce = u2 = 2/2 ,

u3
1/2

(3.16)

which denotes that node v2 is the most central node and nodes v1 and v3 have equal
centrality values.
Example 3.3. For the graph shown in Figure 3.2(b), the adjacency matrix is as
follows:

0 1 0 1 0

1 0 1 1 1

A = 0 1 0 1 0 .
(3.17)
1 1 1 0 0

0 1 0 0 0
The eigenvalues of A are (1.74, 1.27, 0.00, +0.33, +2.68). For eigenvector
centrality, the largest eigenvalue is selected: 2.68. The corresponding eigenvector
is the eigenvector centrality vector and is

Ce =

0.4119
0.5825
0.4119
0.5237
0.2169

(3.18)

Based on eigenvector centrality, node v2 is the most central node.

3.1.3

Katz Centrality

A major problem with eigenvector centrality arises when it considers directed graphs (see Problem 1 in the Exercises). Centrality is only passed
on when we have (outgoing) edges, and in special cases such as when a
node is in a directed acyclic graph, centrality becomes zero, even though
the node can have many edges connected to it. In this case, the problem
can be rectified by adding a bias term to the centrality value. The bias term
is added to the centrality values for all nodes no matter how they are
situated in the network (i.e., irrespective of the network topology). The
78

Figure 3.3: Katz Centrality Example.

resulting centrality measure is called the Katz centrality and is formulated
as
n
X
CKatz (vi ) =
A j,i CKatz (v j ) + .
(3.19)
j=1

The first term is similar to eigenvector centrality, and its effect is controlled by constant . The second term , is the bias term that avoids zero
centrality values. We can rewrite Equation 3.19 in a vector form,
CKatz = AT CKatz + 1,

(3.20)

where 1 is a vector of all 1s. Taking the first term to the left hand side and
factoring CKatz ,
CKatz = (I AT )1 1.
(3.21)
Since we are inverting a matrix here, not all values are acceptable.
When = 0, the eigenvector centrality part is removed, and all nodes get
the same centrality value . However, once gets larger, the effect of
is reduced, and when det(I AT ) = 0, the matrix I AT becomes noninvertible and the centrality values diverge. The det(I AT ) first becomes Divergence in
0 when = 1/, where is the largest eigenvalue2 of AT . In practice, Centrality
Computation
< 1/ is selected so that centralities are computed correctly.
Example 3.4. For the graph shown in Figure 3.3, the adjacency matrix is as
When det(I AT ) = 0, it can be rearranged as det(AT 1 I) = 0, which is basically
the characteristic equation. This equation first becomes zero when the largest eigenvalue
equals 1 , or equivalently = 1/.
2

79

follows:

A =

0
1
1
1
0

1
0
1
1
1

1
1
0
1
1

1
1
1
0
0

0
1
1
0
0

= AT .

(3.22)

The eigenvalues of A are (1.68, 1.0, 1.0, +0.35, +3.32). The largest eigenvalue of A is = 3.32. We assume = 0.25 < 1/ and = 0.2. Then, Katz
centralities are

1.14

1.31

(3.23)
CKatz = (I AT )1 1 = 1.31 .
1.14

0.85
Thus, nodes v2 , and v3 have the highest Katz centralities.

3.1.4

PageRank

Similar to eigenvector centrality, Katz centrality encounters some challenges. A challenge that happens in directed graphs is that, once a node
becomes an authority (high centrality), it passes all its centrality along all
of its out-links. This is less desirable, because not everyone known by
a well known person is well known. To mitigate this problem, one can
divide the value of passed centrality by the number of outgoing links (outdegree) from that node such that each connected neighbor gets a fraction
of the source nodes centrality:
n
X
Cp (v j )
Cp (vi ) =
A j,i out + .
(3.24)
d
j
j=1
This equation is only defined when dout
is nonzero. Thus, assuming
j
out
that all nodes have positive out-degrees (d j > 0)3 , Equation 3.24 can be
reformulated in matrix format,
Cp = AT D1 Cp + 1,
3

(3.25)

When dout
= 0, we know that since the out-degree is zero, i, A j,i = 0. This makes the
j

term inside the summation 00 . We can fix this problem by setting dout
= 1 since the node
j
will not contribute any centrality to any other nodes.

80

Figure 3.4: PageRank Example.

which we can reorganize,
Cp = (I AT D1 )1 1,

(3.26)

where D = diag(dout
, dout
, . . . , dout
n ) is a diagonal matrix of degrees. The
2
1
centrality measure is known as the PageRank centrality measure and is
used by the Google search engine as a measure for ordering webpages. PageRank and Google
Webpages and their links represent an enormous web-graph. PageRank Web Search
defines a centrality measure for the nodes (webpages) in this web-graph.
When a user queries Google, webpages that match the query and have
higher PageRank values are shown first. Similar to Katz centrality, in
practice, < 1 is selected, where is the largest eigenvalue of AT D1 .
In undirected graphs, the largest eigenvalue of AT D1 is = 1; therefore,
< 1.
Example 3.5. For the graph shown in
follows,

0 1

1 0
A = 0 1
1 0

1 1

Figure 3.4, the adjacency matrix is as

0
1
0
1
1

1
0
1
0
0

1
1
1
0
0

We assume = 0.95 < 1 and = 0.1. Then, PageRank values are

2.14

2.13

Cp = (I AT D1 )1 1 = 2.14 .
1.45

2.13
81

(3.27)

(3.28)

Hence, nodes v1 and v3 have the highest PageRank values.

3.1.5

Betweenness Centrality

Another way of looking at centrality is by considering how important

nodes are in connecting other nodes. One approach, for a node vi , is
to compute the number of shortest paths between other nodes that pass
through vi ,
X (v )
st i
,
(3.29)
Cb (vi ) =
st
s,t,v
i

where st is the number of shortest paths from node s to t (also known

as information pathways), and st (vi ) is the number of shortest paths from s
to t that pass through vi . In other words, we are measuring how central
vi s role is in connecting any pair of nodes s and t. This measure is called
betweenness centrality.
Betweenness centrality needs to be normalized to be comparable across
networks. To normalize betweenness centrality, one needs to compute the
maximum value it takes. Betweenness centrality takes its maximum value
when node vi is on all shortest paths from s to t for any pair (s, t); that is,
(v )
(s, t), s , t , vi , stst i = 1. For instance, in Figure 3.1, node v1 is on the
shortest path between all other pairs of nodes. Thus, the maximum value
is
!
X
X (v )
n1
st i
=
1=2
= (n 1)(n 2).
(3.30)
Cb (vi ) =

2
st
s,t,v
s,t,v
i

The betweenness can be divided by its maximum value to obtain the

normalized betweenness,
Cnorm
(vi ) =
b

Cb (vi )

.
2 n1
2

(3.31)

Computing Betweenness
In betweenness centrality (Equation 3.29), we compute shortest paths between all pairs of nodes to compute the betweenness value. If an algorithm
such as Dijkstras is employed, it needs to be run for all nodes, because
82

Figure 3.5: Betweenness Centrality Example.

Dijkstras algorithm will compute shortest paths from a single node to all
other nodes. So, to compute all-pairs shortest paths, Dijkstras algorithm
needs to be run |V| 1 times (with the exception of the node for which
centrality is being computed). More effective algorithms such as the Brandes algorithm [45] have been designed. Interested readers can refer to the
bibliographic notes for further references.
Example 3.6. For Figure 3.1, the (normalized) betweenness centrality of node v1
is
!
8
Cb (v1 ) = 2
,
(3.32)
2
(v1 ) = 1.
(3.33)
Cnorm
b
Since all the paths that

go through any pair (s, t), s , t , v1 pass through node
v1 , the centrality is 2 82 . Similarly, the betweenness centrality for any other node
in this graph is 0.
Example 3.7. Figure 3.5 depicts a sample graph. In this graph, the betweenness
centrality for node v1 is 0, since no shortest path passes through it. For other
nodes, we have
Cb (v2 ) = 2 ( (1/1) + (1/1) + (2/2) + (1/2) + 0 + 0 )
|{z}
|{z}
|{z}
|{z}
|{z}
|{z}
s=v1 ,t=v3

s=v1 ,t=v4

s=v1 ,t=v5

s=v3 ,t=v4

s=v3 ,t=v5

s=v4 ,t=v5

= 2 3.5 = 7,
Cb (v3 ) = 2 (

(3.34)

0 + 0 + (1/2) + 0 + (1/2) + 0 )
|{z}
|{z}
|{z}
|{z}
|{z}
|{z}

s=v1 ,t=v2

s=v1 ,t=v4

s=v1 ,t=v5

= 2 1.0 = 2,

s=v2 ,t=v4

s=v2 ,t=v5

s=v4 ,t=v5

(3.35)
83

Cb (v4 ) = Cb (v3 ) = 2 1.0 = 2,

Cb (v5 ) = 2 (

(3.36)

0 + 0 + 0 + 0 + 0 + (1/2) )
|{z}
|{z}
|{z}
|{z}
|{z}
|{z}

s=v1 ,t=v2

s=v1 ,t=v3

s=v1 ,t=v4

s=v2 ,t=v3

s=v2 ,t=v4

s=v3 ,t=v4

= 2 0.5 = 1,

(3.37)

where centralities are multiplied by 2 because in an undirected graph

P
(v )
2 s,t,vi ,s<t stst i .

3.1.6

P
s,t,vi

st (vi )
st

Closeness Centrality

In closeness centrality, the intuition is that the more central nodes are, the
more quickly they can reach other nodes. Formally, these nodes should
have a smaller average shortest path length to other nodes. Closeness
centrality is defined as
1
Cc (vi ) = ,
(3.38)
lv
i
P
where lv = 1
v ,v li, j is node vis average shortest path length to other
i

n1

nodes. The smaller the average shortest path length, the higher the centrality for the node.
Example 3.8. For nodes in Figure 3.5, the closeness centralities are as follows:
Cc (v1 )
Cc (v2 )
Cc (v3 ) = Cb (v4 )
Cc (v5 )

= 1 / ( (1 + 2 + 2 + 3)/4 ) = 0.5,
= 1 / ( (1 + 1 + 1 + 2)/4 ) = 0.8,
= 1 / ( (1 + 1 + 2 + 2)/4 ) = 0.66,
= 1 / ( (1 + 1 + 2 + 3)/4 ) = 0.57.

(3.39)
(3.40)
(3.41)
(3.42)

Hence, node v2 has the highest closeness centrality.

The centrality measures discussed thus far have different views on what
a central node is. Thus, a central node for one measure may be deemed
unimportant by other measures.
Example 3.9. Consider the graph in Figure 3.6. For this graph, we compute the
top three central nodes based on degree, eigenvector, Katz, PageRank, betweenness,
and closeness centrality methods. These nodes are listed in Table 3.1.
As shown in the table, there is a high degree of similarity between most central
nodes for the first four measures, which utilize eigenvectors or degrees: degree
84

Figure 3.6: Example for All Centrality Measures.

Table 3.1: A Comparison between Centrality Methods
First Node Second Node
Third Node
Degree Centrality
v3 or v6
v6 or v3
v {v4 , v5 , v7 , v8 , v9 }
Eigenvector Centrality
v6
v3
v4 or v5
Katz Centrality: = = 0.3
v6
v3
v4 or v5
PageRank: = = 0.3
v3
v6
v2
Betweenness Centrality
v6
v7
v3
Closeness Centrality
v6
v3 or v7
v7 or v3
centrality, eigenvector centrality, Katz centrality, and PageRank. Betweenness
centrality also generates similar results to closeness centrality because both use
the shortest paths to find most central nodes.

3.1.7

Group Centrality

All centrality measures defined so far measure centrality for a single node.
These measures can be generalized for a group of nodes. In this section,
we discuss how degree centrality, closeness centrality, and betweenness
centrality can be generalized for a group of nodes. Let S denote the set of
nodes to be measured for centrality. Let V S denote the set of nodes not
in the group.
Group Degree Centrality
Group degree centrality is defined as the number of nodes from outside
the group that are connected to group members. Formally,
group

Cd

(S) = |{vi V S|vi is connected to v j S}|.

85

(3.43)

Similar to degree centrality, we can define connections in terms of outdegrees or in-degrees in directed graphs. We can also normalize this value.
In the best case, group members are connected to all other nonmembers.
group
Thus, the maximum value of Cd (S) is |V S|. So dividing group degree
centrality value by |V S| normalizes it.
Group Betweenness Centrality
Similar to betweeness centrality, we can define group betweenness centrality as
X (S)
st
group
,
(3.44)
Cb
(S) =
st
s,t,s<S,t<S
where st (S) denotes the number of shortest paths between s and t that
pass through members of S. In the best case, all shortest paths between s
and t pass through
members of S, and therefore, the maximum value for
group
|VS|

Cb
(S) is 2 2 . Similar to betweenness centrality, we can normalize
group betweenness centrality by dividing it by the maximum value.
Group Closeness Centrality
Closeness centrality for groups can be defined as
group

Cc

(S) =

1
,
lgroup
S

(3.45)

P
group
1
where lS
= |VS|
v j <S lS,v j and lS,v j is the length of the shortest path
between a group S and a nonmember v j V S. This length can be
defined in multiple ways. One approach is to find the closest member in S
to v j :
lS,v j = min lvi ,v j .
(3.46)
vi S

One can also use the maximum distance or the average distance to
compute this value.
Example 3.10. Consider the graph in Figure 3.7. Let S = {v2 , v3 }. Group degree
centrality for S is
group
Cd (S) = 3,
(3.47)
86

Figure 3.7: Group Centrality Example.

since members of the group are connected to all other three members in V S =
{v1 , v4 , v5 }. The normalized
value is 1, since 3/|V S| = 1. Group betweenness
centrality is 6, since for 2 32 shortest paths between any two members of V S, the
path has to pass
through members of S. The normalized group betweenness is 1,
) = 1. Finally, group closeness centrality assuming the distance
since 6/(2 |VS|
2
from nonmembers to members of S is computed using the minimum function is
also 1, since any member of V S is connected to a member of S directly.

3.2

Transitivity and Reciprocity

Often we need to observe a specific behavior in a social media network.

One such behavior is linking behavior. Linking behavior determines how
links (edges) are formed in a social graph. In this section, we discuss
two well-known measures, transitivity and reciprocity, for analyzing this
behavior. Both measures are commonly used in directed networks, and
transitivity can also be applied to undirected networks.

3.2.1

Transitivity

In transitivity, we analyze the linking behavior to determine whether it

demonstrates a transitive behavior. In mathematics, for a transitive relation
R, aRb bRc aRc. The transitive linking behavior can be described as
follows.
87

Figure 3.8: Transitive Linking.

Transitive Linking
Let v1 , v2 , v3 denote three nodes. When edges (v1 , v2 ) and (v2 , v3 ) are formed,
if (v3 , v1 ) is also formed, then we have observed a transitive linking behavior
(transitivity). This is shown in Figure 3.8.
In a less formal setting,
Transitivity is when a friend of my friend is my friend.
As shown in the definition, a transitive behavior needs at least three
edges. These three edges, along with the participating nodes, create a
triangle. Higher transitivity in a graph results in a denser graph, which
in turn is closer to a complete graph. Thus, we can determine how close
graphs are to the complete graph by measuring transitivity. This can be
performed by measuring the [global] clustering coefficient and local clustering
coefficient. The former is computed for the network, whereas the latter is
computed for a node.
Clustering Coefficient
The clustering coefficient analyzes transitivity in an undirected graph.
Since transitivity is observed when triangles are formed, we can measure
it by counting paths of length 2 (edges (v1 , v2 ) and (v2 , v3 )) and checking
whether the third edge (v3 , v1 ) exists (i.e., the path is closed). Thus, clustering coefficient C is defined as
C=

|Closed Paths of Length 2|

.
|Paths of Length 2|
88

(3.48)

Alternatively, we can count triangles

C=

(Number of Triangles) 6
.
|Paths of Length 2|

(3.49)

Since every triangle has six closed paths of length 2, we can rewrite
Equation 3.49 as
C=

(Number of Triangles) 3
.
Number of Connected Triples of Nodes

(3.50)

In this equation, a triple is an ordered set of three nodes, connected by

two (i.e., open triple) or three (closed triple) edges. Two triples are different
when
their nodes are different, or
their nodes are the same, but the triples are missing different edges.
For example, triples vi v j vk and v j vk vi are different, since the first triple
is missing edge e(vk , vi ) and the second triple is missing edge e(vi , v j ), even
though they have the same members. Following the same argument,
triples vi v j vk and vk v j vi are the same, because both are missing edge e(vk , vi )
and have the same members. Since triangles have three edges, one edge
can be missed in each triple; therefore, three different triples can be formed
from one triangle. The number of triangles are therefore multiplied by
a factor of 3 in the numerator of Equation 3.50. Note that the clustering
coefficient is computed for the whole network.
Example 3.11. For the graph in Figure 3.9, the clustering coefficient is
(Number of Triangles) 3
Number of Connected Triples of Nodes
23
=
= 0.75.
23+
2
|{z}

C =

(3.51)

v2 v1 v4 ,v2 v3 v4

The clustering coefficient can also be computed locally. The following

subsection discusses how it can be computed for a single node.
89

Figure 3.9: A Global Clustering Coefficient Example.

Local Clustering Coefficient
The local clustering coefficient measures transitivity at the node level.
Commonly used for undirected graphs, it estimates how strongly neighbors of a node v (nodes adjacent to v) are themselves connected. The
coefficient is defined as
C(vi ) =

Number of Pairs of Neighbors of vi That Are Connected

. (3.52)
Number of Pairs of Neighbors of vi

In an undirected graph, the denominator can be rewritten as

di (di 1)/2, since there are di neighbors for node vi .

di

2

Example 3.12. Figure 3.10 shows how the local clustering coefficient changes for

Figure 3.10: Change in Local Clustering Coefficient for Different Graphs.

Thin lines depict connections to neighbors. Solid lines indicate connected
neighbors, and dashed lines are the missing connections among neighbors.
90

Figure 3.11: A Graph with Reciprocal Edges.

node v1 . Thin lines depict v1 s connections to its neighbors. Dashed lines denote
possible connections among neighbors, and solid lines denote current connections
among neighbors. Note that when none of the neighbors are connected, the local
clustering coefficient is zero, and when all the neighbors are connected, it becomes
maximum, C(vi ) = 1.

3.2.2

Reciprocity

Reciprocity is a simplified version of transitivity, because it considers closed

loops of length 2, which can only happen in directed graphs. Formally, if
node v is connected to node u, u by connecting to v exhibits reciprocity. On
microblogging site Tumblr, for example, these nodes are known as mutual
followers. Informally, reciprocity is
If you become my friend, Ill be yours.
Figure 3.11 shows an example where two nodes (v1 and v2 ) in the graph
demonstrate reciprocal behavior.
Reciprocity counts the number of reciprocal pairs in the graph. Any
directed graph can have a maximum of |E|/2 pairs. This happens when
all edges are reciprocal. Thus, this value can be used as a normalization
factor. Reciprocity can be computed using the adjacency matrix A:
P
i, j,i<j Ai,j A j,i
,
R =
|E|/2
91

2 X
Ai, j A j,i ,
|E| i,j,i<j

2
1
Tr(A2 ),
|E| 2
1
Tr(A2 ),
=
|E|
1
=
Tr(A2 ),
(3.53)
m
P
where Tr(A) = A1,1 + A2,2 + + An,n = ni=1 Ai,i and mPis the number of edges
in the network. Note that the maximum value for i, j Ai, j A j,i is m when all
directed edges are reciprocated.
=

Example 3.13. For the graph shown in Figure 3.11, the adjacency matrix is

A = 1

1
0
1

1
0
0

1
1
1
2
R = Tr(A ) = Tr 0
m
4 1

1
1
0

0
1
0

(3.54)

Its reciprocity is

3.3

2 1
= = .
4 2

(3.55)

Balance and Status

A signed graph can represent the relationships of nodes in a social network,

such as friends or foes. For example, a positive edge from node v1 to v2
denotes that v1 considers v2 as a friend and a negative edge denotes that
v1 assumes v2 is an enemy. Similarly, we can utilize signed graphs to
represent the social status of individuals. A positive edge connecting node
v1 to v2 can also denote that v1 considers v2 s status higher than its own
in the society. Both cases represent interactions that individuals exhibit
about their relationships. In real-world scenarios, we expect some level
of consistency with respect to these interactions. For instance, it is more
plausible for a friend of ones friend to be a friend than to be an enemy.
In signed graphs, this consistency translates to observing triads with three
92

positive edges (i.e., all friends) more frequently than ones with two positive
edges and one negative edge (i.e., a friends friend is an enemy). Assume
we observe a signed graph that represents friends/foes or social status.
Can we measure the consistency of attitudes that individual have toward
one another?
To measure consistency in an individuals attitude, one needs to utilize
theories from social sciences to define what is a consistent attitude. In this
section, we discuss two theories, social balance and social status, that can
help determine consistency in observed signed networks. Social balance
theory is used when edges represent friends/foes, and social status theory
is employed when they represent status.
Social Balance Theory
This theory, also known as structural balance theory, discusses consistency Structural Balance
in friend/foe relationships among individuals. Informally, social balance Theory
theory says friend/foe relationships are consistent when
The friend of my friend is my friend,
The friend of my enemy is my enemy,
The enemy of my enemy is my friend,
The enemy of my friend is my enemy.
We demonstrate a graph representation of social balance theory in Figure 3.12. In this figure, positive edges demonstrate friendships and negative ones demonstrate enemies. Triangles that are consistent based on
this theory are denoted as balanced and triangles that are inconsistent as
unbalanced. Let wi j denote the value of the edge between nodes vi and v j . Balanced and
Then, for a triangle of nodes vi , v j , and vk , it is consistent based on social Unbalanced
Triangles
balance theory; that is, it is balanced if and only if
wi j w jk wki 0.

(3.56)

This is assuming that, for positive edges, wi j = 1, and for negative

edges, wi j = 1. We observe that, for all balanced triangles in Figure
3.12, the value wi j w jk wki is positive, and for all unbalanced triangles, it is
negative. Social balance can also be generalized to subgraphs other than
triangles. In general, for any cycle, if the product of edge values becomes
positive, then the cycle is socially balanced.
93

Figure 3.12: Sample Graphs for Social Balance Theory. In balanced triangles, there are an even number of negative edges.
Social Status Theory
Social status theory measures how consistent individuals are in assigning
status to their neighbors. It can be summarized as follows:
If X has a higher status than Y and Y has a higher status than Z, then
X should have a higher status than Z.
We show this theory using two graphs in Figure 3.13. In this figure,
nodes represent individuals. Positive and negative signs show higher or
lower status depending on the arrow direction. A directed positive edge
from node X to node Y shows that Y has a higher status than X, and a
negative one shows the reverse. In the figure on the left, v2 has a higher
status than v1 and v3 has a higher status than v2 , so based on status theory,
v3 should have a higher status than v1 ; however, we see that v1 has a
higher status in our configuration.4 Based on social status theory, this is
implausible, and thus this configuration is unbalanced. The graph on the
right shows a balanced configuration with respect to social status theory.
4

Here, we start from v1 and follow the edges. One can start from a different node, and
the result should remain the same.

94

Figure 3.13: Sample Graphs for Social Status Theory. The left-hand graph
is an unbalanced configuration, and the right-hand graph is a balanced
configuration.

In the example provided in Figure 3.13, social status is defined for the
most general example: a set of three connected nodes (a triad). However,
social status can be generalized to other graphs. For instance, in a cycle of
n nodes, where n 1 consecutive edges are positive and the last edge is
negative, social status theory considers the cycle balanced.
Note that the identical configuration can be considered balanced by
social balance theory and unbalanced based on social status theory (see
Exercises).

3.4

Similarity

In this section, we review measures used to compute similarity between

two nodes in a network. In social media, these nodes can represent individuals in a friendship network or products that are related. The similarity
between these connected individuals can be computed either based on
the network in which they are embedded (i.e., network similarity) or based
on the similarity of the content they generate (i.e., content similarity). We
discuss content similarity in Chapter 5. In this section, we demonstrate
ways to compute similarity between two nodes using network information regarding the nodes and edges connecting them. When using network
information, the similarity between two nodes can be computed by measuring their structural equivalence or their regular equivalence.
95

3.4.1

Structural Equivalence

To compute structural equivalence, we look at the neighborhood shared by

two nodes; the size of this neighborhood defines how similar two nodes
are. For instance, two brothers have in common sisters, mother, father,
grandparents, and so on. This shows that they are similar, whereas two
random male or female individuals do not have much in common and are
not similar.
The similarity measures detailed in this section are based on the overlap between the neighborhoods of the nodes. Let N(vi ) and N(v j ) be the
neighbors of nodes vi and v j , respectively. In this case, a measure of node
similarity can be defined as follows:
(vi , v j ) = |N(vi ) N(v j )|.

(3.57)

For large networks, this value can increase rapidly, because nodes may
share many neighbors. Generally, similarity is attributed to a value that is
bounded and is usually in the range [0, 1]. Various normalization proceJaccard Similarity and dures can take place such as the Jaccard similarity or the cosine similarity:
Cosine Similarity

|N(vi ) N(v j )|
,
|N(vi ) N(v j )|

(3.58)

|N(vi ) N(v j )|
Cosine (vi , v j ) = p
.
|N(vi )||N(v j )|

(3.59)

Jaccard (vi , v j ) =

In general, the definition of neighborhood N(vi ) excludes the node itself

(vi ). This leads to problems with the aforementioned similarities because
nodes that are connected and do not share a neighbor will be assigned zero
similarity. This can be rectified by assuming nodes to be included in their
neighborhoods.
Example 3.14. Consider the graph in Figure 3.14. The similarity values between
nodes v2 and v5 are
|{v1 , v3 , v4 } {v3 , v6 }|
= 0.25,
|{v1 , v3 , v4 , v6 }|
|{v1 , v3 , v4 } {v3 , v6 }|
Cosine (v2 , v5 ) =
= 0.40.
|{v1 , v3 , v4 }||{v3 , v6 }|

Jaccard (v2 , v5 ) =

96

(3.60)
(3.61)

Figure 3.14: Sample Graph for Computing Similarity.

A more interesting way of measuring the similarity between vi and v j
is to compare (vi , v j ) with the expected value of (vi , v j ) when nodes pick
their neighbors at random. The more distant these two values are, the
more significant the similarity observed between vi and v j ((vi , v j )) is. For
dd

nodes vi and v j with degrees di and d j , this expectation is in j , where n is

the number of nodes. This is because there is a dni chance of becoming vi s
neighbor and, since v j selects d j neighbors, the expected overlap is
can rewrite (vi , v j ) as
X
(vi , v j ) = |N(vi ) N(v j )| =
Ai,k A j,k .

di d j
.
n

We

(3.62)

Hence, a similarity measure can be defined by subtracting the random

dd
expectation in j from Equation 3.62:
di d j
n
k
X
1X
1X
=
Ai,k A j,k n
Ai,k
A j,k
n
n
k
k
k
X
=
Ai,k A j,k nAi A j

significance (vi , v j ) =

k
X

Ai,k A j,k

(Ai,k A j,k Ai A j )

k
X
=
(Ai,k A j,k Ai A j Ai A j + Ai A j )
k
X
=
(Ai,k A j,k Ai,k A j Ai A j,k + Ai A j )
k

97

X
=
(Ai,k Ai )(A j,k A j ),

(3.63)

P
P
where Ai = n1 k Ai,k . The term k (Ai,k Ai )(A j,k A j ) is basically the
covariance between Ai and A j . The covariance can be normalized by the
multiplication of variances,
pearson (vi , v j ) =

Pearson Correlation

significance (vi , v j )
qP
2

2
k (Ai,k Ai )
k (A j,k A j )
P

k (Ai,k Ai )(A j,k A j ),

,
qP
pP
2
2

(A

A
)
(A

A
)
i,k
i
j,k
j
k
k
pP

(3.64)

which is called the Pearson correlation coefficient. Its value, unlike the other
two measures, is in the range [1, 1]. A positive correlation value denotes
that when vi befriends an individual vk , v j is also likely to befriend vk .
A negative value denotes the opposite (i.e., when vi befriends vk , it is
unlikely for v j to befriend vk ). A zero value denotes that there is no linear
relationship between the befriending behavior of vi and v j .

3.4.2

Regular Equivalence

In regular equivalence, unlike structural equivalence, we do not look at

the neighborhoods shared between individuals, but at how neighborhoods
themselves are similar. For instance, athletes are similar not because they
know each other in person, but because they know similar individuals,
such as coaches, trainers, and other players. The same argument holds for
any other profession or industry in which individuals might not know each
other in person, but are in contact with very similar individuals. Regular
equivalence assesses similarity by comparing the similarity of neighbors
and not by their overlap.
One way of formalizing this is to consider nodes vi and v j similar when
they have many similar neighbors vk and vl . This concept is shown in
Figure 3.15(a). Formally,
X
regular (vi , v j ) =
Ai,k A j,l regular (vk , vl ).
(3.65)
k,l

98

Figure 3.15: Regular Equivalence. Solid lines denote edges, and dashed
lines denote similarities between nodes. In regular equivalence, similarity between nodes vi and v j is replaced by similarity between (a) their
neighbors vk and vl or between (b) neighbor vk and node v j .
Unfortunately, this formulation is self-referential because solving for i
and j requires solving for k and l, solving for k and l requires solving for
their neighbors, and so on. So, we relax this formulation and assume that
node vi is similar to node v j when v j is similar to vi s neighbors vk . This is
shown in Figure 3.15(b). Formally,
X
regular (vi , v j ) =
Ai,k regular (vk , v j ).
(3.66)
k

In vector format, we have

regular = Aregular .

(3.67)

A node is highly similar to itself. To make sure that our formulation guarantees this, we can add an identity matrix to this vector format.
Adding an identity matrix will add 1 to all diagonal entries, which represent self-similarities regular (vi , vi ):
regular = Aregular + I.

(3.68)

By rearranging, we get
regular = (I A)1 ,
which we can use to find the regular equivalence similarity.
99

(3.69)

Note the similarity between Equation 3.69 and that of Katz centrality
(Equation 3.21). As with Katz centrality, we must be careful how we choose
for convergence. A common practice is to select an such that < 1/,
where is the largest eigenvalue of A.
Example 3.15. For the graph depicted in Figure 3.14, the adjacency matrix is

A =

0
1
1
0
0

1
0
1
1
0

1
1
0
0
0

0
1
0
0
1

0
0
1
0
1

0
0
0
1
0

(3.70)

The largest eigenvalue of A is 2.43. We set = 0.4 < 1/2.43, and we compute
(I 0.4A)1 , which is the similarity matrix,

regular = (I 0.4A)1

1.43
0.73
0.73
0.26
0.26
0.16

0.73
1.63
0.80
0.56
0.32
0.26

0.73
0.80
1.63
0.32
0.56
0.26

0.26
0.56
0.32
1.31
0.23
0.46

0.26
0.32
0.56
0.23
1.31
0.46

0.16
0.26
0.26
0.46
0.46
1.27

(3.71)

Any row or column of this matrix shows the similarity of a node to other nodes.
We can see that node v1 is the most similar (other than itself) to nodes v2 and v3 .
Furthermore, nodes v2 and v3 have the highest similarity in this graph.

100

3.5

Summary

In this chapter, we discussed measures for a social media network. Centrality measures attempt to find the most central node within a graph.
Degree centrality assumes that the node with the maximum degree is the
most central individual. In directed graphs, prestige and gregariousness
are variants of degree centrality. Eigenvector centrality generalizes degree
centrality and considers individuals who know many important nodes
as central. Based on the Perron-Frobenius theorem, eigenvector centrality is determined by computing the eigenvector of the adjacency matrix.
Katz centrality solves some of the problems with eigenvector centrality
in directed graphs by adding a bias term. PageRank centrality defines
a normalized version of Katz centrality. The Google search engine uses
PageRank as a measure to rank webpages. Betweenness centrality assumes
that central nodes act as hubs connecting other nodes, and closeness centrality implements the intuition that central nodes are close to all other
nodes. Node centrality measures can be generalized to a group of nodes
using group degree centrality, group betweenness centrality, and group
closeness centrality.
Linking between nodes (e.g., befriending in social media) is the most
commonly observed phenomenon in social media. Linking behavior is
analyzed in terms of its transitivity and its reciprocity. Transitivity is
when a friend of my friend is my friend. The transitivity of linking
behavior is analyzed by means of the clustering coefficient. The global
clustering coefficient analyzes transitivity within a network, and the local
clustering coefficient performs that for a node. Transitivity is commonly
considered for closed triads of edges. For loops of length 2, the problem is
simplified and is called reciprocity. In other words, reciprocity is when if
you become my friend, Ill be yours.
To analyze if relationships are consistent in social media, we used various social theories to validate outcomes. Social balance and social status
are two such theories.
Finally, we analyzed node similarity measures. In structural equivalence, two nodes are considered similar when they share neighborhoods.
We discussed cosine similarity and Jaccard similarity in structural equivalence. In regular equivalence, nodes are similar when their neighborhoods
are similar.

101

3.6

Bibliographic Notes

General reviews of different measures in graphs, networks, the web, and

social media can be found in [212, 304, 270, 120, 294].
A more detailed description of the PageRank algorithm can be found in
[224, 174]. In practice, to compute the PageRank values, the power iteration
method is used. Given a matrix A, this method produces an eigenvalue
and an eigenvector v of A. In the case of PageRank, eigenvalue is set to 1.
The iterative algorithm starts with an initial eigenvector v0 and then, vk+1
is computed from vk as follows,
vk+1 = Avk .

(3.72)

The iterative process is continued until vk vk+1 (i.e., convergence occurs).

Other similar techniques to PageRank for computing influential nodes in
a webgraph, such as the HITS [150] algorithm, can be found in [51, 153].
Unlike PageRank, the HITS algorithm5 considers two types of nodes: authority nodes and hub nodes. An authority is a webpage that has many inlinks. A hub is a page with many out-links. Authority pages have in-links
from many hubs. In other words, hubs represent webpages that contain
many useful links to authorities and authorities are influential nodes in the
webgraph. HITS employs an iterative approach to compute authority and
hub scores for all nodes in the graph. Nodes with high authority scores are
classified as authorities and nodes with high hub scores as hubs. Webpage
with high authority scores or hub scores can be recommended to users in
a web search engine.
Betweenness algorithms can be improved using all-pair shortest paths
algorithms [293] or algorithms optimized for computing betweenness, such
as the Brandes algorithm discussed in [45, 278].
A review of node similarity and normalization procedures is provided
in [166]. Jaccard similarity was introduced in [133] and cosine similarity is
introduced by Salton and McGill [244].
REGE [302, 303] and CATREGE [264] are well-known algorithms for
computing regular equivalence.

HITS stands for hypertext-induced topic search.

102

3.7

Exercises

Centrality
1. Come up with an example of a directed connected graph in which
eigenvector centrality becomes zero for some nodes. Describe when
this happens.
2. Does have any effect on the order of centralities? In other words, if
for one value of the centrality value of node vi is greater than that of
v j , is it possible to change in a way such that v j s centrality becomes
larger than that of vi s?
3. In PageRank, what values can we select to guarantee that centrality
values are calculated correctly (i.e., values do not diverge)?
4. Calculate PageRank values for this graph when

= 1, = 0
= 0.85, = 1
= 0, = 1
Discuss the effects of different values of and for this particular
problem.
5. Consider a full n-tree. This is a tree in which every node other than
the leaves has n children. Calculate the betweenness centrality for
the root node, internal nodes, and leaves.
103

6. Show an example where the eigenvector centrality of all nodes in the

graph is the same while betweenness centrality gives different values
for different nodes.

Transitivity and Reciprocity

7. In a directed graph G(V, E),
Let p be the probability that any node vi is connected to any node
v j . What is the expected reciprocity of this graph?
Let m and n be the number of edges and number of nodes,
respectively. What is the maximum reciprocity? What is the
minimum?
8. Given all graphs {G(V, E)|s.t., |E| = m, |V| = n},
(a) When m = 15 and n = 10, find a graph with a minimum average
clustering coefficient (one is enough).
(b) Can you come up with an algorithm to find such a graph for any
m and n?

Balance and Status

9. Find all conflicting directed triad configurations for social balance
and social status. A conflicting configuration is an assignment of
positive/negative edge signs for which one theory considers the triad
balanced and the other considers it unbalanced.

Similarity
10. In Figure 3.6,
Compute node similarity using Jaccard and cosine similarity for
nodes v5 and v4 .
Find the most similar node to v7 using regular equivalence.

104

This chapter is from Social Media Mining: An Introduction.

By Reza Zafarani, Mohammad Ali Abbasi, and Huan Liu.
Cambridge University Press, 2014. Draft version: April 20, 2014.
Complete Draft and Slides Available at: http://dmml.asu.edu/smm

Chapter

Network Models
In May 2011, Facebook had 721 million users, represented by a graph of 721
million nodes. A Facebook user at the time had an average of 190 friends;
that is, all Facebook users, taken into account, had a total of 68.5 billion
friendships (i.e., edges). What are the principal underlying processes that
help initiate these friendships? More importantly, how can these seemingly
independent friendships form this complex friendship network?
In social media, many social networks contain millions of nodes and
billions of edges. These complex networks have billions of friendships,
the reasons for existence of most of which are obscure. Humbled by the
complexity of these networks and the difficulty of independently analyzing
each one of these friendships, we can design models that generate, on a
smaller scale, graphs similar to real-world networks. On the assumption
that these models simulate properties observed in real-world networks
well, the analysis of real-world networks boils down to a cost-efficient
measuring of different properties of simulated networks. In addition,
these models
allow for a better understanding of phenomena observed in realworld networks by providing concrete mathematical explanations
and
allow for controlled experiments on synthetic networks when realworld networks are not available.
We discuss three principal network models in this chapter: the random
graph model, the small-world model, and the preferential attachment model.
105

These models are designed to accurately model properties observed in

real-world networks. Before we delve into the details of these models, we
discuss their properties.

4.1

Properties of Real-World Networks

Real-world networks share common characteristics. When designing network models, we aim to devise models that can accurately describe these
networks by mimicking these common characteristics. To determine these
characteristics, a common practice is to identify their attributes and show
that measurements for these attributes are consistent across networks. In
particular, three network attributes exhibit consistent measurements across
real-world networks: degree distribution, clustering coefficient, and average
path length. As we recall, degree distribution denotes how node degrees
are distributed across a network. The clustering coefficient measures transitivity of a network. Finally, average path length denotes the average
distance (shortest path length) between pairs of nodes. We discuss how
these three attributes behave in real-world networks next.

4.1.1

Degree Distribution

Consider the distribution of wealth among individuals. Most individuals have an average amount of capital, whereas a few are considered
extremely wealthy. In fact, we observe exponentially more individuals
with an average amount of capital than wealthier ones. Similarly, consider
the population of cities. A few metropolitan areas are densely populated,
whereas other cities have an average population size. In social media, we
observe the same phenomenon regularly when measuring popularity or
interestingness for entities. For instance,
Many sites are visited less than a thousand times a month, whereas
a few are visited more than a million times daily.
Most social media users are active on a few sites, whereas a few
individuals are active on hundreds of sites.
There are exponentially more modestly priced products for sale compared to expensive ones.
106

Figure 4.1: Power-Law Degree Distribution and Its Log-Log Plot.

There exist many individuals with a few friends and a handful of
users with thousands of friends.
The last observation is directly related to node degrees in social media.
The degree of a node in social media often denotes the number of friends
an individual has. Thus, the distribution of the number of friends denotes
the degree distribution of the network. It turns out that in all provided
observations, the distribution of values follows a power-law distribution. Power-Law
For instance, let k denote the degree of a node (i.e., the number of friends Distribution
an individual has). Let pk denote the fraction of individuals with degree k
frequency of observing k
). Then, in the power-law distribution
(i.e.,
|V|
pk = akb ,

(4.1)

where b is the power-law exponent and a is the power-law intercept. A

power-law degree distribution is shown in Figure 4.1(a).
Taking the logarithm from both sides of Equation 4.1, we get
ln pk = b ln k + ln a.

(4.2)

Equation 4.2 shows that the log-log plot of a power-law distribution

is a straight line with slope b and intercept ln a (see Figure 4.1(b)). This
also reveals a methodology for checking whether a network exhibits a
power-law distribution.1 We can do the following:
1

For a more detailed approach refer to [58].

107

Figure 4.2: Log-Log Plots for Power-Law Degree Distribution in Social

Media Networks. In these figures, the x-axis represents the logarithm
of the degree, and the y-axis represents the logarithm of the fraction of
individuals with that degree (i.e., log(pk )). The line demonstrates the linear
trend observed in log-log plots of power-law distributions.
Pick a popularity measure and compute it for the whole network.
For instance, we can take the number of friends in a social network
as a measure. We denote the measured value as k.
Compute pk , the fraction of individuals having popularity k.
Plot a log-log graph, where the x-axis represents ln k and the y-axis
represents ln pk .
If a power-law distribution exists, we should observe a straight line
in the plot.
Figure 4.2 depicts some log-log graphs for the number of friends on
real-world networks. In all networks, a linear trend is observed denoting
a power-law degree distribution.
108

Table 4.1: Average Local Clustering Coefficient in Real-World Networks

(from [46, 284, 198])
Web
0.081

Facebook
0.14 (with 100 friends)

Flickr
0.31

LiveJournal
0.33

Orkut
0.17

YouTube
0.13

Networks exhibiting power-law degree distribution are often called

scale-free networks. Since the majority of social networks are scale-free,
we are interested in models that can generate synthetic networks with a
power-law degree distribution.
Scale-free
Networks

4.1.2

Clustering Coefficient

In real-world social networks, friendships are highly transitive. In other

words, friends of an individual are often friends with one another. These
friendships form triads of friendships that are frequently observed in social networks. These triads result in networks with high average [local]
clustering coefficients. In May 2011, Facebook had an average clustering
coefficient of 0.5 for individuals who had two friends; their degree was
2 [284]. This indicates that for 50% of all users with two friends, their two
friends were also friends with each other. Table 4.1 provides the average
clustering coefficient for several real-world social networks and the web.

4.1.3

Average Path Length

In real-world networks, any two members of the network are usually

connected via short paths. In other words, the average path length is
small. This is known as the small-world phenomenon. In the well-known
small-world experiment conducted in the 1960s by Stanley Milgram, Milgram
conjectured that people around the world are connected to one another via
a path of at most six individuals (i.e., the six degrees of separation). Similarly,
we observe small average path lengths in social networks. For example,
in May 2011, the average path length between individuals in the Facebook
graph was 4.7. This average was 4.3 for individuals in the United States
at the same time [284]. Table 4.2 provides the average path length for
real-world social networks and the web.
109

Table 4.2: Average Path Length in Real-World Networks (from [46, 284,
198])
Web
16.12

Facebook
4.7

Flickr
5.67

LiveJournal
5.88

Orkut
4.25

YouTube
5.10

These three properties power-law degree distribution, high clustering

coefficient, and small average path length are consistently observed in
real-world networks. We design models based on simple assumptions on
how friendships are formed, hoping that these models generate scale-free
networks, with high clustering coefficient and small average path lengths.
We start with the simplest network model, the random graph model.

4.2

Random Graphs

We start with the most basic assumption on how friendships can be formed:
Edges (i.e., friendships) between nodes (i.e., individuals) are formed
randomly.
Small-world and Six
Degrees of Separation

G(n, p)

The random graph model follows this basic assumption. In reality

friendships in real-world networks are far from random. By assuming
random friendships, we simplify the process of friendship formation in
real-world networks, hoping that these random friendships ultimately
create networks that exhibit common characteristics observed in real-world
networks.
Formally, we
can assume that for a graph with a fixed number of nodes
n, any of the n2 edges can be formed independently, with probability p.
This graph is called a random graph and we denote it as the G(n, p) model.
This model was first proposed independently by Edgar Gilbert [100] and
Solomonoff and Rapoport [262]. Another way of randomly generating
graphs is to assume that both the number of nodes n and the number of
edges m are fixed. However,
we need to determine which m edges are
n

selected from the set of 2 possible edges. Let denote the set of graphs
with n nodes and m edges. To generate a random graph, we can uniformly
select one of the graphs in . The number of graphs with n nodes and m
110

edges (i.e., ||) is

|| =

n

2

!
.

(4.3)

1
. One can think
The uniform random graph selection probability is ||
of the probability of uniformly selecting a graph as an analog to p, the
probability of selecting an edge in G(n, p).
The second model was introduced by Paul Erdos and Alfred Renyi [83]
and is denoted as the G(n, m) model. In the limit,

both models act
similarly. G(n, m)
The expected number of edges in G(n, p) is n2 p. Now, if we set n2 p = m, in
the limit, both models act the same because they contain the same number
of edges. Note that the G(n, m) model contains a fixed number of edges;
however, the second model G(n, p) is likely to contain none or all possible
edges.
Mathematically, the G(n, p) model is almost always simpler to analyze;
hence the rest of this section deals with properties of this model. Note
that there exist many graphs with n nodes and m edges (i.e., generated by
G(n, m)). The same argument holds for G(n, p), and many graphs can be
generated by the model. Therefore, when measuring properties in random
graphs, the measures are calculated over all graphs that can be generated
by the model and then averaged. This is particularly useful when we are
interested in the average, and not specific, behavior of large graphs.
In G(n, p), the number of edges is not fixed; therefore, we first examine
some mathematical properties regarding the expected number of edges
that are connected to a node, the expected number of edges observed in
the graph, and the likelihood of observing m edges in a random graph
generated by the G(n, p) process.

Proposition 4.1. The expected number of edges connected to a node (expected

degree) in G(n, p) is (n 1)p.
Proof. A node can be connected to at most n 1 nodes (via n 1 edges). All
edges are selected independently with probability p. Therefore, on average
(n 1)p of them are selected. The expected degree is often denoted using
notation c or k in the literature. Since we frequently use k to denote degree
values, we use c to denote the expected degree of a random graph,
c = (n 1)p,
111

(4.4)

or equivalently,
p=

c
.
n1

(4.5)


Proposition 4.2. The expected number of edges in G(n, p) is

n

p.
2

Proof. Following the same line of argument,
because edges are selected
independently
and we have a maximum of n2 edges, the expected number


of edges is n2 p.

Proposition 4.3. In a graph generated by G(n, p) model, the probability of observing m edges is
!
n

n
2
P(|E| = m) =
pm (1 p)(2)m ,
(4.6)
m
which is a binomial distribution.


Proof. m edges are selected from the n2 possible edges. These edges are
formed with probability pm , and other edges are not formed (to guarantee
n
the existence of only m edges) with probability (1 p)(2)m .

Given these basic propositions, we next analyze how random graphs
evolve as we add edges to them.

4.2.1

Giant Component

Evolution of Random Graphs

In random graphs, when nodes form connections, after some time a large
fraction of nodes get connected (i.e., there is a path between any pair of
them). This large fraction forms a connected component, commonly called
the largest connected component or the giant component. We can tune the
behavior of the random graph model by selecting the appropriate p value.
In G(n, p), when p = 0, the size of the largest connected component is 0 (no
two pairs are connected), and when p = 1, the size is n (all pairs are connected). Table 4.3 provides the size of the largest connected component (slc
values in the table) for random graphs with 10 nodes and different p values.
The table also provides information on the average degree c, the diameter
size ds, the size of the largest component slc, and the average path length l
of the random graph.
As shown, in Table 4.3, as p gets larger, the graph gets denser. When p
is very small, the following is found:
112

Table 4.3: Evolution of Random Graphs. Here, p is the random graph

generation probability, c is the average degree, ds is the diameter size, slc
is the size of the largest component, and l is the average path length. The
highlighted column denotes phase transition in the random graph

p
c
ds
slc
l

0.0
0.0
0
0
0.0

0.055
0.8
2
4
1.5

0.11
1
6
7
2.66

1.0
9.0
1
10
1.0

1. No giant component is observed in the graph.

2. Small isolated connected components are formed.
3. The diameter is small because all nodes are in isolated components,
in which they are connected to a handful of other nodes.
As p gets larger, the following occurs:
1. A giant component starts to appear.
2. Isolated components become connected.
3. The diameter values increase.
At this point, nodes are connected to each other via long paths (see
p = 0.11 in Table 4.3). As p continues to get larger, the random graph
properties change again. For larger values, the diameter starts shrinking
as nodes get connected to each other via different paths (that are likely to
be shorter). The point where diameter value starts to shrink in a random
graph is called phase transition. At the point of phase transition, the following Phase Transition
two phenomena are observed:
113

Figure 4.3: Nodes Visited by Moving n-hops away in a Random Graph. c

denotes the expected node degree.
1. The giant component, which just started to appear, starts to grow.
2. The diameter, which just reached its maximum value, starts decreasing.
It is proven that in random graphs phase transition occurs when c = 1;
that is, p = 1/(n 1).
Proposition 4.4. In random graphs, phase transition happens at c = 1.
Proof. (Sketch) Consider a random graph with expected node degree c,
where c = p(n1). In this graph, consider any connected set of nodes S and
consider the complement set S = VS. For the sake of our proof, we assume
Given any node v in S, if we move one hop (edge) away
that |S|  |S|.
from v, we visit approximately c nodes. Following the same argument,
if we move one hop away from nodes in S, we visit approximately |S|c
and
nodes. Assuming |S| is small, the nodes in S only visit nodes in S,
when moving one hop away from S, the set of nodes guaranteed to be
connected gets larger by a factor c (see Figure 4.3). The connected set
of visited nodes gets c2 times larger when moving two hops and so on.
Now, in the limit, if we want this component of visited nodes to become
114

the largest connected component, then after traveling n hops, we must

have
cn 1 or equivalently c 1.
(4.7)
Otherwise (i.e., c < 1), the number of visited nodes dies out exponentially.
Hence, phase transition happens at c = 1.2

Note that this proof sketch provides an intuitive approach to understand the proposition. Interested readers can refer to the bibliographic
notes for a concrete proof.
So far we have discussed the generation and evolution of random
graphs; however, we also need to analyze how random graphs perform in
terms of mimicking properties exhibited by real-world networks. It turns
out that random graphs can model average path length in a real-world
network accurately, but fail to generate a realistic degree distribution or
clustering coefficient. We discuss these properties next.

4.2.2

Properties of Random Graphs

Degree Distribution
When computing degree distribution, we estimate the probability of observing P(dv = d) for node v.
Proposition 4.5. For a graph generated by G(n, p), node v has degree d, d n1,
with probability
!
n1 d
P(dv = d) =
p (1 p)n1d ,
(4.8)
d
which is again a binomial degree distribution.
Proof. The proof is left to the reader.3

This assumes that n is fixed. We can generalize this result by computing

the degree distribution of random graphs in the limit (i.e., n ). In this
Note that for c = 1, the component size is stable, and in the limit, no growth will be
observed. The phase transition happens exactly at c = 1.
3
Hint: The proof is similar to the proof provided for the likelihood of observing m
edges (Proposition 4.3).
2

115

case, using Equation 4.4 and the fact that limx0 ln(1 + x) = x, we can
compute the limit for each term of Equation 4.8:
n1d

lim (1 p)n1d = lim eln(1p)

= lim e(n1d) ln(1p)

n

= lim e(n1d) ln(1 n1 ) = lim e(n1d) n1 = ec .

n

(4.9)
We also have
!
(n 1)!
n1
lim
= lim
n
n (n 1 d)! d!
d
( (n 1) (n 2) (n d) )(n 1 d)!
= lim
n
(n 1 d)! d!
( (n 1) (n 2) (n d) )
= lim
n
d!
(n 1)d

.
(4.10)
d!
We can compute the degree distribution of random graphs in the limit
by substituting Equations 4.10, 4.9, and 4.4 in Equation 4.8,
!
n1 d
lim P(dv = d) = lim
p (1 p)n1d
n
n
d


d
(n 1)d
c d c
c c
=
e =e
,
(4.11)
d!
n1
d!
which is basically the Poisson distribution with mean c. Thus, in the limit,
random graphs generate Poisson degree distribution, which differs from
the power-law degree distribution observed in real-world networks.
Clustering Coefficient
Proposition 4.6. In a random graph generated by G(n, p), the expected local
clustering coefficient for node v is p.
Proof. The local clustering coefficient for node v is
C(v) =

number of connected pairs of vs neighbors

.
number of pairs of vs neighbors
116

(4.12)

However, v can have different degrees depending on the edges that are
formed randomly. Thus, we can compute the expected value for C(v):
E(C(v)) =

n1
X

E(C(v)|dv = d) P(dv = d).

(4.13)

d=0

The first term is basically the local clustering coefficient of a node given
its degree. For a random graph, we have
number of connected pairs of vs d neighbors
number of pairs of vs d neighbors
d

p2
=
= p.
(4.14)
d

E(C(v)|dv = d) =

Substituting Equation 4.14 in Equation 4.13, we get

E(C(v)) = p

d=n1
X

P(dv = d) = p,

(4.15)

d=0

where we have used the fact that all probability distributions sum up
to 1.

Proposition 4.7. The global clustering coefficient of a random graph generated
by G(n, p) is p.
Proof. The global clustering coefficient of a graph defines the probability
of two neighbors of the same node being connected. In random graphs,
for any two nodes, this probability is the same and is equal to the generation probability p that determines the probability of two nodes getting
connected. Note that in random graphs, the expected local clustering
coefficient is equivalent to the global clustering coefficient.

In random graphs, the clustering coefficient is equal to the probability
p; therefore, by appropriately selecting p, we can generate networks with
a high clustering coefficient. Note that selecting a large p is undesirable
because doing so will generate a very dense graph, which is unrealistic, as
in the real-world, networks are often sparse. Thus, random graphs are considered generally incapable of generating networks with high clustering
coefficients without compromising other required properties.
117

Average Path Length

Proposition 4.8. The average path length l in a random graph is
l

ln |V|
,
ln c

(4.16)

Proof. (Sketch) The proof is similar to the proof provided in determining

when phase transition happens (see Section 4.2.1). Let D denote the expected diameter size in the random graph. Starting with any node in a
random graph and its expected degree c, one can visit approximately c
nodes by traveling one edge, c2 nodes by traveling two edges, and cD
nodes by traveling diameter number of edges. After this step, almost all
nodes should be visited. In this case, we have
cD |V|.

(4.17)

In random graphs, the expected diameter size tends to the average path
length l in the limit. This we provide without proof. Interested readers can
refer to the bibliographic notes for pointers to concrete proofs. Using this
fact, we have
cD cl |V|.
(4.18)
. Therefore, the
Taking the logarithm from both sides we get l lnln|V|
c
ln |V|
average path length in a random graph is equal to ln c .


4.2.3

Modeling Real-World Networks with Random Graphs

Given a real-world network, we can simulate it using a random graph

model. We can compute the average degree c in the given network. From
c
c, the connection probability p can be computed (p = n1
). Using p and the
number of nodes in the given network n, a random graph model G(n, p)
can be simulated. Table 4.4 demonstrates the simulation results for various
real-world networks. As observed in the table, random graphs perform
well in modeling the average path lengths; however, when considering
the transitivity, the random graph model drastically underestimates the
clustering coefficient.
To tackle this issue, we study the small-world model.
118

Table 4.4: A Comparison between Real-World Networks and Simulated

Random Graphs. In this table, C denotes the average clustering coefficient.
The last two columns show the average path length and the clustering coefficient for the random graph simulated for the real-world network. Note
that average path lengths are modeled properly, whereas the clustering
coefficient is underestimated
Original Network
Average Average
Degree
Path
Length
225,226 61
3.65
1,520,251 18.1
4.6

Network

Size

Film Actors
Medline
Coauthorship
E.Coli
C.Elegans

282
282

4.3

7.35
14

2.9
2.65

Simulated Random Graph

Average
C
Path
Length
0.79 2.99
0.00027
0.56 4.91
1.8 104
C

0.32 3.04
0.28 2.25

0.026
0.05

Small-World Model

The assumption behind the random graph model is that connections in

real-world networks are formed at random. Although unrealistic, random
graphs can model average path lengths in real-world networks properly,
but underestimate the clustering coefficient. To mitigate this problem,
Duncan J. Watts and Steven Strogatz in 1997 proposed the small-world
model.
In real-world interactions, many individuals have a limited and often
at least, a fixed number of connections. Individuals connect with their parents, brothers, sisters, grandparents, and teachers, among others. Thus,
instead of assuming random connections, as we did in random graph models, one can assume an egalitarian model in real-world networks, where
people have the same number of neighbors (friends). This again is unrealistic; however, it models more accurately the clustering coefficient of
real-world networks. In graph theory terms, this assumption is equiva- Regular Ring Lattice
lent to embedding individuals in a regular network. A regular (ring) lattice
is a special case of regular networks where there exists a certain pattern
for how ordered nodes are connected to one another. In particular, in a
regular lattice of degree c, nodes are connected to their previous c/2 and
following c/2 neighbors. Formally, for node set V = {v1 , v2 , v3 , . . . , vn }, an
119

Figure 4.4: Regular Lattice of Degree 4.

edge exists between node vi and v j if and only if
0 < |i j| c/2.

(4.19)

A regular lattice of degree 4 is shown in Figure 4.4.

The regular lattice can model transitivity well; however, the average
path length is too high. Moreover, the clustering coefficient takes the value
3(c 2) 3
,
4(c 1) 4

(4.20)

which is fixed and not tunable to clustering coefficient values found in

real-world networks. To overcome these problems, the proposed smallworld model dynamically lies between the regular lattice and the random
network.
In the small-world model, we assume a parameter that controls randomness in the model. The model starts with a regular lattice and starts
adding random edges based on . The 0 1 controls how random the
model is. When is 0, the model is basically a regular lattice, and when
= 1, the model becomes a random graph.
120

Algorithm 4.1 Small-World Generation Algorithm

Require: Number of nodes |V|, mean degree c, parameter
1: return A small-world graph G(V, E)
2: G = A regular ring lattice with |V| nodes and degree c
3: for node vi (starting from v1 ), and all edges e(vi , v j ), i < j do
4:
vk = Select a node from V uniformly at random.
5:
if rewiring e(vi , v j ) to e(vi , vk ) does not create loops in the graph or
multiple edges between vi and vk then
6:
rewire e(vi , v j ) with probability : E = E{e(vi , v j )}, E = E{e(vi , vk )};
7:
end if
8: end for
9: Return G(V, E)
The procedure for generating small-world networks is outlined in Algorithm 4.1. The procedure creates new edges by a process called rewiring.
Rewiring replaces an existing edge between nodes vi and v j with a nonexisting edge between vi and vk with probability . In other words, an edge is
disconnected from one of its endpoints v j and connected to a new endpoint
vk . Node vk is selected uniformly.
The network generated using this procedure has some interesting properties. Depending on the value, it can have a high clustering coefficient
and also short average path lengths. The degree distribution, however,
still does not match that of real-world networks.

4.3.1

Properties of the Small-World Model

Degree Distribution
The degree distribution for the small-world model is as follows:
!
min(dc/2,c/2)
X
(c/2)dc/2n c/2
c/2
e
,
P(dv = d) =
(1 )n c/2n
(d

c/2

n)
n
n=0

(4.21)

where P(dv = d) is the probability of observing degree d for node v. We provide this equation without proof due to techniques beyond the scope of this
book (see Bibliographic Notes). Note that the degree distribution is quite
similar to the Poisson degree distribution observed in random graphs (Section 4.2.2). In practice, in the graph generated by the small-world model,
121

most nodes have similar degrees due to the underlying lattice. In contrast,
in real-world networks, degrees are distributed based on a power-law
distribution, where most nodes have small degrees and a few have large
degrees.
Clustering Coefficient
3(c2)

The clustering coefficient for a regular lattice is 4(c1) and for a random
c
. The clustering coefficient for a small-world netgraph model is p = n1
work is a value between these two, depending on . Commonly, the
clustering coefficient for a regular lattice is represented using C(0), and the
clustering coefficient for a small-world model with = p is represented as
C(p). The relation between the two values can be computed analytically; it
has been proven that
C(p) (1 p)3 C(0).
(4.22)
The intuition behind this relation is that because the clustering coefficient enumerates the number of closed triads in a graph, we are interested
in triads that are still left connected after the rewiring process. For a triad
to stay connected, all three edges must not be rewired with probability
(1 p). Since the process is performed independently for each edge, the
probability of observing triads is (1 p)3 times the probability of observing
them in a regular lattice. Note that we also need to take into account new
triads that are formed by the rewiring process; however, that probability
is nominal and hence negligible. The graph in Figure 4.5 depicts the value
C(p)
of C(0) for different values of p.
As shown in the figure, the value for C(p) stays high until p reaches 0.1
(10% rewired) and then decreases rapidly to a value around zero. Since
a high clustering coefficient is required in generated graphs, 0.1 is
preferred.
Average Path Length
The same procedure can be done for the average path length. The average
path length in a regular lattice is
n
.
2c
122

(4.23)

Figure 4.5: Clustering Coefficient and Average Path Length Change in

the Small-World Model (from [298]). In this figure, C(p)/C(0) denotes the
clustering coefficient of a small-world model, with = p, over the regular
lattice. Similarly, L(p)/L(0) denotes the average path length of a smallworld model over the regular lattice. Since models with a high clustering
coefficient and small average path length are desired, values in range
0.01 = p 0.1 are preferred.

We denote this value as L(0). The average path length in a random

n
. We denote L(p) as the average path length for a small-world
graph is ln
ln c
model where = p. Unlike C(p), no analytical formula for comparing
L(p) to L(0) exists; however, the relation can be computed empirically for
L(p)
different values of p. Similar to C(p), we plot L(0) in Figure 4.5. As shown
in the figure, the average path length decays sooner than the clustering
coefficient and becomes stable when around 1% of edges are rewired. Since
we require small average path lengths in the generated graphs, 0.01 is
preferred.
123

Table 4.5: A Comparison between Real-World Networks and Simulated

Graphs Using the Small-World Model. In this table C denotes the average
clustering coefficient. The last two columns show the average path length
and the clustering coefficient for the small-world graph simulated for the
real-world network. Both average path lengths and clustering coefficients
are modeled properly
Original Network
Average Average
Degree
Path
Length
225,226 61
3.65
1,520,251 18.1
4.6

Network

Size

Film Actors
Medline
Coauthorship
E.Coli
C.Elegans

282
282

4.3.2

7.35
14

2.9
2.65

Simulated Graph
Average
C
Path
Length
0.79 4.2
0.73
0.56 5.1
0.52
C

0.32 4.46
0.28 3.49

0.31
0.37

Modeling Real-World Networks with the Small-World

Model

A desirable model for a real-world network should generate graphs with

high clustering coefficients and short average path lengths. As shown
in Figure 4.5, for 0.01 0.10, the small-world network generated is
acceptable, in which the average path length is small and the clustering coefficient is still high. Given a real-world network in which average degree
c and clustering coefficient C are given, we set C(p) = C and determine
using Equation 4.22. Given , c, and n (size of the real-world network), we
can simulate the small-world model.
Table 4.5 demonstrates the simulation results for various real-world
networks. As observed in the table, the small-world model generates a
realistic clustering coefficient and small average path length. Note that the
small-world model is still incapable of generating a realistic degree distribution in the simulated graph. To generate scale-free networks (i.e., with a
power-law degree distribution), we introduce the preferential attachment
model next.
124

Algorithm 4.2 Preferential Attachment

Require: Graph G(V0 , E0 ), where |V0 | = m0 and dv 1 v V0 , number of
expected connections m m0 , time to run the algorithm t
1: return A scale-free network
2: //Initial graph with m0 nodes with degrees at least 1
3: G(V, E) = G(V0 , E0 );
4: for 1 to t do
5:
V = V {vi }; // add new node vi
6:
while di , m do
7:
Connect vi to a random node v j V, i , j ( i.e., E = E {e(vi , v j )} )
with probability P(v j ) =
8:
end while
9: end for
10: Return G(V, E)

4.4

dj
P
.
k dk

Preferential Attachment Model

There exist a variety of scale-free network-modeling algorithms. A wellestablished one is the model proposed by Barabasi and Albert [24]. The
model is called preferential attachment or sometimes the Barabasi-Albert
(BA) model and is as follows:
When new nodes are added to networks, they are more likely to connect
to existing nodes that many others have connected to.
This connection likelihood is proportional to the degree of the node
that the new node is aiming to connect to. In other words, a rich-getricher phenomenon or aristocrat network is observed where the higher the
nodes degree, the higher the probability of new nodes getting connected
to it. Unlike random graphs in which we assume friendships are formed
randomly, in the preferential attachment model we assume that individuals
are more likely to befriend gregarious others. The models algorithm is
provided in Algorithm 4.2.
The algorithm starts with a graph containing a small set of nodes m0
and then adds new nodes one at a time. Each new node gets to connect
to m m0 other nodes, and each connection to existing node vi depends
on the degree of vi (i.e., P(vi ) = Pdid j ). Intrinsically, higher degree nodes get
j

125

more attention from newly added nodes. Note that the initial m0 nodes
must have at least degree 1 for probability P(vi ) = Pdid j to be nonzero.
j

The model incorporates two ingredients (1) the growth element and
(2) the preferential attachment element to achieve a scale-free network.
The growth is realized by adding nodes as time goes by. The preferential attachment is realized by connecting to node vi based on its degree
probability, P(vi ) = Pdid j . Removing any one of these ingredients generj
ates networks that are not scale-free (see Exercises). Next, we show that
preferential attachment models are capable of generating networks with a
power-law degree distribution. They are also capable of generating small
average path length, but unfortunately fail to generate the high clustering
coefficients observed in real-world networks.

4.4.1

Properties of the Preferential Attachment Model

Degree Distribution
We first demonstrate that the preferential attachment model generates
scale-free networks and can therefore model real-world networks. Empirical evidence found by simulating the preferential attachment model
suggests that this model generates a scale-free network with exponent
b = 2.9 0.1 [24]. Theoretically, a mean-field [213] proof can be provided as
follows.
Let di denote the degree for node vi . The probability of an edge connecting from a new node to vi is
di
P(vi ) = P .
j dj

(4.24)

The expected increase to the degree of vi is proportional to di (this is

true on average). Assuming a mean-field setting, the expected temporal
change in di is
ddi
mdi
mdi
di
= mP(vi ) = P
=
= .
(4.25)
dt
2mt 2t
j dj
Note that at each time step, m edges are
P added; therefore, mt edges
are added over time, and the degree sum j d j is 2mt. Rearranging and
126

solving this differential equation, we get

di (t) = m

 0.5
t
.
ti

(4.26)

Here, ti represents the time that vi was added to the network, and
because we set the expected degree to m in preferential attachment, then
di (ti ) = m.
The probability that di is less than d is
P(di (t) < d) = P(ti > m2 t/d2 ).

(4.27)

Assuming uniform intervals of adding nodes,

P(ti > m2 t/d2 ) = 1 P(ti m2 t/d2 ) = 1

m2 t
1
.
2
d (t + m0 )

(4.28)

1
The factor (t+m
shows the probability that one time step has passed
0)
because, at the end of the simulation, t + m0 nodes are in the network. The
probability density for P(d)

P(d) =

P(di (t) < d)

,
d

(4.29)

is what we are interested in, which, when solved, gives

P(d) =

2m2 t
d3 (t + m0 )

and the stationary solution (t ),

P(d) =

2m2
,
d3

(4.30)

which is a power-law degree distribution with exponent b = 3. Note that in

real-world networks, the exponent varies in a range (e.g., [2, 3]); however,
there is no variance in the exponent of the introduced model. To overcome
this issue, several other models are proposed. Interested readers can refer
to the bibliographical notes for further references.
127

Clustering Coefficient
In general, not many triangles are formed by the Barabasi-Albert model,
because edges are created independently and one at a time. Again, using
a mean-field analysis, the expected clustering coefficient can be calculated
as
m0 1 (ln t)2
,
(4.31)
C=
8
t
where t is the time passed in the system during the simulation. We avoid
the details of this calculation due to techniques beyond the scope of this
book. Unfortunately, as time passes, the clustering coefficient gets smaller
and fails to model the high clustering coefficient observed in real-world
networks.
Average Path Length
The average path length of the preferential attachment model increases
logarithmically with the number of nodes present in the network:
l

ln |V|
.
ln(ln |V|)

(4.32)

This indicates that, on average, preferential attachment models generate

shorter path lengths than random graphs. Random graphs are considered
accurate in approximating the average path lengths. The same holds for
preferential attachment models.

4.4.2

Modeling Real-World Networks with the Preferential

Attachment Model

As with random graphs, we can simulate real-world networks by generating a preferential attachment model by setting the expected degree m (see
Algorithm 4.2). Table 4.6 demonstrates the simulation results for various
real-world networks. The preferential attachment model generates a realistic degree distribution and, as observed in the table, small average path
lengths; however, the generated networks fail to exhibit the high clustering
coefficient observed in real-world networks.

128

Table 4.6: A Comparison between Real-World Networks and Simulated

Graphs using Preferential Attachment. C denotes the average clustering
coefficient. The last two columns show the average path length and the
clustering coefficient for the preferential-attachment graph simulated for
the real-world network. Note that average path lengths are modeled properly, whereas the clustering coefficient is underestimated
Original Network
Average Average
Degree
Path
Length
225,226 61
3.65
1,520,251 18.1
4.6

Network

Size

Film Actors
Medline
Coauthorship
E.Coli
C.Elegans

282
282

4.5

7.35
14

2.9
2.65

Simulated Graph
Average
C
Path
Length
0.79 4.90
0.005
0.56 5.36
0.0002
C

0.32 2.37
0.28 1.99

0.03
0.05

Summary

In this chapter, we discussed three well-established models that generate networks with commonly observed characteristics of real-world networks: random graphs, the small-world model, and preferential attachment. Random graphs assume that connections are completely random.
We discussed two variants of random graphs: G(n, p) and G(n, m). Random
graphs exhibit a Poisson degree distribution, a small clustering coefficient
.
p, and a realistic average path length lnln|V|
c
The small-world model assumes that individuals have a fixed number
of connections in addition to random connections. This model generates
networks with high transitivity and short path lengths, both commonly
observed in real-world networks. Small-world models are created through
a process where a parameter controls how edges are randomly rewired
from an initial regular ring lattice. The clustering coefficient of the model
is approximately (1 p)3 times the clustering coefficient of a regular lattice.
No analytical solution to approximate the average path length with respect
to a regular ring lattice has been found. Empirically, when between 1%
to 10% of edges are rewired (0.01 0.1), the model resembles many
real-world networks. Unfortunately, the small-world model generates a
degree distribution similar to the Poisson degree distribution observed in
random graphs.
129

Finally, the preferential attachment model assumes that friendship formation likelihood depends on the number of friends individuals have. The
model generates a scale-free network; that is, a network with a power-law
degree distribution. When k denotes the degree of a node, and pk the fraction of nodes having degree k, then in a power-law degree distribution,
pk = akb .

(4.33)

Networks created using a preferential attachment model have a

power-law degree distribution with exponent b = 2.9 0.1. Using a meanfield approach, we proved that this model has a power-law degree distribution. The preferential attachment model also exhibits realistic average
path lengths that are smaller than the average path lengths in random
graphs. The basic caveat of the model is that it generates a small clustering coefficient, which contradicts high clustering coefficients observed in
real-world networks.

130

4.6

Bibliographic Notes

General reviews of the topics in this chapter can be found in [213, 212, 28,
134].
Initial random graph papers can be found in the works of Paul Erdos
and Alfred Renyi [83, 84, 85] as well as Edgar Gilbert [100] and Solomonoff
and Rapoport [262]. As a general reference, readers can refer to [41, 217,
210]. Random graphs described in this chapter did not have any specific
degree distribution; however, random graphs can be generated with a
specific degree distribution. For more on this refer to [212, 216].
Small-worlds were first noticed in a short story by Hungarian writer F.
Karinthy in 1929. Works of Milgram in 1969 and Kochen and Pool in 1978
treated the subject more systematically. Milgram designed an experiment
in which he asked random participants in Omaha, Nebraska, or Wichita,
Kansas, to help send letters to a target person in Boston. Individuals were
only allowed to send the letter directly to the target person if they knew
the person on a first-name basis. Otherwise, they had to forward it to
someone who was more likely to know the target. The results showed that
the letters were on average forwarded 5.5 to 6 times until they reached the
target in Boston. Other recent research on small-world model dynamics
can be found in [295, 296].
Price [1965, 1976] was among the first who described power laws
observed in citation networks and models capable of generating them.
Power-law distributions are commonly found in social networks and the
web [87, 198]. The first developers of preferential attachment models were
Yule [308], who described these models for generating power-law distributions in plants, and Herbert A. Simon [260], who developed these models
for describing power laws observed in various phenomena: distribution
of words in prose, scientists by citations, and cities by population, among
others. Simon used what is known as the master equation to prove that
preferential attachment models generate power-law degree distributions.
A more rigorous proof for estimating the power-law exponent of the preferential attachment model using the master equation method can be found
in [212]. The preferential attachment model introduced in this chapter has
a fixed exponent b = 3, but, as mentioned, real-world networks have exponents in the range [2, 3]. To solve this issue, extensions have been proposed
in [155, 9].

131

4.7

Exercises

Properties of Real-World Networks

1. A scale invariant function f (.) is one such that, for a scalar ,
f (x) = c f (x),

(4.34)

for some constant c. Prove that the power-law degree distribution is

scale invariant.

Random Graphs
2. Assuming that we are interested in a sparse random graph, what
should we choose as our p value?
3. Construct a random graph as follows. Start with n nodes and a
given k. Generate all the possible combinations of k nodes. For

each combination, create a k-cycle with probability n1

, where is a
(2)
constant.
Calculate the node mean degree and the clustering coefficient.
What is the node mean degree if you create a complete graph
instead of the k-cycle?
4. When does phase transition happen in the evolution of random
graphs? What happens in terms of changes in network properties at
that time?

Small-World Model
5. Show that in a regular lattice the number of connections between
neighbors is given by 38 c(c 2), where c is the average degree.
6. Show how the clustering coefficient can be computed in a regular
lattice of degree k.
132

7. Why are random graphs incapable of modeling real-world graphs?

What are the differences between random graphs, regular lattices,
and small-world models?
8. Compute the average path length in a regular lattice.

Preferential Attachment Model

9. As a function of k, what fraction of pages on the web have k inlinks, assuming that a normal distribution governs the probability
of webpages choosing their links? What if we have a power-law
distribution instead?
10. In the Barabasi-Albert model (BA) two elements are considered:
growth and preferential attachment. The growth (G) is added to
the model by allowing new nodes to connect via m edges. The
preferential attachment (A) is added by weighting the probability of
connection by the degree. For the sake of brevity, we will consider
the model as BA = A + G. Now, consider models that only have one
element: G, or A, and not both. In the G model, the probability of
1
), and in A, the number of nodes reconnection is uniform (P = m0 +t1
main the same throughout the simulation and no new node is added.
In A, at each time step, a node within the network is randomly selected based on degree probability and then connected to another
one within the network.
Compute the degree distribution for these two models.
Determine if these two models generate scale-free networks.
What does this prove?

133

134

This chapter is from Social Media Mining: An Introduction.

By Reza Zafarani, Mohammad Ali Abbasi, and Huan Liu.
Cambridge University Press, 2014. Draft version: April 20, 2014.
Complete Draft and Slides Available at: http://dmml.asu.edu/smm

Chapter

Data Mining Essentials

Mountains of raw data are generated daily by individuals on social media.
Around 6 billion photos are uploaded monthly to Facebook, the blogosphere doubles every five months, 72 hours of video are uploaded every
minute to YouTube, and there are more than 400 million daily tweets on
Twitter. With this unprecedented rate of content generation, individuals
are easily overwhelmed with data and find it difficult to discover content
that is relevant to their interests. To overcome these challenges, we need
tools that can analyze these massive unprocessed sources of data (i.e., raw
data) and extract useful patterns from them. Examples of useful patterns
in social media are those that describe online purchasing habits or individuals website visit duration. Data mining provides the necessary tools
for discovering patterns in data. This chapter outlines the general process
for analyzing social media data and ways to use data mining algorithms
in this process to extract actionable patterns from raw data.
The process of extracting useful patterns from raw data is known as
Knowledge discovery in databases (KDD). It is illustrated in Figure 5.1. The Knowledge Discovery
KDD process takes raw data as input and provides statistically significant in Databases (KDD)
patterns found in the data (i.e., knowledge) as output. From the raw data, a
subset is selected for processing and is denoted as target data. Target data
is preprocessed to make it ready for analysis using data mining algorithm.
Data mining is then performed on the preprocessed (and transformed)
data to extract interesting patterns. The patterns are evaluated to ensure
their validity and soundness and interpreted to provide insights into the
data.
135

Figure 5.1: Knowledge Discovery in Databases (KDD) process.

In social media mining, the raw data is the content generated by individuals, and the knowledge encompasses the interesting patterns observed
in this data. For example, for an online book seller, the raw data is the list
of books individuals buy, and an interesting pattern could describe books
that individuals often buy.
To analyze social media, we can either collect this raw data or use
available repositories that host collected data from social media sites.1
When collecting data, we can either use APIs provided by social media
sites for data collection or scrape the information from those sites. In
either case, these sites are often networks of individuals where one can
perform graph traversal algorithms to collect information from them. In
other words, we can start collecting information from a subset of nodes on
a social network, subsequently collect information from their neighbors,
and so on. The data collected this way needs to be represented in a unified
format for analysis. For instance, consider a set of tweets in which we are
looking for common patterns. To find patterns in these tweets, they need
to be first represented using a consistent data format. In the next section,
we discuss data, its representation, and its types.

See [310] for a repository of network data.

136

5.1

Data

In the KDD process, data is represented in a tabular format. Consider the

example of predicting whether an individual who visits an online book
seller is going to buy a specific book. This prediction can be performed
by analyzing the individuals interests and previous purchase history. For
instance, when John has spent a lot of money on the site, has bought similar
books, and visits the site frequently, it is likely for John to buy that specific
book. John is an example of an instance. Instances are also called points,
data points, or observations. A dataset consists of one or more instances:
Instance,

Name

John
Mary

Attributes
Money Spent Bought Similar
High
Yes
High
Yes

Visits
Frequently
Rarely

Class
Will Buy
?
Yes

A dataset is represented using a set of features, and an instance is represented using values assigned to these features. Features are also known
as measurements or attributes. In this example, the features are Name, Money
Spent, Bought Similar, and Visits; feature values for the first instance
are John, High, Yes, and Frequently. Given the feature values for one
instance, one tries to predict its class (or class attribute) value. In our example, the class attribute is Will Buy, and our class value prediction for first
instance is Yes. An instance such as John in which the class attribute value
is unknown is called an unlabeled instance. Similarly, a labeled instance
is an instance in which the class attribute value in known. Mary in this
dataset represents a labeled instance. The class attribute is optional in a
dataset and is only necessary for prediction or classification purposes. One
can have a dataset in which no class attribute is present, such as a list of
customers and their characteristics.
There are different types of features based on the characteristics of the
feature and the values they can take. For instance, Money Spent can be
represented using numeric values, such as $25. In that case, we have a
continuous feature, whereas in our example it is a discrete feature, which can
take a number of ordered values: {High, Normal, Low}.
Different types of features were first introduced by psychologist Stanley
Smith Stevens [265] as levels of measurement in the theory of scales. He
137

Point,
Data Point, or Observation

Features,
Measurements, or
Attributes

Labeled and
Unlabeled

Levels of
Measurement

claimed that there are four types of features. For each feature type, there
exists a set of permissible operations (statistics) using the feature values
and transformations that are allowed.
Nominal (categorical). These features take values that are often
represented as strings. For instance, a customers name is a nominal
feature. In general, a few statistics can be computed on nominal
features. Examples are the chi-square statistic (2 ) and the mode (most
common feature value). For example, one can find the most common
first name among customers. The only possible transformation on
the data is comparison. For example, we can check whether our
customers name is John or not. Nominal feature values are often
presented in a set format.
Ordinal. Ordinal features lay data on an ordinal scale. In other
words, the feature values have an intrinsic order to them. In our
example, Money Spent is an ordinal feature because a High value for
Money Spent is more than a Low one.
Interval. In interval features, in addition to their intrinsic ordering, differences are meaningful whereas ratios are meaningless. For
interval features, addition and subtraction are allowed, whereas multiplications and division are not. Consider two time readings: 6:16
PM and 3:08 PM. The difference between these two time readings is
meaningful (3 hours and 8 minutes); however, there is no meaning
PM
, 2.
to 6:16
3:08 PM
Ratio. Ratio features, as the name suggests, add the additional properties of multiplication and division. An individuals income is an
example of a ratio feature where not only differences and additions
are meaningful but ratios also have meaning (e.g., an individuals
income can be twice as much as Johns income).
In social media, individuals generate many types of nontabular data,
such as text, voice, or video. These types of data are first converted to tabular data and then processed using data mining algorithms. For instance,
voice can be converted to feature values using approximation techniques
such as the fast Fourier transform (FFT) and then processed using data
mining algorithms. To convert text into the tabular format, we can use a
138

process denoted as vectorization. A variety of vectorization methods exist.

A well-known method for vectorization is the vector-space model introduced
by Salton, Wong, and Yang [243].
Vectorization
Vector Space Model
In the vector space model, we are given a set of documents D. Each
document is a set of words. The goal is to convert these textual documents
to [feature] vectors. We can represent document i with vector di ,
di = (w1,i , w2,i , . . . , wN,i ),

(5.1)

where w j,i represents the weight for word j that occurs in document i and
N is the number of words used for vectorization.2 To compute w j,i , we can
set it to 1 when the word j exists in document i and 0 when it does not.
We can also set it to the number of times the word j is observed in document i. A more generalized approach is to use the term frequency-inverse
document frequency (TF-IDF) weighting scheme. In the TF-IDF scheme, w j,i
is calculated as
w j,i = t f j,i id f j ,
(5.2)
where t f j,i is the frequency of word j in document i. id f j is the inverse TF-IDF
frequency of word j across all documents,
id f j = log2

|D|
,
|{document D | j document}|

(5.3)

which is the logarithm of the total number of documents divided by the

number of documents that contain word j. TF-IDF assigns higher weights
to words that are less frequent across documents and, at the same time,
have higher frequencies within the document they are used. This guarantees that words with high TF-IDF values can be used as representative
examples of the documents they belong to and also, that stop words,
such as the, which are common in all documents, are assigned smaller
weights.
Example 5.1. Consider the words apple and orange that appear 10 and 20
times in document d1 . Let |D| = 20 and assume the word apple only appears in
2

One can use all unique words in all documents (D) or a more frequent subset of words
in the documents for vectorization.

139

document d1 and the word orange appears in all 20 documents. Then, TF-IDF
values for apple and orange in document d1 are
t f id f (apple, d1 ) = 10 log2

20
= 43.22,
1

20
= 0.
20
Example 5.2. Consider the following three documents:
t f id f (orange, d1 ) = 20 log2

(5.4)
(5.5)

d1 = social media mining

d2 = social media data
d3 = financial market data

(5.6)
(5.7)
(5.8)

The t f values are as follows:

d1
d2
d3

social
1
1
0

media
1
1
0

mining
1
0
0

data
0
1
1

financial
0
0
1

market
0
0
1

The id f values are

id fsocial
id fmedia
id fmining
id fdata
id ffinancial
id fmarket

=
=
=
=
=
=

log2 (3/2) = 0.584

log2 (3/2) = 0.584
log2 (3/1) = 1.584
log2 (3/2) = 0.584
log2 (3/1) = 1.584
log2 (3/1) = 1.584.

(5.9)
(5.10)
(5.11)
(5.12)
(5.13)
(5.14)

The TF-IDF values can be computed by multiplying t f values with the id f

values:
d1
d2
d3

social
0.584
0.584
0

media
0.584
0.584
0

mining
1.584
0
0

data
0
0.584
0.584

financial
0
0
1.584

market
0
0
1.584

After vectorization, documents are converted to vectors, and common

data mining algorithms can be applied. However, before that can occur,
the quality of data needs to be verified.
140

5.1.1

Data Quality

When preparing data for use in data mining algorithms, the following four
data quality aspects need to be verified:
1. Noise is the distortion of the data. This distortion needs to be removed or its adverse effect alleviated before running data mining
algorithms because it may adversely affect the performance of the algorithms. Many filtering algorithms are effective in combating noise
effects.
2. Outliers are instances that are considerably different from other instances in the dataset. Consider an experiment that measures the
average number of followers of users on Twitter. A celebrity with
many followers can easily distort the average number of followers
per individuals. Since the celebrities are outliers, they need to be removed from the set of individuals to accurately measure the average
number of followers. Note that in special cases, outliers represent
useful patterns, and the decision to removing them depends on the
context of the data mining problem.
3. Missing Values are feature values that are missing in instances. For
example, individuals may avoid reporting profile information on social media sites, such as their age, location, or hobbies. To solve
this problem, we can (1) remove instances that have missing values,
(2) estimate missing values (e.g., replacing them with the most common value), or (3) ignore missing values when running data mining
algorithms.
4. Duplicate data occurs when there are multiple instances with the
exact same feature values. Duplicate blog posts, duplicate tweets,
or profiles on social media sites with duplicate information are all
instances of this phenomenon. Depending on the context, these instances can either be removed or kept. For example, when instances
need to be unique, duplicate instances should be removed.
After these quality checks are performed, the next step is preprocessing
or transformation to prepare the data for mining.
141

5.2

Data Preprocessing

Often, the data provided for data mining is not immediately ready. Data
preprocessing (and transformation in Figure 5.1) prepares the data for
mining. Typical data preprocessing tasks are as follows:
1. Aggregation. This task is performed when multiple features need
to be combined into a single one or when the scale of the features
change. For instance, when storing image dimensions for a social
media website, one can store by image width and height or equivalently store by image area (width height). Storing image area saves
storage space and tends to reduce data variance; hence, the data has
higher resistance to distortion and noise.
2. Discretization. Consider a continuous feature such as money spent
in our previous example. This feature can be converted into discrete
values High, Normal, and Low by mapping different ranges to
different discrete values. The process of converting continuous features to discrete ones and deciding the continuous range that is being
assigned to a discrete value is called discretization.
3. Feature Selection. Often, not all features gathered are useful. Some
may be irrelevant, or there may be a lack of computational power
to make use of all the features, among many other reasons. In these
cases, a subset of features are selected that could ideally enhance the
performance of the selected data mining algorithm. In our example,
customers name is an irrelevant feature to the value of the class
attribute and the task of predicting whether the individual will buy
the given book or not.
4. Feature Extraction. In contrast to feature selection, feature extraction
converts the current set of features to a new set of features that can
perform the data mining task better. A transformation is performed
on the data, and a new set of features is extracted. The example
we provided for aggregation is also an example of feature extraction
where a new feature (area) is constructed from two other features
(width and height).
5. Sampling. Often, processing the whole dataset is expensive. With
the massive growth of social media, processing large streams of data
142

is nearly impossible. This motivates the need for sampling. In sampling, a small random subset of instances are selected and processed
instead of the whole data. The selection process should guarantee
that the sample is representative of the distribution that governs the
data, thereby ensuring that results obtained on the sample are close
to ones obtained on the whole dataset. The following are three major
sampling techniques:
Random sampling. In random sampling, instances are selected
uniformly from the dataset. In other words, in a dataset of size
n, all instances have equal probability n1 of being selected. Note
that other probability distributions can also be used to sample
the dataset, and the distribution can be different from uniform.
Sampling with or without replacement. In sampling with replacement, an instance can be selected multiple times in the sample. In sampling without replacement, instances are removed
from the selection pool once selected.
Stratified sampling. In stratified sampling, the dataset is first
partitioned into multiple bins; then a fixed number of instances
are selected from each bin using random sampling. This technique is particularly useful when the dataset does not have a
uniform distribution for class attribute values (i.e., class imbalance). For instance, consider a set of 10 females and 5 males. A
sample of 5 females and 5 males can be selected using stratified
sampling from this set.
In social media, a large amount of information is represented in
network form. These networks can be sampled by selecting a subset
of their nodes and edges. These nodes and edges can be selected
using the aforementioned sampling methods. We can also sample
these networks by starting with a small set of nodes (seed nodes) and
sample
(a) the connected components they belong to;
(b) the set of nodes (and edges) connected to them directly; or
(c) the set of nodes and edges that are within n-hop distance from
them.
143

After preprocessing is performed, the data is ready to be mined. Next,

we discuss two general categories of data mining algorithms and how each
can be evaluated.

5.3

Data Mining Algorithms

Data mining algorithms can be divided into several categories. Here, we

discuss two well-established categories: supervised learning and unsupervised learning. In supervised learning, the class attribute exists, and the
task is to predict the class attribute value. Our previous example of predicting the class attribute will buy is an example of supervised learning.
In unsupervised learning, the dataset has no class attribute, and our task
is to find similar instances in the dataset and group them. By grouping
these similar instances, one can find significant patterns in a dataset. For
example, unsupervised learning can be used to identify events on Twitter, because the frequency of tweeting is different for various events. By
using unsupervised learning, tweets can be grouped based on the times
at which they appear and hence, identify the tweets corresponding realworld events. Other categories of data mining algorithms exist; interested
readers can refer to the bibliographic notes for pointers to these categories.

5.4

Supervised Learning

The first category of algorithms, supervised learning algorithms, are those

for which the class attribute values for the dataset are known before running the algorithm. This data is called labeled data or training data. Instances in this set are tuples in the format (x, y), where x is a vector and
y is the class attribute, commonly a scalar. Supervised learning builds a
model that maps x to y. Roughly, our task is to find a mapping m(.) such
that m(x) = y. We are also given an unlabeled dataset or test dataset, in
which instances are in the form (x, ?) and y values are unknown. Given m(.)
learned from training data and x of an unlabeled instance, we can compute m(x), the result of which is prediction of the label for the unlabeled
instance.
Consider the task of detecting spam emails. A set of emails is given
where users have manually identified spam versus non-spam (training
144

Figure 5.2: Supervised Learning.

data). Our task is to use a set of features such as words in the email (x) to
identify the spam/non-spam status (y) of unlabeled emails (test data). In
this case, y = {spam, non-spam}.
Supervised learning can be divided into classification and regression.
When the class attribute is discrete, it is called classification; when the
class attribute is continuous, it is regression. We introduce classification
methods such as decision tree learning, naive Bayes classifier, k-nearest neighbor
classifier, and classification with network information and regression methods
such as linear regression and logistic regression. We also introduce how
supervised learning algorithms are evaluated. Before we delve into supervised learning techniques, we briefly discuss the systematic process of a
supervised learning algorithm.
This process is depicted in Figure 5.2. It starts with a training set
(i.e., labeled data) where both features and labels (class attribute values)
are known. A supervised learning algorithm is run on the training set
in a process known as induction. In the induction process, the model is
generated. The model maps the feature values to the class attribute values.
The model is used on a test set in which the class attribute value is unknown
to predict these unknown class attribute values (deduction process).

5.4.1

Decision Tree Learning

Consider the dataset shown in Table 5.1. The last attribute represents the
class attribute, and the other attributes represent the features. In decision
tree classification, a decision tree is learned from the training dataset, and
145

Table 5.1: A Sample Dataset. In this dataset, features are characteristics

of individuals on Twitter, and the class attribute denotes whether they are
influential or not
ID
Celebrity
Verified Account
# Followers
Influential?
1
Yes
No
1.25 M
No
2
No
Yes
1M
No
3
No
Yes
600 K
No
4
Yes
Unknown
2.2 M
No
5
No
No
850 K
Yes
6
No
Yes
750 K
No
7
No
No
900 K
Yes
8
No
No
700 K
No
9
Yes
Yes
1.2 M
No
10
No
Unknown
950 K
Yes
that tree is later used to predict the class attribute value for instances
in the test dataset. As an example, two learned decision trees from the
dataset shown in Table 5.1 are provided in Figure 5.3. As shown in this
figure, multiple decision trees can be learned from the same dataset, and
these decision trees can both correctly predict the class attribute values
for all instances in the dataset. Construction of decision trees is based on
heuristics, as different heuristics generate different decision trees from the
same dataset.
Decision trees classify examples based on their feature values. Each

Figure 5.3: Decision Trees Learned from Data Provided in Table 5.1.
146

nonleaf node in a decision tree represents a feature, and each branch represents a value that the feature can take. Instances are classified by following
a path that starts at the root node and ends at a leaf by following branches
based on instance feature values. The value of the leaf determines the class
attribute value predicted for the instance (see Figure 5.3).
Decision trees are constructed recursively from training data using a
top-down greedy approach in which features are sequentially selected. In
Figure 5.3(a), the feature selected for the root node is Celebrity. After
selecting a feature for each node, based on its values, different branches
are created: For Figure 5.3(a), since the Celebrity feature can only take
either Yes or No, two branches are created: one labeled Yes and one labeled
No. The training set is then partitioned into subsets based on the feature
values, each of which fall under the respective feature value branch; the
process is continued for these subsets and other nodes. In Figure 5.3(a),
instances 1, 4, and 9 from Table 5.1 represent the subset that falls under the
Celebrity=Yes branch, and the other instances represent the subset that
falls under the Celebrity=No branch.
When selecting features, we prefer features that partition the set of
instances into subsets that are more pure. A pure subset has instances
that all have the same class attribute value. In Figure 5.3(a), the instances
that fall under the left branch of the root node (Celebrity=Yes) form a
pure subset in which all instances have the same class attribute value
Influential?=No. When reaching pure subsets under a branch, the decision tree construction process no longer partitions the subset, creates a leaf
under the branch, and assigns the class attribute value for subset instances
as the leafs predicted class attribute value. In Figure 5.3(a), the instances
that fall under the right branch of the root node form an impure dataset;
therefore, further branching is required to reach pure subsets. Purity of
subsets can be determined with different measures. A common measure
of purity is entropy. Over a subset of training instances, T, with a binary
class attribute (values {+, }), the entropy of T is defined as
entropy(T) = p+ log p+ p log p ,

(5.15)

where p+ is the proportion of instances with + class attribute value in T

and p is the proportion of instances with class attribute value.
Example 5.3. Assume that there is a subset T, containing 10 instances. Seven
instances have a positive class attribute value, and three instances have a negative
147

class attribute value [7+, 3]. The entropy for subset T is

entropy(T) =

7
3
3
7
log

log
= 0.881.
10
10 10
10

(5.16)

Note that if the number of positive and negative instances in the set are equal
(p+ = p = 0.5), then the entropy is 1.
In a pure subset, all instances have the same class attribute value and
the entropy is 0. If the subset being measured contains an unequal number
of positive and negative instances, the entropy is between 0 and 1.

5.4.2

Naive Bayes Classifier

Among many methods that use the Bayes theorem, the naive Bayes classifier (NBC) is the simplest. Given two random variables X and Y, Bayes
theorem states that
P(X|Y)P(Y)
.
(5.17)
P(Y|X) =
P(X)
In NBC, Y represents the class variable and X represents the instance
features. Let X be (x1 , x2 , x3 , . . . , xm ), where xi represents the value of feature
i. Let {y1 , y2 , . . . , yn } represent the values the class attribute Y can take.
Then, the class attribute value of instance X can be calculated by measuring
arg max P(yi |X).
yi

(5.18)

Based on the Bayes theorem,

P(yi |X) =

P(X|yi )P(yi )
.
P(X)

(5.19)

Note that P(X) is constant and independent of yi , so we can ignore the

denominator of Equation 5.19 when maximizing Equation 5.18. The NBC
also assumes conditional independence to make the calculations easier;
that is, given the class attribute value, other feature attributes become
conditionally independent. This condition, though unrealistic, performs
well in practice and greatly simplifies calculation.
P(X|yi ) = mj=1 P(x j |yi ).
148

(5.20)

No.
1
2
3
4
5
6
7
8

Table 5.2: Naive Bayes Classifier (NBC) Toy Dataset

Outlook (O) Temperature (T) Humidity (H) Play Golf (PG)
sunny
hot
high
N
sunny
mild
high
N
overcast
hot
high
Y
rain
mild
high
Y
sunny
cool
normal
Y
rain
cool
normal
N
overcast
cool
normal
Y
sunny
mild
high
?

Substituting P(X|yi ) from Equation 5.20 in Equation 5.19, we get

P(yi |X) =

(mj=1 P(x j |yi ))P(yi )

P(X)

(5.21)

We clarify how the naive Bayes classifier works with an example.

Example 5.4. Consider the dataset in Table 5.2.
We predict the label for instance 8 (i8 ) using the naive Bayes classifer and the
given dataset. We have
P(i8 |PG = Y)P(PG = Y)
P(i8 )
= P(O = Sunny, T = mild, H = high|PG = Y)
P(PG = Y)

P(i8 )
= P(O = Sunny|PG = Y) P(T = mild|PG = Y)
P(PG = Y)
P(H = high|PG = Y)
P(i8 )

P(PG = Y|i8 ) =

1 1 2
1
=
7 =
.
4 4 4 P(i8 ) 28P(i8 )
Similarly,
P(PG = N|i8 ) =

P(i8 |PG = N)P(PG = N)

P(i8 )
149

(5.22)

Algorithm 5.1 k-Nearest Neighbor Classifier

Require: Instance i, A Dataset of Real-Value Attributes, k (number of
neighbors), distance measure d
1: return Class label for instance i
2: Compute k nearest neighbors of instance i based on distance mea- sure
d.
3: l = the majority class label among neighbors of instance i. If more than
one majority label, select one randomly.
4: Classify instance i as class l
= P(O = Sunny, T = mild, H = high|PG = N)
P(PG = N)

P(i8 )
= P(O = Sunny|PG = N) P(T = mild|PG = N)
P(PG = N)
P(H = high|PG = N)
P(i8 )
3

2 1 2
4
=
7 =
.
3 3 3 P(i8 ) 63P(i8 )
4
Since 63P(i
>
8)
Play Golf = N.

5.4.3

1
,
28P(i8 )

(5.23)

for instance i8 , and based on NBC calculations, we have

Nearest Neighbor Classifier

As the name suggests, k-nearest neighbor or kNN uses the k nearest instances, called neighbors, to perform classification. The instance being
classified is assigned the label (class attribute value) that the majority of
its k neighbors are assigned. The algorithm is outlined in Algorithm 5.1.
When k = 1, the closest neighbors label is used as the predicted label for
the instance being classified. To determine the neighbors of an instance,
we need to measure its distance to all other instances based on some distance metric. Commonly, Euclidean distance is employed; however, for
higher dimensional spaces, Euclidean distance becomes less meaningful
and other distance measures can be used.
Example 5.5. Consider the example depicted in Figure 5.4. As shown, depending
on the value of k, different labels can be predicted for the circle. In our example,
150

Figure 5.4: k-Nearest Neighbor Example. In this figure, our goal is to

predict the label for the instance shown using a circle. When k = 5, the
predicted label is N and when k = 9 the predicted label is .
k = 5 and k = 9 generate different labels for the instance (triangle and square,
respectively).
As shown in our example, an important issue with the k-nearest neighbor algorithm is the choice of k. The choice of k can easily change the label
of the instance being predicted. In general, we are interested in a value of
k that maximizes the performance of the learning algorithm.

5.4.4

Classification with Network Information

Consider a friendship network on social media and a product being marketed to this network. The product seller wants to know who the potential
buyers are for this product. Assume we are given the network with the
list of individuals who decided to buy or not buy the product. Our goal
is to predict the decision for the undecided individuals. This problem
can be formulated as a classification problem based on features gathered
from individuals. However, in this case, we have additional friendship
information that may be helpful in building more accurate classification
models. This is an example of classification with network information.
151

Assume we are not given any profile information, but only connections
and class labels (i.e., the individual bought/will not buy). By using the
rows of the adjacency matrix of the friendship network for each node as
features and the decision (e.g., buy/not buy) as a class label, we can predict
the label for any unlabeled node using its connections; that is, its row in the
adjacency matrix. Let P(yi = 1|N(vi )) denote the probability of node vi having class attribute value 1 given its neighbors. Individuals decisions are
often highly influenced by their immediate neighbors. Thus, we can approximate P(yi = 1) using the neighbors of the individual by assuming that
P(yi = 1) P(yi = 1|N(vi )).

Weighted-Vote
Relational-Neighbor
Classifier

(5.24)

We can estimate P(yi = 1|N(vi )) via different approaches. The weightedvote relational-neighbor (wvRN) classifier is one such approach. It estimates P(yi = 1|N(vi )) as
P(yi = 1|N(vi )) =

X
1
P(y j = 1|N(v j )).
|N(vi )| v N(v )
j

(5.25)

In other words, the probability of node vi having class attribute value 1

is the average probability of its neighbors having this class attribute value.
Note that P(yi = 1|N(vi )) is only calculated for vi s that are unlabeled. For
node vk , which is labeled 1, p(yk = 1|N(vk )) = 1 and the probability is never
estimated. Similarly, if vk will not buy the product, p(yk = 0|N(vk )) = 1.
Since the probability of a node having class attribute value 1 depends on
the probability of its neighbors having the same value, the probability
of the node is affected if the probabilities of its neighbors change. Thus,
the order of updating nodes can affect the estimated probabilities. In
practice, one follows an order sequence for estimating node probabilities.
Starting with an initial probability estimate for all unlabeled nodes and
following this order, we estimate probabilities until probabilities no longer
change (i.e., converge). We can assume the initial probability estimate to
be P(yi = 1|N(vi )) = 0.5 for all unlabeled nodes.3 We show how the wvRN
classifier learns probabilities using the following example.
3

Note that in our example, the class attribute can take two values; therefore, the initial
guess of P(yi = 1|N(vi )) = 21 = 0.5 is reasonable. When a class attribute takes n values, we
can set our initial guess to P(yi = 1|N(vi )) = n1 .

152

Figure 5.5: Weighted-Vote Relational-Neighbor (wvRN) Example. Labeled

nodes have their class attribute values next to them. The goal is to predict
labels for other nodes in the network.
Example 5.6. Consider the network given in Figure 5.5. Labeled nodes have their
class attribute values next to them. Therefore,
P(y1 = 1|N(v1 )) = 1,
P(y2 = 1|N(v2 )) = 1,
P(y5 = 1|N(v5 )) = 0.

(5.26)
(5.27)
(5.28)

We have three unlabeled nodes {v3 , v4 , v6 }. We choose their natural order to

update their probabilities. Thus, we start with v3 :
P(y3 |N(v3 ))
X
1
P(y j = 1|N(v j ))
=
|N(v3 )| v N(v )
j

1
= (P(y1 = 1|N(v1 )) + P(y2 = 1|N(v2 )) + P(y5 = 1|N(v5 )))
3
1
= (1 + 1 + 0) = 0.67.
(5.29)
3
P(y3 |N(v3 )) does not need to be computed again because its neighbors are all
labeled (thus, this probability estimation has converged). Similarly,
1
P(y4 |N(v4 )) = (1 + 0.5) = 0.75,
2
1
P(y6 |N(v6 )) = (0.75 + 0) = 0.38.
2

(5.30)
(5.31)

We need to recompute both P(y4 |N(v4 )) and P(y6 |N(v6 )) until convergence.
153

Let P(t) (yi |N(vi )) denote the estimated probability after t computations. Then,
1
P(1) (y4 |N(v4 )) = (1 + 0.38) = 0.69,
2
1
P(1) (y6 |N(v6 )) = (0.69 + 0) = 0.35,
2
1
P(2) (y4 |N(v4 )) = (1 + 0.35) = 0.68,
2
1
P(2) (y6 |N(v6 )) = (0.68 + 0) = 0.34,
2
1
P(3) (y4 |N(v4 )) = (1 + 0.34) = 0.67,
2
1
P(3) (y6 |N(v6 )) = (0.67 + 0) = 0.34,
2
1
P(4) (y4 |N(v4 )) = (1 + 0.34) = 0.67,
2
1
P(4) (y6 |N(v6 )) = (0.67 + 0) = 0.34.
2

(5.32)
(5.33)
(5.34)
(5.35)
(5.36)
(5.37)
(5.38)
(5.39)

After four iterations, both probabilities converge. So, from these probabilities
(Equations 5.29, 5.38, and 5.39), we can tell that nodes v3 and v4 will likely have
class attribute value 1 and node v6 will likely have class attribute value 0.

5.4.5

Regression

In classification, class attribute values are discrete. In regression, class

attribute values are real numbers. For instance, we wish to predict the
stock market value (class attribute) of a company given information about
the company (features). The stock market value is continuous; therefore,
regression must be used to predict it. The input to the regression method is
a dataset where attributes are represented using x1 , x2 , . . . , xm (also known
as regressors) and class attribute is represented using Y (also known as the
dependent variable), where the class attribute is a real number. We want to
find the relation between Y and the vector X = (x1 , x2 , . . . , xm ). We discuss
two basic regression techniques: linear regression and logistic regression.
154

Linear Regression
In linear regression, we assume that the class attribute Y has a linear relation
with the regressors (feature set) X by considering a linear error . In other
words,
Y = XW + ,
(5.40)
where W represents the vector of regression coefficients. The problem of
regression can be solved by estimating W using the training dataset and
its labels Y such that fitting error  is minimized. A variety of methods
have been introduced to solve the linear regression problem, most of which
use least squares or maximum-likelihood estimation. We employ the least
squares technique here. Interested readers can refer to the bibliographic
notes for more detailed analyses. In the least square method, we find W
using regressors X and labels Y such that the square of fitting error epsilon
is minimized.
2 = ||2 || = ||Y XW||2 .
(5.41)
To minimize , we compute the gradient and set it to zero to find the
optimal W:
||Y XW||2
= 0.
(5.42)
W
We know that for any X, ||X||2 = (XT X); therefore,
(Y XW)T (Y XW)
||Y XW||2
=
W
W
T
(Y W T XT )(Y XW)
=
W
(YT Y YT XW W T XT Y + W T XT XW)
=
W
= 2XT Y + 2XT XW = 0.

(5.43)

Therefore,
XT Y = XT XW.

(5.44)

Since XT X is invertible for any X, by multiplying both sides by (XT X)1 ,

we get
W = (XT X)1 XT Y.
(5.45)
155

Alternatively, one can compute the singular value decomposition (SVD)

of X = UV T :
W =
=
=
=
=

(XT X)1 XT Y
(VUT UV T )1 VUT Y
(V2 V T )1 VUT Y
V2 V T VUT Y
V1 UT Y,

(5.46)

and since we can have zero singular values,

W = V+ UT Y,

(5.47)

where + is the submatrix of with nonzero singular values.

Logistic Regression
Logistic regression provides a probabilistic view of regression. For simplicity, let us assume that the class attribute can only take values of 0 and
1. Formally, logistic regression finds probability p such that
P(Y = 1|X) = p,

(5.48)

where X is the vector of features and Y is the class attribute. We can use
linear regression to approximate p. In other words, we can assume that
probability p depends on X; that is,
p = X,

(5.49)

where is a vector of coefficients. Unfortunately, X can take unbounded

values because X can take on any value and there are no constraints on how
s are chosen. However, probability p must be in range [0, 1]. Since X
is unbounded, we can perform a transformation g(.) on p such that it also
becomes unbounded. Then, we can fit g(p) to X. One such transformation
g(.) for p is
p
g(p) = ln
,
(5.50)
1p
which for any p between [0, 1] generates a value in range [, +]. The
function g(.) is known as the logit function. The transformed p can be
approximated using a linear function of feature vector X,
g(p) = X.
156

(5.51)

Figure 5.6: Logistic Function.

Combining Equations 5.50 and 5.51 and solving for p, we get
p=

1
eX
= X
.
X
e +1 e
+1

(5.52)

This function is known as the logistic function and is plotted in Figure 5.6. An interesting property of this function is that, for any real value
(negative to positive infinity), it will generate values between 0 and 1. In
other words, it acts as a probability function.
Our task is to find s such that P(Y|X) is maximized. Unlike linear
regression models, there is no closed form solution to this problem, and it
is usually solved using iterative maximum likelihood methods (See Bibliographic Notes).
After s are found, similar to the Naive Bayes Classifier (NBC), we
compute the probability P(Y|X) using Equation 5.52. In a situation where
the class attribute takes two values, when this probability is larger than
0.5, the class attribute is predicted 1; otherwise, 0 is predicted.

5.4.6

Supervised Learning Evaluation

Supervised learning algorithms often employ a training-testing framework

in which a training dataset (i.e., the labels are known) is used to train a
model and then the model is evaluated on a test dataset. The performance
157

Leave-one-out

k-fold
Cross
Validation

of the supervised learning algorithm is measured by how accurate it is in

predicting the correct labels of the test dataset. Since the correct labels of
the test dataset are unknown, in practice, the training set is divided into
two parts, one used for training and the other used for testing. Unlike the
original test set, for this test set the labels are known. Therefore, when
testing, the labels from this test set are removed. After these labels are
predicted using the model, the predicted labels are compared with the
masked labels (ground truth). This measures how well the trained model
is generalized to predict class attributes. One way of dividing the training set into train/test sets is to divide the training set into k equally sized
partitions, or folds, and then using all folds but one to train, with the one
left out for testing. This technique is called leave-one-out training. Another
way is to divide the training set into k equally sized sets and then run
the algorithm k times. In round i, we use all folds but fold i for training
and fold i for testing. The average performance of the algorithm over k
rounds measures the generalization accuracy of the algorithm. This robust
technique is known as k-fold cross validation.
To compare the masked labels with the predicted labels, depending on
the type of supervised learning algorithm, different evaluation techniques
can be used. In classification, the class attribute is discrete so the values it
can take are limited. This allows us to use accuracy to evaluate the classifier.
The accuracy is the fraction of labels that are predicted correctly. Let n be
the size of the test dataset and let c be the number of instances from the
test dataset for which the labels were predicted correctly using the trained
model. Then the accuracy of this model is
accuracy =

c
.
n

(5.53)

In the case of regression, however, it is unreasonable to assume that

the label can be predicted precisely because the labels are real values. A
small variation in the prediction would result in extremely low accuracy.
For instance, if we train a model to predict the temperature of a city in
a given day and the model predicts the temperature to be 71.1 degrees
Fahrenheit and the actual observed temperature is 71, then the model is
highly accurate; however, using the accuracy measure, the model is 0%
accurate. In general, for regression, we check if the predictions are highly
correlated with the ground truth using correlation analysis, or we can fit
lines to both ground truth and prediction results and check if these lines
158

Measure Name
Mahalanobis

Table 5.3: Distance Measures

Formula
p
d(X, Y) = (X Y)T 1 (X Y)

d(X, Y) =

Manhattan (L1 norm)

Lp -norm

P
i

|xi yi |

1
P
d(X, Y) = ( i |xi yi |n ) n

Description
X, Y are features vectors and
is the covariance matrix of the
dataset
X, Y are features
vectors
X, Y are features
vectors

are close. The smaller the distance between these lines, the more accurate
the models learned from the data.

5.5

Unsupervised Learning

Unsupervised learning is the unsupervised division of instances into groups

of similar objects. In this topic, we focus on clustering. In clustering, the Clustering
data is often unlabeled. Thus, the label for each instance is not known to
the clustering algorithm. This is the main difference between supervised
and unsupervised learning.
Any clustering algorithm requires a distance measure. Instances are
put into different clusters based on their distance to other instances. The
most popular distance measure for continuous features is the Euclidean
distance:
q
(x1 y1 )2 + (x2 y2 )2 + + (xn yn )2
d(X, Y) =
v
t n
X
=
(xi yi )2 ,
(5.54)
i=1

where X = (x1 , x2 , . . . , xn ) and Y = (y1 , y2 , . . . , yn ) are n-dimensional feature

vectors in Rn . A list of some commonly used distance measures is provided
in Table 5.3.
159

Once a distance measure is selected, instances are grouped using it.

Clusters are usually represented by compact and abstract notations. Cluster centroids are one common example of this abstract notation. Finally,
clusters are evaluated. There is still a large debate on the issue of evaluating clustering because of the lack of cluster labels in unsupervised
learning. Clustering validity and the definition of valid clusters are two of
the challenges in the ongoing research.

5.5.1

Clustering Algorithms

There are many clustering algorithm types. In this section, we discuss

partitional clustering algorithms, which are the most frequently used clustering algorithms. In Chapter 6, we discuss two other types of clustering
algorithms: spectral clustering and hierarchical clustering.
Partitional Algorithms
Partitional clustering algorithms partition the dataset into a set of clusters.
In other words, each instance is assigned to a cluster exactly once, and no
instance remains unassigned to clusters. k-means [135] is a well-known
example of a partitional algorithm. The output of the k-means algorithm
(k = 6) on a sample dataset is shown in Figure 5.7. In this figure, the dataset
has two features, and instances can be visualized in a two-dimensional
space. The instances are shown using symbols that represent the cluster to
which they belong. The pseudocode for k-means algorithm is provided in
Algorithm 5.2.
The algorithm starts with k initial centroids. In practice, these centroids
are randomly chosen instances from the dataset. These initial instances
form the initial set of k clusters. Then, we assign each instance to one of
these clusters based on its distance to the centroid of each cluster. The
calculation of distances from instances to centroids depends on the choice
of distance measure. Euclidean distance is the most widely used distance
measure. After assigning all instances to a cluster, the centroids, are recomputed by taking the average (mean) of all instances inside the clusters
(hence, the name k-means). This procedure is repeated using the newly
computed centroids. Note that this procedure is repeated until convergence. The most common criterion to determine convergence is to check
whether centroids are no longer changing. This is equivalent to clustering
160

Figure 5.7: k-Means Output on a Sample Dataset. Instances are twodimensional vectors shown in the 2-D space. k-means is run with k = 6,
and the clusters found are visualized using different symbols.
assignments of the data instances stabilizing. In practice, the algorithm
execution can be stopped when the Euclidean distance between the centroids in two consecutive steps is bounded above by some small positive
. As an alternative, k-means implementations try to minimize an objective
function. A well-known objective function in these implementations is the
squared distance error:
n(i)
k X
X
2
||xij ci || ,
(5.55)
i=1 j=1

where xij is the jth instance of cluster i, n(i) is the number of instances
in cluster i, and ci is the centroid of cluster i. The process stops when
the difference between the objective function values of two consecutive
iterations of the k-means algorithm is bounded by some small value .
Note that k-means is highly sensitive to the initial k centroids, and different clustering results can be obtained on a single dataset depending on
161

Algorithm 5.2 k-Means Algorithm

Require: A Dataset of Real-Value Attributes, k (number of Clusters)
1: return A Clustering of Data into k Clusters
2: Consider k random instances in the data space as the initial cluster
centroids.
3: while centroids have not converged do
4:
Assign each instance to the cluster that has the closest cluster centroid.
5:
If all instances have been assigned then recalculate the cluster centroids by averaging instances inside each cluster
6: end while

the initial k centroids. This problem can be mitigated by running k-means

multiple times and selecting the clustering assignment that is observed
most often or is more desirable based on an objective function, such as
the squared error. Since k-means assumes that instances that belong to
the same cluster are the ones that found the clusters centroid closer than
other centroids in the dataset, one can safely assume that all the instances
inside a cluster fall into a hyper-sphere, with the centroid being its center.
The radius for this hyper-sphere is defined based on the farthest instance
inside this cluster. If clusters that need to be extracted are nonspherical
(globular), k-means has problems detecting them. This problem can be
addressed by a preprocessing step in which a transformation is performed
on the dataset to solve this issue.

5.5.2

Unsupervised Learning Evaluation

When clusters are found, there is a need to evaluate how accurately the task
has been performed. When ground truth is available, we have prior knowledge of which instances should belong to which cluster, as discussed in
Chapter 6 in detail. However, evaluating clustering is a challenge because
ground truth is often not available. When ground truth is unavailable,
we incorporate techniques that analyze the clusters found and describe
the quality of clusters found. In particular, we can use techniques that
measure cohesiveness or separateness of clusters.
162

Figure 5.8: Unsupervised Learning Evaluation.

Cohesiveness
In evaluating clustering, we are interested in clusters that exhibit cohesiveness. In cohesive clusters, instances inside the clusters are close to each
other. In statistical terms, this is equivalent to having a small standard
deviation (i.e., being close to the mean value). In clustering, this translates
to being close to the centroid of the cluster. So cohesiveness is defined as
the distance from instances to the centroid of their respective clusters,
cohesiveness =

n(i)
k X
X

||xij ci || ,

(5.56)

i=1 j=1

which is the squared distance error (also known as SSE) discussed previously. Small values of cohesiveness denote highly cohesive clusters in
which all instances are close to the centroid of the cluster.
Example 5.7. Figure 5.8 shows a dataset of four one-dimensional instances. The
instances are clustered into two clusters. Instances in cluster 1 are x11 and x12 , and
instances in cluster 2 are x21 and x22 . The centroids of these two clusters are denoted
as c1 and c2 . For these two clusters, the cohesiveness is
cohesiveness = | 10 (7.5)|2 + | 5 (7.5)|2 + |5 7.5)|2
+|10 7.5|2 = 25.
(5.57)
Separateness
We are also interested in clustering of the data that generates clusters that
are well separated from one another. To measure this distance between
clusters, we can use the separateness measure. In statistics, separateness
163

can be measured by standard deviation. Standard deviation is maximized

when instances are far from the mean. In clustering terms, this is equivalent
to cluster centroids being far from the mean of the entire dataset:
separateness =

k
X

||c ci ||2 ,

(5.58)

i=1

P
where c = n1 ni=1 xi is the centroid of all instances and ci is the centroid of
cluster i. Large values of separateness denote clusters that are far apart.
Example 5.8. For the dataset shown in Figure 5.8, the centroid for all instances
is denoted as c. For this dataset, the separateness is
separateness = | 7.5 0|2 + |7.5 0|2 = 112.5.

(5.59)

In general, we are interested in clusters that are both cohesive and

separate. The silhouette index combines both these measures.
Silhouette Index
The silhouette index combines both cohesiveness and separateness. It compares the average distance value between instances in the same cluster
and the average distance value between instances in different clusters. In
a well-clustered dataset, the average distance between instances in the
same cluster is small (cohesiveness), and the average distance between
instances in different clusters is large (separateness). Let a(x) denote the
average distance between instance x of cluster C and all other members of
C:
X
1
||x y||2 .
(5.60)
a(x) =
|C| 1 yC,y,x
Let G , C denote the cluster that is closest to x in terms of the average
distance between x and members of G. Let b(x) denote the average distance
between instance x and instances in cluster G:
b(x) = min
G,C

1 X
||x y||2 .
|G| yG
164

(5.61)

Since we want distance between instances in the same cluster to be

smaller than distance between instances in different clusters, we are interested in a(x) < b(x). The silhouette clustering index is formulated as
b(x) a(x)
,
max(b(x), a(x))
1X
silhouette =
s(x).
n x
s(x) =

(5.62)
(5.63)

The silhouette index takes values between [1, 1]. The best clustering
happens when x a(x)  b(x). In this case, silhouette 1. Similarly when
silhouette < 0, that indicates that many instances are closer to other clusters
than their assigned cluster, which shows low-quality clustering.
Example 5.9. In Figure 5.8, the a(.), b(.), and s(.) values are
a(x11 ) = | 10 (5)|2 = 25
1
(| 10 5|2 + | 10 10|2 ) = 312.5
b(x11 ) =
2
312.5 25
s(x11 ) =
= 0.92
312.5
a(x12 ) = | 5 (10)|2 = 25
1
b(x12 ) =
(| 5 5|2 + | 5 10|2 ) = 162.5
2
162.5 25
s(x12 ) =
= 0.84
162.5
a(x21 ) = |5 10|2 = 25
1
b(x21 ) =
(|5 (10)|2 + |5 (5)|2 ) = 162.5
2
162.5 25
s(x21 ) =
= 0.84
162.5
a(x22 ) = |10 5|2 = 25
1
b(x22 ) =
(|10 (5)|2 + |10 (10)|2 ) = 312.5
2
312.5 25
s(x22 ) =
= 0.92.
312.5
Given the s(.) values, the silhouette index is
1
silhouette = (0.92 + 0.84 + 0.84 + 0.92) = 0.88.
4
165

(5.64)
(5.65)
(5.66)
(5.67)
(5.68)
(5.69)
(5.70)
(5.71)
(5.72)
(5.73)
(5.74)
(5.75)

(5.76)

5.6

Summary

This chapter covered data mining essentials. The general process for analyzing data is known as knowledge discovery in databases (KDD). The first
step in the KDD process is data representation. Data instances are represented in tabular format using features. These instances can be labeled or
unlabeled. There exist different feature types: nominal, ordinal, interval,
and ratio. Data representation for text data can be performed using the
vector space model. After having a representation, quality measures need
to be addressed and preprocessing steps completed before processing the
data. Quality measures include noise removal, outlier detection, missing
values handling, and duplicate data removal. Preprocessing techniques
commonly performed are aggregation, discretization, feature selection,
feature extraction, and sampling.
We covered two categories of data mining algorithms: supervised and
unsupervised learning. Supervised learning deals with mapping feature
values to class labels, and unsupervised learning is the unsupervised division of instances into groups of similar objects.
When labels are discrete, the supervised learning is called classification,
and when labels are real numbers, it is called regression. We covered, these
classification methods: decision tree learning, naive Bayes classifier (NBC),
nearest neighbor classifier, and classifiers that use network information.
We also discussed linear and logistic regression.
To evaluate supervised learning, a training-testing framework is used
in which the labeled dataset is partitioned into two parts, one for training
and the other for testing. Different approaches for evaluating supervised
learning such as leave-one-out or k-fold cross validation were discussed.
Any clustering algorithm requires the selection of a distance measure.
We discussed partitional clustering algorithms and k-means from these
algorithms, as well as methods of evaluating clustering algorithms. To
evaluate clustering algorithms, one can use clustering quality measures
such as cohesiveness, which measures how close instances are inside clusters, or separateness, which measures how separate different clusters are
from one another. Silhouette index combines the cohesiveness and separateness into one measure.

166

5.7

Bibliographic Notes

A general review of data mining algorithms can be found in the machine

learning and pattern recognition [40, 75, 199, 234, 235, 163], data mining
[92, 120, 304, 270, 120], and pattern recognition [39, 75] literature.
Among preprocessing techniques, feature selection and feature extraction have gained much attention due to their importance. General references for feature selection and extraction can be found in [175, 64, 65,
117, 313, 175, 176]. Feature selection has also been discussed in social media data in [275, 276, 277]. Although not much research is dedicated to
sampling in social media, it plays an important role in the experimental
outcomes of social media research. Most experiments are performed using
sampled social media data, and it is important for these samples to be representative samples of the site that is under study. For instance, Morstatter
et al. [203] studied whether Twitters heavily sampled Streaming API,
a free service for social media data, accurately portrays the true activity
on Twitter. They show that the bias introduced by the Streaming API is
significant.
In addition to the data mining categories covered in this chapter, there
are other important categories in the area of data mining and machine
learning. In particular, an interesting category is semi-supervised learning.
In semi-supervised learning, the label is available for some instances, but
not all. The model uses the labeled information and the feature distribution
of the unlabeled data to learn a model. Expectation maximization (EM) is
a well-established technique from this area. In short, EM learns a model
from the data that is partially labeled (expectation step). Then, it uses this
model to predict labels for the unlabeled instances (maximization step).
The predicted labels for instances are used once again to refine the learned
model and revise predictions for unlabeled instances in an iterative fashion
until convergence in reached. In addition to supervised methods covered,
neural networks deserve mention [123]. More on regression techniques in
available in [208, 40].
Clustering is one of the most popular areas in the field of machine
learning research. A taxonomy of clustering algorithms can be found
in [35, 136, 305, 197]. Among clustering algorithms, some of which use
data density of cluster data, DBSCAN [86], GDBSCAN [245], CLARANS
[218], and OPTICS [14] are some of the most well known and practiced
algorithms. Most of the previous contributions in the area of clustering
167

consider the number of clusters as an input parameter. Early literature in

clustering had attempted to solve this by running algorithms for several
Ks (number of clusters) and selecting the best K that optimizes some coefficients [196, 35]. For example, the distance between two cluster centroids
normalized by a clusters standard deviation could be used as a coefficient. After the coefficient is selected, the coefficient values are plotted as
a function of K (number of clusters) and the best K is selected.
An interesting application of data mining is sentiment analysis in which
the level of subjective content in information is quantified; for example,
identifying the polarity (i.e., being positive/negative) of a digital camera
review. General references for sentiment analysis can be found in [227, 174],
and examples of recent developments in social media are available in
[131, 132].

168

5.8

Exercises

1. Describe how methods from this chapter can be applied in social

media.
2. Outline a framework for using the supervised learning algorithm for
unsupervised learning.

Data
3. Describe methods that can be used to deal with missing data.
4. Given a continuous attribute, how can we convert it to a discrete
attribute? How can we convert discrete attributes to continuous
ones?
5. If you had the chance of choosing either instance selection or feature
selection, which one would you choose? Please justify.
6. Given two text documents that are vectorized, how can we measure
document similarity?
7. In the example provided for TF-IDF (Example 5.1), the word orange
received zero score. Is this desirable? What does a high TF-IDF value
show?

Supervised Learning
8. Provide a pseudocode for decision tree induction.
9. How many decision trees containing n attributes and a binary class
can be generated?
10. What does zero entropy mean?
11.

What is the time complexity for learning a naive Bayes classifer?

What is the time complexity for classifying using the naive Bayes
classifier?
169

Figure 5.9: Nonlinearity of XOR Function.

Linear separability: Two sets of two-dimensional instances are
linearly separable if they can be completely separated using one
line. In n-dimensional space, two set of instances are linearly
separable if one can separate them by a hyper-plane. A classical
example of nonlinearity is the XOR function. In this function, the
two instance sets are the black-and-white instances (see Figure
5.9), which cannot be separated using a single line. This is an
example of a nonlinear binary function. Can a naive Bayes
classifier learn nonlinear binary functions? Provide details.
What about linear separability and K-NN? Are K-NNs capable
of solving such problems?
12. Describe how the least square solution can be determined for regression.

Unsupervised Learning
13. (a) Given k clusters and their respective cluster sizes s1 , s2 , . . . , sk ,
what is the probability that two random (with replacement) data
vectors (from the clustered dataset) belong to the same cluster?
(b) Now, assume you are given this probability (you do not have
si s and k), and the fact that clusters are equally sized, can you
find k? This gives you an idea how to predict the number of
clusters in a dataset.
170

14. Give an example of a dataset consisting of four data vectors where

there exist two different optimal (minimum SSE) 2-means (k-means,
k = 2) clusterings of the dataset.
Calculate the optimal SSE value for your example.
In general, how should datasets look like geometrically so that
we have more than one optimal solution?
What defines the number of optimal solutions?
Perform two iterations of the k-means algorithm in order to obtain
two clusters for the input instances given in Table 5.4. Assume that
the first centroids are instances 1 and 3. Explain if more iterations are
needed to get the final clusters.
Table 5.4: Dataset
Instance
X
Y
1
12.0 15.0
2
12.0 33.0
3
18.0 15.0
4
18.0 27.0
5
24.0 21.0
6
36.0 42.0

16.
15. What is the usual shape of clusters generated by k-means? Give
an example of cases where k-means has limitations in detecting the
patterns formed by the instances.
17. Describe a preprocessing strategy that can help detect nonspherical
clusters using k-means.

171

172

Part II
Communities and Interactions

173

This chapter is from Social Media Mining: An Introduction.

By Reza Zafarani, Mohammad Ali Abbasi, and Huan Liu.
Cambridge University Press, 2014. Draft version: April 20, 2014.
Complete Draft and Slides Available at: http://dmml.asu.edu/smm

Chapter

Community Analysis
In November 2010, a team of Dutch law enforcement agents dismantled
a community of 30 million infected computers across the globe that were
sending more than 3.6 billion daily spam mails. These distributed networks
of infected computers are called botnets. The community of computers in
a botnet transmit spam or viruses across the web without their owners
permission. The members of a botnet are rarely known; however, it is
vital to identify these botnet communities and analyze their behavior to
enhance internet security. This is an example of community analysis. In this
chapter, we discuss community analysis in social media.
Also known as groups, clusters, or cohesive subgroups, communities have
been studied extensively in many fields and, in particular, the social sciences. In social media mining, analyzing communities is essential. Studying communities in social media is important for many reasons. First,
individuals often form groups based on their interests, and when studying individuals, we are interested in identifying these groups. Consider
the importance of finding groups with similar reading tastes by an online book seller for recommendation purposes. Second, groups provide a
clear global view of user interactions, whereas a local-view of individual
behavior is often noisy and ad hoc. Finally, some behaviors are only observable in a group setting and not on an individual level. This is because
the individuals behavior can fluctuate, but group collective behavior is
more robust to change. Consider the interactions between two opposing
political groups on social media. Two individuals, one from each group,
can hold similar opinions on a subject, but what is important is that their
175

communities can exhibit opposing views on the same subject.

In this chapter, we discuss communities and answer the following three
questions in detail:
1. How can we detect communities? This question is discussed in different disciplines, and in diverse forms. In particular, quantization in
electrical engineering, discretization in statistics, and clustering in
machine learning tackle a similar challenge. As discussed in Chapter 5, in clustering, data points are grouped together based on a
similarity measure. In community detection, data points represent
actors in social media, and similarity between these actors is often
defined based on the interests these users share. The major difference
between clustering and community detection is that in community
detection, individuals are connected to others via a network of links,
whereas in clustering, data points are not embedded in a network.
2. How do communities evolve and how can we study evolving communities?
Social media forms a dynamic and evolving environment. Similar
to real-world friendships, social media interactions evolve over time.
People join or leave groups; groups expand, shrink, dissolve, or
split over time. Studying the temporal behavior of communities is
necessary for a deep understanding of communities in social media.
3. How can we evaluate detected communities? As emphasized in our
botnet example, the list of community members (i.e., ground truth)
is rarely known. Hence, community evaluation is a challenging task
and often means to evaluating detected communities in the absence
of ground truth.

Social Communities
Broadly speaking, a real-world community is a body of individuals with
common economic, social, or political interests/characteristics, often living
in relatively close proximity. A virtual community comes into existence
when like-minded users on social media form a link and start interacting
with each other. In other words, formation of any community requires (1)
a set of at least two nodes sharing some interest and (2) interactions with
respect to that interest.
176

Figure 6.1: Zacharys Karate Club. Nodes represent karate club members
and edges represent friendships. A conflict in the club divided the members into two groups. The color of the nodes denotes which one of the two
groups the nodes belong to.
As a real-world community example, consider the interactions of a college karate club collected by Wayne Zachary in 1977. The example is often
referred to as Zacharys Karate Club [309] in the literature. Figure 6.1 depicts Zacharys Karate Club
the interactions in a college karate club over two years. The links show
friendships between members. During the observation period, individuals split into two communities due to a disagreement between the club
administrator and the karate instructor, and members of one community
left to start their own club. In this figure, node colors demonstrate the
communities to which individuals belong. As observed in this figure, using graphs is a convenient way to depict communities because color-coded
nodes can denote memberships and edges can be used to denote relations.
Furthermore, we can observe that individuals are more likely to be friends
with members of their own group, hence, creating tightly knit components
in the graph.
Zacharys Karate Club is an example of two explicit communities. An ex- Explicit (emic)
plicit community, also known as an emic community, satisfies the following Communities
three criteria:
177

1. Community members understand that they are its members.

2. Nonmembers understand who the community members are.
3. Community members often have more interactions with each other
than with nonmembers.
Implicit (etic)
Communities

In contrast to explicit communities, in implicit communities, also known

as etic communities, individuals tacitly interact with others in the form of
an unacknowledged community. For instance, individuals calling Canada
from the United States on a daily basis need not be friends and do not consider each other as members of the same explicit community. However,
from the phone operators point of view, they form an implicit community that needs to be marketed the same promotions. Finding implicit
communities is of major interest, and this chapter focuses on finding these
communities in social media.
Communities in social media are more or less representatives of communities in the real world. As mentioned, in the real world, members of
communities are often geographically close to each other. The geographical
location becomes less important in social media, and many communities
on social media consist of highly diverse people from all around the planet.
In general, people in real-world communities tend to be more similar than
those of social media. People do not need to share language, location, and
the like to be members of social media communities. Similar to real-world
communities, communities in social media can be labeled as explicit or
implicit. Examples of explicit communities in well-known social media
sites include the following:
Facebook. In Facebook, there exist a variety of explicit communities,
such as groups and communities. In these communities, users can post
messages and images, comment on other messages, like posts, and
view activities of others.
Yahoo! Groups. In Yahoo! groups, individuals join a group mailing
list where they can receive emails from all or a selection of group
members (administrators) directly.
LinkedIn. LinkedIn provides its users with a feature called Groups
and Associations. Users can join professional groups where they can
post and share information related to the group.
178

Because these sites represent explicit communities, individuals have

an understanding of when they are joining them. However, there exist
implicit communities in social media as well. For instance, consider individuals with the same taste for certain movies on a movie rental site.
These individuals are rarely all members of the same explicit community.
However, the movie rental site is particularly interested in finding these
implicit communities so it can better market to them by recommending
movies similar to their tastes. We discuss techniques to find these implicit
communities next.

6.1

Community Detection

As mentioned earlier, communities can be explicit (e.g., Yahoo! groups), or

implicit (e.g., individuals who write blogs on the same or similar topics).
In contrast to explicit communities, in many social media sites, implicit
communities and their members are obscure to many people. Community
detection finds these implicit communities.
In the simplest form, similar to the graph shown in Figure 6.1, community detection algorithms are often provided with a graph where nodes
represent individuals and edges represent friendships between individual.
This definition can be generalized. Edges can also be used to represent
contents or attributes shared by individuals. For instance, we can connect
individuals at the same location, with the same gender, or who bought the
same product using edges. Similarly, nodes can also represent products,
sites, and webpages, among others. Formally, for a graph G(V, E), the task
of community detection is to find a set of communities {Ci }ni=1 in a G such
that ni=1 Ci V.

6.1.1

Community Detection Algorithms

There are a variety of community detection algorithms. When detecting

communities, we are interested in detecting communities with either (1)
specific members or (2) specific forms of communities. We denote the former
as member-based community detection and the latter as group-based community detection. Consider the network of 10 individuals shown in Figure
6.2 where 7 are wearing black t-shirts and 3 are wearing white ones. If
we group individuals based on their t-shirt color, we end up having a
179

Figure 6.2: Community Detection Algorithms Example. Member-based

community detection groups members based on their characteristics. Here,
we divide the network based on color. In group-based community detection, we find communities based on group properties. Here, groups are
formed based on the density of interactions among their members.

community of three and a community of seven. This is an example of

member-based community detection, where we are interested in specific
members characterized by their t-shirts color. If we group the same set
based on the density of interactions (i.e., internal edges), we get two other
communities. This is an instance of group-based community detection,
where we are interested in specific communities characterized by their interactions density.
Member-based community detection uses community detection algorithms that group members based on attributes or measures such as similarity, degree, or reachability. In group-based community detection, we
are interested in finding communities that are modular, balanced, dense,
robust, or hierarchical.
180

Figure 6.3: A 4-Cycle.

6.1.2

Member-Based Community Detection

The intuition behind member-based community detection is that members with the same (or similar) characteristics are more often in the same
community. Therefore, a community detection algorithm following this
approach should assign members with similar characteristics to the same
community. Let us consider a simple example. We can assume that nodes
that belong to a cycle form a community. This is because they share the
same characteristic: being in the cycle. Figure 6.3 depicts a 4-cycle. For
instance, we can search for all n-cycles in the graph and assume that they
represent a community. The choice for n can be based on empirical evidence or heuristics, or n can be in a range [1 , 2 ] for which all cycles are
found. A well-known example is the search for 3-cycles (triads) in graphs.
In theory, any subgraph can be searched for and assumed to be a community. In practice, only subgraphs that have nodes with specific characteristics are considered as communities. Three general node characteristics
that are frequently used are node similarity, node degree (familiarity), and node
reachability.
When employing node degrees, we seek subgraphs, which are often
connected, such that each node (or a subset of nodes) has a certain node
degree (number of incoming or outgoing edges). Our 4-cycle example
follows this property, the degree of each node being two. In reachability, we
seek subgraphs with specific properties related to paths existing between
nodes. For instance, our 4-cycle instance also follows the reachability
characteristic where all pairs of nodes can be reached via two independent
paths. In node similarity, we assume nodes that are highly similar belong
181

Figure 6.4: First Four Complete Graphs.

to the same community.
Node Degree
The most common subgraph searched for in networks based on node degrees is a clique. A clique is a maximum complete subgraph in which all
pairs of nodes inside the subgraph are connected. In terms of the node
degree characteristic, a clique of size k is a subgraph of k nodes where all
node degrees in the induced subgraph are k 1. The only difference between cliques and complete graphs is that cliques are subgraphs, whereas
complete graphs contain the whole node set V. The simplest four complete
graphs (or cliques, when these are subgraphs) are represented in Figure
6.4.
To find communities, we can search for the maximum clique (the one
with the largest number of vertices) or for all maximal cliques (cliques
that are not subgraphs of a larger clique; i.e., cannot be expanded further).
However, both problems are NP-hard, as is verifying whether a graph
contains a clique larger than size k. To overcome these theoretical barriers,
for sufficiently small networks or subgraphs, we can (1) use brute force, (2)
add some constraints such that the problem is relaxed and polynomially
solvable, or (3) use cliques as the seed or core of a larger community.
Brute-force clique identification. The brute force method can find all
maximal cliques in a graph. For each vertex vx , we try to find the maximal
clique that contains node vx . The brute-force algorithm is detailed in
Algorithm 6.1.
The algorithm starts with an empty stack of cliques. This stack is
initialized with the node vx that is being analyzed (a clique of size 1). Then,
from the stack, a clique is popped (C). The last node added to clique C
182

Algorithm 6.1 Brute-Force Clique Identification

Require: Adjacency Matrix A, Vertex vx
1: return Maximal Clique C containing vx
2: CliqueStack = {{vx }}, Processed = {};
3: while CliqueStack not empty do
4:
C=pop(CliqueStack); push(Processed,C);
5:
vlast = Last node added to C;
6:
N(vlast ) = {vi |Avlast ,vi = 1}.
7:
for all S
vtemp N(vlast ) do
8:
if C {vtemp } is a clique then
S
9:
push(CliqueStack, C {vtemp });
10:
end if
11:
end for
12: end while
13: Return the largest clique from Processed
is selected (vlast ). All the neighbors of vlast are added to the popped clique
C sequentially, and if the new set of nodes creates a larger clique (i.e., the
newly added node is connected to all of the other members), then the new
clique is pushed back into the stack. This procedure is followed until nodes
can no longer be added.
The brute-force algorithm becomes impractical for large networks. For
instance, for a complete graph of only 100 nodes, the algorithm will generate at least 299 1 different cliques starting from any node in the graph
(why?).
The performance of the brute-force algorithm can be enhanced by pruning specific nodes and edges. If the cliques being searched for are of size
k or larger, we can simply assume that the clique, if found, should contain
nodes that have degrees equal to or more than k 1. We can first prune all
nodes (and edges connected to them) with degrees less than k 1. Due to
the power-law distribution of node degrees, many nodes exist with small
degrees (1, 2, etc.). Hence, for a large enough k many nodes and edges will
be pruned, which will reduce the computation drastically. This pruning
works for both directed and undirected graphs.
Even with pruning, there are intrinsic properties with cliques that make
them a less desirable means for finding communities. Cliques are rarely
observed in the real world. For instance, consider a clique of 1,000 nodes.
183

Figure 6.5: Maximal k-plexes for k = 1, 2, and 3.

This subgraph has 999 21000 = 499,500 edges. A single edge removal from
this many edges results in a subgraph that is no longer a clique. That
represents less than 0.0002% of the edges, which makes finding cliques a
challenging task.
In practice, to overcome this challenge, we can either relax the clique
structure or use cliques as a seed or core of a community.

k-plex

Relaxing cliques. A well-known clique relaxation that comes from sociology is the k-plex concept. In a clique of size k, all nodes have the degree
of k 1; however, in a k-plex, all nodes have a minimum degree that is not
necessarily k 1 (as opposed to cliques of size k). For a set of vertices V,
the structure is called a k-plex if we have
dv |V| k, v V,

(6.1)

where dv is the degree of v in the induced subgraph (i.e., the number of

nodes from the set V that are connected to v).
Clearly, a clique of size k is a 1-plex. As k gets larger in a k-plex, the
structure gets increasingly relaxed, because we can remove more edges
from the clique structure. Finding the maximum k-plex in a graph still
tends to be NP-hard, but in practice, finding it is relatively easier due to
smaller search space. Figure 6.5 shows maximal k-plexes for k = 1, 2, and
3. A k-plex is maximal if it is not contained in a larger k-plex (i.e., with
more nodes).

Clique
Percolation
Method (CPM)

Using cliques as a seed of a community. When using cliques as a seed

or core of a community, we assume communities are formed from a set of
cliques (small or large) in addition to edges that connect these cliques. A
well-known algorithm in this area is the clique percolation method (CPM)
[225]. The algorithm is provided in Algorithm 6.2. Given parameter k, the
184

Algorithm 6.2 Clique Percolation Method (CPM)

Require: parameter k
1: return Overlapping Communities
2: Cliquesk = find all cliques of size k
3: Construct clique graph G(V, E), where |V| = |Cliquesk |
4: E = {ei j | clique i and clique j share k 1 nodes}
5: Return all connected components of G
method starts by finding all cliques of size k. Then a graph is generated
(clique graph) where all cliques are represented as nodes, and cliques that
share k 1 vertices are connected via edges. Communities are then found
by reporting the connected components of this graph. The algorithm
searches for all cliques of size k and is therefore computationally intensive.
In practice, when using the CPM algorithm, we often solve CPM for a small
k. Relaxations discussed for cliques are desirable to enable the algorithm
to perform faster. Lastly, CPM can return overlapping communities.
Example 6.1. Consider the network depicted in Figure 6.6(a). The corresponding
clique graph generated by the CPM algorithm for k = 3 is provided in Figure
6.6(b). All cliques of size k = 3 have been identified and cliques that share
k1 = 2 nodes are connected. Connected components are returned as communities
({v1 , v2 , v3 }, {v8 , v9 , v10 }, and {v3 , v4 , v5 , v6 , v7 , v8 }). Nodes v3 and v8 belong to two
communities, and these communities are overlapping.
Node Reachability
When dealing with reachability, we are seeking subgraphs where nodes
are reachable from other nodes via a path. The two extremes of reachability are achieved when nodes are assumed to be in the same community
if (1) there is a path between them (regardless of the distance) or (2) they
are so close as to be immediate neighbors. In the first case, any graph
traversal algorithm such as BFS or DFS can be used to identify connected
components (communities). However, finding connected components is
not very useful in large social media networks. These networks tend to
have a large-scale connected component that contains most nodes, which
are connected to each other via short paths. Therefore, finding connected
components is less powerful for detecting communities in them. In the second case, when nodes are immediate neighbors of all other nodes, cliques
185

Figure 6.6: Clique Percolation Method (CPM) Example for k = 3.

are formed, and as discussed previously, finding cliques is considered a
very challenging process.
To overcome these issues, we can find communities that are in between
cliques and connected components in terms of connectivity and have small
shortest paths between their nodes. There are predefined subgraphs, with
roots in social sciences, with these characteristics. Well-known ones include the following:

k-Clique, k-Club, and

k-Clan

k-Clique is a maximal subgraph where the shortest path between

any two nodes is always less than or equal to k. Note that in kcliques, nodes on the shortest path should not necessarily be part of
the subgraph.
186

Figure 6.7: Examples of 2-Cliques, 2-Clubs, and 2-Clans.

k-Club is a more restricted definition; it follows the same definition
as k-cliques with the additional constraint that nodes on the shortest
paths should be part of the subgraph.
k-Clan is a k-clique where, for all shortest paths within the subgraph,
the distance is less than or equal to k. All k-clans are k-cliques and
k-clubs, but not vice versa. In other words,
k-Clans = k-Cliques k-Clubs.
Figure 6.7 depicts an example of the three discussed models.
Node Similarity
Node similarity attempts to determine the similarity between two nodes vi
and v j . Similar nodes (or most similar nodes) are assumed to be in the same
community. Often, once the similarities between nodes are determined,
a classical clustering algorithm (see Chapter 5) is applied to find communities. Determining similarity between two nodes has been addressed in
different fields; in particular, the problem of structural equivalence in the Structural
field of sociology considers the same problem. In structural equivalence, Equivalence
similarity is based on the overlap between the neighborhood of the vertices. Let N(vi ) and N(v j ) be the neighbors of vertices vi and v j , respectively.
In this case, a measure of vertex similarity can be defined as follows:
(vi , v j ) = |N(vi ) N(v j )|.

(6.2)

For large networks, this value can increase rapidly, because nodes may
share many neighbors. Generally, similarity is attributed to a value that is
bounded and usually in the range [0, 1]. For that to happen, various normalization procedures such as the Jaccard similarity or the cosine similarity
187

can be done:
|N(vi ) N(v j )|
,
|N(vi ) N(v j )|

(6.3)

|N(vi ) N(v j )|
Cosine (vi , v j ) = p
.
|N(vi )||N(v j )|

(6.4)

Jaccard (vi , v j ) =

Example 6.2. Consider the graph in Figure 6.7. The similarity values between
nodes v2 and v5 are
|{v1 , v3 , v4 } {v3 , v6 }|
= 0.25,
|{v1 , v3 , v4 , v6 }|
|{v1 , v3 , v4 } {v3 , v6 }|
= 0.40.
Cosine (v2 , v5 ) =
|{v1 , v3 , v4 }||{v3 , v6 }|

Jaccard (v2 , v5 ) =

(6.5)
(6.6)

In general, the definition of neighborhood N(vi ) excludes the node itself

(vi ). This, however, leads to problems with the aforementioned similarity
values because nodes that are connected and do not share a neighbor will
be assigned zero similarity. This can be rectified by assuming that nodes
are included in their own neighborhood.
A generalization of structural equivalence is known as regular equivalence. Consider the situation of two basketball players in two different
countries. Though sharing no neighborhood overlap, the social circles of
these players (coach, players, fans, etc.) might look quite similar due to
their social status. In other words, nodes are regularly equivalent when
they are connected to nodes that are themselves similar (a self-referential
definition). For more details on regular equivalence, refer to Chapter 3.

6.1.3

Group-Based Community Detection

When considering community characteristics for community detection, we

are interested in communities that have certain group properties. In this
section, we discuss communities that are balanced, robust, modular, dense,
or hierarchical.
Balanced Communities
As mentioned before, community detection can be thought of as the problem of clustering in data mining and machine learning. Graph-based
188

Figure 6.8: Minimum Cut (A) and Two More Balanced Cuts (B and C) in a
Graph.
clustering techniques have proven to be useful in identifying communities in social networks. In graph-based clustering, we cut the graph into
several partitions and assume these partitions represent communities.
Formally, a cut in a graph is a partitioning (cut) of the graph into two
(or more) sets (cutsets). The size of the cut is the number of edges that are
being cut and the summation of weights of edges that are being cut in a
weighted graph. A minimum cut (min-cut) is a cut such that the size of the
cut is minimized. Figure 6.8 depicts several cuts in a graph. For example,
cut B has size 4, and A is the minimum cut.
Minimum Cut
Based on the well-known max-flow min-cut theorem, the minimum cut
of a graph can be computed efficiently. However, minimum cuts are not always preferred for community detection. Often, they result in cuts where
a partition is only one node (singleton), and the rest of the graph is in
the other. Typically, communities with balanced sizes are preferred. Figure 6.8 depicts an example where the minimum cut (A) creates unbalanced
partitions, whereas, cut C is a more balanced cut.
To solve this problem, variants of minimum cut define an objective
function, minimizing (or maximizing) that during the cut-finding procedure, results in a more balanced and natural partitioning of the data.
Consider a graph G(V, E). A partitioning of G into k partitions is a tuple
S
P = (P1 , P2 , P3 , . . . , Pk ), such that Pi V, Pi P j = and ki=1 Pi = V. Then,
the objective function for the ratio cut and normalized cut are defined as Ratio cut and
Normalized Cut
follows:
k
1 X cut(Pi , P i )
,
Ratio Cut(P) =
k i=1
|Pi |

189

(6.7)

k
1 X cut(Pi , P i )
Normalized Cut(P) =
,
k i=1 vol(Pi )

(6.8)

where P i = V Pi is the P
complement cut set, cut(Pi , P i ) is the size of the
cut, and volume vol(Pi ) = vPi dv . Both objective functions provide a more
balanced community size by normalizing the cut size either by the number
of vertices in the cutset or the volume (total degree).
Both the ratio cut and normalized cut can be formulated in a matrix
format. Let matrix X {0, 1}|V|k denote the community membership matrix,
where Xi,j = 1 if node i is in community j; otherwise, Xi,j = 0. Let D =
diag(d1 , d2 , . . . , dn ) represent the diagonal degree matrix. Then the ith entry
on the diagonal of XT AX represents the number of edges that are inside
community i. Similarly, the ith element on the diagonal of XT DX represents
the number of edges that are connected to members of community i. Hence,
the ith element on the diagonal of XT (D A)X represents the number of
edges that are in the cut that separates community i from all other nodes. In
fact, the ith diagonal element of XT (DA)X is equivalent to the summation
term cut(Pi , P i ) in both the ratio and normalized cut. Thus, for ratio cut, we
have
k
1 X cut(Pi , P i )
Ratio Cut(P) =
k i=1
|Pi |
k

(6.9)

1 X Xi (D A)Xi
=
k i=1
XiT Xi

(6.10)

1 X T
X (D A)X i ,
k i=1 i

(6.11)

where X i = Xi /(XiT Xi )1/2 . A similar approach can be followed to formulate

the normalized cut and to obtain a different X i . To formulate the summation
=
in both the ratio and normalized cut, we can use the trace of matrix (tr(X)
P
n

i=1 Xii ). Using the trace, the objectives for both the ratio and normalized
cut can be formulated as trace-minimization problems,

min Tr(X T LX),

X

190

(6.12)

where L is the (normalized) graph Laplacian, defined as follows:

Ratio Cut Laplacian (Unnormalized Laplacian);

DA
L=

1/2
1/2
ID
AD
Normalized Cut Laplacian (Normalized Laplacian).

(6.13)
It has been shown that both ratio cut and normalized cut minimization Normalized and
are NP-hard; therefore, approximation algorithms using relaxations are Unnormalized Graph
Laplacian
desired. Spectral clustering is one such relaxation:

min Tr(X T LX),

(6.14)

s.t. X T X = Ik .

(6.15)

The solution to this problem is the top eigenvectors of L.1 Given L, the Spectral Clustering
top k eigenvectors corresponding to the smallest eigen values are computed
and then k-means is run on X to extract communities
and used as X,
memberships (X). The first eigenvector is meaningless (why?); hence, the
rest of the eigenvectors (k 1) are used as k-means input.
Example 6.3. Consider the graph in Figure 6.8. We find two communities in this
graph using spectral clustering (i.e., k = 2). Then, we have
D = diag(2, 2, 4, 4, 4, 4, 4, 3, 1).
The adjacency matrix A and the unnormalized laplacian L are

0 1 1 0 0 0 0 0 0

1 0 1 0 0 0 0 0 0

1 1 0 1 1 0 0 0 0

0 0 1 0 1 1 1 0 0

A = 0 0 1 1 0 1 1 0 0 ,

0 0 0 1 1 0 1 1 0

0 0 0 1 1 1 0 1 0

0 0 0 0 0 1 1 0 1

0 0 0 0 0 0 0 1 0
1

For more details refer to [56].

191

(6.16)

(6.17)

L = D A = 0

1 1

2 1

4 1

0 1

1 1

0 1

1 1

0 1

4 1

0 1

0 1 1

4
0

0
0

0
0

0
0

0
0 .

1
0

1
0

3 1

1
1
0

(6.18)
We aim to find two communities; therefore, we get two eigenvectors corresponding to the two smallest eigenvalues from L:
1
2
3
4
X = 5
6
7
8
9

0.33 0.46

0.33 0.46

0.33 0.26

0.33 0.01

0.33 0.01 .

0.33
0.13

0.33
0.13

0.33
0.33

0.33
0.59

(6.19)

As mentioned, the first eigenvector is meaningless, because it assigns all nodes

to the same community. The second is used with k-means; based on the vector
signs, we get communities {1, 2, 3} and {4, 5, 6, 7, 8, 9}.
Robust Communities

k-connected

When seeking robust communities, our goal is to find subgraphs robust

enough such that removing some edges or nodes does not disconnect the
subgraph. A k-vertex connected graph (or k-connected) is an example of
such a graph. In this graph, k is the minimum number of nodes that must
192

be removed to disconnect the graph (i.e., there exist at least k independent

paths between any pair of nodes). A similar subgraph is the k-edge graph,
where at least k edges must be removed to disconnect the graph. An upperbound analysis on k-edge connectivity shows that the minimum degree for
any node in the graph should not be less than k (why?). For example, a
complete graph of size n is a unique n-connected graph, and a cycle is a
2-connected graph.
Modular Communities
Modularity is a measure that defines how likely the community structure
found is created at random. Clearly, community structures should be far
from random. Consider an undirected graph G(V, E), |E| = m where the
degrees are known beforehand, but edges are not. Consider two nodes vi
and v j , with degrees di and d j , respectively. What is the expected number of
edges between these two nodes? Consider node vi . For any edge going out
of vi randomly, the probability of this edge getting connected to node v j is
dj
P
i di

dj
.
2m

Because the degree for vi is di , we have di number of such edges;

dd

i j
hence, the expected number of edges between vi and v j is 2m
. So, given
a degree distribution, the expected number of edges between any pair of
vertices can be computed. Real-world communities are far from random;
therefore, the more distant they are from randomly generated communities, the more structure they exhibit. Modularity defines this distance, Modularity
and modularity maximization tries to maximize this distance. Consider
a partitioning of the graph G into k partitions, P = (P1 , P2 , P3 , . . . , Pk ). For
partition Px , this distance can be defined as

Ai j

vi ,v j Px

di d j
.
2m

(6.20)

This distance can be generalized for partitioning P with k partitions,

k
X
X

Ai j

x=1 vi ,v j Px

di d j
.
2m

(6.21)

The summation is over all edges (m), and because all edges are counted
twice (Ai j = A ji ), the normalized version of this distance is defined as
193

modularity [211]:

k
di d j
1 X X
.
Ai j
Q=

2m x=1 v ,v P
2m
i

(6.22)

We define the modularity matrix as B = A ddT /2m, where d Rn1

is the degree vector for all nodes. Similar to spectral clustering matrix
formulation, modularity can be reformulated as
Q=

1
Tr(XT BX),
2m

(6.23)

where X Rnk is the indicator (partition membership) function; that

is, Xi j = 1 iff. vi P j . This objective can be maximized such that the best
membership function is extracted with respect to modularity. The problem
is NP-hard; therefore, we relax X to X that has an orthogonal structure
(X T X = Ik ). The optimal X can be computed using the top k eigenvectors
of B corresponding to the largest positive eigenvalues. Similar to spectral
Note that this requires that
clustering, to find X, we can run k-means on X.
B has at least k positive eigenvalues.
Dense Communities

Graph Density

Often, we are interested in dense communities, which have sufficiently frequent interactions. These communities are of particular interest in social
media where we would like to have enough interactions for analysis to
make statistical sense. When we are measuring density in communities,
the community may or may not be connected as long as it satisfies the properties required, assuming connectivity is not one such property. Cliques,
clubs, and clans are examples of connected dense communities. Here,
we focus on subgraphs that have the possibility of being disconnected.
Density-based community detection has been extensively discussed in the
field of clustering (see Chapter 5, Bibliographic Notes).
The density of a graph defines how close a graph is to a clique. In
other words, the density is the ratio of the

number of edges |E| that graph
:
G has over the maximum it can have |V|
2
=

|E|
.
|V|

2

194

(6.24)



A graph G = (V, E) is -dense if |E| |V|
. Note that a 1-dense graph
2
is a clique. Here, we discuss the interesting scenario of connected dense
graphs (i.e., quasi-cliques). A quasi-clique (or -clique) is a connected Quasi-Clique
-dense graph. Quasi-cliques can be searched for using approaches previously discussed for finding cliques. We can utilize the brute-force clique
identification algorithm (Algorithm 6.1) for finding quasi-cliques as well.
The only part of the algorithm that needs to be changed is the part where
the clique condition is checked (Line 8). This can be replaced with a quasiclique checking condition. In general, because there is less regularity in
quasi-cliques, searching for them becomes harder. Interested readers can
refer to the bibliographic notes for faster algorithms.
Hierarchical Communities
All previously discussed methods have considered communities at a single
level. In reality, it is common to have hierarchies of communities, in which
each community can have sub/super communities. Hierarchical clustering
deals with this scenario and generates community hierarchies. Initially, n
nodes are considered as either 1 or n communities in hierarchical clustering.
These communities are gradually merged or split (agglomerative or divisive
hierarchical clustering algorithms), depending on the type of algorithm,
until the desired number of communities are reached. A dendrogram
is a visual demonstration of how communities are merged or split using
hierarchical clustering. The Girvan-Newman [101] algorithm is specifically
designed for finding communities using divisive hierarchical clustering.
The assumption underlying this algorithm is that, if a network has a
set of communities and these communities are connected to one another
with a few edges, then all shortest paths between members of different
communities should pass through these edges. By removing these edges
(at times referred to as weak ties), we can recover (i.e., disconnect) communities in a network. To find these edges, the Girvan-Newman algorithm
uses a measure called edge betweenness and removes edges with higher edge
betweenness. For an edge e, edge betweenness is defined as the number Edge
of shortest paths between node pairs (vi , v j ) such that the shortest path Betweenness
between vi and v j passes through e. For instance, in Figure 6.9(a), edge
betweenness for edge e(1, 2) is 6/2 + 1 = 4, because all the shortest paths
from 2 to {4, 5, 6, 7, 8, 9} have to either pass e(1, 2) or e(2, 3), and e(1, 2) is the
shortest path between 1 and 2. Formally, the Girvan-Newman algorithm
195

Figure 6.9: An Example of Girvan-Newman Algorithm Example: (a) graph

and (b) its hierarchical clustering dendrogram based on edge betweenness.
Girvan-Newman
Algorithm

is as follows:
1. Calculate edge betweenness for all edges in the graph.
2. Remove the edge with the highest betweenness.
3. Recalculate betweenness for all edges affected by the edge removal.
4. Repeat until all edges are removed.
Example 6.4. Consider the graph depicted in Figure 6.9(a). For this graph, the
edge-betweenness values are as follows:

1 2 3 4 5 6 7 8 9

1
0 4 1 9 0 0 0 0 0

2
4 0 4 0 0 0 0 0 0

3
1 4 0 9 0 0 0 0 0

4
9 0 9 0 10 10 0 0 0

.
(6.25)
5

0
0
0
10
0
1
6
3
0

6
0 0 0 10 1 0 6 3 0

7
0 0 0 0 6 6 0 2 8

8
0 0 0 0 3 3 2 0 0

9
0 0 0 0 0 0 8 0 0
Therefore, by following the algorithm, the first edge that needs to be removed
is e(4, 5) (or e(4, 6)). By removing e(4, 5), we compute the edge betweenness once
196

Figure 6.10: Community Detection Algorithms.

again; this time, e(4, 6) has the highest betweenness value: 20. This is because
all shortest paths between nodes {1,2,3,4} to nodes {5,6,7,8,9} must pass e(4, 6);
therefore, it has betweenness 4 5 = 20. By following the first few steps of
the algorithm, the dendrogram shown in Figure 6.9(b) and three disconnected
communities ({1, 2, 3, 4}, {5, 6, 7, 8}, {9}) can be obtained.
We discussed various community detection algorithms in this section.
Figure 6.10 summarizes the two categories of community detection algorithms.

6.2

Community Evolution

Community detection algorithms discussed so far assume that networks

are static; that is, their nodes and edges are fixed and do not change
over time. In reality, with the rapid growth of social media, networks and
their internal communities change over time. Earlier community detection
algorithms have to be extended to deal with evolving networks. Before
197

analyzing evolving networks, we need to answer the question, How do

networks evolve? In this section, we discuss how networks evolve in general
and then how communities evolve over time. We also demonstrate how
communities can be found in these evolving networks.

6.2.1

How Networks Evolve

Large social networks are highly dynamic, where nodes and links appear
or disappear over time. In these evolving networks, many interesting patterns are observed; for instance, when distances (in terms of shortest path
distance) between two nodes increase, their probability of getting connected decreases.2 We discuss three common patterns that are observed in
evolving networks: segmentation, densification, and diameter shrinkage.
Network Segmentation
Often, in evolving networks, segmentation takes place, where the large
network is decomposed over time into three parts:
1. Giant Component: As network connections stabilize, a giant component of nodes is formed, with a large proportion of network nodes
and edges falling into this component.
2. Stars: These are isolated parts of the network that form star structures. A star is a tree with one internal node and n leaves.
3. Singletons: These are orphan nodes disconnected from all nodes in
the network.
Figure 6.11 depicts a segmented network and these three components.
Graph Densification
It is observed in evolving graphs that the density of the graph increases as
the network grows. In other words, the number of edges increases faster
than the number of nodes. This phenomenon is called densification. Let
V(t) denote nodes at time t and let E(t) denote edges at time t,
|E(t)| |V(t)| .
2

See [154] for details.

198

(6.26)

Figure 6.11: Network Segmentation. The network is decomposed into a

giant component (dark gray), star components (medium gray), and singletons (light gray).
If densification happens, then we have 1 2. There is linear
growth when = 1, and we get clique structures when = 2 (why?).
Networks exhibit values between 1 and 2 when evolving. Figure 6.12
depicts a log-log graph for densification for a physics citation network and
a patent citation network. During the evolution process in both networks,
the number of edges is recorded as the number of nodes grows. These
recordings show that both networks have 1.6 (i.e., the log-log graph
of |E| with respect to |V| is a straight line with slope 1.6). This value also
implies that when V is given, to realistically model a social network, we
should generate O(|V|1.6 ) edges.

Figure 6.12: Graph Densification (from [167]).

199

Figure 6.13: Diameter Shrinkage over Time for a Patent Citation Network
(from [167]).

Diameter Shrinkage
Another property observed in large networks is that the network diameter
shrinks in time. This property has been observed in random graphs as well
(see Chapter 4). Figure 6.13 depicts the diameter shrinkage for the same
patent network discussed in Figure 6.12.
In this section we discussed three phenomena that are observed in
evolving networks. Communities in evolving networks also evolve. They
appear, grow, shrink, split, merge, or even dissolve over time. Figure 6.14
depicts different situations that can happen during community evolution.
Both networks and their internal communities evolve over time. Given
evolution information (e.g., when edges or nodes are added), how can
we study evolving communities? And can we adapt static (nontemporal)
methods to use this temporal information? We discuss these questions
next.
200

Figure 6.14: Community Evolution (reproduced from [226]).

6.2.2

Community Detection in Evolving Networks

Consider an instant messaging (IM) application in social media. In these

IM systems, members become available or offline frequently. Consider individuals as nodes and messages between them as edges. In this
example, we are interested in finding a community of individuals who
send messages to one another frequently. Clearly, community detection at
any time stamp is not a valid solution because interactions are limited at
any point in time. A valid solution to this problem needs to use temporal
information and interactions between users over time. In this section, we
present community detection algorithms that incorporate temporal information. To incorporate temporal information, we can extend previously
discussed static methods as follows:
1. Take t snapshots of the network, G1 , G2 , . . . , Gt , where Gi is a snapshot
at time i.
201

2. Perform a static community detection algorithm on all snapshots

independently.
3. Assign community members based on communities found in all t
different time stamps. For instance, we can assign nodes to communities based on voting. In voting, we assign nodes to communities
they belong to the most over time.
Unfortunately, this method is unstable in highly dynamic networks
because community memberships are always changing. An alternative is
to use evolutionary clustering.
Evolutionary Clustering
In evolutionary clustering, it is assumed that communities do not change
most of the time; hence, it tries to minimize an objective function that
considers both communities at different time stamps (snapshot cost or
SC) and how they evolve throughout time (temporal cost or TC). Then,
the objective function for evolutionary clustering is defined as a linear
combination of the snapshot cost and temporal cost (SC and TC),
Cost = SC + (1 ) TC,

(6.27)

where 0 1. Let us assume that spectral clustering (discussed in

Section 6.1.3) is used to find communities at each time stamp. We know
that the objective for spectral clustering is Tr(XT LX) s.t. XT X = Im , so we
will have the objective function at time t as
Costt = SC + (1 ) TC,
= Tr(XtT LXt ) + (1 ) TC,

(6.28)
(6.29)

where Xt is the community membership matrix at time t. To define TC,

we can compute the distance between the community assignments of two
snapshots:
TC = ||Xt Xt1 ||2 .
(6.30)
Unfortunately, this requires both Xt and Xt1 to have the same number
of columns (number of communities). Moreover, Xt is not unique and
202

can change by orthogonal transformations;3 therefore, the distance value

||Xt Xt1 ||2 can change arbitrarily. To remove the effect of orthogonal
transformations and allow different numbers of columns, TC is defined as
TC =
=
=
=
=

1
T
||Xt XtT Xt1 Xt1
||2 ,
2
1
T T
T
) (Xt XtT Xt1 Xt1
)),
Tr((Xt XtT Xt1 Xt1
2
1
T
T
T
),
Tr(Xt XtT Xt XtT 2Xt XtT Xt1 Xt1
+ Xt1 Xt1
Xt1 Xt1
2
T
),
Tr(I Xt XtT Xt1 Xt1
T
T
Tr(I Xt Xt1 Xt1 Xt ),
(6.31)

where 12 is for mathematical convenience, and Tr(AB) = Tr(BA) is used.

Therefore, evolutionary clustering objective can be stated as
1
T
Costt = Tr(XtT LXt ) + (1 ) ||Xt XtT Xt1 Xt1
||2 ,
2
T
= Tr(XtT LXt ) + (1 ) Tr(I XtT Xt1 Xt1
Xt ),
T
T
T
T
= Tr(Xt LXt ) + (1 ) Tr(Xt IXt Xt Xt1 Xt1
Xt ),
T
= Tr(XtT LXt ) + Tr(XtT (1 )IXt XtT (1 )Xt1 Xt1
Xt ).
(6.32)
Assuming the normalized Laplacian is used in spectral clustering, L =
I D1/2
At D1/2
,
t
t
) Xt )
Costt = Tr(XtT (I D1/2
At D1/2
t
t
T
T
T
+ Tr(Xt (1 ) I Xt Xt (1 ) Xt1 Xt1
Xt ),
T
) Xt ),
= Tr(XtT (I D1/2
At D1/2
(1 ) Xt1 Xt1
t
t
t ),
= Tr(Xt LX

(6.33)

T
where L = I D1/2
At D1/2
(1 )Xt1 Xt1
. Similar to spectral clustering,
t
t

Xt can be obtained by taking the top eigenvectors of L.

3

Let X be the solution to spectral clustering. Consider an orthogonal matrix Q

(i.e., QQT = I). Let Y = XQ. In spectral clustering, we are maximizing Tr(XT LX) =
Tr(XT LXQQT ) = Tr(QT XT LXQ) = Tr((XQ)T L(XQ)) = Tr(YT LY). In other words, Y is
another answer to our trace-maximization problem. This proves that the solution X to
spectral clustering is non-unique under orthogonal transformations Q.

203

Figure 6.15: Community Evaluation Example. Circles represent communities, and items inside the circles represent members. Each item is represented using a symbol, +, , or 4, that denotes the items true label.
Note that at time t, we can obtain Xt directly by solving spectral clustering for the laplacian of the graph at time t, but then we are not employing
any temporal information. Using evolutionary clustering and the new
we incorporate temporal information into our community delaplacian L,
tection algorithm and disallow user memberships in communities at time
t: Xt to change dramatically from Xt1 .

6.3

Community Evaluation

When communities are found, one must evaluate how accurately the detection task has been performed. In terms of evaluating communities, the
task is similar to evaluating clustering methods in data mining. Evaluating
clustering is a challenge because ground truth may not be available. We
consider two scenarios: when ground truth is available and when it is not.

6.3.1

Evaluation with Ground Truth

When ground truth is available, we have at least partial knowledge of

what communities should look like. Here, we assume that we are given
the correct community (clustering) assignments. We discuss four measures: precision and recall, F-measure, purity, and normalized mutual
information (NMI). Consider Figure 6.15, where three communities are
found and the points are shown using their true labels.
204

Precision and Recall

Community detection can be considered a problem of assigning all similar
nodes to the same community. In the simplest case, any two similar nodes
should be considered members of the same community. Based on our
assignments, four cases can occur:
1. True Positive (TP) Assignment: when similar members are assigned
to the same community. This is a correct decision.
2. True Negative (TN) Assignment: when dissimilar members are assigned to different communities. This is a correct decision.
3. False Negative (FN) Assignment: when similar members are assigned to different communities. This is an incorrect decision.
4. False Positive (FP) Assignment: when dissimilar members are assigned to the same community. This is an incorrect decision.
Precision (P) and Recall (R) are defined as follows,
TP
,
TP + FP
TP
R =
.
TP + FN
P =

(6.34)
(6.35)

Precision defines the fraction of pairs that have been correctly assigned
to the same community. Recall defines the fraction of pairs that the community detection algorithm assigned to the same community of all the
pairs that should have been in the same community.
Example 6.5. We compute these values for Figure 6.15. For TP, we need to
compute the number of pairs with the same label that are

in the same community.
For instance, for label and community 1, we have 52 such pairs. Therefore,
!
!
!
!
5
6
4
2
TP =
+
+ (
+
) = 32.
2
2
2
2
|{z}
|{z}
| {z }
Community 1 Community 2 Community 3
205

(6.36)

For FP, we need to compute dissimilar pairs that are in the same community.
For instance, for community 1, this is (5 1 + 5 1 + 1 1). Therefore,
FP = (5 1 + 5 1 + 1 1) +
(6 1)
+
(4 2)
= 25.
|
{z
}
| {z }
| {z }
Community 1
Community 2 Community 3

(6.37)

FN computes similar members that are in different communities. For instance,

for label +, this is (6 1 + 6 2 + 2 1). Similarly,
FN = (5 1) + (6 1 + 6 2 + 2 1) + (4 1) = 29.
| {z } |
{z
} | {z }
+

(6.38)

Finally, TN computes the number of dissimilar pairs in dissimilar communities:

+,

,+

4,+

4,

z}|{ z}|{ z}|{ z}|{

TN = ( 5 6 + 1 1 + 1 6 + 1 1 )
|
{z
}
Communities 1 and 2
,4

,+

+,4

4,+

z}|{ z}|{ z}|{ z}|{

+ (5 4 + 5 2 + 1 4 + 1 2)
|
{z
}
Communities 1 and 3
+,4

,+

,4

z}|{ z}|{ z}|{

+ ( 6 4 + 1 2 + 1 4 = 104.
|
{z
}
Communities 2 and 3

(6.39)

Hence,
32
= 0.56
32 + 25
32
R=
= 0.52.
32 + 29
P=

(6.40)
(6.41)

F-Measure
To consolidate precision and recall into one measure, we can use the harmonic mean of precision and recall:
F=2

PR
.
P+R

206

(6.42)

Computed for the same example, we get F = 0.54.

Purity
In purity, we assume that the majority of a community represents the community. Hence, we use the label of the majority of the community against
the label of each member of the community to evaluate the algorithm. For
instance, in Figure 6.15, the majority in Community 1 is ; therefore, we
assume majority label for that community. The purity is then defined
as the fraction of instances that have labels equal to their communitys
majority label. Formally,
k

Purity =

1 X
max |Ci L j |,
N i=1 j

(6.43)

where k is the number of communities, N is the total number of nodes, L j

is the set of instances with label j in all communities, and Ci is the set of
members in community i. In the case of our example, purity is 5+6+4
= 0.75.
20
Normalized Mutual Information
Purity can be easily manipulated to generate high values; consider when
nodes represent singleton communities (of size 1) or when we have very
large pure communities (ground truth = majority label). In both cases,
purity does not make sense because it generates high values.
A more precise measure to solve problems associated with purity is
the normalized mutual information (NMI) measure, which originates in
information theory. Mutual information (MI) describes the amount of
information that two random variables share. In other words, by knowing one of the variables, MI measures the amount of uncertainty reduced
regarding the other variable. Consider the case of two independent variables; in this case, the mutual information is zero because knowing one
does not help in knowing more information about the other. Mutual information of two variables X and Y is denoted as I(X, Y). We can use mutual
information to measure the information one clustering carries regarding
the ground truth. It can be calculated using Equation 6.44, where L and H
are labels and found communities; nh and nl are the number of data points
207

in community h and with label l, respectively; nh,l is the number of nodes

in community h and with label l; and n is the number of nodes.
X X nh,l
n nh,l
log
(6.44)
MI = I(X, Y) =
n
nh nl
hH lL

Unfortunately, mutual information is unbounded; however, it is common for measures to have values in range [0,1]. To address this issue, we
can normalize mutual information. We provide the following equation,
without proof, which will help us normalize mutual information,
MI min(H(L), H(H)),

(6.45)

where H() is the entropy function,

H(L) =

Xn

lL

H(H) =

Xn

hH

log

nl
n

(6.46)

log

nh
.
n

(6.47)

From Equation 6.45, we have MI H(L) and MI H(H); therefore,

(MI)2 H(H)H(L).

(6.48)

Equivalently,
MI

p
H(H) H(L).

(6.49)

Equation 6.49 can be used to normalize mutual information. Thus, we

introduce the NMI as
MI
NMI = p
.
p
H(L) H(H)
By plugging Equations 6.47, 6.46, and 6.44 into 6.50,
P P
nnh,l
hH
lL nh,l log nh nl
NMI = q
.
P
nh P
nl
( hH nh log n )( lL nl log n )

(6.50)

(6.51)

An NMI value close to one indicates high similarity between communities found and labels. A value close to zero indicates a long distance
between them.
208

Figure 6.16: Tag Clouds for Two Communities.

6.3.2

Evaluation without Ground Truth

When no ground truth is available, we can incorporate techniques based on

semantics or clustering quality measures to evaluate community detection
algorithms.
Evaluation with Semantics
A simple way of analyzing detected communities is to analyze other attributes (posts, profile information, content generated, etc.) of community
members to see if there is a coherency among community members. The
coherency is often checked via human subjects. For example, the Amazon Mechanical Turk platform4 allows defining this task on its platform
for human workers and hiring individuals from all around the globe to
perform tasks such as community evaluation. To help analyze these communities, one can use word frequencies. By generating a list of frequent
keywords for each community, human subjects determine whether these
keywords represent a coherent topic. A more focused and single-topic set
of keywords represents a coherent community. Tag clouds are one way of
demonstrating these topics. Figure 6.16 depicts two coherent tag clouds
for a community related to the U.S. Constitution and another for sports.
Larger words in these tag clouds represent higher frequency of use.
Evaluation Using Clustering Quality Measures
When experts are not available, an alternative is to use clustering quality
measures. This approach is commonly used when two or more commu4

http://www.mturk.com.

209

nity detection algorithms are available. Each algorithm is run on the target
network, and the quality measure is computed for the identified communities. The algorithm that yields a more desirable quality measure value
is considered a better algorithm. SSE (sum of squared errors) and intercluster distance are some of the quality measures. For other measures refer
to Chapter 5.
We can also follow this approach for evaluating a single community
detection algorithm; however, we must ensure that the clustering quality
measure used to evaluate community detection is different from the measure used to find communities. For instance, when using node similarity
to group individuals, a measure other than node similarity should be used
to evaluate the effectiveness of community detection.

210

6.4

Summary

In this chapter, we discussed community analysis in social media, answering three general questions: (1) how can we detect communities, (2) how
do communities evolve and how can we study evolving communities, and
(3) how can we evaluate detected communities? We started with a description of communities and how they are formed. Communities in social
media are either explicit (emic) or implicit (etic). Community detection
finds implicit communities in social media.
We reviewed member-based and group-based community detection
algorithms. In member-based community detection, members can be
grouped based on their degree, reachability, and similarity. For example, when using degrees, cliques are often considered as communities.
Brute-force clique identification is used to identify cliques. In practice, due
to the computational complexity of clique identifications, cliques are either
relaxed or used as seeds of communities. k-Plex is an example of relaxed
cliques, and the clique percolation algorithm is an example of methods that
use cliques as community seeds. When performing member-based community detection based on reachability, three frequently used subgraphs
are the k-clique, k-club, and k-clan. Finally, in member-based community
detection based on node similarity, methods such as Jaccard and Cosine
similarity help compute node similarity. In group-based community detection, we described methods that find balanced, robust, modular, dense,
or hierarchical communities. When finding balanced communities, one
can employ spectral clustering. Spectral clustering provides a relaxed
solution to the normalized cut and ratio cut in graphs. For finding robust communities, we search for subgraphs that are hard to disconnect.
k-edge and k-vertex graphs are two examples of these robust subgraphs.
To find modular communities, one can use modularity maximization and
for dense communities, we discussed quasi-cliques. Finally, we provided
hierarchical clustering as a solution to finding hierarchical communities,
with the Girvan-Newman algorithm as an example.
In community evolution, we discussed when networks and, on a lower
level, communities evolve. We also discussed how communities can be
detected in evolving networks using evolutionary clustering. Finally, we
presented how communities are evaluated when ground truth exists and
when it does not.

211

6.5

Bibliographic Notes

A general survey of community detection in social media can be found

in [91] and a review of heterogeneous community detection in [278]. In
related fields, [35, 305, 136] provide surveys of clustering algorithms and
[294] provides a sociological perspective. Comparative analysis of community detection algorithms can be found in [162] and [168]. The description
of explicit communities in this chapter is due to Kadushin [140].
For member-based algorithms based on node degree, refer to [158],
which provides a systematic approach to finding clique-based communities with pruning. In algorithms based on node reachability, one can find
communities by finding connected components in the network. For more
information on finding connected components of a graph refer to [130]. In
node similarity, we discussed structural equivalence, similarity measures,
and regular equivalence. More information on structural equivalence can
be found in [178, 166], on Jaccard similarity in [133], and on regular equivalence in [264].
In group-based methods that find balanced communities, we are often interested in solving the max-flow min-cut theorem. Linear programming and Ford-Fulkerson [61], Edmonds-Karp [80], and Push-Relabel [103]
methods are some established techniques for solving the max-flow min-cut
problem. We discussed quasi-cliques that help find dense communities.
Finding the maximum quasi-clique is discussed in [231]. A well-known
greedy algorithm for finding quasi-cliques is introduced by [2]. In their
approach a local search with a pruning strategy is performed on the graph
to enhance the speed of quasi-clique detection. They define a peel strategy,
in which vertices that have some degree k along with their incident edges
are recursively removed. There are a variety of algorithms to find dense
subgraphs, such as the one discussed in [99] where the authors propose
an algorithm that recursively fingerprints the graph (shingling algorithm)
and creates dense subgraphs. In group-based methods that find hierarchical communities, we described hierarchical clustering. Hierarchical
clustering algorithms are usually variants of single link, average link, or
complete link algorithms [135]. In hierarchical clustering, COBWEB [88]
and CHAMELEON [142] are two well-known algorithms.
In group-based community detection, latent space models [121, 129]
are also very popular, but are not discussed in this chapter. In addition to
the topics discussed in this chapter, community detection can also be per212

formed for networks with multiple types of interaction (edges) [279]; [280].
We also restricted our discussion to community detection algorithms that
use graph information. One can also perform community detection based
on the content that individuals share on social media. For instance, using
tagging relations (i.e., individuals who shared the same tag) [292], instead
of connections between users, one can discover overlapping communities,
which provides a natural summarization of the interests of the identified
communities.
In network evolution analysis, network segmentation is discussed in
[157]. Segment-based clustering [269] is another method not covered in
this chapter.
NMI was first introduced in [267] and in terms of clustering quality
measures, the Davies-Bouldin [67] measure, Rand index [236], C-index
[76], Silhouette index [241], and Goodman-Kruskal index [106] can be
used.

213

6.6

Exercises

1. Provide an example to illustrate how community detection can be

subjective.

Community Detection
2. Given a complete graph Kn , how many nodes will the clique percolation method generate for the clique graph for value k? How many
edges will it generate?
3. Find all k-cliques, k-clubs, and k-clans in a complete graph of size 4.
4. For a complete graph of size n, is it m-connected? What possible
values can m take?
5. Why is the smallest eigenvector meaningless when using an unnormalized laplacian matrix?
6. Modularity can be defined as
#
"
di d j
1 X
(ci , c j ),
Q=
Ai j
2m i j
2m

(6.52)

where ci and c j are the communities for vi and v j , respectively.

(ci , c j ) (Kronecker delta) is 1 when vi and v j both belong to the same
community (ci = c j ), and 0 otherwise.
What is the range [1 , 2 ] for Q values? Provide examples for
both extreme values of the range and cases where modularity
becomes zero.
What are the limitations for modularity? Provide an example
where modularity maximization does not seem reasonable.
Find three communities in Figure 6.8 by performing modularity
maximization.
7. For Figure 6.8:
214

 Compute Jaccard and Cosine similarity between nodes v4 and

v8 , assuming that the neighborhood of a node excludes the node
itself.
Compute Jaccard and Cosine similarity when the node is included in the neighborhood.

Community Evolution
8. What is the upper bound on densification factor ? Explain.

Community Evaluation
9. Normalized mutual information (NMI) is used to evaluate community detection results when the actual communities (labels) are
known beforehand.
What are the maximum and minimum values for the NMI?
Provide details.
Explain how NMI works (describe the intuition behind it).
10. Compute NMI for Figure 6.15.
11. Why is high precision not enough? Provide an example to show that
both precision and recall are important.
12. Discuss situations where purity does not make sense.
13. Compute the following for Figure 6.17:

Figure 6.17: Community Evaluation Example.

215

 precision and recall

F-measure
NMI
purity

216

This chapter is from Social Media Mining: An Introduction.

By Reza Zafarani, Mohammad Ali Abbasi, and Huan Liu.
Cambridge University Press, 2014. Draft version: April 20, 2014.
Complete Draft and Slides Available at: http://dmml.asu.edu/smm

Chapter

Information Diffusion in Social

Media
In February 2013, during the third quarter of Super Bowl XLVII, a power
outage stopped the game for 34 minutes. Oreo, a sandwich cookie company, tweeted during the outage: Power out? No Problem, You can still
dunk it in the dark. The tweet caught on almost immediately, reaching
nearly 15,000 retweets and 20,000 likes on Facebook in less than two days.
A simple tweet diffused into a large population of individuals. It helped
the company gain fame with minimum cost in an environment where
companies spent as much as $4 million to run a 30-second ad. This is an
example of information diffusion.
Information diffusion is a field encompassing techniques from a plethora
of sciences. In this chapter, we discuss methods from fields such as sociology, epidemiology, and ethnography, which can help social media mining.
Our focus is on techniques that can model information diffusion.
Societies provide means for individuals to exchange information through
various channels. For instance, people share knowledge with their immediate network (friends) or broadcast it via public media (TV, newspapers,
etc.) throughout the society. Given this flow of information, different
research fields have disparate views of what is an information diffusion
process. We define information diffusion as the process by which a piece of
information (knowledge) is spread and reaches individuals through interactions.
The diffusion process involves the following three elements:
1. Sender(s). A sender or a small set of senders initiate the information
217

diffusion process.
2. Receiver(s). A receiver or a set of receivers receive diffused information. Commonly, the set of receivers is much larger than the set of
senders and can overlap with the set of senders.
3. Medium. This is the medium through which the diffusion takes
place. For example, when a rumor is spreading, the medium can be
the personal communication between individuals.

Intervention

Local and Global

Dependence

This definition can be generalized to other domains. In a diseasespreading process, the disease is the analog to the information, and infection can be considered a diffusing process. The medium in this case is the
air shared by the infecter and the infectee. An information diffusion can
be interrupted. We define the process of interfering with information diffusion by expediting, delaying, or even stopping diffusion as intervention.
Individuals in online social networks are situated in a network where
they interact with others. Although this network is at times unavailable
or unobservable, the information diffusion process takes place in it. Individuals facilitate information diffusion by making individual decisions
that allow information to flow. For instance, when a rumor is spreading,
individuals decide if they are interested in spreading it to their neighbors.
They can make this decision either dependently (i.e., depending on the
information they receive from others) or independently. When they make
dependent decisions, it is important to gauge the level of dependence that
individuals have on others. It could be local dependence, where an individuals decision is dependent on all of his or her immediate neighbors
(friends) or global dependence, where all individuals in the network are
observed before making decisions.
In this chapter, we present in detail four general types of information
diffusion: herd behavior, information cascades, diffusion of innovation, and
epidemics.
Herd behavior takes place when individuals observe the actions of all
others and act in an aligned form with them. An information cascade
describes the process of diffusion when individuals merely observe their
immediate neighbors. In information cascades and herd behavior, the network of individuals is observable; however, in herding, individuals decide
based on global information (global dependence); whereas, in information
218

Figure 7.1: Information Diffusion Types.

cascades, decisions are made based on knowledge of immediate neighbors

(local dependence).
Diffusion of innovations provides a birds-eye view of how an innovation (e.g., a product, music video, or fad) spreads through a population. It
assumes that interactions among individuals are unobservable and that the
sole available information is the rate at which products are being adopted
throughout a certain period of time. This information is particularly interesting for companies performing market research, where the sole available
information is the rate at which their products are being bought. These
companies have no access to interactions among individuals. Epidemic
models are similar to diffusion of innovations models, with the difference
that the innovations analog is a pathogen and adoption is replaced by infection. Another difference is that in epidemic models, individuals do not
decide whether to become infected or not and infection is considered a random natural process, as long as the individual is exposed to the pathogen.
Figure 7.1 summarizes our discussion by providing a decision tree of the
information diffusion types.
219

7.1

Herd Behavior

Consider people participating in an online auction. Individuals are connected via the auctions site where they cannot only observe the bidding
behaviors of others but can also often view profiles of others to get a feel
for their reputation and expertise. Individuals often participate actively in
online auctions, even bidding on items that might otherwise be considered
unpopular. This is because they trust others and assume that the high
number of bids that the item has received is a strong signal of its value. In
this case, herd behavior has taken place.
Herd behavior, a term first coined by British surgeon Wilfred [283],
describes when a group of individuals performs actions that are aligned
without previous planning. It has been observed in flocks, herds of animals, and in humans during sporting events, demonstrations, and religious gatherings, to name a few examples. In general, any herd behavior
requires two components:
1. connections between individuals
2. a method to transfer behavior among individuals or to observe their
behavior

Solomon Asch
Conformity
Experiment

Individuals can also make decisions that are aligned with others (mindless decisions) when they conform to social or peer pressure. A well-known
example is the set of experiments performed by Solomon Asch during the
1950s [17]. In one experiment, he asked groups of students to participate
in a vision test where they were shown two cards (Figure 7.2), one with
a single line segment and one with three lines, and told to match the line
segments with the same length.
Each participant was put into a group where all the other group members were actually collaborators with Asch, although they were introduced
as participants to the subject. Asch found that in control groups with no
pressure to conform, in which the collaborators gave the correct answer,
only 3% of the subjects provided an incorrect answer. However, when participants were surrounded by individuals providing an incorrect answer,
up to 32% of the responses were incorrect.
In contrast to this experiment, we refer to the process in which individuals consciously make decisions aligned with others by observing the
decisions of other individuals as herding or herd behavior. In theory, there is
220

Figure 7.2: Solomon Asch Experiment. Participants were asked to match

the line on the left card to the line on the right card that has the exact same
length.
no need to have a network of people. In practice, there is a network, and
this network is close to a complete graph, where nodes can observe at least
most other nodes. Consider this example of herd behavior.
Example 7.1. Diners Example [23]. Assume you are visiting a metropolitan
area that you are not familiar with. Planning for dinner, you find restaurant A
with excellent reviews online and decide to go there. When arriving at A, you see
that A is almost empty and that restaurant B, which is next door and serves the
same cuisine, is almost full. Deciding to go to B, based on the belief that other
diners have also had the chance of going to A, is an example of herd behavior.
In this example, when B is getting more and more crowded, herding is
taking place. Herding happens because we consider crowd intelligence trustworthy. We assume that there must be private information not known to
us, but known to the crowd, that resulted in the crowd preferring restaurant
B over A. In other words, we assume that, given this private information,
we would have also chosen B over A.
In general, when designing a herding experiment, the following four
conditions need to be satisfied:
1. There needs to be a decision made. In this example, the decision
involves going to a restaurant.
2. Decisions need to be in sequential order.
221

3. Decisions are not mindless, and people have private information that
helps them decide.
4. No message passing is possible. Individuals do not know the private
information of others, but can infer what others know from what
they observe from their behavior.
Anderson and Holt [11, 12] designed an experiment satisfying these
four conditions, in which students guess whether an urn containing red
and blue marbles is majority red or majority blue. Each student had access
to the guesses of students beforehand. Anderson and Holt observed a herd
behavior where students reached a consensus regarding the majority color
over time. It has been shown [78] that Bayesian modeling is an effective
technique for demonstrating why this herd behavior occurs. Simply put,
computing conditional probabilities and selecting the most probable majority color result in herding over time. We detail this experiment and how
conditional probabilities can explain why herding takes place next.

7.1.1

Bayesian Modeling of Herd Behavior

In this section, we show how Bayesian modeling can be used to explain

herd behavior by describing in detail the urn experiment devised by Anderson and Holt [11, 12]. In front of a large class of students, there is an urn
that has three marbles in it. These marbles are either blue (B) or red (R), and
we are guaranteed to have at least one of each color. So, the urn is either
majority blue (B,B,R) or majority red (R,R,B). We assume the probability of
being either majority blue or majority red is 50%. During the experiment,
each student comes to the urn, picks one marble, and checks its color in
private. The student predicts majority blue or red, writes the prediction on
the blackboard (which was blank initially), and puts the marble back in the
urn. Other students cannot see the color of the marble taken out, but can
see the predictions made by the students regarding the majority color and
written on the blackboard. Let the BOARD variable denote the sequence
of predictions written on the blackboard. So, before the first student, it is

222

We start with the first student. If the marble selected is red, the prediction will be majority red; if blue, it will be majority blue. Assuming it was
blue, on the board we have

The second student can pick a blue or a red marble. If blue, he also
predicts majority blue because he knows that the previous student must
have picked blue. If red, he knows that because he has picked red and
the first student has picked blue, he can randomly assume majority red or
blue. So, after the second student we either have

Assume we end up with BOARD: {B, B}. In this case, if the third student
takes out a red ball, the conditional probability is higher for majority blue,
although she observed a red marble. Hence, a herd behavior takes place,
and on the board, we will have BOARD: {B,B,B}. From this student and
onward, independent of what is being observed, everyone will predict
majority blue. Let us demonstrate why this happens based on conditional
probabilities and our problem setting. In our problem, we know that the
first student predicts majority blue if P(majority blue|students obervation) >
1/2 and majority red otherwise. We also know from the experiments setup
that
P(majority blue) = P(majority red) = 1/2,
P(blue|majority blue) = P(red|majority red) = 2/3.

(7.1)
(7.2)

Let us assume that the first student observes blue; then,

P(blue|majority blue)P(majority blue)
P(blue)
P(blue) = P(blue|majority blue)P(majority blue)
+ P(blue|majority red)P(majority red)
= 2/3 1/2 + 1/3 1/2 = 1/2.

P(majority blue|blue) =

223

(7.3)
(7.4)
(7.5)

= 2/3. So, if the first student

Therefore, P(majority blue|blue) = 2/31/2
1/2
picks blue, she will predict majority blue, and if she picks red, she will
predict majority red. Assuming the first student picks blue, the same
argument holds for the second student; if blue is picked, he will also
predict majority blue. Now, in the case of the third student, assuming she
has picked red, and having BOARD: {B,B} on the blackboard, then,
P(blue,blue,red|majority blue)
P(blue,blue,red)
P(majority blue)
(7.6)
P(blue, blue, red|majority blue) = 2/3 2/3 1/3 = 4/27
(7.7)
P(blue, blue, red) = P(blue, blue, red|majority blue)
P(majority blue)
+ P(blue, blue, red|majority red)
P(majority red)
(7.8)
= (2/3 2/3 1/3) 1/2
+ (1/3 1/3 2/3) 1/2 = 1/9.
P(majority blue|blue, blue, red) =

= 2/3. So, the third

Therefore, P(majority blue|blue,blue,red) = 4/271/2
1/9
student predicts majority blue even though she picks red. Any student
after the third student also predicts majority blue regardless of what is
being picked because the conditional remains above 1/2. Note that the
urn can in fact be majority red. For instance, when blue, blue, red is picked,
there is a 1 2 /3 =1 /3 chance that it is majority red; however, due to
herding, the prediction could become incorrect. Figure 7.3 depicts the
herding process. In the figure, rectangles represent the board status, and
edge values represent the observations. Dashed arrows depict transitions
between states that contain the same statistical information that is available
to the students.

7.1.2

Intervention

As herding converges to a consensus over time, it is interesting how one can

intervene with this process. In general, intervention is possible by providing private information to individuals that was not previously available.
Consider an urn experiment where individuals decide on majority red over
time. Either (1) a private message to individuals informing them that the
224

Figure 7.3: Urn Experiment. Rectangles represent student predictions

written on the blackboard, and edge values represent what the students
observe. Rectangles are filled with the most likely majority, computed
from conditional probabilities.
urn is majority blue or (2) writing the observations next to predictions on
the board stops the herding and changes decisions.

7.2

Information Cascades

In social media, individuals commonly repost content posted by others

in the network. This content is often received via immediate neighbors
(friends). An information cascade occurs as information propagates through
friends.
Formally, an information cascade is defined as a piece of information
or decision being cascaded among a set of individuals, where (1) individuals are connected by a network and (2) individuals are only observing
225

decisions of their immediate neighbors (friends). Therefore, cascade users

have less information available to them compared to herding users, where
almost all information about decisions are available.
There are many approaches to modeling information cascades. Next,
we introduce a basic model that can help explain information cascades.

7.2.1

Independent Cascade Model (ICM)

In this section, we discuss the independent cascade model (ICM) [146]

that can be utilized to model information cascades. Variants of this model
have been discussed in the literature. Here, we discuss the one detailed by
Kempe et al. [146]. Interested readers can refer to the bibliographic notes
for further references. Underlying assumptions for this model include the
following:
The network is represented using a directed graph. Nodes are actors
and edges depict the communication channels between them. A
node can only influence nodes that it is connected to.
Decisions are binary nodes can be either active or inactive. An active
nodes means that the node decided to adopt the behavior, innovation,
or decision.
A node, once activated, can activate its neighboring nodes.
Activation is a progressive process, where nodes change from inactive
to active, but not vice versa.1

Sender-Centric
Model

Considering nodes that are active as senders and nodes that are being
activated as receivers, in the independent cascade model (ICM) senders
activate receivers. Therefore, ICM is denoted as a sender-centric model. In
this model, the node that becomes active at time t has, in the next time step
t + 1, one chance of activating each of its neighbors. Let v be an active node
at time t. Then, for any neighbor w, there is a probability pv,w that node w
gets activated at t + 1. A node v that has been activated at time t has a single
chance of activating its neighbor w and that activation can only happen at
t + 1. We start with a set of active nodes and we continue until no further
activation is possible. Algorithm 7.1 details the process of the ICM model.
1

This assumption can be lifted [146].

226

Algorithm 7.1 Independent Cascade Model (ICM)

Require: Diffusion graph G(V, E), set of initial activated nodes A0 , activation probabilities pv,w
1: return Final set of activated nodes A
2: i = 0;
3: while Ai , {} do
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:

i = i + 1;
Ai = {};
for all v Ai1 do
for all w neighbor of v, w < ij=0 A j do
rand = generate a random number in [0,1];
if rand < pv,w then
activate w;
Ai = Ai {w};
end if
end for
end for
end while
A = ij=0 A j ;
Return A ;

Example 7.2. Consider the network in Figure 7.4 as an example. The network is
undirected; therefore, we assume pv,w = pw,v . Since it is undirected, for any two
vertices connected via an edge, there is an equal chance of one activating the other.
Consider the network in step 1. The values on the edges denote pv,w s. The ICM
procedure starts with a set of nodes activated. In our case, it is node v1 . Each
activated node gets one chance of activating its neighbors. The activated node
generates a random number for each neighbor. If the random number is less than
the respective pv,w of the neighbor (see Algorithm 7.1, lines 911), the neighbor
gets activated. The random numbers generated are shown in Figure 7.4 in the
form of inequalities, where the left-hand side is the random number generated and
the right-hand side is the pv,w . As depicted, by following the procedure after five
steps, five nodes get activated and the ICM procedure converges.
227

Figure 7.4: Independent Cascade Model (ICM) Simulation. The numbers

on the edges represent the weights pv,w . When there is an inequality, the
activation condition is checked. The left number denotes the random
number generated, and the right number denotes weight pv,w .

Clearly, the ICM characterizes an information diffusion process.2 It

is sender-centered, and once a node is activated, it aims to activate all
its neighboring nodes. Node activation in ICM is a probabilistic process.
Thus, we might get different results for different runs.
One interesting question when dealing with the ICM model is that
given a network, how to activate a small set of nodes initially such that the
final number of activated nodes in the network is maximized. We discuss
this next.
2

See [112] for an application in the blogosphere.

228

7.2.2

Maximizing the Spread of Cascades

Consider a network of users and a company that is marketing a product. The company is trying to advertise its product in the network. The
company has a limited budget; therefore, not all users can be targeted.
However, when users find the product interesting, they can talk with their
friends (immediate neighbors) and market the product. Their neighbors,
in turn, will talk about it with their neighbors, and as this process progresses, the news about the product is spread to a population of nodes in
the network. The company plans on selecting a set of initial users such
that the size of the final population talking about the product is maximized.
Formally, let S denote a set of initially activated nodes (seed set) in ICM.
Let f (S) denote the number of nodes that get ultimately activated in the
network if nodes in S are initially activated. For our ICM example depicted
in Figure 7.4, |S| = 1 and f (S) = 5. Given a budget k, our goal is to find a
set S such that its size is equal to our budget |S| = k and f (S) is maximized.
Since the activations in ICM depend on the random number generated
for each node (see line 9, Algorithm 7.1), it is challenging to determine
the number of nodes that ultimately get activated f (S) for a given set S.
In other words, the number of ultimately activated individuals can be
different depending on the random numbers generated. ICM can be made
deterministic (nonrandom) by generating these random numbers in the
beginning of the ICM process for the whole network. In other words, we
can generate a random number ru,w for any connected pair of nodes. Then,
whenever node v has a chance of activating u, instead of generating the
random number, it can compare ru,w with pv,w . Following this approach,
ICM becomes deterministic, and given any set of initially activated nodes
S, we can compute the number of ultimately activated nodes f (S).
Before finding S, we detail properties of f (S). The function f (S) is nonnegative because for any set of nodes S, in the worst case, no node gets
activated. It is also monotone:
f (S {v}) f (S).

(7.9)

This is because when a node is added to the set of initially activated nodes,
it either increases the number of ultimately activated nodes or keeps them
the same. Finally, f (S) is submodular. A set function f is submodular if Submodular
function
for any finite set N,
S T N, v N \ T, f (S {v}) f (S) f (T {v}) f (T).
229

(7.10)

The proof that function f is submodular is beyond the scope of this

book, but interested readers are referred to [146] for the proof. So, f is nonnegative, monotone, and submodular. Unfortunately, for a submodular
non-negative monotone function f , finding a k element set S such that f (S)
is maximized is an NP-hard problem [146]. In other words, we know no
efficient algorithm for finding this set.3 Often, when a computationally
challenging problem is at hand, approximation algorithms come in handy.
In particular, the following theorem helps us approximate S.
Theorem 7.1 (Kempe et al. [146]). Let f be a (1) non-negative, (2) monotone,
and (3) submodular set function. Construct k-element set S, each time by adding
node v, such that f (S {v}) (or equivalently, f (S {v}) f (s)) is maximized.
Let SOptimal be the k-element set such that f is maximized. Then f (S) (1
1
) f (SOptimal ).
e
This theorem states that by constructing the set S greedily one can get at
least a (1 1/e) 63% approximation of the optimal value. Algorithm 7.2
details this greedy approach. The algorithm starts with an empty set S and
adds node v1 , which ultimately activates most other nodes if activated.
Formally, v1 is selected such that f ({v1 }) is the maximum. The algorithm
then selects the second node v2 such that f ({v1 , v2 }) is maximized. The process is continued until the kth node vk is selected. Following this algorithm,
we find an approximately reasonable solution for the problem of cascade
maximization.
Example 7.3. For the following graph, assume that node i activates node j when
|i j| 2 (mod 3). Solve cascade maximization for k = 2.

To find the first node v, we compute f ({v}) for all v. We start with node 1. At
time 0, node 1 can only activate node 6, because
|1 6| 2 (mod 3),
3

(7.11)

Formally, assuming P , NP, there is no polynomial time algorithm for this problem.

230

Algorithm 7.2 Maximizing the spread of cascades Greedy algorithm

Require: Diffusion graph G(V, E), budget k
1: return Seed set S (set of initially activated nodes)
2: i = 0;
3: S = {};
4: while i , k do
5:
v = arg maxvV\S f (S {v});
or equivalently arg maxvV\S f (S {v}) f (s)
6:
S = S {v};
7:
i = i + 1;
8: end while
9: Return S;
|1 5| . 2 (mod 3).

(7.12)

At time 1, node 1 can no longer activate others, but node 6 is active and can
activate others. Node 6 has outgoing edges to nodes 4 and 5. From 4 and 5, node
6 can only activate 4:
|6 4| 2 (mod 3)
|6 5| . 2 (mod 3).

(7.13)
(7.14)

At time 2, node 4 is activated. It has a single out-link to node 2 and since

|4 2| 2 (mod 3), 2 is activated. Node 2 cannot activate other nodes; therefore,
f ({1}) = 4. Similarly, we find that f ({2}) = 1, f ({3}) = 1, f ({4}) = 2, f ({5}) = 1,
and f ({6}) = 4. So, 1 or 6 can be chosen for our first node. Let us choose 6. If 6
is initially activated, nodes 1, 2, 4, and 6 will become activated at the end. Now,
from the set {1, 2, 3, 4, 5, 6} \ {1, 2, 4, 6} = {3, 5}, we need to select one more node.
This is because in the setting for this example, f ({6, 1}) = f ({6, 2}) = f ({6, 4}) =
f ({6}) = 4. In general, one needs to compute f (S {v}) for all v V \ S (see
Algorithm 7.2, line 5). We have f ({6, 3}) = f ({6, 5}) = 5, so we can select one
node randomly. We choose 3. So, S = {6, 3} and f (S) = 5.

7.2.3

Intervention

Consider a false rumor spreading in social media. This is an example

where we are interested in stopping an information cascade in social media.
231

Intervention in the independent cascade model can be achieved using three

methods:
1. By limiting the number of out-links of the sender node and potentially
reducing the chance of activating others. Note that when the sender
node is not connected to others via directed edges, no one will get
activated by the sender.
2. By limiting the number of in-links of receiver nodes and therefore
reducing their chance of getting activated by others.
3. By decreasing the activation probability of a node (pv,w ) and therefore
reducing the chance of activating others.

7.3

Diffusion of Innovations

Diffusion of innovations is a phenomenon observed regularly in social media. A music video going viral or a piece of news being retweeted many
times are examples of innovations diffusing across social networks. As
defined by Rogers [239], an innovation is an idea, practice, or object that
is perceived as new by an individual or other unit of adoption. Innovations are created regularly; however, not all innovations spread through
populations. The theory of diffusion of innovations aims to answer why
and how these innovations spread. It also describes the reasons behind
the diffusion process, the individuals involved, and the rate at which ideas
spread. In this section, we review characteristics of innovations that are
likely to be diffused through populations and detail well-known models
in the diffusion of innovations. Finally, we provide mathematical models
that can model the process of diffusion of innovations and describe how
we can intervene with these models.

7.3.1

Innovation Characteristics

For an innovation to be adopted, the individual adopting it (adopter) and

the innovation must have certain qualities.
Innovations must be highly observable, should have a relative advantage over current practices, should be compatible with the sociocultural
232

paradigm to which it is being presented, should be observable under various trials (trialability), and should not be highly complex.
In terms of individual characteristics, many researchers [239, 127] claim
that the adopter should adopt the innovation earlier than other members
of his or her social circle (innovativeness).

7.3.2

Diffusion of Innovations Models

Some of the earliest models for diffusion of innovations were provided by

Gabriel Tarde in the early 20th century [281]. In this section, we review
basic diffusion of innovations models. Interested readers may refer to the
bibliographical notes for further study.

Ryan and Gross: Adopter Categories

Ryan and Gross [242] studied the adoption of hybrid seed corn by farmers
in Iowa [266]. The hybrid seed corn was highly resistant to diseases and
other catastrophes such as droughts. However, farmers did not adopt it
because of its high price and the seeds inability to reproduce. Their study
showed that farmers received information through two main channels:
mass communications from companies selling the seeds and interpersonal
communications with other farmers. They found that although farmers received information from the mass channel, the influence on their behavior
was coming from the interpersonal channel. They argued that adoption
depended on a combination of information from both channels. They also
observed that the adoption rate follows an S-shaped curve and that there
are five different types of adopters based on the order in which they adopt
the innovations: (1) Innovators (top 2.5%), (2) Early Adopters (13.5%), (3)
Early Majority (34%), (4) Late Majority (34%), and (5) Laggards (16%). Figure 7.5 depicts the distribution of these adopters as well as the cumulative
adoption S-shaped curve. As shown in the figure, the adoption rate is slow
when innovators or early adopters adopt the product. Once early majority
individuals start adopting, the adoption curve becomes linear, and the rate
is constant until all late majority members adopt the product. After the late
majority adopts the product, the adoption rate becomes slow once again
as laggards start adopting, and the curve slowly approaches 100%.
233

Figure 7.5: Types of Adopters and S-Shaped Cumulative Adoption Curve.

Katz: Two-Step Flow Model

Elihu Katz, a professor of communication at the University of Pennsylvania, is a well-known figure in the study of the flow of information. In
addition to a study similar to the adoption of hybrid corn seed on how
physicians adopted the new tetracycline drug [59], Katz also developed a
two-step flow model (also known as the multistep flow model) [143] that describes how information is delivered through mass communication. The
basic idea is depicted in Figure 7.6. Most information comes from mass
media and is then directed toward influential figures called opinion leaders.
These leaders then convey the information (or form opinions) and act as
hubs for other members of the society.
234

Figure 7.6: Katz Two-Step Flow Model.

Rogers: Diffusion of Innovations Process
Rogers in his well-known book, Diffusion of Innovations [239], discusses various theories regarding the diffusion of innovations process. In particular,
he describes a five stage process of adoption:
1. Awareness: In this stage, the individual becomes aware of the innovation, but her information about the product is limited.
2. Interest: The individual shows interest in the product and seeks more
information.
3. Evaluation: The individual imagines using the product and decides
whether or not to adopt it.
4. Trial: The individual performs a trial use of the product.
5. Adoption: The individual decides to continue the trial and adopts
the product for full use.
235

7.3.3

Modeling Diffusion of Innovations

To effectively make use of the theories regarding the diffusion of innovations, we demonstrate a mathematical model for it in this section. The
model incorporates basic elements discussed so far and can be used to
effectively model a diffusion of innovations process. It can be concretely
described as
dA(t)
= i(t)[P A(t)].
(7.15)
dt
Here, A(t) denotes the total population that adopted the innovation
until time t. i(t) denotes the coefficient of diffusion, which describes the
innovativeness of the product being adopted, and P denotes the total number of potential adopters (until time t). This equation shows that the rate at
which the number of adopters changes throughout time depends on how
innovative is the product being adopted. The adoption rate only affects
the potential adopters who have not yet adopted the product. Since A(t)
is the total population of adopters until time t, it is a cumulative sum and
can be computed as follows:
Z t
A(t) =
a(t)dt,
(7.16)
t0

where a(t) defines the adopters at time t. Let A0 denote the number of
adopters at time t0 . There are various methods of defining the diffusion
coefficient [185]. One way is to define i(t) as a linear combination of the
cumulative number of adopters at different times A(t),
i(t) = + 0 A0 + + t A(t) = +

t
X

i A(i),

(7.17)

i=t0

where i s are the weights for each time step. Often a simplified version of
this linear combination is used. In particular, the following three models
for computing i(t) are considered in the literature:
i(t) = ,
i(t) = A(t),
i(t) = + A(t),
External Influence
Factor

External-Influence Model
Internal-Influence Model
Mixed-Influence Model

(7.18)
(7.19)
(7.20)

where is the external-influence factor and is the imitation factor. Equation

236

7.18 describes i(t) in terms of only and is independent of the current

number of adopters A(t); therefore, in this model, the adoption only depends on the external influence. In the second model, i(t) depends on the
number of adopters at any time and is therefore dependent on the internal
factors of the diffusion process. defines how much the current adopter
population is going to affect the adoption and is therefore denoted as the
imitation factor. The mixed-influence model is a model between the two Imitation
Factor
that uses a linear combination of both previous models.
External-Influence Model
In the external-influence model, the adoption coefficient only depends on
an external factor. One such example of external influence in social media is
when important news goes viral. Often, people who post or read the news
do not know each other; therefore, the importance of the news determines
whether it goes viral. The external-influence model can be formulated as
dA(t)
= [P A(t)].
dt

(7.21)

By solving Equation 7.21,

A(t) = P(1 et ),

(7.22)

when A(t = t0 = 0) = 0. The A(t) function is shown in Figure 7.7. The

number of adopters increases exponentially and then saturates near P.
Internal-Influence Model
In the internal-influence model, adoption depends on how many have
adopted the innovation in the current time step.4 In social media there is
internal influence when a group of friends join a site due to peer pressure.
Think of a group of individuals where the likelihood of joining a social
networking site increases as more group members join the site. The internal
influence model can be described as follows:
dA(t)
= A(t)[P A(t)].
dt
237

(7.23)

Figure 7.7: External-Influence Model for P = 100 and = 0.01.

Since the diffusion rate in this model depends on A(t), it is called the pure Pure Imitation
Model
imitation model. The solution to this model is defined as
P
A(t) =
,
(7.24)
PA0 P(tt0 )
1 + A0 e
where A(t = t0 ) = A0 . The A(t) function is shown in Figure 7.8.
Mixed-Influence Model
As discussed, the mixed influence model is situated in between the internaland external-influence models. The mixed-influence model is defined as
dA(t)
= ( + A(t))[P A(t)].
(7.25)
dt
By solving the differential equation, we arrive at
A(t) =

(PA0 ) (+P)(tt0 )
e
+A0

1+

(PA0 ) (+P)(tt )
0
e
+A0

(7.26)

The internal-influence model is similar to the SI model discussed later in the section
on epidemics. For the sake of completeness, we provide solutions to both. Readers are
encouraged to refer to that model in Section 7.4 for further insight.

238

Figure 7.8: Internal-Influence Model for A0 = 30, = 105 , and P = 200.

where A(t = t0 ) = A0 . The A(t) function for the mixed-influence model is
depicted in Figure 7.9.
We discussed three models in this section: internal, external, and mixed
influence. Depending on the model used to describe the diffusion of innovations process, the respective equation for A(t) (Equations 7.22, 7.24, or
7.26) should be employed to model the system.

7.3.4

Intervention

Consider a faulty product being adopted. The product company is planning to stop or delay adoptions until the product is fixed and re-released.
This intervention can be performed by doing the following:
Limiting the distribution of the product or the audience that can adopt the
product. In our mathematical model, this is equivalent to reducing
the population P that can potentially adopt the product.
Reducing interest in the product being sold. For instance, the company
can inform adopters of the faulty status of the product. In our models,
this can be achieved by tampering : setting to a very small value
in Equation 7.22 results in a slow adoption rate.
239

Figure 7.9: Mixed-Influence Model for P = 200, = 105 , A0 = 30, and

= 103 .
Reducing interactions within the population. Reduced interactions result
in less imitation of product adoptions and a general decrease in the
trend of adoptions. In our models, this can be achieved by setting
to a small value.

7.4

Epidemics

In an epidemic, a disease spreads widely within a population. This process

consists of a pathogen (the disease being spread), a population of hosts
(humans, animals, and plants, among others), and a spreading mechanism
(breathing, drinking, sexual activity, etc.). Unlike information cascades
and herding, but similar to diffusion of innovations models, epidemic
models assume an implicit network and unknown connections among
individuals. This makes epidemic models more suitable when we are
interested in global patterns, such as trends and ratios of people getting
infected, and not in who infects whom.
In general, a complete understanding of the epidemic process requires
substantial knowledge of the biological process within each host and the im240

mune system process, as well as a comprehensive analysis of interactions

among individuals. Other factors such as social and cultural attributes also
play a role in how, when, and where epidemics happen. Large epidemics,
also known as pandemics, have spread through human populations and
include the Black Death in the 13th century (killing more than 50% of Europes population), the Great Plague of London (100,000 deaths), the smallpox epidemic, in the 17th century (killing more than 90% of Massachusetts
Bay Native Americans) and recent pandemics such as HIV/AIDS, SARS,
H5N1 (Avian flu), and influenza. These pandemics motivated the introduction of epidemic models in the early 20th century and the establishment
of the epidemiology field.
There are various ways of modeling epidemics. For instance, one can
look at how hosts contact each other and devise methods that describe
how epidemics happen in networks. These networks are called contact
networks. A contact network is a graph where nodes represent the hosts Contact Networks
and edges represent the interactions between these hosts. For instance, in
the case of the HIV/AIDS epidemic, edges represent sexual interactions,
and in the case of influenza, nodes that are connected represent hosts that
breathe the same air. Nodes that are close in a contact network are not
necessarily close in terms of real-world proximity. Real-world proximity
might be true for plants or animals, but diseases such as SARS or avian
flu travel between continents because of the traveling patterns of hosts.
This spreading pattern becomes clearer when the science of epidemics is
employed to understand the propagation of computer viruses in cell phone
networks or across the internet [229, 214].
Another way of looking at epidemic models is to avoid considering
network information and to analyze only the rates at which hosts get infected, recover, and the like. This analysis is known as the fully mixed Fully Mixed
technique, assuming that each host has an equal chance of meeting other Technique
hosts. Through these interactions, hosts have random probabilities of
getting infected. Though simplistic, the technique reveals several useful
methods of modeling epidemics that are often capable of describing various real-world outbreaks. In this section, we concentrate on the fully
mixed models that avoid the use of contact networks.5
Note that the models of information diffusion that we have already
discussed, such as the models in diffusion of innovations or information
5

A generalization of these techniques over networks can be found in [126, 125, 212].

241

cascades, are more or less related to epidemic models. However, what

makes epidemic models different is that, in the other models of information diffusion, actors decide whether to adopt the innovation or take the
decision and the system is usually fully observable. In epidemics, however, the system has a high level of uncertainty, and individuals usually
do not decide whether to get infected or not. The models discussed in this
section assume that (1) no contact network information is available and
(2) the process by which hosts get infected is unknown. These models can
be applied to situations in social media where the decision process has a
certain uncertainty to it or is ambiguous to the analyst.

7.4.1

Definitions

Since there is no network, we assume that we have a population where the

disease is being spread. Let N define the size of this crowd. Any member
of the crowd can be in either one of three states:

Closed-world
Assumption

1. Susceptible: When an individual is in the susceptible state, he or she

can potentially get infected by the disease. In reality, infections can
come from outside the population where the disease is being spread
(e.g., by genetic mutation, contact with an animal, etc.); however,
for simplicity, we make a closed-world assumption, where susceptible
individuals can only get infected by infected people in the population.
We denote the number of susceptibles at time t as S(t) and the fraction
of the population that is susceptible as s(t) = S(t)/N.
2. Infected: An infected individual has the chance of infecting susceptible parties. Let I(t) denote the number of infected individuals at
time t, and let i(t) denote the fraction of individuals who are infected,
i(t) = I(t)/N.
3. Recovered (or Removed): These are individuals who have either
recovered from the disease and hence have complete or partial immunity against the infection or were killed by the infection. Let R(t)
denote the size of this set at time t and r(t) the fraction recovered,
r(t) = R(t)/N .
Clearly, N = S(t) + I(t) + R(t) for all t. Since we are assuming that there is
some level of randomness associated with the values of S(t), I(t), and R(t),
242

Figure 7.10: SI Model.

we try to deal with expected values and assume S, I, and R represent these
at time t.

7.4.2

SI Model

We start with the most basic model. In this model, the susceptible individuals get infected, and once infected, they will never get cured. Denote
as the contact probability. In other words, the probability of a pair of
people meeting in any time step is . So, if = 1, everyone comes into
contact with everyone else, and if = 0, no one meets another individual.
Assume that when an infected individual meets a susceptible individual
the disease is being spread with probability 1 (this can be generalized to
other values). Figure 7.10 demonstrates the SI model and the transition
between states that happens in this model for individuals. The value over
the arrow shows that each susceptible individual meets at least I infected
individuals during the next time step.
Given this situation, infected individuals will meet N people on average. We know from this set that only the fraction S/N will be susceptible
and that the rest are infected already. So, each infected individual will
infect NS/N = S others. Since I individuals are infected, IS will be
infected in the next time step. This means that the number of susceptible individuals will be reduced by this factor as well. So, to get different
values of S and I at different times, we can solve the following differential
equations:
dS
= IS,
(7.27)
dt
dI
= IS.
(7.28)
dt
Since S + I = N at all times, we can eliminate one equation by replacing
S with N I:
dI
= I(N I).
(7.29)
dt
243

The solution to this differential equation is called the logistic growth

function,
I(t) =

NI0 et
,
N + I0 (et 1)

(7.30)

where I0 is the number of individuals infected at time 0. In general, analyzing epidemics in terms of the number of infected individuals has nominal
generalization power. To address this limitation, we can consider infected
I0
in the previous equation,
fractions. We therefore substitute i0 = N
i(t) =

i0 et
.
1 + i0 (et 1)

(7.31)

Note that in the limit, the SI model infects all the susceptible population
because there is no recovery in the model. Figure 7.11(a) depicts the
logistic growth function (infected individuals) and susceptible individuals
for N = 100, I0 = 1, and = 0.003. Figure 7.11(b) depicts the infected
population for HIV/AIDS for the past 20 years. As observed, the infected
population can be approximated well with the logistic growth function
and follows the SI model. Note that in the HIV/AIDS graph, not everyone
is getting infected. This is because not everyone in the United States is in
the susceptible population, so not everyone will get infected in the end.
Moreover, there are other factors that are far more complex than the details
of the SI model that determine how people get infected with HIV/AIDS.

Figure 7.11: SI model simulation compared to the HIV/AIDS growth in the

United States.
244

Figure 7.12: SIR Model.

7.4.3

SIR Model

The SIR model, first introduced by Kermack, and McKendrick [148], adds
more detail to the standard SI model. In the SIR model, in addition to the
I and S states, a recovery state R is present. Figure 7.12 depicts the model.
In the SIR model, hosts get infected, remain infected for a while, and then
recover. Once hosts recover (or are removed), they can no longer get
infected and are no longer susceptible. The process by which susceptible
individuals get infected is similar to the SI model, where a parameter
defines the probability of contacting others. Similarly, a parameter
in the SIR model defines how infected people recover, or the recovering
probability of an infected individual in a time period t.
In terms of differential equations, the SIR model is
dS
= IS,
dt
dI
= IS I,
dt
dR
= I.
dt

(7.32)
(7.33)
(7.34)

Equation 7.32 is identical to that of the SI model (Equation 7.27). Equation 7.33 is different from Equation 7.28 of the SI model by the addition
of the term I, which defines the number of infected individuals who recovered. These are removed from the infected set and are added to the
recovered ones in Equation 7.34. Dividing Equation 7.32 by Equation 7.34,
we get

dS
= S,
dR

(7.35)

and by assuming the number of recovered at time 0 is zero (R0 = 0),

log

S0
= R.
S

245

(7.36)

S0 = Se R

S = S0 e

(7.37)
(7.38)

Since I + S + R = N, we replace I in Equation 7.34,

dR
= (N S R).
dt

(7.39)

Now combining Equations 7.38 and 7.39,

dR
= (N S0 e R R).
dt

(7.40)

If we solve this equation for R, then we can determine S from 7.38 and
I from I = N R S. The solution for R can be computed by solving the
following integration:
Z
1 R
dx
.
(7.41)
t=
0 N S e x x
0
However, there is no closed-form solution to this integration, and only
numerical approximation is possible. Figure 7.13 depicts the behavior of
the SIR model for a set of initial parameters.
The two models in the next two subsections are generalized versions
of the two models discussed thus far: SI and SIR. These models allow
individuals to have temporary immunity and to get reinfected.

7.4.4

SIS Model

The SIS model is the same as the SI model, with the addition of infected
nodes recovering and becoming susceptible again (see Figure 7.14). The
differential equations describing the model are
dS
= I IS,
dt
dI
= IS I.
dt

(7.42)
(7.43)

By replacing S with N I in Equation 7.43, we arrive at

dI
= I(N I) I = I(N ) I2 .
dt
246

(7.44)

Figure 7.13: SIR Model Simulated with S0 = 99, I0 = 1, R0 = 0, = 0.01,

and = 0.1.
When N , the first term will be negative or zero at most; hence,
the whole term becomes negative. Therefore, in the limit, the value I(t)
will decrease exponentially to zero. However, when N > , we will have
a logistic growth function as in the SI model. Having said this, as the
simulation of the SIS model shows in Figure 7.15, the model will never
infect everyone. It will reach a steady state, where both susceptibles and
infecteds reach an equilibrium (see the epidemics exercises).

Figure 7.14: SIS Model.

247

Figure 7.15: SIS Model Simulated with S0 = 99, I0 = 1, = 0.01, and = 0.1.

7.4.5

SIRS Model

The final model analyzed in this section is the SIRS model. Just as the
SIS model extends the SI, the SIRS model extends the SIR, as shown in
Figure 7.16. In this model, the assumption is that individuals who have
recovered will lose immunity after a certain period of time and will become
susceptible again. A new parameter has been added to the model that
defines the probability of losing immunity for a recovered individual. The

Figure 7.16: SIRS Model.

248

set of differential equations that describe this model is

dS
= R IS,
dt
dI
= IS I,
dt
dR
= I R.
dt

(7.45)
(7.46)
(7.47)

Like the SIR model, this model has no closed-form solution, so numerical integration can be used. Figure 7.17 demonstrates a simulation of the
SIRS model with given parameters of choice. As observed, the simulation
outcome is similar to the SIR model simulation (see Figure 7.13). The major
difference is that in the SIRS, the number of susceptible and recovered individuals changes non-monotonically over time. For example, in SIRS, the
number of susceptible individuals decreases over time, but after reaching
the minimum count, starts increasing again. On the contrary, in the SIR,
both susceptible individuals and recovered individuals change monotonically, with the number of susceptible individuals decreasing over time and
that of recovered individuals increasing over time. In both SIR and SIRS,
the infected population changes non-monotonically.

7.4.6

Intervention

A pressing question in any pandemic or epidemic outbreak is how to stop

the process. In this section, we discuss epidemic intervention based on
a recent discovery [55]. In any epidemic outbreak, infected individuals
infect susceptible individuals. Although in this chapter we discussed random infection in the real world, what actually takes place is quite different.
Infected individuals have a limited number of contacts and can only infect
them if said contacts are susceptible. A well-connected infected individual
is more dangerous to the epidemic outbreak than someone who has no
contacts. In other words, the epidemic takes place in a network. Unfortunately, it is often difficult to trace these contacts and outline the contact
network. If this was possible, the best way to intervene with the epidemic
outbreak would be to vaccinate the highly connected nodes and stop the
epidemic. This would result in what is known as herd immunity and would
stop the epidemic outbreak. Herd immunity entails vaccinating a population inside a herd such that the pathogen cannot initiate an outbreak
249

Figure 7.17: SIRS Model Simulated with S0 = 99, I0 = 1, R0 = 0, = 0.1,

= 0.01, and = 0.02.
inside the herd. In general, creating herd immunity requires at least a
random sample of 96% of the population to be vaccinated. Interestingly,
we can achieve the same herd immunity by making use of friends in a network. In general, people know which of their friends have more friends.
So, they know or have access to these higher-degree and more-connected
nodes. Researchers found that if a random population of 30% of the herd
is selected and then these 30% are asked for their highest degree friends,
one can achieve herd immunity by vaccinating these friends. Of course,
older intervention techniques such as separating those infected from those
susceptible (quarantining them) or removing those infected (killing cows
with mad cow disease) still work.

250

7.5

Summary

In this chapter, we discussed the concept of information diffusion in social

networks. In the herd behavior, individuals observe the behaviors of
others and act similarly to them based on their own benefit. We reviewed
the well-known diners example and urn experiment and demonstrated
how conditional probabilities can be used to determine why herding takes
place. We discussed how herding experiments should be designed and
ways to intervene with it.
Next, we discussed the information cascade problem with the constraint
of sequential decision making. The independent cascade model (ICM) is
a sender-centric model and has a level of stochasticity associated with it.
The spread of cascades can be maximized in a network given a budget
on how many initial nodes can be activated. Unfortunately, the problem
is NP-hard; therefore, we introduced a greedy approximation algorithm
that has guaranteed performance due to the submodularity of ICMs activation function. Finally, we discussed how to intervene with information
cascades.
Our next topic was the diffusion of innovations. We discussed the characteristics of adoption both from the individual and innovation point of
view. We reviewed well-known theories such as the models introduced by
Ryan and Gross, Katz, and Rogers, in addition to experiments in the field,
and different types of adopters. We also detailed mathematical models that
account for internal, external, and mixed influences and their intervention
procedures.
Finally, we moved on to epidemics, an area where decision making is
usually performed unconsciously. We discussed four epidemic models:
SI, SIR, SIS, and SIRS; the two last models allow for reinfected individuals.
For each model we provided differential equations, numerical solutions,
and closed-form solutions, when available. We concluded the chapter
with intervention approaches to epidemic outbreaks and a review of herd
immunity in epidemics. Although a 96% random vaccination is required
for achieving herd immunity, it is also possible to achieve it by selecting
a random population of 30% and then vaccinating their highest degree
friends.

251

7.6

Bibliographic Notes

The concept of the herd has been well studied in psychology by Freud
(crowd psychology), Carl Gustav Jung (the collective unconscious), and
Gustave Le Bon (the popular mind). It has also been observed in economics
by Veblen [288] and in studies related to the bandwagon effect [240, 259, 165].
The behavior is also discussed in terms of sociability [258] in sociology.
Herding, first coined by Banerjee [23], at times refers to a slightly different concept. In herd behaviour discussed in this chapter, the crowd does
not necessarily start with the same decision, but will eventually reach one,
whereas in herding the same behavior is usually observed. Moreover, in
herd behavior, individuals decide whether the action they are taking has
some benefits to themselves or is rational, and based on that, they will
align with the population. In herding, some level of uncertainty is associated with the decision, and the individual does not know why he or she is
following the crowd.
Another confusion is that the terms herd behavior/herding is often
used interchangebly with information cascades [37, 299]. To avoid this
problem, we clearly define both in the chapter and assume that in herd
behavior, decisions are taken based on global information, whereas in
information cascades, local information is utilized.
Herd behavior has been studied in the context of financial markets
[60, 74, 38, 69] and investment [250]. Gale analyzes the robustness of
different herd models in terms of different constraints and externalities [93],
and Shiller discusses the relation between information, conversation, and
herd behavior [256]. Another well-known social conformity experiment
was conducted in Manhattan by Milgram et al. [195].
Other recent applications of threshold models can be found in [307,
295, 296, 285, 286, 252, 232, 202, 184, 183, 108, 34]. Bikhchandani et al.
[1998] review conformity, fads, and information cascades and describe
how observing past human decisions can help explain human behavior.
Hirshleifer [128] provides information cascade examples in many fields,
including zoology and finance.
In terms of diffusion models, Robertson [238] describes the process and
Hagerstrand et al. [118] introduce a model based on the spatial stages
of the diffusion of innovations and Monte Carlo simulation models for
diffusion of innovations. Bass [30] discusses a model based on differential
equations. Mahajan and Peterson [187] extend the Bass model.
252

Instances of external-influence models can be found in [119, 59] and

internal-influence models are applied in [188, 111, 110]. The Gompertz
function [189], widely used in forecasting, has a direct relationship with
the internal-influence diffusion curve. Mixed-influence model examples
include the work of Mahajan and Muller [186] and Bass model [30].
Midgley and Dowling [193] introduce the contingency model. Abrahamson and Rosenkopf [3] mathematically analyze the bandwagon effect and
diffusion of innovations. Their model predicts whether the bandwagon
effect will occur and how many organizations will adopt the innovation.
Network models of diffusion and thresholds for diffusion of innovations
models are discussed by Valente [286, 287]. Diffusion through blogspace
and in general, social networks, has been analyzed by [112, 169, 306, 310].
For information on different pandemics, refer to [220, 31, 230, 68, 77, 53,
113, 206]. To review some early and in-depth analysis of epidemic models,
refer to [21, 13]. Surveys of epidemics can be found in [124, 125, 126, 72].
Epidemics in networks have been discussed [212, 201, 144] extensively.
Other general sources include [171, 78, 212]; [28]. A generalized model
for contagion is provided by Dodds and Watts [73] and, in the case of best
response dynamics, in [202].
Other topics related to this chapter include wisdom of crowd models
[104] and swarm intelligence [79, 82, 42, 147]. One can also analyze information provenance, which aims to identify the sources from which information
has diffused. Barbier at al. [25] provide an overview of information provenance in social media in their book.

253

7.7

Exercises

1. Discuss how different information diffusion modeling techniques differ. Name applications on social media that can make use of methods
in each area.

Herd Effect
2. What are the minimum requirements for a herd behavior experiment?
Design an experiment of your own.

Diffusion of Innovation
3. Simulate internal-, external-, and mixed-influence models in a program. How are the saturation levels different for each model?
4. Provide a simple example of diffusion of innovations and suggest a
specific way of intervention to expedite the diffusion.

Information Cascades
5. Briefly describe the independent cascade model (ICM).
6. What is the objective of cascade maximization? What are the usual
constraints?
7. Follow the ICM procedure until it converges for the following graph.
Assume that node i activates node j when i j 1 (mod 3) and node
5 is activated at time 0.

254

Epidemics
8. Discuss the mathematical relationship between the SIR and the SIS
models.
9. Based on our assumptions in the SIR model, the probability that an
individual remains infected follows a standard exponential distribution. Describe why this happens.
10. In the SIR model, what is the most likely time to recover based on the
value of ?
11. In the SIRS model, compute the length of time that an infected individual is likely to remain infected before he or she recovers.
12. After the model saturates, how many are infected in the SIS model?

255

256

Part III
Applications

257

This chapter is from Social Media Mining: An Introduction.

By Reza Zafarani, Mohammad Ali Abbasi, and Huan Liu.
Cambridge University Press, 2014. Draft version: April 20, 2014.
Complete Draft and Slides Available at: http://dmml.asu.edu/smm

Chapter

Influence and Homophily

Social forces connect individuals in different ways. When individuals get
connected, one can observe distinguishable patterns in their connectivity
networks. One such pattern is assortativity, also known as social similarity.
In networks with assortativity, similar nodes are connected to one another
more often than dissimilar nodes. For instance, in social networks, a high
similarity between friends is observed. This similarity is exhibited by
similar behavior, similar interests, similar activities, and shared attributes
such as language, among others. In other words, friendship networks are
assortative. Investigating assortativity patterns that individuals exhibit on
social media helps one better understand user interactions. Assortativity
is the most commonly observed pattern among linked individuals. This
chapter discusses assortativity along with principal factors that result in
assortative networks.
Many social forces induce assortative networks. Three common forces
are influence, homophily, and confounding. Influence is the process by which
an individual (the influential) affects another individual such that the influenced individual becomes more similar to the influential figure. Homophily is observed in already similar individuals. It is realized when
similar individuals become friends due to their high similarity. Confounding is the environments effect on making individuals similar. For instance,
individuals who live in Russia speak Russian fluently because of the environment and are therefore similar in language. The confounding force is
an external factor that is independent of inter-individual interactions and
is therefore not discussed further.
259

Assortativity

Influence,
Homophily, and
Confounding

Figure 8.1: Influence and Homophily.

Note that both influence and homophily social forces give rise to assortative networks. After either of them affects a network, the network
exhibits more similar nodes; however, when friends become similar, we
denote that as influence, and when similar individuals become friends,
we call it homophily. Figure 8.1 depicts how both influence and homophily
affect social networks.
In particular, when discussing influence and homophily in social media,
we are interested in asking the following questions:
How can we measure influence or homophily?
How can we model influence or homophily?
How can we distinguish between the two?
Because both processes result in assortative networks, we can quantify
their effect on the network by measuring the assortativity of the network.
260

8.1

Measuring Assortativity

Measuring assortativity helps quantify how much influence and homophily,

among other factors, have affected a social network. Assortativity can be
quantified by measuring how similar connected nodes are to one another.
Figure 8.2 depicts the friendship network in a U.S. high school in 1994.1
In the figure, races are represented with different colors: whites are white,
blacks are gray, Hispanics are light gray, and others are black. As we observe, there is a high assortativity between individuals of the same race,
particularly among whites and among blacks. Hispanics have a high tendency to become friends with whites.

Figure 8.2: A U.S. High School Friendship Network in 1994 between Races.
Eighty percent of the links exist between members of the same race (from
[62]).
To measure assortativity, we measure the number of edges that fall in
between the nodes of the same race. This technique works for nominal
attributes, such as race, but does not work for ordinal ones such as age.
Consider a network where individuals are friends with people of different
ages. Unlike races, individuals are more likely to be friends with others
close in age, but not necessarily with ones of the exact same age. Hence,
1

From ADD health data: http://www.cpc.unc.edu/projects/addhealth.

261

we discuss two techniques: one for nominal attributes and one for ordinal
attributes.

8.1.1

Measuring Assortativity for Nominal Attributes

Consider a scenario where we have nominal attributes assigned to nodes.

As in our example, this attribute could be race or nationality, gender, or
the like. One simple technique to measure assortativity is to consider the
number of edges that are between nodes of the same type. Let t(vi ) denote
the type of node vi . In an undirected graph2, G(V, E), with adjacency matrix
A, this measure can be computed as follows,
1 X
1 X
( t(vi ), t(v j ) ) =
Ai j ( t(vi ), t(v j ) ),
m (v ,v )E
2m i j
i

(8.1)

where m is the number of edges in the graph, m1 is applied for normalization,

and the factor 12 is added because G is undirected. (., .) is the Kronecker
delta function:
(
0, if x , y;
(x, y) =
(8.2)
1, if x = y.

Assortativity
Significance

This measure has its limitations. Consider a school of Hispanic students. Obviously, all connections will be between Hispanics, and assortativity value 1 is not a significant finding. However, consider a school where
half the population is white and half the population is Hispanic. It is statistically expected that 50% of the connections will be between members of
different race. If connections in this school were only between whites and
Hispanics and not within groups, then our observation is significant. To
account for this limitation, we can employ a common technique where we
measure the assortativity significance by subtracting the measured assortativity by the statistically expected assortativity. The higher this value, the
more significant the assortativity observed.
Consider a graph G(V, E), |E| = m, where the degrees are known beforehand (how many friends an individual has), but the edges are not.
Consider two nodes vi and v j , with degrees di and d j , respectively. What is
the expected number of edges between these two nodes? Consider node
2

The directed case is left to the reader.

262

vi . For any edge going out of vi randomly, the probability of this edge
d
d
getting connected to node v j is P jdi = 2mj . Since the degree for vi is di , we
i
have di such edges; hence, the expected number of edges between vi and
di d j
v j is 2m
. Now, the expected number of edges between vi and v j that are of
dd

i j
the same type is 2m
( t(vi ), t(v j ) ) and the expected number of edges of the
same type in the whole graph is

1 X di d j
1 X di d j
( t(vi ), t(v j ) ) =
( t(vi ), t(v j ) ).
m (v ,v )E 2m
2m i j 2m
i

(8.3)

We are interested in computing the distance between the assortativity

observed and the expected assortativity:
Q =
=

1 X
1 X di d j
Ai j ( t(vi ), t(v j ) )
( t(vi ), t(v j ) )
2m i j
2m i j 2m

(8.4)

di d j
1 X
) ( t(vi ), t(v j ) ).
( Ai j
2m i j
2m

(8.5)

This measure is called modularity [211]. The maximum modularity Modularity

value for a network depends on the number of nodes of the same type
and degree. The maximum occurs when all edges are connecting nodes
of the same type (i.e., when Ai j = 1, ( t(vi ), t(v j ) ) = 1). We can normalize
modularity by dividing it by the maximum it can take:
Qnormalized =

Q
Qmax

(8.6)
dd

i j
Ai j 2m
) ( t(vi ), t(v j ) )
=
P
P di d j
1
1
max[ 2m
i j Ai j ( t(vi ), t(v j ) ) 2m
i j 2m ( t(vi ), t(v j ) )]

1
2m

i j(

(8.7)
=

di d j
i j ( Ai j 2m )
P di d j
1
1
2m 2m
i j 2m
2m
1
2m

( t(vi ), t(v j ) )
( t(vi ), t(v j ) )

i j(

(8.8)

dd

i j
Ai j 2m
) ( t(vi ), t(v j ) )
.
P di d j
2m i j 2m
( t(vi ), t(v j ) )

P
=

263

(8.9)

Figure 8.3: A Modularity Example for a Bipartite Graph.

Modularity can be simplified using a matrix format. Let Rnk
denote the indicator matrix and let k denote the number of types,
(
1, if t(x) = k;
x,k =
(8.10)
0, if t(x) , k
Note that function can be reformulated using the indicator matrix:
X
( t(vi ), t(v j ) ) =
vi ,k v j ,k .
(8.11)
k

Therefore, (T )i,j = (t(vi ), t(v j )). Let B = A ddT /2m denote the
modularity matrix where d Rn1 is the degree vector for all nodes. Given
that the trace of multiplication of two matrices X and YT is Tr(XYT ) =
P
i,j Xi,j Yi,j and Tr(XY) = Tr(YX), modularity can be reformulated as
Q =

di d j
1
1 X
( Ai j
) ( t(vi ), t(v j ) ) =
Tr(BT )
2m i j
2m | {z } 2m
| {z }
(T )i,j

Bij

1
Tr(T B).
(8.12)
2m
Example 8.1. Consider the bipartite graph in Figure 8.3. For this bipartite graph,
=

0
0

A =
1

0
0
1
1

1
1
0
0

1
1
0
0

1
1

=
0

0
264

0
0
1
1

2
2

d = , m = 4.
2

2

(8.13)

Figure 8.4: A Correlation Example.

Therefore, matrix B is

0.5
0.5

B = A ddT /2m =
0.5

0.5

0.5
0.5
0.5
0.5
0.5
0.5
0.5 0.5 0.5
0.5 0.5 0.5

(8.14)

The modularity value Q is

1
Tr(T B) = 0.5.
2m

(8.15)

In this example, all edges are between nodes of different color. In other words,
the number of edges between nodes of the same color is less than the expected
number of edges between them. Therefore, the modularity value is negative.

8.1.2

Measuring Assortativity for Ordinal Attributes

A common measure for analyzing the relationship between two variables

with ordinal values is covariance. Covariance describes how two variables Covariance
change with respect to each other. In our case, we are interested in how
correlated, the attribute values of nodes connected via edges are. Let xi be
the ordinal attribute value associated with node vi . In Figure 8.4, for node
c, the value associated is xc = 21.
We construct two variables XL and XR , where for any edge (vi , v j ) we
assume that xi is observed from variable XL and x j is observed from variable
265

XR . For Figure 8.4,

18
21

XL =
,
21

20

21
18

XR =
.
20

21

(8.16)

In other words, XL represents the ordinal values associated with the left
node of the edges, and XR represents the values associated with the right
node of the edges. Our problem is therefore reduced to computing the covariance between variables XL and XR . Note that since we are considering
an undirected graph, both edges (vi , v j ) and (v j , vi ) exist; therefore, xi and
x j are observed in both XL and XR . Thus, XL and XR include the same set of
values but in a different order. This implies that XL and XR have the same
mean and standard deviation.
E(XL ) = E(XR ),
(XL ) = (XR ).

(8.17)
(8.18)

Since we have m edges and each edge appears twice for the undirected
graph, then XL and XR have 2m elements. Each value xi appears di times
since it appears as endpoints of di edges. The covariance between XL and
XR is
(XL , XR ) =
=
=
=

E[(XL E[XL ])(XR E[XR ])]

E[XL XR XL E[XR ] E[XL ]XR + E[XL ]E[XR ]]
E[XL XR ] E[XL ]E[XR ] E[XL ]E[XR ] + E[XL ]E[XR ]
E[XL XR ] E[XL ]E[XR ].
(8.19)

E(XL ) is the mean (expected value) of variable XL , and E(XL XR ) is the

mean of the multiplication of XL and XR . In our setting and following
Equation 8.17, these expectations are as follows:
P
P
di xi
i (XL )i
E(XL ) = E(XR ) =
= i
(8.20)
2m P
2m
1 X
i j Ai j xi x j
E(XL XR ) =
(XL )i (XR )i =
.
(8.21)
2m i
2m
By plugging Equations 8.20 and 8.21 into Equation 8.19, the covariance
between XL and XR is
(XL , XR ) = E[XL XR ] E[XL ]E[XR ]
266

P
=

ij

Ai j xi x j

ij

di d j xi x j

2m
(2m)2
X
di d j
1
=
( Ai j
)xi x j .
2m i j
2m

(8.22)

Similar to modularity (Section 8.1.1), we can normalize covariance.

Pearson correlation (XL , XR ) is the normalized version of covariance:
Pearson
(XL , XR )
(XL , XR ) =
.
(XL )(XR )

Correlation

(8.23)

From Equation 8.18, (XL ) = (XR ); thus,

(XL , XR ) =
=
=

(XL , XR )
,
(XL )2
P
1
i j ( Ai j
2m

di d j
2m

)xi x j

E[(XL )2 ] (E[XL ])2

P
di d j
1
i j ( Ai j 2m )xi x j
2m
.
P
P di d j
1
1
2
i j Ai j xi 2m
i j 2m xi x j
2m

(8.24)

Note the similarity between Equations 8.9 and 8.24. Although modularity is used for nominal attributes and correlation for ordinal attributes,
the major difference between the two equations is that the function in
modularity is replaced by xi x j in the correlation equation.
Example 8.2. Consider Figure 8.4 with values demonstrating the attributes associated with each node. Since this graph is undirected, we have the following
edges:
E = {(a, c), (c, a), (c, b), (b, c)}.
(8.25)
The correlation is between the values associated with the endpoints of the edges.
Consider XL as the value of the left end of an edge and XR as the value of the right
end of an edge:

18
21
21
18

XL =
(8.26)
, XR =

20
21

20
21
The correlation between these two variables is (XL , XR ) = 0.67.
267

8.2

Influence

Influence3 is the act or power of producing an effect without apparent

exertion of force or direct exercise of command. In this section, we discuss
influence and, in particular, how we can (1) measure influence in social
media and (2) design models that detail how individuals influence one
another in social media.

8.2.1
Prediction-Based
Influence Measures

Measuring Influence

Influence can be measured based on (1) prediction or (2) observation.

Prediction-Based Measures. In prediction-based measurement, we assume that an individuals attribute or the way she is situated in the network predicts how influential she will be. For instance, we can assume that
the gregariousness (e.g., number of friends) of an individual is correlated
with how influential she will be. Therefore, it is natural to use any of the
centrality measures discussed in Chapter 3 for prediction-based influence
measurements. Examples of such centrality measures include PageRank
and degree centrality. In fact, many of these centrality measures were
introduced as influence-measuring techniques. For instance, on Twitter,
in-degree (number of followers) is a common attribute for measuring influence. Since these methods were covered in-depth in that chapter, in this
section we focus on observational techniques.

Observation-based
Influence Measures

Observation-Based Measures. In observation-based measures, we quantify the influence of an individual by measuring the amount of influence
attributed to him. An individual can influence differently in diverse settings, and so, depending on the context, the observation-based measuring
of influence changes. We next describe three different settings and how
influence can be measured in each.
1. When an individual is the role model. This happens in the case of
individuals in the fashion industry, teachers, and celebrities. In this
case, the size of the audience that has been influenced due to that
fashion, charisma, or the like could act as an accurate measure. A
3

As defined by the Merriam-Webster dictionary.

268

local grade-school teacher has a tremendous influence over a class of

students, whereas Gandhi influenced millions.
2. When an individual spreads information. This scenario is more
likely when a piece of information, an epidemic, or a product is being spread in a network. In this case, the size of the cascade that is,
the number of hops the information traveled or the population affected, or the rate at which population gets influenced is considered
a measure.
3. When an individuals participation increases the value of an item
or action. As in the case of diffusion of innovations (see Chapter 7),
often when individuals perform actions such as buying a product,
they increase the value of the product for other individuals. For
example, the first individual who bought a fax machine had no one
to send faxes to. The second individual who bought a fax machine
increased its value for the first individual. So, the increase (or rate of
increase) in the value of an item or action (such as buying a product)
is often used as a measure.
Case Studies for Measuring Influence in Social Media
This section provides examples of measuring influence in the blogosphere
and on the microblogging site Twitter. These techniques can be adapted to
other social media sites, as well.
Measuring Social Influence in the Blogosphere
The goal of measuring influence in the blogosphere is to identify influential
bloggers. Due to the limited time that individuals have, following the
influentials is often necessary for fast access to interesting news. One
common measure for quantifying the influence of bloggers is to use indegree centrality: the number of (in-)links that point to the blog. However,
because of the sparsity of in-links, more detailed analysis is required to
measure influence in the blogosphere.
In their book, The Influentials: One American in Ten Tells the Other Nine
How to Vote, Where to Eat, and What to Buy. Keller and Berry [145] argue
that the influentials are individuals who (1) are recognized by others, (2)
269

whose activities result in follow-up activities, (3) have novel perspectives,

and (4) are eloquent.
To address these issues, Agarwal et al. [6] proposed the iFinder system
to measure influence of blogposts and to identify influential bloggers. In
particular, for each one of these four characteristics and a blogpost p, they
approximate the characteristic by collecting specific blogposts attributes:
1. Recognition. Recognition for a blogpost can be approximated by the
links that point to the blogpost (in-links). Let Ip denote the set of
in-links that point to blogpost p.
2. Activity Generation. Activity generated by a blogpost can be estimated using the number of comments that p receives. Let cp denote
the number of comments that blogpost p receives.
3. Novelty. The blogposts novelty is inversely correlated with the
number of references a blogpost employs. In particular the more
citations a blogpost has, the less novel it is. Let Op denote the set of
out-links for blogpost p.
4. Eloquence. Eloquence can be estimated by the length of the blogpost.
Given the informal nature of blogs and the bloggers tendency to
write short blogposts, longer blogposts are commonly believed to be
more eloquent. So, the length of a blogpost lp can be employed as a
measure of eloquence.

Influence Flow

Given these approximations for each one of these characteristics, we

can design a measure of influence for each blogpost. Since the number
of out-links inversely affects the influence of a blogpost and the number
of in-links increases it, we construct an influence graph, or i-graph, where
blogposts are nodes and influence flows through the nodes. The amount
of this influence flow for each post p can be characterized as
InfluenceFlow(p) = win

|Ip |
X
m=1

I(Pm ) wout

|Op |
X

I(Pn ),

(8.27)

n=1

where I(.) denotes the influence of a blogpost and win and wout are the
weights that adjust the contribution of in- and out-links, respectively. In
this equation, Pm s are blogposts that point to post p, and Pn s are blogposts
270

that are referred to in post p. Influence flow describes a measure that only
accounts for in-links (recognition) and out-links (novelty). To account for
the other two factors, we design the influence of a blogpost p as
I(p) = wlength lp (wcomment cp + InfluenceFlow(p)).

(8.28)

Here, wlength is the weight for the length of blogpost4 . wcomment describes
how the number of comments is weighted. Note that the four weights win ,
wout , wcomments , and wlength need to be tuned to make the model more accurate. This tuning can be done by a variety of techniques. For instance, we
can use a test system where the influential posts are already known (labeled
data) to tune them.5 Finally, a bloggers influence index (iIndex) can be defined as the maximum influence value among all his or her N blogposts,
iIndex = max I(pn ).
pn N

(8.29)

Computing iIndex for a set of bloggers over all their blogposts can help
identify and rank influential bloggers in a system.
Measuring Social Influence on Twitter. On Twitter, a microblogging platform, users receive tweets from other users by following them. Intuitively,
we can think of the number of followers as a measure of influence (indegree centrality). In particular, three measures are frequently used to
quantify influence in Twitter,
1. In-degree: the number of users following a person on Twitter. As
discussed, the number of individuals who are interested in someones
tweets (i.e., followers) is commonly used as an influence measure on
Twitter. In-degree denotes the audience size of an individual.
2. Number of mentions: the number of times an individual is mentioned in tweets. Mentioning an individual with a username handle
is performed by including @username in a tweet. The number of
times an individual is mentioned can be used as an influence measure. The number of mentions denotes the ability in engaging others
in conversation [50].
4

In the original paper, the authors utilize a weight function instead. Here, for clarity,
we use coefficients for all parameters.
5
Note that Equation 8.28 is defined recursively, because I(p) depends on InfluenceFlow
and that, in turn, depends on I(p) (Equation 8.27). Therefore, to estimate I(p), we can
use iterative methods where we start with an initial value for I(p) and compute until
convergence.

271

Table 8.1: Rank Correlation between Top 10% of Influentials for Different
Measures on Twitter
Measures
Correlation Value
In-degree vs. retweets
0.122
In-degree vs. mentions
0.286
Retweets vs. mentions
0.638
3. Number of retweets: the number of times tweets of a user are
retweeted. Individuals on Twitter have the opportunity to forward
tweets to a broader audience via the retweet capability. Clearly, the
more ones tweets are retweeted, the more likely one is influential.
The number of retweets indicates an individuals ability to generate
content that is worth being passed along.

Spearmans Rank
Correlation

Each one of these measures by itself can be used to identify influential

users in Twitter. This can be done by utilizing the measure for each individual and then ranking individuals based on their measured influence
value. Contrary to public perception, the number of followers is considered an inaccurate measure compared to the other two. This is shown
in [50], where the authors ranked individuals on Twitter independently
based on these three measures. To see if they are correlated or redundant,
they compared ranks of individuals across three measures using rank correlation measures. One such measure is the Spearmans rank correlation
coefficient,
P
6 ni=1 (mi1 mi2 )2
,
(8.30)
=1
n3 n
where mi1 and mi2 are ranks of individual i based on measures m1 and m2 ,
and n is the total number of users. Spearmans rank correlation is the
Pearson correlation coefficient for ordinal variables that represent ranks
(i.e., takes values between 1. . . n); hence, the value is in range [1,1]. Their
findings suggest that popular users (users with high in-degree) do not
necessarily have high ranks in terms of number of retweets or mentions.
This can be observed in Table 8.1, which shows the Spearmans correlation
between the top 10% influentials for each measure.
272

8.2.2

Modeling Influence

In influence modeling, the goal is to design models that can explain how
individuals influence one another. Given the nature of social media, it is
safe to assume that influence takes place among connected individuals.
At times, this network is observable (explicit networks), and at others
times, it is unobservable (implicit networks). For instance, in referral
networks, where people refer others to join an online service on social
media, the network of referrals is often observable. In contrast, people are
influenced to buy products, and in most cases, the seller has no information
on who referred the buyer, but does have approximate estimates on the
number of products sold over time. In the observable (explicit) network,
we resort to threshold models such as the linear threshold model (LTM) Linear
to model influence; in implicit networks, we can employ methods such as Threshold
the linear influence model (LIM) that take the number of individuals who Model (LTM)
get influenced at different times as input (e.g., the number of buyers per
week).
Modeling Influence in Explicit Networks
Threshold models are simple yet effective methods for modeling influence
in explicit networks. In these models, nodes make decision based on the
number or the fraction (the threshold) of their neighbors (or incoming
neighbors in a directed graph) who have already decided to make the
same decision. Threshold models were employed in the literature as early
as the 1970s in the works of Granovetter [109] and Schelling [251]. Using
a threshold model, Schelling demonstrated that minor local preferences in
having neighbors of the same color leads to global racial segregation.
A linear threshold model (LTM) is an example of a threshold model.
Assume a weighted directed graph where nodes v j and vi are connected
with weight w j,i 0. This weight denotes how much node v j can affect
node vi s decision. We also assume
X
w j,i 1,
(8.31)
v j Nin (vi )

where Nin (vi ) denotes the incoming neighbors of node vi . In a linear threshold model, each node vi is assigned a threshold i such that when the
amount of influence exerted toward vi by its active incoming neighbors is
273

Algorithm 8.1 Linear Threshold Model (LTM)

Require: Graph G(V, E), set of initial activated nodes A0
1: return Final set of activated nodes A
2: i=0;
3: Uniformly assign random thresholds v from the interval [0, 1];
4: while i = 0 or (Ai1 , Ai , i 1) do
5:
Ai+1 = Ai
6:
inactive = V Ai ;
7:
for all
P v inactive do
8:
if j connected to v,jAi w j,v v . then
9:
activate v;
10:
Ai+1 = Ai+1 {v};
11:
end if
12:
end for
13:
i = i + 1;
14: end while
15: A = Ai ;
16: Return A ;
more than i , then vi becomes active, if still inactive. Thus, for vi to become
active at time t, we should have
X
w j,i i ,
(8.32)
v j Nin (vi ),v j At1

where At1 denotes the set of active nodes at the end of time t 1. The
threshold values are generally assigned uniformly at random to nodes
from the interval [0,1]. Note that the threshold i defines how resistant to
change node vi is: a very small i value might indicate that a small change
in the activity of vi s neighborhood results in vi becoming active and a large
i shows that vi resists changes.
Provided a set of initial active nodes A0 and a graph, the LTM algorithm
is shown in Algorithm 8.1. In each step, for all inactive nodes, the condition
in Equation 8.32 is checked, and if it is satisfied, the node becomes active.
The process ends when no more nodes can be activated. Once thresholds
are fixed, the process is deterministic and will always converge to the same
state.
274

Figure 8.5: Linear Threshold Model (LTM) Simulation. The values attached to nodes denote thresholds i , and the values on the edges represent
weights wi, j .
Example 8.3. Consider the graph in Figure 8.5. Values attached to nodes represent the LTM thresholds, and edge values represent the weights. At time 0, node
v1 is activated. At time 2, both nodes v2 and v3 receive influence from node v1 .
Node v2 is not activated since 0.5 < 0.8 and node v3 is activated since 0.8 > 0.7.
Similarly, the process continues and then stops with five activated nodes.

Modeling Influence in Implicit Networks

An implicit network is one where the influence spreads over edges in the
network; however, unlike the explicit model, we cannot observe the individuals (the influentials) who are responsible for influencing others, but
only those who get influenced. In other words, the information available
275

Figure 8.6: The Size of the Influenced Population as a Summation of Individuals Influenced by Activated Individuals (from [306]).
is the set of influenced population P(t) at any time and the time tu , when
each individual u gets initially influenced (activated). We assume that any
influenced individual u can influence I(u, t) number of non-influenced (inactive) individuals after t time steps. We call I(., .) the influence function.
Assuming discrete time steps, we can formulate the size of the influenced
population |P(t)|:
X
|P(t)| =
I(u, t tu ).
(8.33)
uP(t)

Figure 8.6 shows how the model performs. Individuals u, v, and w are
activated at time steps tu , tv , and tw , respectively. At time t, the total number
of influenced individuals is a summation of influence functions Iu , Iv , and
Iw at time steps t tu , t tv , and t tw , respectively. Our goal is to estimate
I(., .) given activation times and the number of influenced individuals at all
times. A simple approach is to utilize a probability distribution to estimate
I function. For instance, we can employ the power-law distribution to
276

estimate influence. In this case, I(u, t) = cu (t tu )u , where we estimate

coefficients cu and u for any u by methods such as maximum likelihood
estimation (see [205] for more details).
This is called the parametric estimation, and the method assumes that
all users influence others in the same parametric form. A more flexible
approach is to assume a nonparametric function and estimate the influence
functions form. This approach was first introduced as the linear influence
model (LIM) [306].
Linear
In LIM, we extend our formulation by assuming that nodes get de- Influence
Model (LIM)
activated over time and then no longer influence others. Let A(u, t) = 1
denote that node u is active at time t, and A(u, t) = 0 denote that node u
is either deactivated or still not influenced. Following a network notation
and assuming that |V| is the total size of the population and T is the last
time step, we can reformulate Equation 8.33 for |P(t)| as
|P(t)| =

|V| X
T
X

A(u, t)I(u, t),

(8.34)

u=1 t=1

or equivalently in matrix form,

P = AI.

(8.35)

It is common to assume that individuals can only activate other individuals and cannot stop others from becoming activated. Hence, negative
values for influence do not make sense; therefore, we would like measured
influence values to be positive I 0,
minimize ||P AI||22
subject to
I 0.

(8.36)
(8.37)

This formulation is similar to regression coefficients computation outlined in Chapter 5, where we compute a least square estimate of I; however,
this formulation cannot be solved using regression techniques studied earlier because, in regression, computed I values can become negative. In
practice, this formulation can be solved using non-negative least square
methods (see [164] for details).
277

8.3

Homophily

Homophily is the tendency of similar individuals to become friends. It

happens on a daily basis in social media and is clearly observable in social
networking sites where befriending can explicitly take place. The wellknown saying, birds of a feather flock together, is frequently quoted
when discussing homophily. Unlike influence, where an influential influences others, in homophily, two similar individuals decide to get connected.

8.3.1

Measuring Homophily

Homophily is the linking of two individuals due to their similarity and

leads to assortative networks over time. To measure homophily, we measure how the assortativity of the network has changed over time.6 Consider
two snapshots of a network Gt1 (V, Et1 ) and Gt2 (V, Et2 ) at times t1 and t2 , respectively, where t2 > t1 . Without loss of generality, we assume that the
number of nodes is fixed and only edges connecting these nodes change
(i.e., are added or removed).
When dealing with nominal attributes, the homophily index is defined
as
2
1
H = Qtnormalized
Qtnormalized
,
(8.38)
where Qnormalized is defined in Equation 8.9. Similarly, for ordinal attributes,
the homophily index can be defined as the change in the Pearson correlation
(Equation 8.24):
H = t2 t1 .
(8.39)

8.3.2

Modeling Homophily

Homophily can be modeled using a variation of the independent cascade

model discussed in Chapter 7. In this variation, at each time step a single
node gets activated, and the activated node gets a chance of getting connected to other nodes due to homophily. In other words, if the activated
node finds other nodes in the network similar enough (i.e., their similarity
6

Note that we have assumed that homophily is the leading social force in the network
and that it leads to its assortativity change. This assumption is often strong for social
networks because other social forces act in these networks.

278

Algorithm 8.2 Homophily Model

Require: Graph G(V, E), E = , similarities sim(v, u)
1: return Set of edges E
2: for all v V do
3:
v = generate a random number in [0,1];
4:
for all (v, u) < E do
5:
if v < sim(v, u) then
6:
E = E (v, u);
7:
end if
8:
end for
9: end for
10: Return E;
is higher than some tolerance value), it connects to them via an edge. A
node once activated has no chance of getting activated again.
Modeling homophily is outlined in Algorithm 8.2. Let sim(u, v) denote
the similarity between nodes u and v. When a node gets activated, we
generate a random tolerance value for the node v between 0 and 1. Alternatively, we can set this tolerance to some predefined value. The tolerance
value defines the minimum similarity that node v tolerates for connecting
to other nodes. Then, for any likely edge (v, u) that is still not in the edge
set, if the similarity is more than the tolerance: sim(v, u) > v , the edge
(v, u) is added. The process continues until all nodes are activated.
The model can be used in two different scenarios. First, given a network
in which assortativity is attributed to homophily, we can estimate tolerance values for all nodes. To estimate tolerance values, we can simulate
the homophily model in Algorithm 8.2 on the given network (by removing
all its edges) with different tolerance values. We can then compare the
assortativity of the simulated network and the given network. By finding
the simulated network that best fits the given network (i.e., has the closest
assortativity value to the given networks assortativity), we can determine
the tolerance values for individuals. Second, when a network is given and
the source of assortativity is unknown, we can estimate how much of the
observed assortativity can be attributed to homophily. To measure assortativity due to homophily, we can simulate homophily on the given network
by removing edges. The distance between the assortativity measured on
the simulated network and the given network explains how much of the
279

observed assortativity is due to homophily. The smaller this distance, the

higher the effect of homophily in generating the observed assortativity.

8.4

Distinguishing Influence and Homophily

We are often interested in understanding which social force (influence or

homophily) resulted in an assortative network. To distinguish between an
influence-based assortativity or homophily-based one, statistical tests can
be used. In this section, we discuss three tests: the shuffle test, the edgereversal test, and the randomization test. The first two can detect whether
influence exists in a network or not, but are incapable of detecting homophily. The last one, however, can distinguish influence and homophily.
Note that in all these tests, we assume that several temporal snapshots
of the dataset are available (like the LIM model) where we know exactly
when each node is activated, when edges are formed, or when attributes
are changed.

8.4.1

Social
Correlation

Shuffle Test

The shuffle test was originally introduced by Anagnostopoulos et al. [10].

The basic idea behind the shuffle test comes from the fact that influence is
temporal. In other words, when u influences v, then v should have been
activated after u. So, in the shuffle test, we define a temporal assortativity
measure. We assume that if there is no influence, then a shuffling of
the activation time stamps should not affect the temporal assortativity
measurement.
In this temporal assortativity measure, called social correlation, the probability of activating a node v depends on a, the number of already active
friends it has. This activation probability is calculated using a logistic
function,7
ea+
,
(8.40)
p(a) =
1 + ea+
In the original paper, instead of a, the authors use ln(a + 1) as the variable. This helps
remove the effect of a power-law distribution in the number of activated friends. Here,
for simplicity, we use the nonlogarithmic form.
7

280

or equivalently,
p(a)
) = a + ,
(8.41)
1 p(a)
where measures the social correlation and denotes the activation bias.
For computing the number of already active nodes of an individual, we
need to know the activation time stamps of the nodes.
Let ya,t denote the number of individuals who became activated at time
t and had a active friends and let na,t denote the ones
active
P who had a P
friends but did not get activated at time t. Let ya = t ya,t and na = t na,t .
We define the likelihood function as
Y
p(a) ya (1 p(a))na .
(8.42)
ln(

To estimate and , we find their values such that the likelihood function denoted in Equation 8.42 is maximized. Unfortunately, there is no
closed-form solution, but there exist software packages that can efficiently
compute the solution to this optimization.8
Let tu denote the activation time (when a node is first influenced) of
node u. When activated node u influences nonactivated node v, and v
is activated, then we have tu < tv . Hence, when temporal information is
available about who activated whom, we see that influenced nodes are
activated at a later time than those who influenced them. Now, if there is
no influence in the network, we can randomly shuffle the activation time
stamps, and the predicted should not change drastically. So, if we shuffle
activation time stamps and compute the correlation coefficient 0 and its
value is close to the computed in the original unshuffled dataset (i.e.,
| 0 | is small), then the network does not exhibit signs of social influence.

8.4.2

Edge-Reversal Test

The edge-reversal test introduced by Christakis and Fowler [54] follows

a similar approach as the shuffle test. If influence resulted in activation,
then the direction of edges should be important (who influenced whom).
So, we can reverse the direction of edges, and if there is no social influence
in the network, then the value of social correlation , as defined in Section
8.4.1, should not change dramatically.
8

Note that maximizing this term is equivalent to maximizing the logarithm; this is
where Equation 8.41 comes into play.

281

Figure 8.7: The Effect of Influence and Homophily on Attributes and Links
over Time (reproduced from [161]).

8.4.3

Randomization Test

Unlike the other two tests, the randomization test [161] is capable of detecting both influence and homophily in networks. Let X denote the attributes
associated with nodes (age, gender, location, etc.) and Xt denote the attributes at time t. Let Xi denote attributes of node vi . As mentioned before,
in influence, individuals already linked to one another change their attributes (e.g., a user changes habits), whereas in homophily, attributes do
not change but connections are formed due to similarity. Figure 8.7 demonstrates the effect of influence and homophily in a network over time.
The assumption is that, if influence or homophily happens in a network, then networks become more assortative. Let A(Gt , Xt ) denote the
assortativity of network G and attributes X at time t. Then, the network
becomes more assortative at time t + 1 if
A(Gt+1 , Xt+1 ) A(Gt , Xt ) > 0.

Influence Gain and

Homophily Gain

(8.43)

Now, we can assume that part of this assortativity is due to influence if

the influence gain GInfluence is positive,
GInfluence (t) = A(Gt , Xt+1 ) A(Gt , Xt ) > 0,

(8.44)

and part is due to homophily if we have positive homophily gain GHomophily :

GHomophily (t) = A(Gt+1 , Xt ) A(Gt , Xt ) > 0.
282

(8.45)

Algorithm 8.3 Influence Significance Test

Require: Gt , Gt+1 , Xt , Xt+1 , number of randomized runs n,
1: return Significance
2: g0 = GInfluence (t);
3: for all 1 i n do
4:
XRit+1 = randomizeI (Xt , Xt+1 );
5:
gi = A(Gt , XRit+1 ) A(Gt , Xt );
6: end for
7: if g0 larger than (1 /2)% of values in {gi }ni=1 then
8:
return significant;
9: else if g0 smaller than /2% of values in {gi }ni=1 then
10:
return significant;
11: else
12:
return insignificant;
13: end if
Note that Xt+1 denotes the changes in attributes, and Gt+1 denotes the
changes in links in the network (new friendships formed). In randomization tests, one determines whether changes in A(Gt , Xt+1 ) A(Gt , Xt )
(influence), or A(Gt+1 , Xt ) A(Gt , Xt ) (homophily), are significant or not.
To detect change significance, we use the influence significance test and
homophily significance test algorithms outlined in Algorithms 8.3 and 8.4,
respectively. The influence significance algorithm starts with computing Influence
influence gain, which is the assortativity difference observed due to influ- Significance Test
ence (g0 ). It then forms a random attribute set at time t+1 (null-hypotheses),
assuming that attributes changed randomly at t + 1 and not due to influence. This random attribute set XRit+1 is formed from Xt+1 by making sure
that effects of influence in changing attributes are removed.
For instance, assume two users u and v are connected at time t, and u
has hobby movies at time t and v does not have this hobby listed at time
t. Now, assuming there is an influence of u over v, so that at time t + 1, v
adds movies to her set of hobbies. In other words, movies < Xtv and movies
v
Xt+1
. To remove this influence, we can construct XRit+1 by removing
movies from the hobbies of v at time t + 1 and adding some random hobby
such as reading, which is < Xtu and < Xtv , to the list of hobbies of v at time
t + 1 in XRit+1 . This guarantees that the randomized XRit+1 constructed has
no sign of influence. We construct this randomized set n times; this set is
283

Algorithm 8.4 Homophily Significance Test

Require: Gt , Gt+1 , Xt , Xt+1 , number of randomized runs n,
1: return Significance
2: g0 = GHomophily (t);
3: for all 1 i n do
4:
GRit+1 = randomizeH (Gt , Gt+1 );
5:
gi = A(GRit+1 , Xt ) A(Gt , Xt );
6: end for
7: if g0 larger than (1 /2)% of values in {gi }ni=1 then
8:
return significant;
9: else if g0 smaller than /2% of values in {gi }ni=1 then
10:
return significant;
11: else
12:
return insignificant;
13: end if

Homophily
Significance Test

then used to compute influence gains {gi }ni=1 . Obviously, the more distant
g0 is from these gains, the more significant influence is. We can assume
that whenever g0 is smaller than /2% (or larger than 1 /2%) of {gi }ni=1
values, it is significant. The value of is set empirically.
Similarly, in the homophily significance test, we compute the original
homophily gain and construct random graph links GRit+1 at time t + 1, such
that no homophily effect is exhibited in how links are formed. To perform
this for any two (randomly selected) links ei j and ekl formed in the original
Gt+1 graph, we form edges eil and ek j in GRit+1 . This is to make sure that the
homophily effect is removed and that the degrees in GRit+1 are equal to that
of Gt+1 .

284

8.5

Summary

Individuals are driven by different social forces across social media. Two
such important forces are influence and homophily.
In influence, an individuals actions induce her friends to act in a similar
fashion. In other words, influence makes friends more similar. Homophily
is the tendency for similar individuals to befriend each other. Both influence and homophily result in networks where similar individuals are
connected to each other. These are assortative networks. To estimate the
assortativity of networks, we use different measures depending on the
attribute type that is tested for similarity. We discussed modularity for
nominal attributes and correlation for ordinal ones.
Influence can be quantified via different measures. Some are predictionbased, where the measure assumes that some attributes can accurately predict how influential an individual will be, such as with in-degree. Others
are observation-based, where the influence score is assigned to an individual based on some history, such as how many individuals he or she has
influenced. We also presented case studies for measuring influence in the
blogosphere and on Twitter.
Influence is modeled differently depending on the visibility of the network. When network information is available, we employ threshold models such as the linear threshold model (LTM), and when network information is not available, we estimate influence rates using the linear influence
model (LIM). Similarly, homophily can be measured by computing the assortativity difference in time and modeled using a variant of independent
cascade models.
Finally, to determine the source of assortativity in social networks, we
described three statistical tests: the shuffle test, the edge-reversal test, and
the randomization test. The first two can determine if influence is present
in the data, and the last one can determine both influence and homophily.
All tests require temporal data, where activation times and changes in
attributes and links are available.

285

8.6

Bibliographic Notes

Indications of assortativity observed in the real world can be found in [62].

General reviews of the assortativity measuring methods discussed in this
chapter can be found in [209, 212, 215].
Influence and homophily are extensively discussed in the social sciences
literature (see [57, 192]). Interesting experiments in this area can be found
in Milgrams seminal experiment on obedience to authority [194]. In his
controversial study, Milgram showed many individuals, because of fear
or their desire to appear cooperative, are willing to perform acts that are
against their better judgment. He recruited participants in what seemingly
looked like a learning experiment. Participants were told to administer
increasingly severe electric shocks to another individual (the learner) if
he answered questions incorrectly. These shocks were from 15450 volts
(lethal level). In reality, the learner was an actor, a confederate of Milgram,
and never received any shocks. However, the actor shouted loudly to
demonstrate the painfulness of the shocks. Milgram found that 65% of
participants in his experiments were willing to give lethal electric shocks
up to 450 volts to the learner, after being given assurance statements such
as Although the shocks may be painful, there is no permanent tissue
damage, so please go on, or given direct orders, such as the experiment
requires that you continue. Another study is the 32-year longitudinal study
on the spread of obesity in social networks [54]. In this study, Christakis
et al. analyzed a population of 12,067 individuals. The body mass index
for these individuals was available from 19712003. They showed that an
individuals likelihood of becoming obese over time increased by almost
60% if he or she had an obese friend. This likelihood decreased to around
40% for those with an obese sibling or spouse.
The analysis of influence and homophily is also an active topic in social
media mining. For studies regarding influence and homophily online,
refer to [297, 255, 62, 223, 300, 22]. The effect of influence and homophily
on the social network has also been used for prediction purposes. For
instance, Tang et al. [271] use the effect of homophily for trust prediction.
Modeling influence is challenging. For a review of threshold models,
similar techniques, and challenges, see [107, 296, 108, 146].
In addition to tests discussed for identifying influence or homophily,
we refer readers to the works of Aral et al. [15] and Snijders et al. [261].

286

8.7

Exercises

1. State two common factors that explain why connected people are
similar or vice versa.

Measuring Assortativity
2.

What is the range [1 , 2 ] for modularity Q values? Provide

examples for both extreme values of the range, as well as cases
where modularity becomes zero.
What are the limitations for modularity?
Compute modularity in the following graph. Assume that {ai }4i=0
nodes are category a, {bi }4i=0 nodes are category b, and {ci }4i=0 nodes
are category c.

Influence
3. Does the linear threshold model (LTM) converge? Why?
4. Follow the LTM procedure until convergence for the following graph.
Assume all the thresholds are 0.5 and node v1 is activated at time 0.

287

5. Discuss a methodology for identifying the influentials given multiple

influence measures using the following scenario: on Twitter, one can
use in-degree and number of retweets as two independent influence
measures. How can you find the influentials by employing both
measures?

Homophily
6. Design a measure for homophily that takes into account assortativity
changes due to influence.

Distinguishing Influence and Homophily

7. What is a shuffle test designed for in the context of social influence?
Describe how it is performed.
8. Describe how the edge-reversal test works. What is it used for?

288

This chapter is from Social Media Mining: An Introduction.

By Reza Zafarani, Mohammad Ali Abbasi, and Huan Liu.
Cambridge University Press, 2014. Draft version: April 20, 2014.
Complete Draft and Slides Available at: http://dmml.asu.edu/smm

Chapter

Recommendation in Social Media

Individuals in social media make a variety of decisions on a daily basis.
These decisions are about buying a product, purchasing a service, adding
a friend, and renting a movie, among others. The individual often faces
many options to choose from. These diverse options, the pursuit of optimality, and the limited knowledge that each individual has create a desire
for external help. At times, we resort to search engines for recommendations; however, the results in search engines are rarely tailored to our
particular tastes and are query-dependent, independent of the individuals
who search for them.
Applications and algorithms are developed to help individuals decide easily, rapidly, and more accurately. These algorithms are tailored
to individuals tastes such that customized recommendations are available for them. These algorithms are called recommendation algorithms or Recommender System
recommender systems.
Recommender systems are commonly used for product recommendation. Their goal is to recommend products that would be interesting to
individuals. Formally, a recommendation algorithm takes a set of users U
and a set of items I and learns a function f such that
f :UI R

(9.1)

In other words, the algorithm learns a function that assigns a real value
to each user-item pair (u, i), where this value indicates how interested user
u is in item i. This value denotes the rating given by user u to item i. The
recommendation algorithm is not limited to item recommendation and
289

can be generalized to recommending people and material, such as, ads or

content.
Recommendation vs. Search
When individuals seek recommendations, they often use web search engines. However, search engines are rarely tailored to individuals needs
and often retrieve the same results as long as the search query stays the
same. To receive accurate recommendation from a search engine, one
needs to send accurate keywords to the search engine. For instance, the
query best 2013 movie to watch issued by an 8-year old and an
adult will result in the same set of movies, whereas their individual tastes
dictate different movies.
Recommendation systems are designed to recommend individual-based
choices. Thus, the same query issued by different individuals should result
in different recommendations. These systems commonly employ browsing
history, product purchases, user profile information, and friends information to make customized recommendations. As simple as this process may
look, a recommendation system algorithm actually has to deal with many
challenges.

9.1

Challenges

Recommendation systems face many challenges, some of which are presented next:
Cold-Start Problem. Many recommendation systems use historical data or information provided by the user to recommend items,
products, and the like. However, when individuals first join sites,
they have not yet bought any product: they have no history. This
makes it hard to infer what they are going to like when they start
on a site. The problem is referred to as the cold-start problem. As
an example, consider an online movie rental store. This store has no
idea what recently joined users prefer to watch and therefore cannot
recommend something close to their tastes. To address this issue,
these sites often ask users to rate a couple of movies before they begin recommend others to them. Other sites ask users to fill in profile
290

information, such as interests. This information serves as an input to

the recommendation algorithm.
Data Sparsity. Similar to the cold-start problem, data sparsity occurs
when not enough historical or prior information is available. Unlike the cold-start problem, data sparsity relates to the system as a
whole and is not specific to an individual. In general, data sparsity
occurs when a few individuals rate many items while many other
individuals rate only a few items. Recommender systems often use
information provided by other users to help offer better recommendations to an individual. When this information is not reasonably
available, then it is said that a data sparsity problem exists. The problem is more promi- nent in sites that are recently launched or ones
that are not popular.
Attacks. The recommender system may be attacked to recommend
items otherwise not recommended. For instance, consider a system
that recommends items based on similarity between ratings (e.g.,
lens A is recommended for camera B because they both have rating
4). Now, an attacker that has knowledge of the recommendation
algorithm can create a set of fake user accounts and rate lens C (which
is not as good as lens A) highly such that it can get rating 4. This
way the recommendation system will recommend C with camera B
as well as A. This attack is called a push attack, because it pushes Nuke Attack
the ratings up such that the system starts recommending items that and
would otherwise not be recommended. Other attacks such as nuke Push Attack
attacks attempt to stop the whole recommendation system algorithm
and make it unstable. A recommendation system should have the
means to stop such attacks.
Privacy. The more information a recommender system has about
the users, the better the recommendations it provides to the users.
However, users often avoid revealing information about themselves
due to privacy concerns. Recommender systems should address this
challenge while protecting individuals privacy.
Explanation. Recommendation systems often recommend items
without having an explanation why they did so. For instance, when
291

several items are bought together by many users, the system recommends these to new users items together. However, the system does
not know why these items are bought together. Individuals may
prefer some reasons for buying items; therefore, recommendation
algorithms should provide explanation when possible.

9.2

Classical Recommendation Algorithms

Classical recommendation algorithms have a long history on the web. In

recent years, with the emergence of social media sites, these algorithms
have been provided new information, such as friendship information, interactions, and so on. We review these algorithms in this section.

9.2.1

Content-Based Methods

Content-based recommendation systems are based on the fact that a users

interest should match the description of the items that are recommended
by the system. In other words, the more similar the items description
to the users interest, the higher the likelihood that the user is going to
find the items recommendation interesting. Content-based recommender
systems implement this idea by measuring the similarity between an items
description and the users profile information. The higher this similarity,
the higher the chance that the item is recommended.
To formalize a content-based method, we first represent both user profiles and item descriptions by vectorizing (see Chapter 5) them using a
set of k keywords. After vectorization, item j can be represented as a kdimensional vector I j = (i j,1 , i j,2 , . . . , i j,k ) and user i as Ui = (ui,1 , ui,2 , . . . , ui,k ).
To compute the similarity between user i and item j, we can use cosine
similarity between the two vectors Ui and I j :
Pk
sim(Ui , I j ) = cos(Ui , I j ) = q
Pk
l=1

ui,l i j,l
q
Pk
2

l=1

ui,l

l=1 i j,l

(9.2)
2

In content-based recommendation, we compute the topmost similar

items to a user j and then recommend these items in the order of similarity.
Algorithm 9.1 shows the main steps of content-based recommendation.
292

Algorithm 9.1 Content-based recommendation

Require: User is Profile Information, Item descriptions for items j
{1, 2, . . . , n}, k keywords, r number of recommendations.
1: return r recommended items.
2: Ui = (u1 , u2 , . . . , uk ) = user is profile vector;
3: {I j }nj=1 = {(i j,1 , i j,2 , . . . , i j,k ) = item js description vector}nj=1 ;
4: si, j = sim(Ui , I j ), 1 j n;
5: Return top r items with maximum similarity si,j .

Table 9.1: User-Item Matrix

John
Joe
Jill
Jane
Jorge

9.2.2

Lion King
3
5
1
3
2

Aladdin
0
4
2
?
2

Mulan
3
0
4
1
0

Anastasia
3
2
2
0
1

Collaborative Filtering (CF)

Collaborative filtering is another set of classical recommendation techniques. In collaborative filtering, one is commonly given a user-item matrix where each entry is either unknown or is the rating assigned by the
user to an item. Table 9.1 is an user-item matrix where ratings for some cartoons are known and unknown for others (question marks). For instance,
on a review scale of 5, where 5 is the best and 0 is the worst, if an entry (i, j)
in the user-item matrix is 4, that means that user i liked item j.
In collaborative filtering, one aims to predict the missing ratings and
possibly recommend the cartoon with the highest predicted rating to the
user. This prediction can be performed directly by using previous ratings
in the matrix. This approach is called memory-based collaborative filtering
because it employs historical data available in the matrix. Alternatively,
one can assume that an underlying model (hypothesis) governs the way
users rate items. This model can be approximated and learned. After
the model is learned, one can use it to predict other ratings. The second
approach is called model-based collaborative filtering.
293

Memory-Based Collaborative Filtering

In memory-based collaborative filtering, one assumes one of the following
(or both) to be true:
Users with similar previous ratings for items are likely to rate future
items similarly.
Items that have received similar ratings previously from users are
likely to receive similar ratings from future users.
If one follows the first assumption, the memory-based technique is a
user-based CF algorithm, and if one follows the latter, it is an item-based
CF algorithm. In both cases, users (or items) collaboratively help filter
out irrelevant content (dissimilar users or items). To determine similarity
between users or items, in collaborative filtering, two commonly used
similarity measures are cosine similarity and Pearson correlation. Let ru,i
denote the rating that user u assigns to item i, let ru denote the average
rating for user u, and let ri be the average rating for item i. Cosine similarity
between users u and v is
P
ru,i rv,i
Uu Uv
= pP i pP
.
(9.3)
sim(Uu , Uv ) = cos(Uu , Uv ) =
2
2
||Uu || ||Uv ||
i ru,i
i rv,i
And the Pearson correlation coefficient is defined as
P
(ru,i ru )(rv,i rv )
sim(Uu , Uv ) = pP i
.
pP
2
2
i (ru,i ru )
i (rv,i rv )

Neighborhood

(9.4)

Next, we discuss user- and item-based collaborative filtering.

User-Based Collaborative Filtering. In this method, we predict the rating
of user u for item i by (1) finding users most similar to u and (2) using
a combination of the ratings of these users for item i as the predicted
rating of user u for item i. To remove noise and reduce computation, we
often limit the number of similar users to some fixed number. These most
similar users are called the neighborhood for user u, N(u). In user-based
collaborative filtering, the rating of user u for item i is calculated as
P
vN(u) sim(u, v)(rv,i rv )
P
,
(9.5)
ru,i = ru +
vN(u) sim(u, v)
where the number of members of N(u) is predetermined (e.g., top 10 most
similar members).
294

Example 9.1. In Table 9.1, rJane,Aladdin is missing. The average ratings are the
following:
rJohn =
rJoe =
rJill =
rJane =
rJorge =

3+3+0+3
= 2.25
4
5+4+0+2
= 2.75
4
1+2+4+2
= 2.25
4
3+1+0
= 1.33
3
2+2+0+1
= 1.25.
4

(9.6)
(9.7)
(9.8)
(9.9)
(9.10)

Using cosine similarity (or Pearson correlation), the similarity between Jane
and others can be computed:
33+13+03

10 27
35+10+02
sim(Jane, Joe) =

10 29
31+14+02
sim(Jane, Jill) =

10 21
32+10+01
sim(Jane, Jorge) =

10 5
sim(Jane, John) =

= 0.73

(9.11)

= 0.88

(9.12)

= 0.48

(9.13)

= 0.84.

(9.14)

Now, assuming that the neighborhood size is 2, then Jorge and Joe are the two
most similar neighbors. Then, Janes rating for Aladdin computed from user-based
collaborative filtering is
rJane,Aladdin = rJane +
+

sim(Jane, Joe)(rJoe,Aladdin rJoe )

sim(Jane, Joe) + sim(Jane, Jorge)

sim(Jane, Jorge)(rJorge,Aladdin rJorge )

sim(Jane, Joe) + sim(Jane, Jorge)

= 1.33 +

0.88(4 2.75) + 0.84(2 1.25)

= 2.33
0.88 + 0.84
295

(9.15)

Item-based Collaborative Filtering. In user-based collaborative filtering,

we compute the average rating for different users and find the most similar
users to the users for whom we are seeking recommendations. Unfortunately, in most online systems, users do not have many ratings; therefore,
the averages and similarities may be unreliable. This often results in a different set of similar users when new ratings are added to the system. On
the other hand, products usually have many ratings and their average and
the similarity between them are more stable. In item-based CF, we perform
collaborative filtering by finding the most similar items. The rating of user
u for item i is calculated as
P
jN(i) sim(i, j)(ru,j r j )
P
,
(9.16)
ru,i = ri +
jN(i) sim(i, j)
where ri and r j are the average ratings for items i and j, respectively.
Example 9.2. In Table 9.1, rJane,Aladdin is missing. The average ratings for items
are
3+5+1+3+2
rLion King =
= 2.8.
(9.17)
5
0+4+2+2
rAladdin =
= 2.
(9.18)
4
3+0+4+1+0
rMulan =
= 1.6.
(9.19)
5
3+2+2+0+1
rAnastasia =
= 1.6.
(9.20)
5
Using cosine similarity (or Pearson correlation), the similarity between Aladdin and others can be computed:
sim(Aladdin, Lion King) =

0 3+4 5+2 1+2 2

= 0.84.

24 39
(9.21)

sim(Aladdin, Mulan) =

0 3+4 0+2 4+2 0

= 0.32.

24 25
(9.22)

sim(Aladdin, Anastasia) =

0 3+4 2+2 2+2 1

= 0.67.

24 18
(9.23)
296

Now, assuming that the neighborhood size is 2, then Lion King and Anastasia
are the two most similar neighbors. Then, Janes rating for Aladdin computed
from item-based collaborative filtering is
sim(Aladdin, Lion King)(rJane,Lion King rLion King )
sim(Aladdin, Lion King) + sim(Aladdin, Anastasia)
sim(Aladdin, Anastasia)(rJane,Anastasia rAnastasia )
+
sim(Aladdin, Lion King) + sim(Aladdin, Anastasia)
0.84(3 2.8) + 0.67(0 1.6)
= 2+
= 1.40.
(9.24)
0.84 + 0.67

rJane,Aladdin = rAladdin +

Model-Based Collaborative Filtering

In memory-based methods (either item-based or user-based), one aims to
predict the missing ratings based on similarities between users or items. In
model-based collaborative filtering, one assumes that an underlying model
governs the way users rate. We aim to learn that model and then use that
model to predict the missing ratings. Among a variety of model-based
techniques, we focus on a well-established model-based technique that is
based on singular value decomposition (SVD).
Singular
mn Value
SVD is a linear algebra technique that, given a real matrix X R ,
Decomposition
m n, factorizes it into three matrices,
X = UV T ,

(9.25)

where U Rmm and V Rnn are orthogonal matrices and Rmn is a

Lossless
diagonal matrix. The product of these matrices is equivalent to the original Matrix
matrix; therefore,
no information is lost. Hence, the process is lossless.
Factorization
qP
m Pn
2
Let kXkF =
i=1
j=1 Xi j denote the Frobenius norm of matrix X. A
low-rank matrix approximation of matrix X is another matrix C Rmn . C
approximates X, and Cs rank (the maximum number of linearly indepen- Frobenius
dent columns) is a fixed number k  min(m, n):
Norm
rank(C) = k.

(9.26)

The best low-rank matrix approximation is a matrix C that minimizes

kX CkF . Low-rank approximations of matrices remove noise by assuming
that the matrix is not generated at random and has an underlying structure.
297

SVD can help remove noise by computing a low-rank approximation of a

matrix. Consider the following matrix Xk , which we construct from matrix
X after computing the SVD of X = UV T :
1. Create k from by keeping only the first k elements on the diagonal.
This way, k Rkk .
2. Keep only the first k columns of U and denote it as Uk Rmk , and
keep only the first k rows of V T and denote it as Vk T Rkn .
3. Let Xk = Uk k Vk T , Xk Rmn .
Eckart-Young-Mirsky
Theorem

As it turns out, Xk is the best low-rank approximation of a matrix X. The

following Eckart-Young-Mirsky theorem outlines this result.
Theorem 9.1 (Eckart-Young-Mirsky Low-Rank Matrix Approximation).
Let X be a matrix and C be the best low-rank approximation of X; if kX CkF is
minimized, and rank(C) = k, then C = Xk .
To summarize, the best rank-k approximation of the matrix can be easily
computed by calculating the SVD of the matrix and then taking the first k
columns of U, truncating to the the first k entries, and taking the first k
rows of V T .
As mentioned, low-rank approximation helps remove noise from a
matrix by assuming that the matrix is low rank. In low-rank approximation
using SVD, if X Rmn , then Uk Rmk , k Rkk , and VkT Rkn .
Hence, Uk has the same number of rows as X, but in a k-dimensional space.
Therefore, Uk represents rows of X, but in a transformed k-dimensional
space. The same holds for VkT because it has the same number of columns
as X, but in a k-dimensional space. To summarize, Uk and VkT can be
thought of as k-dimensional representations of rows and columns of X. In
this k-dimensional space, noise is removed and more similar points should
be closer.
Now, given the user-item matrix X, we can remove its noise by computing Xk from X and getting the new k-dimensional user space Uk or the
k-dimensional item space VkT . This way, we can compute the most similar
neighbors based on distances in this k-dimensional space. The similarity
in the k-dimensional space can be computed using cosine similarity or
Pearson correlation. We demonstrate this via Example 9.3.
298

Table 9.2: An User-Item Matrix

Lion King Aladdin Mulan
John
3
0
3
Joe
5
4
0
Jill
1
2
4
Jorge
2
2
0
Example 9.3. Consider the user-item matrix, in Table 9.2. Assuming this matrix
is X, then by computing the SVD of X = UV T ,1 we have

0.4151 0.4754 0.7679

0.7437
0.5278
0.0169

U =
0.4110 0.6626
0.6207

0.3251
0.2373
0.1572

0
0
8.0265

0
4.3886
0

0
0
2.0777

0
0
0

0.7506 0.5540 0.3600

T
0.2872 0.9290
V = 0.2335

0.6181
0.7814
0.0863

0.1093
0.4099
0.0820
0.9018

(9.27)

(9.28)

(9.29)

Considering a rank 2 approximation (i.e., k = 2), we truncate all three matrices:

Uk

k
VkT

"
=
"
=

0.4151 0.4754
0.7437
0.5278

0.4110 0.6626

0.3251
0.2373
#
8.0265
0
0
4.3886
#
0.7506 0.5540 0.3600
.
0.2335
0.2872 0.9290

(9.30)

(9.31)
(9.32)

The rows of Uk represent users. Similarly the columns of VkT (or rows of Vk )
represent items. Thus, we can plot users and items in a 2-D figure. By plotting
1

In Matlab, this can be performed using the svd command.

299

Figure 9.1: Users and Items in the 2-D Space.

user rows or item columns, we avoid computing distances between them and can
visually inspect items or users that are most similar to one another. Figure 9.1
depicts users and items depicted in a 2-D space. As shown, to recommend for Jill,
John is the most similar individual to her. Similarly, the most similar item to Lion
King is Aladdin.
After most similar items or users are found in the lower k-dimensional
space, one can follow the same process outlined in user-based or itembased collaborative filtering to find the ratings for an unknown item. For
instance, we showed in Example 9.3 (see Figure 9.1) that if we are predicting
the rating rJill,Lion King and assume that neighborhood size is 1, item-based
CF uses rJill,Aladdin , because Aladdin is closest to Lion King. Similarly, userbased collaborative filtering uses rJohn,Lion King , because John is the closest
user to Jill.

9.2.3

Extending Individual Recommendation to Groups of

Individuals

All methods discussed thus far are used to predict a rating for item i
for an individual u. Advertisements that individuals receive via email
marketing are examples of this type of recommendation on social media.
However, consider ads displayed on the starting page of a social media site.
These ads are shown to a large population of individuals. The goal when
showing these ads is to ensure that they are interesting to the individuals
300

who observe them. In other words, the site is advertising to a group of

individuals.
Our goal in this section is to formalize how existing methods for recommending to a single individual can be extended to a group of individuals.
Consider a group of individuals G = {u1 , u2 , . . . , un } and a set of products
I = {i1 , i2 , . . . , im }. From the products in I, we aim to recommend products to our group of individuals G such the recommendation satisfies the
group being recommended to as much as possible. One approach is to first
consider the ratings predicted for each individual in the group and then
devise methods that can aggregate ratings for the individuals in the group.
Products that have the highest aggregated ratings are selected for recommendation. Next, we discuss these aggregation strategies for individuals
in the group.
Aggregation Strategies for a Group of Individuals
We discuss three major aggregation strategies for individuals in the group.
Each aggregation strategy considers an assumption based on which ratings are aggregated. Let ru,i denote the rating of user u G for item i I.
Denote Ri as the group-aggregated rating for item i.
Maximizing Average Satisfaction. We assume that products that satisfy
each member of the group on average are the best to be recommended
to the group. Then, Ri group rating based on the maximizing average
satisfaction strategy is given as
Ri =

1X
ru,i .
n uG

(9.33)

After we compute Ri for all items i I, we recommend the items that

have the highest Ri s to members of the group.
Least Misery. This strategy combines ratings by taking the minimum of
them. In other words, we want to guarantee that no individuals is being
recommended an item that he or she strongly dislikes. In least misery, the
aggregated rating Ri of an item is given as
Ri = min ru,i .
uG

301

(9.34)

Similar to the previous strategy, we compute Ri for all items i I

and recommend the items with the highest Ri values. In other words, we
prefer recommending items to the group such that no member of the group
strongly dislikes them.
Most Pleasure. Unlike the least misery strategy, in the most pleasure
approach, we take the maximum rating in the group as the group rating:
Ri = max ru,i .

(9.35)

uG

Since we recommend items that have the highest Ri values, this strategy
guarantees that the items that are being recommended to the group are
enjoyed the most by at least one member of the group.
Example 9.4. Consider the user-item matrix in Table 9.3. Consider group G =
{John, Jill, Juan}. For this group, the aggregated ratings for all products using
average satisfaction, least misery, and most pleasure are as follows.
Table 9.3: User-Item Matrix
Soda Water Tea Coffee
John
1
3
1
1
Joe
4
3
1
2
Jill
2
2
4
2
Jorge
1
1
3
5
Juan
3
3
4
5
Average Satisfaction:
1+2+3
3
3+2+3
=
3
1+4+4
=
3
1+2+5
=
3

RSoda =

= 2.

(9.36)

= 2.66.

(9.37)

= 3.

(9.38)

= 2.66.

(9.39)

RSoda = min{1, 2, 3} = 1.

(9.40)

RWater
RTea
RCoffee
Least Misery:

302

RWater = min{3, 2, 3} = 2.
RTea = min{1, 4, 4} = 1.
RCoffee = min{1, 2, 5} = 1.

(9.41)
(9.42)
(9.43)

Most Pleasure:
RSoda
RWater
RTea
RCoffee

=
=
=
=

max{1, 2, 3} = 3.
max{3, 2, 3} = 3.
max{1, 4, 4} = 4.
max{1, 2, 5} = 5.

(9.44)
(9.45)
(9.46)
(9.47)

Thus, the first recommended items are tea, water, and coffee based on average
satisfaction, least misery, and most pleasure, respectively.

9.3

Recommendation Using Social Context

In social media, in addition to ratings of products, there is additional information available, such as the friendship network among individuals.
This information can be used to improve recommendations, based on the
assumption that an individuals friends have an impact on the ratings ascribed to the individual. This impact can be due to homophily, influence,
or confounding, discussed in Chapter 8. When utilizing this social information (i.e., social context) we can (1) use friendship information alone, (2)
use social information in addition to ratings, or (3) constrain recommendations using social information. Figure 9.2 compactly represents these three
approaches.

9.3.1

Using Social Context Alone

Consider a network of friendships for which no user-item rating matrix is

provided. In this network, we can still recommend users from the network
to other users for friendship. This is an example of friend recommendation
in social networks. For instance, in social networking sites, users are often
provided with a list of individuals they may know and are asked if they
wish to befriend them. How can we recommend such friends?
There are many methods that can be used to recommend friends in
social networks. One such method is link prediction, which we discuss
303

Figure 9.2: Recommendation using Social Context. When utilizing social

information, we can 1) utilize this information independently, 2) add it to
user-rating matrix, or 3) constrain recommendations with it.
in detail in Chapter 10. We can also use the structure of the network
to recommend friends. For example, it is well known that individuals
often form triads of friendships on social networks. In other words, two
friends of an individual are often friends with one another. A triad of three
individuals a, b, and c consists of three edges e(a, b), e(b, c), and e(c, a). A
triad that is missing one of these edges is denoted as an open triad. To
recommend friends, we can find open triads and recommend individuals
who are not connected as friends to one another.

9.3.2

Extending Classical Methods with Social Context

Social information can also be used in addition to a user-item rating matrix to improve recommendation. Addition of social information can be
performed by assuming that users that are connected (i.e., friends) have
similar tastes in rating items. We can model the taste of user Ui using
a k-dimensional vector Ui Rk1 . We can also model items in the kdimensional space. Let V j Rk1 denote the item representation in kdimensional space. We can assume that rating Ri j given by user i to item j
304

can be computed as
Ri j = UiT Vi .

(9.48)

To compute Ui and Vi , we can use matrix factorization. We can rewrite

Equation 9.48 in matrix format as
R = UT V,

(9.49)

where R Rnm , U Rkn , V Rkm , n is the number of users, and m is

the number of items. Similar to model-based CF discussed in Section 9.2.2,
matrix factorization methods can be used to find U and V, given user-item
rating matrix R. In mathematical terms, in this matrix factorization, we are
finding U and V by solving the following optimization problem:
1
min ||R UT V||2F .
U,V 2

(9.50)

Users often have only a few ratings for items; therefore, the R matrix
is very sparse and has many missing values. Since we compute U and V
only for nonmissing ratings, we can change Equation 9.50 to
n

1 XX
Ii j (Ri j UiT V j )2 ,
min
U,V 2
i=1 j=1

(9.51)

where Ii j {0, 1} and Ii j = 1 when user i has rated item j and is equal to
0 otherwise. This ensures that nonrated items do not contribute to the
summations being minimized in Equation 9.51. Often, when solving this
optimization problem, the computed U and V can estimate ratings for the
already rated items accurately, but fail at predicting ratings for unrated
items. This is known as the overfitting problem. The overfitting problem can Overfitting
be mitigated by allowing both U and V to only consider important features
required to represent the data. In mathematical terms, this is equivalent to
both U and V having small matrix norms. Thus, we can change Equation
9.51 to
n
m
1
2
1 XX
Ii j (Ri j UiT V j )2 + ||U||2F + ||V||2F ,
(9.52)
2 i=1 j=1
2
2
where 1 , 2 > 0 are predetermined constants that control the effects of
matrix norms. The terms 21 ||U||2F and 22 ||V||2F are denoted as regularization
305

terms. Note that to minimize Equation 9.52, we need to minimize all terms
in the equation, including the regularization terms. Thus, whenever one Regularization
Term
needs to minimize some other constraint, it can be introduced as a new
additive term in Equation 9.52. Equation 9.52 lacks a term that incorporates
the social network of users. For that, we can add another regularization
term,
n X
X
sim(i, j)||Ui U j ||2F ,
(9.53)
i=1 jF(i)

where sim(i, j) denotes the similarity between user i and j (e.g., cosine
similarity or Pearson correlation between their ratings) and F(i) denotes
the friends of i. When this term is minimized, it ensures that the taste for
user i is close to that of all his friends j F(i). As we did with previous
regularization terms, we can add this term to Equation 9.51. Hence, our
final goal is to solve the following optimization problem:
n

XX
1 XX
Ii j (Ri j UiT V j )2 +
sim(i, j)||Ui U j ||2F
min
U,V 2
i=1 j=1
i=1 jF(i)
+

2
1
||U||2F + ||V||2F ,
2
2

(9.54)

where is the constant that controls the effect of social network regularization. A local minimum for this optimization problem can be obtained
using gradient-descent-based approaches. To solve this problem, we can
compute the gradient with respect to Ui s and Vi s and perform a gradientdescent-based method.

9.3.3

Recommendation Constrained by Social Context

In classical recommendation, to estimate ratings of an item, one determines

similar users or items. In other words, any user similar to the individual
can contribute to the predicted ratings for the individual. We can limit the
set of individuals that can contribute to the ratings of a user to the set of
friends of the user. For instance, in user-based collaborative filtering, we
determine a neighborhood of most similar individuals. We can take the
intersection of this neighborhood with the set of friends of the individual
306

Table 9.4: User-Item Matrix

Lion King Aladdin Mulan Anastasia
John
4
3
2
2
Joe
5
2
1
5
Jill
2
5
?
0
Jane
1
3
4
3
Jorge
3
1
1
2
to attract recommendations only from friends who are similar enough:
P
vN(u)F(u) sim(u, v)(rv,i rv )
P
ru,i = ru +
.
(9.55)
vN(u)F(u) sim(u, v)
This approach has its own shortcomings. When there is no intersection
between the set of friends and the neighborhood of most similar individuals, the ratings cannot be computed. To mitigate this, one can use the set
of k most similar friends of an individual S(i) to predict the ratings,
P
vS(u) sim(u, v)(rv,i rv )
P
.
(9.56)
ru,i = ru +
vS(u) sim(u, v)
Similarly, when friends are not very similar to the individual, the predicted rating can be different from the rating predicted using most similar
users. Depending on the context, both equations can be utilized.
Example 9.5. Consider the user-item matrix in Table 9.4 and the following adjacency matrix denoting the friendship among these individuals.

John Joe Jill Jane Jorge

John
0
1
0
0
1

Joe
1
0
1
0
0

,
A =
(9.57)
0
1
0
1
1
Jill

Jane
0
0
1
0
0

Jorge
1
0
1
0
0
We wish to predict rJill,Mulan . We compute the average ratings and similarity
between Jill and other individuals using cosine similarity:
rJohn =

4+3+2+2
= 2.75.
4
307

(9.58)

5+2+1+5
= 3.25.
4
2+5+0
=
= 2.33.
3
1+3+4+3
=
= 2.75.
4
3+1+1+2
=
= 1.75.
4

rJoe =

(9.59)

rJill

(9.60)

rJane
rJorge

(9.61)
(9.62)

The similarities are

24+53+02

29 29
25+52+05
sim(Jill, Joe) =

29 54
21+53+03
sim(Jill, Jane) =

29 19
23+51+02
sim(Jill, Jorge) =

29 14
sim(Jill, John) =

= 0.79.

(9.63)

= 0.50.

(9.64)

= 0.72.

(9.65)

= 0.54.

(9.66)

Considering a neighborhood of size 2, the most similar users to Jill are John
and Jane:
N(Jill) = {John, Jane}.
(9.67)
We also know that friends of Jill are
F(Jill) = {Joe, Jane, Jorge}.

(9.68)

We can use Equation 9.55 to predict the missing rating by taking the intersection of friends and neighbors:
sim(Jill, Jane)(rJane,Mulan rJane )
sim(Jill, Jane)
= 2.33 + (4 2.75) = 3.58.

rJill,Mulan = rJill +

(9.69)

Similarly, we can utilize Equation 9.56 to compute the missing rating. Here,
we take Jills two most similar neighbors: Jane and Jorge.
rJill,Mulan = rJill +

sim(Jill, Jane)(rJane,Mulan rJane )

sim(Jill, Jane) + sim(Jill, Jorge)
308

sim(Jill, Jorge)(rJorge,Mulan rJorge )

sim(Jill, Jane) + sim(Jill, Jorge)
0.72(4 2.75) + 0.54(1 1.75)
= 2.33 +
= 2.72
0.72 + 0.54
+

9.4

(9.70)

Evaluating Recommendations

When a recommendation algorithm predicts ratings for items, one must

evaluate how accurate its recommendations are. One can evaluate the (1)
accuracy of predictions, (2) relevancy of recommendations, or (3) rankings
of recommendations.

9.4.1

Evaluating Accuracy of Predictions

When evaluating the accuracy of predictions, we measure how close predicted ratings are to the true ratings. Similar to the evaluation of supervised
learning, we often predict the ratings of some items with known ratings
(i.e., true ratings) and compute how close the predictions are to the true
ratings. One of the simplest methods, mean absolute error (MAE), computes the average absolute difference between the predicted ratings and
true ratings,
P
ri j ri j |
i j |
,
(9.71)
MAE =
n
where n is the number of predicted ratings, ri j is the predicted rating, and
ri j is the true rating. Normalized mean absolute error (NMAE) normalizes
MAE by dividing it by the range ratings can take,
NMAE =

MAE
,
rmax rmin

(9.72)

where rmax is the maximum rating items can take and rmin is the minimum.
In MAE, error linearly contributes to the MAE value. We can increase this
contribution by considering the summation of squared errors in the root
mean squared error (RMSE):
s
1X
RMSE =
(ri j ri j )2 .
(9.73)
n i, j
309

Example 9.6. Consider the following table with both the predicted ratings and
true ratings of five items:
Item
1
2
3
4
5

Predicted Rating
1
2
3
4
4

True Rating
3
5
3
2
1

The MAE, NMAE, and RMSE values are

|1 3| + |2 5| + |3 3| + |4 2| + |4 1|
= 2.
5
MAE
NMAE =
= 0.5.
5

1
r
(1 3)2 + (2 5)2 + (3 3)2 + (4 2)2 + (4 1)2
RMSE =
.
5
= 2.28.
MAE =

9.4.2

(9.74)
(9.75)
(9.76)

Evaluating Relevancy of Recommendations

When evaluating recommendations based on relevancy, we ask users if

they find the recommended items relevant to their interests. Given a set
of recommendations to a user, the user describes each recommendation
as relevant or irrelevant. Based on the selection of items for recommendations and their relevancy, we can have the four types of items outlined
in Table 9.5. Given this table, we can define measures that use relevancy
Table 9.5: Partitioning of Items with Respect to Their Selection for Recommendation and Their Relevancy
Selected Not Selected Total
Relevant
Nrs
Nrn
Nr
Irrelevant
Nis
Nin
Ni
Total
Ns
Nn
N
310

information provided by users. Precision is one such measure. It defines

the fraction of relevant items among recommended items:
P=

Nrs
.
Ns

(9.77)

Similarly, we can use recall to evaluate a recommender algorithm, which

provides the probability of selecting a relevant item for recommendation:
R=

Nrs
.
Nr

(9.78)

We can also combine both precision and recall by taking their harmonic
mean in the F-measure:
2PR
.
(9.79)
F=
P+R
Example 9.7. Consider the following recommendation relevancy matrix for a set
of 40 items. For this table, the precision, recall, and F-measure values are

Relevant
Irrelevant
Total

Selected
9
3
12

Not Selected
15
13
28

Total
24
16
40

9
= 0.75.
12
9
R =
= 0.375.
24
2 0.75 0.375
F =
= 0.5.
0.75 + 0.375
P =

9.4.3

(9.80)
(9.81)
(9.82)

Evaluating Ranking of Recommendations

Often, we predict ratings for multiple products for a user. Based on the
predicted ratings, we can rank products based on their levels of interestingness to the user and then evaluate this ranking. Given the true ranking
of interestingness of items, we can compare this ranking with it and report
a value. Rank correlation measures the correlation between the predicted
ranking and the true ranking. One such technique is the Spearmans rank
311

correlation discussed in Chapter 8. Let xi , 1 xi n, denote the rank

predicted for item i, 1 i n. Similarly, let yi , 1 yi n, denote the true
rank of item i from the users perspective. Spearmans rank correlation is
defined as
P
6 ni=1 (xi yi )2
,
(9.83)
=1
n3 n

Kendalls Tau

where n is the total number of items.

Here, we discuss another rank correlation measure: Kendalls tau. We
say that the pair of items (i, j) are concordant if their ranks {xi , yi } and {x j , y j }
are in order:
xi > x j , yi > y j or xi < x j , yi < y j .
(9.84)
A pair of items is discordant if their corresponding ranks are not in order:
xi > x j ,

yi < y j

or xi < x j ,

yi > y j .

(9.85)

When xi = x j or yi = y j , the pair is neither concordant nor discordant.

Let c denote the total number of concordant item pairs and d the total
number of discordant item pairs. Kendalls tau computes the difference


between the two, normalized by the total number of item pairs n2 :
=

cd
n
.

(9.86)

Kendalls tau takes value in range [1, 1]. When the ranks completely
agree, all pairs are concordant and Kendalls tau takes value 1, and when
the ranks completely disagree, all pairs are discordant and Kendalls tau
takes value 1.
Example 9.8. Consider a set of four items I = {i1 , i2 , i3 , i4 } for which the predicted
and true rankings are as follows:

i1
i2
i3
i4

Predicted Rank
1
2
3
4
312

True Rank
1
4
2
3

The pair of items and their status {concordant/discordant} are

(i1 , i2 )
(i1 , i3 )
(i1 , i4 )
(i2 , i3 )
(i2 , i4 )
(i3 , i4 )

:
:
:
:
:
:

concordant
concordant
concordant
discordant
discordant
concordant

(9.87)
(9.88)
(9.89)
(9.90)
(9.91)
(9.92)

Thus, Kendalls tau for the rankings is

=

42
= 0.33.
6

313

(9.93)

9.5

Summary

In social media, recommendations are constantly being provided. Friend

recommendation, product recommendation, and video recommendation,
among others, are all examples of recommendations taking place in social
media. Unlike web search, recommendation is tailored to individuals
interests and can help recommend more relevant items. Recommendation
is challenging due to the cold-start problem, data sparsity, attacks on these
systems, privacy concerns, and the need for an explanation for why items
are being recommended.
In social media, sites often resort to classical recommendation algorithms to recommend items or products. These techniques can be divided into content-based methods and collaborative filtering techniques.
In content-based methods, we use the similarity between the content (e.g.,
item description) of items and user profiles to recommend items. In collaborative filtering (CF), we use historical ratings of individuals in the form of
a user-item matrix to recommend items. CF methods can be categorized
into memory-based and model-based techniques. In memory-based techniques, we use the similarity between users (user-based) or items (itembased) to predict missing ratings. In model-based techniques, we assume
that an underlying model describes how users rate items. Using matrix
factorization techniques we approximate this model to predict missing
ratings. Classical recommendation algorithms often predict ratings for
individuals. We discussed ways to extend these techniques to groups of
individuals.
In social media, we can also use friendship information to give recommendations. These friendships alone can help recommend (e.g., friend
recommendation), can be added as complementary information to classical techniques, or can be used to constrain the recommendations provided
by classical techniques.
Finally, we discussed the evaluation of recommendation techniques.
Evaluation can be performed in terms of accuracy, relevancy, and rank of
recommended items. We discussed MAE, NMAE, and RMSE as methods
that evaluate accuracy, precision, recall, and F-measure from relevancybased methods, and Kendalls tau from rank-based methods.

314

9.6

Bibliographic Notes

General references for the content provided in this chapter can be found in
[138, 237, 249, 5]. In social media, recommendation is utilized for various
items, including blogs [16], news [177, 63], videos [66], and tags [257]. For
example, YouTube video recommendation system employs co-visitation
counts to compute the similarity between videos (items). To perform
recommendations, videos with high similarity to a seed set of videos are
recommended to the user. The seed set consists of the videos that users
watched on YouTube (beyond a certain threshold), as well as videos that
are explicitly favorited, liked, rated, or added to playlists.
Among classical techniques, more on content-based recommendation
can be found in [226], and more on collaborative filtering can be found
in [268, 246, 248]. Content-based and CF methods can be combined into
hybrid methods, which are not discussed in this chapter. A survey of hybrid
methods is available in [48]. More details on extending classical techniques
to groups are provided in [137].
When making recommendations using social context, we can use additional information such as tags [116, 254] or trust [102, 222, 190, 180]. For
instance, in [272], the authors discern multiple facets of trust and apply
multifaceted trust in social recommendation. In another work, Tang et
al. [273] exploit the evolution of both rating and trust relations for social
recommendation. Users in the physical world are likely to ask for suggestions from their local friends while they also tend to seek suggestions from
users with high global reputations (e.g., reviews by vine voice reviewers
of Amazon.com). Therefore, in addition to friends, one can also use global
network information for better recommendations. In [274], the authors
exploit both local and global social relations for recommendation.
When recommending people (potential friends), we can use all these
types of information. A comparison of different people recommendation
techniques can be found in the work of Chen et al. [52]. Methods that
extend classical techniques with social context are discussed in [181, 182,
152].

315

9.7

Exercises

Classical Recommendation Algorithm

1. Discuss one difference between content-based recommendation and
collaborative filtering.
2. Compute the missing rating in this table using user-based collaborative filtering (CF). Use cosine similarity to find the nearest neighbors.

Newton
Einstein
Gauss
Aristotle
Euclid

God
3
5
1
3
2

Le Cercle
Rouge
0
4
2
?
2

Cidade
de Deu
3
0
4
1
0

La vita
e bella
2
3
0
2
5

Rashomon
3
2
2
0
1

ru

1.5

Assuming that you have computed similarity values in the following

table, calculate Aristotles rating by completing these four tasks:
Aristotle

Newton
0.76

Einstein
?

Gauss
0.40

Euclid
0.78

Calculate the similarity value between Aristotle and Einstein.

Identify Aristotles two nearest neighbors.
Calculate ru values for everyone (Aristotles is given).
Calculate Aristotles rating for Le Cercle Rouge.
3. In an item-based CF recommendation, describe how the recommender finds and recommends items to the given user.

Recommendation Using Social Context

4. Provide two examples where social context can help improve classical
recommendation algorithms in social media.
316

P P
5. In Equation 9.54, the term ni=1 jF(i) sim(i, j)||Ui U j ||2F is added to
model the similarity between friends tastes. Let T Rnn denote
the pairwise trust matrix, in which 0 Ti j 1 denotes how much
user i trusts user j. Using your intuition on how trustworthiness
of individuals should affect recommendations received from them,
modify Equation 9.54 using trust matrix T.

Evaluating Recommendation Algorithms

6. What does high precision mean? Why is precision alone insufficient to measure performance under normal circumstances? Provide
an example to show that both precision and recall are important.
7. When is Kendalls tau equal to 1? In other words, how is the
predicted ranking different from the true ranking?

317

318

This chapter is from Social Media Mining: An Introduction.

By Reza Zafarani, Mohammad Ali Abbasi, and Huan Liu.
Cambridge University Press, 2014. Draft version: April 20, 2014.
Complete Draft and Slides Available at: http://dmml.asu.edu/smm

Chapter

10

Behavior Analytics
What motivates individuals to join an online group? When individuals
abandon social media sites, where do they migrate to? Can we predict
box office revenues for movies from tweets posted by individuals? These
questions are a few of many whose answers require us to analyze or predict
behaviors on social media.
Individuals exhibit different behaviors in social media: as individuals
or as part of a broader collective behavior. When discussing individual behavior, our focus is on one individual. Collective behavior emerges when a
population of individuals behave in a similar way with or without coordination or planning.
In this chapter we provide examples of individual and collective behaviors and elaborate techniques used to analyze, model, and predict these
behaviors.

10.1

Individual Behavior

We read online news; comment on posts, blogs, and videos; write reviews
for products; post; like; share; tweet; rate; recommend; listen to music;
and watch videos, among many other daily behaviors that we exhibit on
social media. What are the types of individual behavior that leave a trace
on social media?
We can generally categorize individual online behavior into three categories (shown in Figure 10.1):
319

Figure 10.1: Individual Behavior.

1. User-User Behavior. This is the behavior individuals exhibit with respect to other individuals. For instance, when befriending someone,
sending a message to another individual, playing games, following,
inviting, blocking, subscribing, or chatting, we are demonstrating a
user-user behavior.
2. User-Community Behavior. The target of this type of behavior is a
community. For example, joining or leaving a community, becoming
a fan of a community, or participating in its discussions are forms of
user-community behavior.
3. User-Entity Behavior. The target of this behavior is an entity in
social media. For instance, it includes writing a blogpost or review
or uploading a photo to a social media site.
As we know, link data and content data are frequently available on social
media. Link data represents the interactions users have with other users,
and content data is generated by users when using social media. One can
320

think of user-user behavior as users linking to other users and user-entity

behavior as users generating and consuming content. Users interacting
with communities is a blend of linking and content-generation behavior,
in which one can simply join a community (linking), read or write content
for a community (content consumption and generation), or can do a mix
of both activities. Link analysis and link prediction are commonly used to
analyze links, and text analysis is designed to analyze content. We use these
techniques to analyze, model, and predict individual behavior.

10.1.1

Individual Behavior Analysis

Individual behavior analysis aims to understand how different factors

affect individual behaviors observed online. It aims to correlate those
behaviors (or their intensity) with other measurable characteristics of users,
sites, or contents that could have possibly resulted in those behaviors.
First we discuss an example of behavior analysis on social media and
demonstrate how this behavior can be analyzed. After that, we outline the
process that can be followed to analyze any behavior on social media.
Community Membership in Social Media
Users often join different communities in social media; the act of becoming a community member is an example of user-community behavior.
Why do users join communities? In other words, what factors affect the
community-joining behavior of individuals?
To analyze community-joining behavior, we can observe users who join
communities and determine the factors that are common among them.
Hence, we require a population of users U = {u1 , u2 , . . . , un }, a community C, and community membership information (i.e., users ui U who
are members of C). The community need not be explicitly defined. For
instance, one can think of individuals buying a product as a community,
and people buying the product for the first time as individuals joining the
community. To distinguish between users who have already joined the
community and those who are now joining it, we need community memberships at two different times: t1 and t2 , with t2 > t1 . At t2 , we determine
users such as u who are currently members of the community, but were not
members at t1 . These new users form the subpopulation that is analyzed
for community-joining behavior.
321

Figure 10.2: Probability of Joining a Community (with Error Bars) as a

Function of the Number of Friends m Already in the Community (from
[19]).
To determine factors that affect community-joining behavior, we can design hypotheses based on different factors that describe when communityjoining behavior takes place. We can verify these hypotheses by using data
available on social media. The factors used in the validated hypotheses
describe the behavior under study most accurately.

Diminishing
Returns

One such hypothesis is that individuals are inclined toward an activity

when their friends are engaged in the same activity. Thus, if the hypothesis
is valid, a factor that plays a role in users joining a community is the number
of their friends who are already members of the community. In data
mining terms, this translates to using the number of friends of an individual
in a community as a feature to predict whether the individual joins the
community (i.e., class attribute). Figure 10.2 depicts the probability of
joining a community with respect to the number of friends an individual
has who are already members of the community. The probability increases
as more friends are in a community, but a diminishing returns property is also
observed, meaning that when enough friends are inside the community,
more friends have no or only marginal effects on the likelihood of the
individuals act of joining the community.
322

Figure 10.3: User Community-Joining Behavior Features (from [19]).

Thus far we have defined only one feature. However, one can go
beyond a single feature. Figure 10.3 lists the comprehensive features that
can be used to analyze community-joining behavior.
As discussed, these features may or may not affect the joining behavior;
thus, a validation procedure is required to understand their effect on the
joining behavior. Which one of these features is more relevant to the joining
behavior? In other words, which feature can help best determine whether
individuals will join or not?
To answer this question, we can use any feature selection algorithm. Feature selection algorithms determine features that contribute the most to the
prediction of the class attribute. Alternatively, we can use a classification
algorithm, such as decision tree learning, to identify the relationship between features and the class attribute (i.e., joined={Yes, No}). The earlier a
feature is selected in the learned tree (i.e., is closer to the root of the tree),
the more important to the prediction of the class attribute value.
By performing decision tree learning for a large dataset of users and the
features listed in Figure 10.3, one finds that not only the number of friends
inside a community but also how these friends are connected to each other
affect the joining probability. In particular, the denser the subgraph of
friends inside a community, the higher the likelihood of a user joining the
community. Let S denote the set of friends inside community C, and let ES
denote the set of edges between these
|S| friends. The maximum number of


|S|

edges between these S friends is 2 . So, the edge density is (S) = Es / |S|2 .
323

Figure 10.4: Decision Tree Learned for Community-Joining Behavior (from

[19]).
One finds that the higher this density, the more likely that one is going to
join a community. Figure 10.4 shows the first two levels of the decision tree
learned for this task using features described in Figure 10.3. Higher level
features are more discriminative in decision tree learning, and in our case,
the most important feature is the density of edges for the friends subgraph
inside the community.
To analyze community-joining behavior, one can design features that
are likely to be related to community joining behavior. Decision tree
learning can help identify which features are more predictive than others. However, how can we evaluate if these features are designed well
and whether other features are not required to accurately predict joining
behavior? Since classification is used to learn the relation between features
and behaviors one can always use classification evaluation metrics such as
accuracy to evaluate the performance of the learned model. An accurate
model translates to an accurate learning of feature-behavior association.
A Behavior Analysis Methodology
The analysis of community-joining behavior can be summarized via a
four-step methodology for behavioral analysis. The same approach can
be followed as a general guideline for analyzing other behaviors in social
media.
Commonly, to perform behavioral analysis, one needs the following
four components:
324

1. An observable behavior. The behavior that is analyzed needs to be

observable. For instance, to analyze community-joining behavior, it
is necessary to be able to accurately observe the joining of individuals
(and possibly their joining times).
2. Features. One needs to construct relevant data features (covariates)
that may or may not affect (or be affected by) the behavior. Anthropologists and sociologists can help design these features. The
intrinsic relation between these features and the behavior should be
clear from the domain experts point of view. In community-joining
behavior, we used the number of friends inside the community as
one feature.
3. Feature-Behavior Association. This step aims to find the relationship between features and behavior, which describes how changes
in features result in the behavior (or changes its intensity). We used
decision tree learning to find features that are most correlated with
community-joining behavior.
4. Evaluation Strategy. The final step evaluates the findings. This evaluation guarantees that the findings are due to the features defined
and not to externalities. We use classification accuracy to verify the
quality of features in community-joining behavior. Various evaluation techniques can be used, such as randomization tests discussed
in Chapter 8. In randomization tests, we measure a phenomenon in
a dataset and then randomly generate subsamples from the dataset
in which the phenomenon is guaranteed to be removed. We assume
the phenomenon has happened when the measurements on the subsamples are different from the ones on the original dataset. Another
approach is to use causality testing methods. Causality testing methods measure how a feature can affect a phenomenon. A well-known
causality detection technique is called granger causality due to Clive
W. J. Granger, the Nobel laureate in economics.
Definition 10.1. Granger Causality. Assume we are given two temporal
variables X = {X1 , X2 , . . . , Xt , Xt+1 , . . .} and Y = {Y1 , Y2 , . . . , Yt , Yt+1 , . . .}.
Variable X Granger causes variable Y when historical values of X can
help better predict Y than just using the historical values of Y.
325

Causality
Testing

Granger
Causality

Consider a linear regression model outlined in Chapter 5. We can

predict Yt+1 by using either Y1 , . . . , Yt or a combination of X1 , . . . , Xt
and Y1 , . . . , Yt .
Yt+1 =
Yt+1 =

t
X
i=1
t
X

ai Yi + 1 ,
ai Yi +

i=1

t
X

(10.1)
bi Xi + 2 ,

(10.2)

i=1

where 1 and 2 are the regression model errors. Now, if 2 < 1 , it

indicates that using X helps reduce the error. In this case, X Granger
causes Y.

10.1.2

Individual Behavior Modeling

Similar to network models, models of individual behavior can help concretely describe why specific individual behaviors are observed in social
media. In addition, they allow for controlled experiments and simulations
that can help study individuals in social media.
As with other modeling approaches (see Chapter 4), in behavior modeling, one must make a set of assumptions. Behavior modeling can be
performed via a variety of techniques, including those from economics,
game theory, or network science. We discussed some of these techniques
in earlier chapters. We review them briefly here, and refer interested readers to the respective chapters for more details.
Threshold models (Chapter 8). When a behavior diffuses in a network, such as the behavior of individuals buying a product and
referring it to others, one can use threshold models. In threshold
models, the parameters that need to be learned are the node activation threshold i and the influence probabilities wi j . Consider
the following methodology for learning these values. Consider a
merchandise store where the store knows the connections between
individuals and their transaction history (e.g., the items that they
have bought). Then, wi j can be defined as the
fraction of times user i buys a product and
user j buys the same product soon after that
326

The definition of soon requires clarification and can be set based on

a sites preference and the average time between friends buying the
same product. Similarly, i can be estimated by taking into account
the average number of friends who need to buy a product before user
i decides to buy it. Of course, this is only true when the products
bought by user i are also bought by her friends. When this is not
the case, methods from collaborative filtering (see Chapter 9) can be
used to find out the average number of similar items that are bought
by user is friends before user i decides to buy a product.
Cascade Models (Chapter 7). Cascade models are examples of
scenarios where an innovation, product, or information cascades
through a network. The discussion with respect to cascade models is similar, to the threshold models with the exception that cascade
models are sender-centric. That is, the sender decides to activate
the receiver, whereas threshold models are receiver-centric, in which
receivers get activated by multiple senders. Therefore, the computation of the ICM parameters needs to be done from the senders point
of view in cascade models. Note that both threshold and cascade
models are examples of individual behavior modeling.

10.1.3

Individual Behavior Prediction

As discussed previously, most behaviors result in newly formed links in

social media. It can be a link to a user, as in befriending behavior; a link
to an entity, as in buying behavior; or a link to a community, as in joining
behavior. Hence, one can formulate many of these behaviors as a link
prediction problem. Next, we discuss link prediction in social media.
Link Prediction
Link prediction assumes a graph G(V, E). Let e(u, v) E represent an
interaction (edge) between nodes u and v, and let t(e) denote the time
of the interaction. Let G[t1 , t2 ] represent the subgraph of G such that all
edges are created between t1 and t2 (i.e., for all edges e in this subgraph,
t1 < t(e) < t2 ). Now given four time stamps t1 < t01 < t2 < t02 , a link
prediction algorithm is given the subgraph G[t1 , t01 ] (training interval) and
is expected to predict edges in G[t2 , t02 ] (testing interval). Note that, just
327

like new edges, new nodes can be introduced in social networks; therefore,
G[t2 , t02 ] may contain nodes not present in G[t1 , t01 ]. Hence, a link prediction
algorithm is generally constrained to predict edges only for pairs of nodes
that are present during the training period. One can add extra constraints
such as predicting links only for nodes that are incident to at least k edges
(i.e., have degree greater or equal to k) during both testing and training
intervals.
Let G(Vtrain , Etrain ) be our training graph. Then, a link prediction algorithm generates a sorted list of most probable edges in Vtrain Vtrain Etrain .
The first edge in this list is the one the algorithm considers the most likely
to soon appear in the graph. The link prediction algorithm assigns a score
(x, y) to every edge e(x, y) in Vtrain Vtrain Etrain . Edges sorted by this value
in decreasing order will create our ranked list of predictions. (x, y) can be
predicted based on different techniques. Note that any similarity measure
between two nodes can be used for link prediction; therefore, methods
discussed in Chapter 3 are of practical use here. We outline some of the
most well-established techniques for computing (x, y) here.
Node Neighborhood-Based Methods
The following methods take advantage of neighborhood information to
compute the similarity between two nodes.
Common Neighbors. In this method, one assumes that the more
common neighbors that two nodes share, the more similar they are.
Let N(x) denote the set of neighbors of node x. This method is
formulated as
(x, y) = |N(x) N(y)|.
(10.3)
Jaccard Similarity. This commonly used measure calculates the likelihood of a node that is a neighbor of either x or y to be a common
neighbor. It can be formulated as the number of common neighbors
divided by the total number of neighbors of either x or y:
(x, y) =

|N(x) N(y)|
.
|N(x) N(y)|

(10.4)

Adamic and Adar Measure. A similar measure to Jaccard, this measure was introduced by Lada Adamic and Eytan Adar [2003]. The
328

Figure 10.5: Neighborhood-Based Link Prediction Example.

intuition behind it measure is that if two individuals share a neighbor
and that neighbor is a rare neighbor, it should have a higher impact
on their similarity. For instance, we can define the rareness of a node
based on its degree (i.e., the smaller the nodes degree, the higher
its rareness). The original version of the measure is defined based
on webpage features. A modified version based on neighborhood
information is
X
1
.
(10.5)
(x, y) =
log |N(z)|
zN(x)N(y)
Preferential Attachment. In the preferential attachment model discussed in Chapter 4, one assumes that nodes of higher degree have a
higher chance of getting connected to incoming nodes. Therefore, in
terms of connection probability, higher degree nodes are similar. The
preferential attachment measure is defined to capture this similarity:
(x, y) = |N(x)| |N(y)|.

(10.6)

Example 10.1. For the graph depicted in Figure 10.5, the similarity between
nodes 5 and 7 based on different neighborhood-based techniques is
(Common Neighbor) (5, 7) = |{4, 6} {4}| = 1
|{4, 6} {4}| 1
(Jaccard) (5, 7) =
=
|{4, 6} {4}| 2
1
1
=
(Adamic and Adar) (5, 7) =
log |{5, 6, 7}| log 3
(Preferential Attachment) (5, 7) = |{4}| |{4, 6}| = 1 2 = 2
329

(10.7)
(10.8)
(10.9)
(10.10)

Table 10.1: A Comparison between Link Prediction Methods

First edge
Second edge
Third edge
Common Neighbors
(6, 7) = 1
(1, 3) = 1
(5, 7) = 1
Jaccard Similarity
(1, 3) = 1
(6, 7) = 1/2
(5, 7) = 1/2
Adamic and Adar
(1, 3) = 1/ log 2 (6, 7) = 1/ log 3 (5, 7) = 1/ log 3
Preferential Attachment
(2, 4) = 6
(2, 5) = 4
(2, 6) = 4
In Figure 10.5, there are eight nodes; therefore, we can have a maximum of
= 28 edges. We already have six edges in the graph; hence, there are 286 = 22
other edges that are not in the graph. For all these edges, we can compute the
similarity between their endpoints using the aforementioned neighborhood-based
techniques and identify the top three most likely edges that are going to appear
in the graph based on each technique. Table 10.1 shows the top three edges based
on each technique and the corresponding values for each edge. As shown in this
table, different methods predict different edges to be most important; therefore, the
method of choice depends on the application.
8

2

Methods Based on Paths between Nodes

Similarity between nodes can simply be computed from the shortest path
distance between them. The closer the nodes are, the higher their similarity.
This similarity measure can be extended by considering multiple paths
between nodes and their neighbors. The following measures can be used
to calculate similarity.
Katz measure. Similar to the Katz centrality defined in Chapter 3,
one can define the similarity between nodes x and y as
(x, y) =

l |paths<l>
x,y |,

(10.11)

l=1

where |paths<l>
x,y | denotes the number of paths of length l between x and
y. is a constant that exponentially damps longer paths. Note that a
very small results in a common neighbor measure (see Exercises).
Similar to our finding in Chapter 3, one can find the Katz similarity
measure in a closed form by (I A)1 I. The Katz measure can also
330

be weighted or unweighted. In the unweighted format, |paths<1>

x,y | = 1
if there is an edge between x and y. The weighted version is more
suitable for multigraphs, where multiple edges can exist between the
same pair of nodes. For example, consider two authors x and y who
have collaborated c times. In this case, |paths<1>
x,y | = c.
Hitting and Commute Time. Consider a random walk that starts
at node x and moves to adjacent nodes uniformly. Hitting time Hx,y
is the expected number of random walk steps needed to reach y
starting from x. This is a distance measure. In fact, a smaller hitting
time implies a higher similarity; therefore, a negation can turn it into
a similarity measure:
(x, y) = Hx,y .
(10.12)
Note that if node y is highly connected to other nodes in the network
(i.e., has a high stationary probability y ), then a random walk starting from any x likely ends up visiting y early. Hence, all hitting times
to y are very short, and all nodes become similar to y. To account
for this, one can normalize hitting time by multiplying it with the
stationary probability y :
(x, y) = Hx,y y .

(10.13)

Hitting time is not symmetric, and in general, Hx,y , H y,x . Thus, one
can introduce the commute time to mitigate this issue:
(x, y) = (Hx,y + H y,x ).

(10.14)

Similarly, commute time can also be normalized,

(x, y) = (Hx,y y + H y,x x ).

(10.15)

Rooted PageRank. A modified version of the PageRank algorithm

can be used to measure similarity between two nodes x and y. In
rooted PageRank, we measure the stationary probability of y: y
given the condition that during each random walk step, we jump to
x with probability or a random neighbor with probability 1 .
The matrix format discussed in Chapter 3 can be used to solve this
problem.
331

 SimRank. One can define similarity between two nodes recursively

based on the similarity between their neighbors. In other words, similar nodes have similar neighbors. SimRank performs the following:
P
P
0
0
x0 N(x)
y0 N(y) (x , y )
,
(10.16)
(x, y) =
|N(x)||N(y)|
where is some value in range [0, 1]. We set (x, x) = 1, and by
finding the fixed point of this equation, we can find the similarity
between node x and node y
After one of the aforementioned measures is selected, a list of the top
most similar pairs of nodes are selected. These pairs of nodes denote edges
predicted to be the most likely to soon appear in the network. Performance
(precision, recall, or accuracy) can be evaluated using the testing graph and
by comparing the number of the testing graphs edges that the link prediction algorithm successfully reveals. Note that the performance is usually
very low, since many edges are created due to reasons not solely available
in a social network graph. So, a common baseline is to compare the performance with random edge predictors and report the factor improvements
over random prediction.

10.2

Collective Behavior

Collective behavior, first defined by sociologist Robert Park, refers to a

population of individuals behaving in a similar way. This similar behavior
can be planned and coordinated, but is often spontaneous and unplanned.
For instance, individuals stand in line for a new product release, rush
into stores for a sale event, and post messages online to support their
cause or show their support for an individual. These events, though
formed by independent individuals, are observed as a collective behavior
by outsiders.

10.2.1

Collective Behavior Analysis

Collective behavior analysis is often performed by analyzing individuals

performing the behavior. In other words, one can divide collective behavior into many individual behaviors and analyze them independently.
332

The result, however, when all these analyses are put together would be an
expected behavior for a large population. The user migration behavior we
discuss in this section is an example of this type of analysis of collective
behavior.
One can also analyze the population as a whole. In this case, an individuals opinion or behavior is rarely important. In general, the approach
is the same as analyzing an individual, with the difference that the content
and links are now considered for a large community. For instance, if we are
analyzing 1,000 nodes, one can combine these nodes and edges into one
hyper-node, where the hyper-node is connected to all other nodes in the
graph to which its members are connected and has an internal structure
(subgraph) that details the interaction among its members. This approach
is unpopular for analyzing collective behavior because it does not consider
specific individuals and at times, interactions within the population. Interested readers can refer to the bibliographic notes for further references
that use this approach to analyze collective behavior. On the contrary, this
approach is often considered when predicting collective behavior, which
is discussed later in this chapter.
User Migration in Social Media
Users often migrate from one site to another for different reasons. The main
rationale behind it is that users have to select some sites over others due
to their limited time and resources. Moreover, social medias networking
often dictates that one cannot freely choose a site to join or stay. An
individuals decision is heavily influenced by his or her friends, and vice
versa. Sites are often interested in keeping their users, because they are
valuable assets that help contribute to their growth and generate revenue
by increased traffic. There are two types of migration that take place in
social media sites: site migration and attention migration.
1. Site Migration. For any user who is a member of two sites s1 and s2
at time ti , and is only a member of s2 at time t j > ti , then the user is
said to have migrated from site s1 to site s2 .
2. Attention Migration. For any user who is a member of two sites s1
and s2 and is active at both at time ti , if the user becomes inactive on
s1 and remains active on s2 at time t j > ti , then the users attention is
said to have migrated away from site s1 and toward site s2 .
333

Activity (or inactivity) of a user can be determined by observing the

users actions performed on the site. For instance, we can consider a user
active in interval [ti , ti + ], if the user has performed at least one action on
the site during this interval. Otherwise, the user is considered inactive.
The interval could be measured at different granularity, such as days,
weeks, months, and years. It is common to set = 1 month. To analyze
the migration of populations across sites, we can analyze migrations of
individuals and then measure the rate at which the population of these
individuals is migrating across sites. Since this method analyzes migrations at the individual level, we can use the methodology outlined in
Section 10.1.1 for individual behavior analysis as follows.
The Observable Behavior
Site migration is rarely observed since users often abandon their accounts
rather than closing them. A more observable behavior is attention migration, which is clearly observable on most social media sites. Moreover,
when a user commits site migration, it is often too late to perform preventive measures. However, when attention migration is detected, it is
still possible to take actions to retain the user or expedite his or her attention migration to guarantee site migration. Thus, we focus on individuals
whose attention has migrated.
To observe attention migrations, several steps need to be taken. First,
users are required to be identified on multiple networks so that their activity on multiple sites can be monitored simultaneously. For instance,
username huan.liu1 on Facebook is username liuhuan on Twitter. This
identification can be done by collecting information from sites where individuals list their multiple identities on social media sites. On social
networking sites such as Google+ or Facebook, this happens regularly.
The second step is collecting multiple snapshots of social media sites. At
least two snapshots are required to observe migrations. After these two
steps, we can observe whether attention migrations have taken place or
not. In other words, we can observe if users have become inactive on one
of the sites over time. Figure 10.6 depicts these migrations for some wellknown social media sites. In this figure, each radar chart shows migrations
from a site to multiple sites. Each target site is shown as a vertex, and the
longer the spokes toward that site, the larger the migrating population to
it.
334

Figure 10.6: Pairwise Attention Migration among Social Media Sites.

Features
Three general features can be considered for user migration: (1) user activity
on one site, (2) user network size, and (3) user rank. User activity is important,
because we can conjecture that a more active user on one site is less likely
to migrate. User network size is important, because a user with more social
ties (i.e., friends) in a social network is less likely to move. Finally, user
rank is important. The rank is the value of a user as perceived by others.
A user with high status in a network is less likely to move to a new one
where he or she must spend more time getting established.
User activity can be measured differently for different sites. On Twitter,
it can be the number of tweets posted by the user; on Flickr, the number
photos uploaded by the user; and on YouTube, the number of videos the
335

user has uploaded. One can normalize this value by its maximum in the
site (e.g., the maximum number of videos any user has uploaded) to get
an activity measure in the range [0,1]. If a user is allowed to have multiple
activities on a site, as in posting comments and liking videos, then a linear
com- bination of these measures can be used to describe user activity on a
site.
User network size can be easily measured by taking the number of
friends a user has on the site. It is common for social media sites to
facilitate the addition of friends. The number of friends can be normalized
in the range [0,1] by the maximum number of friends one can have on the
site.
Finally, user rank is how important a user is on the site. Some sites
explicitly provide their users prestige rank list (e.g., top 100 bloggers),
whereas for others, one needs to approximate a users rank. One way
of approximating it is to count the number of citations (in-links) an individual is receiving from others. A practical technique is to perform this
via web search engines. For instance, user test on StumbleUpon has
http://test.stumpleupon.com as his profile page. A Google search for
link:http://test.stumbleupon.com provides us with the number of inlinks to the profile on StumbleUpon and can be considered as a ranking
measure for user test.
These three features are correlated with the site attention migration
behavior and one expects changes in them when migrations happen.

Feature-Behavior Association
Given two snapshots of a network, we know if users migrated or not. We
can also compute the values for the aforementioned features. Hence, we
can determine the correlation between features and migration behavior.
Let vector Y Rn indicate whether any of our n users have migrated or
not. Let Xt R3n be the features collected (activity, friends, rank) for any
one of these users at time stamp t. Then, the correlation between features
Xt and labels Y can be computed via logistic regression. How can we verify
that this correlation is not random? Next, we discuss how we verify that
this correlation is statistically significant.
336

Evaluation Strategy
To verify if the correlation between features and the migration behavior is
not random, we can construct a random set of migrating users and compute
XRandom and YRandom for them as well. This can be obtained by shuffling
the rows of the original Xt and Y. Then, we perform logistic regression
on these new variables. This approach is very similar to the shuffle test
presented in Chapter 8. The idea is that if some behavior creates a change
in features, then other random behaviors should not create that drastic a
change. So, the observed correlation between features and the behavior
should be significantly different in both cases. The correlation can be
described in terms of logistic regression coefficients, and the significance
can be measured via any significance testing methodology. For instance,
we can employ the 2 -statistic,
2 -statistic
n
X
(Ai Ri )2
,
=
R
i
i=1
2

(10.17)

where n is the number of logistic regression coefficients, Ai s are the coefficients determined using the original dataset, and Ri s are the coefficients
obtained from the random dataset.

10.2.2

Collective Behavior Modeling

Consider a hypothetical model that can simulate voters who cast ballots in
elections. This effective model can help predict an elections turnout rate
as an outcome of the collective behavior of voting and help governments
prepare logistics accordingly. This is an example of collective behavior
modeling, which improves our understanding of the collective behaviors
that take place by providing concrete explanations.
Collective behavior can be conveniently modeled using some of the
techniques discussed in Chapter 4, Network Models. Similar to collective behavior, in network models, we express models in terms of characteristics observable in the population. For instance, when a power-law degree
distribution is required, the preferential attachment model is preferred, and
when the small average shortest path is desired, the small-world model is
the method of choice. In network models, node properties rarely play a
role; therefore, they are reasonable for modeling collective behavior.
337

10.2.3

Collective Behavior Prediction

Collective behavior can be predicted using methods we discussed in Chapters 7 and 8. For instance, epidemics can predict the effect of a disease on
a population and the behavior that the population will exhibit over time.
Similarly, implicit influence models such as the LIM model discussed in
Chapter 8 can estimate the influence of individuals based on collective behavior attributes, such as the size of the population adopting an innovation
at any time.
As noted earlier, collective behavior can be analyzed either in terms of
individuals performing the collective behavior or based on the population
as a whole. When predicting collective behavior, it is more common to
consider the population as a whole and aim to predict some phenomenon.
This simplifies the challenges and reduces the computation dramatically,
since the number of individuals who perform a collective behavior is often
large and analyzing them one at a time is cumbersome.
In general, when predicting collective behavior, we are interested in
predicting the intensity of a phenomenon, which is due to the collective
behavior of the population (e.g., how many of them will vote?) To perform this prediction, we utilize a data mining approach where features
that describe the population well are used to predict a response variable
(i.e., the intensity of the phenomenon). A training-testing framework or
correlation analysis is used to determine the generalization and the accuracy of the predictions. We discuss this collective behavior prediction
strategy through the following example. This example demonstrates how
the collective behavior of individuals on social media can be utilized to
predict real-world outcomes.
Predicting Box Office Revenue for Movies
Can we predict opening-weekend revenue for a movie from its prerelease
chatter among fans? This tempting goal of predicting the future has been
around for many years. The goal is to predict the collective behavior
of watching a movie by a large population, which in turn determines the
revenue for the movie. One can design a methodology to predict box office
revenue for movies that uses Twitter and the aforementioned collective
behavior prediction strategy. To summarize, the strategy is as follows:
1. Set the target variable that is being predicted. In this case, it is the
338

revenue that a movie produces. Note that the revenue is the direct
result of the collective behavior of going to the theater to watch the
movie.
2. Determine the features in the population that may affect the target
variable.
3. Predict the target variable using a supervised learning approach,
utilizing the features determined in step 2.
4. Measure performance using supervised learning evaluation.
One can use the population that is discussing the movie on Twitter
before its release to predict its opening-weekend revenue. The target variable is the amount of revenue. In fact, utilizing only eight features, one
can predict the revenue with high accuracy. These features are the average
hourly number of tweets related to the movie for each of the seven days
prior to the movie opening (seven features) and the number of opening
theaters for the movie (one feature). Using only these eight features, training data for some movies (their seven-day tweet rates, their number of
opening theaters, and their revenue), and a linear regression model, one
can predict the movie opening-weekend revenue with high correlation. It
has been shown by researchers (see Bibliographic Notes) that the predictions using this approach are closer to reality than that of the Hollywood
Stock Exchange (HSX), which is the gold standard for predicting revenues
for movies.
This simple model for predicting movie revenue can be easily extended
to other domains. For instance, assume we are planning to predict another
collective behavior outcome, such as the number of individuals who aim
to buy a product. In this case, the target variable y is the number of
individuals who will buy the product. Similar to tweet rate, we require
some feature A that denotes the attention the product is receiving. We also
need to model the publicity of the product P. In our example, this was
the number of theaters for the movie; for a product, it could represent the
number of stores that sell it. A simple linear regression model can help
learn the relation between these features and the target variable:
y = w1 A + w2 P + ,

(10.18)

where  is the regression error. Similar to our movie example, one attempts
to extract the values for A and P from social media.
339

10.3

Summary

Individuals exhibit different behaviors in social media, which can be categorized into individual and collective behavior. Individual behavior
is the behavior that an individual targets toward (1) another individual
(individual-individual behavior), (2) an entity (individual-entity behavior), or
(3) a community (individual-community behavior). We discussed how to analyze and predict individual behavior. To analyze individual behavior,
there is a four-step procedure, outlined as a guideline. First, the behavior observed should be clearly observable on social media. Second, one
needs to design meaningful features that are correlated with the behavior
taking place in social media. The third step aims to find correlations and
relationships between features and the behavior. The final step is to verify
these relationships that are found. We discussed community joining as
an example of individual behavior. Modeling individual behavior can be
performed via cascade or threshold models. Behaviors commonly result in
interactions in the form of links; therefore, link prediction techniques are
highly efficient in predicting behavior. We discussed neighborhood-based
and path-based techniques for link prediction.
Collective behavior is when a group of individuals with or without
coordination act in an aligned manner. Collective behavior analysis is either done via individual behavior analysis and then averaged or analyzed
collectively. When analyzed collectively, one commonly looks at the general patterns of the population. We discussed user migrations in social
media as an example of collective behavior analysis. Modeling collective
behavior can be performed via network models, and prediction is possible
by using population properties to predict an outcome. Predicting movie
box-office revenues was given as an example, which uses population properties such as the rate at which individuals are tweeting to demonstrate
the effectiveness of this approach.
It is important to evaluate behavior analytics findings to ensure that
these finding are not due to externalities. We discussed causality testing, randomization tests, and supervised learning evaluation techniques
for evaluating behavior analytics findings. However, depending on the
context, researchers may need to devise other informative techniques to
ensure the validity of the outcomes.

340

10.4

Bibliographic Notes

In addition to methods discussed in this chapter, game theory and theories

from economics can be used to analyze human behavior [78]. Communityjoining behavior analysis was first introduced by Backstrom et al. [19]. The
approach discussed in this chapter is a brief summary of their approach
for analyzing community-joining behavior. Among other individual behaviors, tie formation is analyzed in detail. In [290], the authors analyze
tie formation behavior on Facebook and investigate how visual cues influence individuals with no prior interaction to form ties. The features
used are gender (i.e., male or female), and visual conditions (attractive,
nonattractive, and no photo). Their analyses show that individuals have a
tendency to connect to attractive opposite-sex individuals when no other
information is available. Analyzing individual information-sharing behavior helps understand how individuals disseminate information on social media. Gundecha et al. [114] analyze how the information-sharing
behavior of individuals results in vulnerabilties and how one can exploit
such vulnerabilities to secure user privacy on a social networking site. Finally, most social media mining research is dedicated to analyzing a single
site; however, users are often members of different sites and hence, current
studies need to be generalized to cover multiple sites. Zafarani and Liu
[311, 312] were the first to design methods that help connect user identities across social media sites using behavioral modeling. A study of user
tagging behavior across sites is available in [291].
General surveys on link prediction can be found in [4, 172, 8, 179]. Individual behavior prediction is an active area of research. Location prediction
is an active area of individual behavior analysis that has been widely studied over a long period in the realm of mobile computing. Researchers
analyze human mobility patterns to improve location prediction services,
thereby exploiting their potential power on various applications such as
mobile marketing [29, 27], traffic planning [33, 70], and even disaster relief [98, 94, 105, 95, 289, 26, 160]. Other general references can be found in
[20, 200, 263, 282, 247, 96, 97].
Kumar et al. [159] first analyzed migration in social media. Other
collective behavior analyses can be found in Leskovec et al. [170]. The
movie revenue prediction was first discussed by Asur and Huberman [18].
Another example of collective behavior prediction can be found in the work
of OConnor et al. [221], which proposed using Twitter data for opinion
341

polls. Their results are highly correlated with Gallup opinion polls for
presidential job approval. In [1], the authors analyzed collective social
media data and show that by carefully selecting data from social media,
it is possible to use social media as a lens to analyze and even predict
real-world events.

342

10.5

Exercises

Individual Behavior
1.

Name five real-world behaviors that are commonly difficult to

observe in social media (e.g., your daily schedule or where you
eat lunch are rarely available in social media).
Select one behavior that is most likely to leave traces online. Can
you think of a methodology for identifying that behavior using
these traces?

2. Consider the commenting under a blogpost behavior in social media. Follow the four steps of behavior analysis to analyze this behavior.
3. We emphasized selecting meaningful features for analyzing a behavior. Discuss a methodology to verify if the selected features carry
enough information with respect to the behavior being analyzed.
4. Correlation does not imply causality. Discuss how this fact relates to
most of the datasets discussed in this chapter being temporal.
5. Using a neighborhood-based link prediction method compute the
top two most likely edges for the following figure.

6. Compute the most likely edge for the following figure for each pathbased link prediction technique.

343

7. In a link prediction problem, show that for small , the Katz similarity measure ((u, v) =
` |path<`>
u,v |) becomes Common neighbors
`=1
((u, v) = |N(u) N(v)|).
8. Provide the matrix format for rooted PageRank and SimRank techniques.

Collective Behavior
9. Recent research has shown that social media can help replicate survey
results for elections and ultimately predict presidential election outcomes. Discuss what possible features can help predict a presidential
election.

344

Bibliography

[1] Mohammad Ali Abbasi, Sun-Ki Chai, Huan Liu, and Kiran. Sagoo,
Real-world behavior analysis through a social media lens., Social Computing, Behavioral-Cultural Modeling and Prediction, Springer, 2012,
pp. 1826.
[2] J. Abello, M. Resende, and S. Sudarsky, Massive quasi-clique detection.,
LATIN 2002: Theoretical Informatics (2002), 598612.
[3] E. Abrahamson and L. Rosenkopf, Institutional and competitive bandwagons: Using mathematical modeling as a tool to explore innovation
diffusion., Academy of management review (1993), 487517.
[4] L.A. Adamic and E. Adar, Friends and neighbors on the web., Social
Networks 25 (2003), no. 3, 211230.
[5] Gediminas Adomavicius and Alexander. Tuzhilin, Toward the next
generation of recommender systems: a survey of the state-of-the-art and
possible extensions., IEEE Transactions on Knowledge and Data Engineering 17 (2005), no. 6, 734749.
[6] N. Agarwal, H. Liu, L. Tang, and P.S. Yu, Identifying the influential
bloggers in a community, Proceedings of the International Conference
on Web Search and Web Data Mining, ACM, 2008, pp. 207218.
[7] R.K. Ahuja, T.L. Magnanti, J.B. Orlin, and K. Weihe, Network flows:
theory, algorithms and applications., ZOR-Methods and Models of Operations Research 41 (1995), no. 3, 252254.
345

[8] Mohammad Al Hasan, Vineet Chaoji, Saeed Salem, and Mohammed.

Zaki, Link prediction using supervised learning., SDM06: Workshop on
Link Analysis, Counter-terrorism and Security, 2006.
[9] R. Albert and A.L. Barabasi, Topology of evolving networks: local events
and universality., Physical review letters 85 (2000), no. 24, 52345237.
[10] A. Anagnostopoulos, R. Kumar, and M. Mahdian, Influence and correlation in social networks, Proceedings of the 14th ACM SIGKDD
international conference on Knowledge Discovery and Data Mining,
ACM, 2008, pp. 715.
[11] L.R. Anderson and C.A. Holt, Classroom games: information cascades.,
The Journal of Economic Perspectives 10 (1996), no. 4, 187193.
[12]

, Information cascades in the laboratory., The American Economic Review (1997), 847862.

[13] Roy M Anderson, Robert M May, and B Anderson, Infectious diseases

of humans: dynamics and control., vol. 26, Wiley Online Library, 1992.
[14] M. Ankerst, M.M. Breunig, H.P. Kriegel, and J. Sander, OPTICS:
ordering points to identify the clustering structure., Proceedings of the
1999 ACM SIGMOD international conference on Management of
Data (1999), 4960.
[15] S. Aral, L. Muchnik, and A. Sundararajan, Distinguishing influencebased contagion from homophily-driven diffusion in dynamic networks.,
Proceedings of the National Academy of Sciences 106 (2009), no. 51,
2154421549.
[16] Jaime Arguello, Jonathan Elsas, Jamie Callan, and Jaime. Carbonell,
Document representation and query expansion models for blog recommendation., Proceedings of the second international conference on Weblogs and Social Media (ICWSM), 2008.
[17] S.E. Asch, Studies of independence and conformity: I. a minority of one
against a unanimous majority., Psychological Monographs: General
and Applied 70 (1956), no. 9, 170.
346

[18] Sitaram Asur and Bernardo A. Huberman, Predicting the future with
social media. ieee international conference on, Web Intelligence and Intelligent Agent Technology, vol. 1, IEEE, 2010, pp. 492499.
[19] L. Backstrom, D. Huttenlocher, J.M. Kleinberg, and X. Lan, Group
formation in large social networks: membership, growth, and evolution,
Proceedings of the 12th ACM SIGKDD international conference on
Knowledge Discovery and Data Mining, ACM, 2006, pp. 4454.
[20] L. Backstrom, E. Sun, and C. Marlow, Find me if you can: improving
geographical prediction with social and spatial proximity., Proceedings of
the 19th International conference on World Wide Web, ACM, 2010,
pp. 6170.
[21] N.T.J. Bailey, The mathematical theory of infectious diseases and its applications., Charles Griffin & Company Ltd, 1975.
[22] E. Bakshy, J.M. Hofman, W.A. Mason, and D.J. Watts, Everyones an
influencer: quantifying influence on twitter, Proceedings of the fourth
ACM international conference on Web Search and Data Mining,
ACM, 2011, pp. 6574.
[23] A.V. Banerjee, A simple model of herd behavior., The Quarterly Journal
of Economics 107 (1992), no. 3, 797817.
[24] A.L. Barabasi and R. Albert, Emergence of scaling in random networks.,
science 286 (1999), no. 5439, 509512.
[25] Geoffrey Barbier, Zhuo Feng, Pritam Gundecha, and Huan. Liu, Morgan & Claypool Publishers, 2013.
[26] Geoffrey Barbier, Reza Zafarani, Huiji Gao, Gabriel Fung, and Huan.
Liu, Maximizing benefits from crowdsourced data., Computational and
Mathematical Organization Theory 18 (2012), no. 3, 257279.
[27] S.J. Barnes and E. Scornavacca, Mobile marketing: the role of permission
and acceptance., International Journal of Mobile Communications 2
(2004), no. 2, 128139.
[28] Alain Barrat, Marc Barthelemy, and Alessandro. Vespignani, Dynamical processes on complex networks., vol. 1, Cambridge University
Press, 2008.
347

[29] P. Barwise and C. Strong, Permission-based mobile advertising., Journal

of Interactive Marketing 16 (2002), no. 1, 1424.
[30] F. Bass, A new product growth model for product diffusion., Management
Science 15 (1969), 215227.
[31] W.G. Bell, The great plague in london in 1665., The Great Plague in
London in 1665. (1924).
[32] R. Bellma, On a routing problem., NOTES 16 (1956), no. 1.
[33] M. Ben-Akiva, M. Bierlaire, H. Koutsopoulos, and R. Mishalani, Dynamit: a simulation-based system for traffic prediction., DACCORS Short
Term Forecasting Workshop, The Netherlands, 1998.
[34] E. Berger, Dynamic monopolies of constant size., Journal of Combinatorial Theory, Series B 83 (2001), no. 2, 191200.
[35] P. Berkhin, A survey of clustering data mining techniques., Grouping
Multidimensional Data (2006), no. c, 2571.
[36] H. Russell. Bernard, Social research methods: qualitative and quantitative
approaches., Sage, 2012.
[37] S. Bikhchandani, D. Hirshleifer, and I. Welch, A theory of fads, fashion, custom, and cultural change as informational cascades., Journal of
Political Economy (1992), 9921026.
[38] S. Bikhchandani and S. Sharma, Herd behavior in financial markets.,
IMF Staff Papers (2000), 279310.
[39] C.M. Bishop, Neural networks for pattern recognition., (1995).
, Pattern recognition and machine learning., vol. 4, Springer,

[40]
2006.

[41] B. Bollobas, Random graphs., vol. 73, Cambridge University Press,

2001.
[42] E. Bonabeau, M. Dorigo, and G. Theraulaz, Swarm intelligence: from
natural to artificial systems., no. 1, Oxford University Press, 1999.
348

[43] J.A. Bondy and U.S.R. Murty, Graph theory with applications., vol. 290,
MacMillan London, 1976.
[44] Stephen Poythress Boyd and Lieven. Vandenberghe, Convex optimization., Cambridge University Press, 2004.
[45] Ulrik. Brandes, A faster algorithm for betweenness centrality., Journal of
Mathematical Sociology 25 (2001), no. 2, 163177.
[46] Andrei Broder, Ravi Kumar, Farzin Maghoul, Prabhakar Raghavan,
Sridhar Rajagopalan, Raymie Stata, Andrew Tomkins, and Janet.
Wiener, Graph structure in the web., Computer networks 33 (2000),
no. 1, 309320.
[47] Alan. Bryman, Social research methods., Oxford University Press, 2012.
[48] Robin. Burke, Hybrid recommender systems: Survey and experiments.,
User Modeling and User-Adapted Interaction 12 (2002), no. 4, 331
370.
[49] K. Selcuk Candan and Maria Luisa. Sapino, Data management for
multimedia retrieval., Cambridge University Press, 2010.
[50] M. Cha, H. Haddadi, F. Benevenuto, and K.P. Gummadi, Measuring
user influence in twitter: The million follower fallacy, AAAI Conference
on Weblogs and Social Media, vol. 14, 2010, p. 8.
[51] Soumen. Chakrabarti, Mining the Web: discovering knowledge from
hypertext data., Morgan Kaufmann, 2003.
[52] Jilin Chen, Werner Geyer, Casey Dugan, Michael Muller, and Ido.
Guy, Make new friends, but keep the old: recommending people on social
networking sites., Proceedings of the SIGCHI Conference on Human
Factors in Computing Systems, ACM, 2009, pp. 201210.
[53] S. Chinese and et al., Molecular evolution of the sars coronavirus during
the course of the sars epidemic in china., Science 303 (2004), no. 5664,
1666.
[54] N.A. Christakis and J.H. Fowler, The spread of obesity in a large social
network over 32 years., New England Journal of Medicine 357 (2007),
no. 4, 370379.
349

[55] Nicholas A. Christakis and James H. Fowler, Connected: The surprising power of our social networks and how they shape our lives., Norsk
epidemiologi= Norwegian Journal of Epidemiology 19 (2009), no. 1,
5.
[56] F.R.K. Chung, Spectral graph theory., no. 92, American Mathematical
Society, 1997.
[57] R. B. Cialdini and M. R. Trost., Social influence: Social norms, conformity
and compliance., (1998), 151.
[58] Aaron Clauset, Cosma Rohilla Shalizi, and Mark EJ. Newman, Powerlaw distributions in empirical data., SIAM Review 51 (2009), no. 4, 661
703.
[59] J.S. Coleman, E. Katz, and H. Menzel, Medical innovation: a diffusion
study., Bobbs-Merrill Company, 1966.
[60] R. Cont and J.P. Bouchaud, Herd behavior and aggregate fluctuations in
financial markets., Macroeconomic Dynamics 4 (2000), no. 02, 170196.
[61] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest, and
Clifford. Stein, Introduction to algorithms., 2009.
[62] S. Currarini, M.O. Jackson, and P. Pin, An economic model of friendship:
Homophily, minorities, and segregation., Econometrica 77 (2009), no. 4,
10031045.
[63] Abhinandan S. Das, Mayur Datar, Ashutosh Garg, and Shyam. Rajaram, Google news personalization: scalable online collaborative filtering.,
Proceedings of the 16th international conference on World Wide Web,
ACM, 2007, pp. 271280.
[64] Manoranjan Dash and Huan. Liu, Feature selection for classification.,
Intelligent Data Analysis 1 (1997), no. 3, 131156.
[65]

, Feature selection for clustering., Knowledge Discovery and

Data Mining. Current Issues and New Applications, Springer, 2000,
pp. 110121.
350

[66] James Davidson, Benjamin Liebald, Junning Liu, Palash Nandy, Taylor Van Vleet, Ullas Gargi, Sujoy Gupta, Yu He, Mike Lambert, Blake
Livingston, and et al., The youtube video recommendation system., Proceedings of the fourth ACM conference on Recommender Systems,
ACM, 2010, pp. 293296.
[67] D.L. Davies and D.W. Bouldin, A cluster separation measure., IEEE
Transactions on Pattern Analysis and Machine Intelligence (1979),
no. 2, 224227.
[68] D.C. Des Jarlais, S.R. Friedman, J.L. Sotheran, J. Wenston, M. Yancovitz Marmor, Frank S.R., Beatrice B., and Mildvan D. S., Continuity
and change within an hiv epidemic., JAMA: the journal of the American
Medical Association 271 (1994), no. 2, 121127.
[69] A. Devenow and I. Welch, Rational herding in financial economics.,
European Economic Review 40 (1996), no. 3, 603615.
[70] H. Dia, An object-oriented neural network approach to short-term traffic
forecasting., European Journal of Operational Research 131 (2001),
no. 2, 253261.
[71] R. Diestel, Graph theory. 2005., Graduate Texts in Math (2005).
[72] K. Dietz, Epidemics and rumours: a survey., Journal of the Royal Statistical Society. Series A (General) (1967), 505528.
[73] P.S. Dodds and D.J. Watts, Universal behavior in a generalized model of
contagion., Physical Review Letters 92 (2004), no. 21, 218701.
[74] M. Drehmann, J. Oechssler, and A. Roider, Herding and contrarian
behavior in financial markets an internet experiment., (2005).
[75] Richard O. Duda, Peter E. Hart, and David G. Stork, Pattern classification., Wiley-interscience, 2012.
[76] J.C. Dunn, Well-separated clusters and optimal fuzzy partitions., Journal
of cybernetics 4 (1974), no. 1, 95104.
[77] C. Dye and N. Gay, Modeling the sars epidemic., Science 300 (2003),
no. 5627, 18841885.
351

[78] D. Easley and J.M. Kleinberg, Networks, crowds, and markets., Cambridge University Press, 2010.
[79] R.C. Eberhart, Y. Shi, and J. Kennedy, Swarm intelligence., Elsevier,
2001.
[80] J. Edmonds and R.M. Karp, Theoretical improvements in algorithmic
efficiency for network flow problems., Journal of the ACM (JACM) 19
(1972), no. 2, 248264.
[81] Nicole B. Ellison and et al., Social network sites: definition, history,
and scholarship., Journal of Computer-Mediated Communication 13
(2007), no. 1, 210230.
[82] A.P. Engelbrecht, Fundamentals of computational swarm intelligence.,
Recherche 67 (2005), 02.
[83] P. Erdos and A. Renyi, On random graphs., Publicationes Mathematicae Debrecen 6 (1959), 290297.
[84]

, On the evolution of random graphs., Akad. Kiado,

1960.

[85]

, On the strength of connectedness of a random graph., Acta Mathematica Hungarica 12 (1961), no. 1, 261267.

[86] M. Ester, H.P. Kriegel, J. Sander, and X. Xu, A density-based algorithm

for discovering clusters in large spatial databases with noise., Proceedings
of the second international conference on Knowledge Discovery and
Data Mining, Portland, OR, AAAI Press (1996), 226231.
[87] M. Faloutsos, P. Faloutsos, and C. Faloutsos, On power-law relationships of the internet topology., ACM SIGCOMM Computer Communication Review, vol. 29, ACM, 1999, pp. 251262.
[88] D. Fisher, Improving inference through conceptual clustering., Proceedings of the 1987 AAAI conference (1987), 461465.
[89] R.W. Floyd, Algorithm 97: shortest path., Communications of the ACM
5 (1962), no. 6, 345.
[90] L.R. Ford and D.R. Fulkerson, Maximal flow through a network., Canadian Journal of Mathematics 8 (1956), no. 3, 399404.
352

[91] S. Fortunato, Community detection in graphs., Physics Reports 486

(2010), no. 3-5, 75174.
[92] J. Friedman, T. Hastie, and R. Tibshirani, The elements of statistical
learning., vol. 1, Springer Series in Statistics, 2001.
[93] D. Gale, What have we learned from social learning?, European Economic Review 40 (1996), no. 3, 617628.
[94] H. Gao, G. Barbier, and R. Goolsby, Harnessing the crowdsourcing
power of social media for disaster relief., Intelligent Systems, IEEE 26
(2011), no. 3, 1014.
[95] H. Gao, J. Tang, and H. Liu, Exploring social-historical ties on locationbased social networks., Proceedings of the Sixth International Conference on Weblogs and Social Media, 2012.
[96]

, Mobile location prediction in spatio-temporal context., Nokia

Mobile Data Challenge Workshop (2012).

[97] Huiji Gao, Jiliang Tang, and Huan. Liu, gscorr: modeling geo-social
correlations for new check-ins on location-based social networks., Proceedings of the 21st ACM international conference on Information and
Knowledge Management, ACM, 2012, pp. 15821586.
[98] Huiji Gao, Xufei Wang, Geoffrey Barbier, and Huan. Liu, Promoting coordination for disaster relief from crowdsourcing to coordination.,
Social Computing, Behavioral-Cultural Modeling and Prediction,
Springer, 2011, pp. 197204.
[99] D. Gibson, R. Kumar, and A. Tomkins, Discovering large dense subgraphs in massive graphs, Discovering large dense subgraphs in massive graphs., VLDB Endowment, 2005, pp. 721732.
[100] E.N. Gilbert, Random graphs., The Annals of Mathematical Statistics
30 (1959), no. 4, 11411144.
[101] M. Girvan and M.E.J. Newman, Community structure in social and
biological networks., Proceedings of the National Academy of Sciences
99 (2002), no. 12, 7821.
353

[102] Jennifer Golbeck and James. Hendler, Filmtrust: movie recommendations using trust in web-based social networks., Proceedings of the IEEE
Consumer Communications and Networking Conference, vol. 96,
Citeseer, 2006.
[103] A.V. Goldberg and R.E. Tarjan, A new approach to the maximum-flow
problem., Journal of the ACM (JACM) 35 (1988), no. 4, 921940.
[104] B. Golub and M.O. Jackson, Naive learning in social networks and the
wisdom of crowds., American Economic Journal: Microeconomics 2
(2010), no. 1, 112149.
[105] M.F. Goodchild and J.A. Glennon, Crowdsourcing geographic information for disaster response: a research frontier., International Journal of
Digital Earth 3 (2010), no. 3, 231241.
[106] L. Goodman and W. Kruskal, Measures of associations for crossvalidations., Journal of the American Statistical Association 49, 732
764.
[107] A. Goyal, F. Bonchi, and L.V.S. Lakshmanan, Learning influence probabilities in social networks, Proceedings of the Third ACM international
conference on Web Search and Data Mining, ACM, 2010, pp. 241250.
[108] M. Granovetter, Threshold models of collective behavior., American Journal of Sociology (1978), 14201443.
[109] M.S. Granovetter, The strength of weak ties., American Journal of Sociology (1973), 13601380.
[110] V. Gray, Innovation in the states: a diffusion study., The American
Political Science Review 67 (1973), no. 4, 11741185.
[111] Z. Griliches, Hybrid corn: an exploration in the economics of technological
change., Econometrica, Journal of the Econometric Society (1957),
501522.
[112] D. Gruhl, R. Guha, D. Liben-Nowell, and A. Tomkins, Information
diffusion through blogspace, Proceedings of the 13th international conference on the World Wide Web, ACM, 2004, pp. 491501.
354

[113] Y. Guan, H. Chen, K.S. Li, S. Riley, G.M. Leung, R. Webster, J.S.M.
Peiris, and K.Y. Yuen, A model to control the epidemic of h5n1 influenza
at the source., BMC Infectious Diseases 7 (2007), no. 1, 132.
[114] Pritam Gundecha, Geoffrey Barbier, and Huan. Liu, Exploiting Vulnerability to Secure User Privacy on a Social Networking Site., Proceedings of the 17th ACM SIGKDD international conference on Knowledge Discovery and Data Mining, KDD, 2011, pp. 511519.
[115] Pritam Gundecha and Huan. Liu, Mining social media: a brief introduction., Tutorials in Operations Research 1 (2012), no. 4.
[116] Ido Guy, Naama Zwerdling, Inbal Ronen, David Carmel, and Erel.
Uziel, Social media recommendation based on people and tags., Proceedings of the 33rd international ACM SIGIR conference on Research
and Development in Information Retrieval, ACM, 2010, pp. 194201.
[117] Isabelle. Guyon, Feature extraction: foundations and applications., vol.
207, Springer, 2006.
[118] T. Hagerstrand and et al., Innovation diffusion as a spatial process.,
Innovation diffusion as a spatial process. (1968).
[119] R.L. Hamblin, R.B. Jacobsen, and J.L.L. Miller, A mathematical theory
of social change., Wiley, 1973.
[120] Jiawei Han, Micheline Kamber, and Jian. Pei, Data mining: concepts
and techniques., Morgan Kaufmann, 2006.
[121] M.S. Handcock, A.E. Raftery, and J.M. Tantrum, Model-based clustering for social networks., Journal of the Royal Statistical Society: Series
A (Statistics in Society) 170 (2007), no. 2, 301354.
[122] P.E. Hart, N.J. Nilsson, and B. Raphael, A formal basis for the heuristic
determination of minimum cost paths., Systems Science and Cybernetics, IEEE Transactions on 4 (1968), no. 2, 100107.
[123] Simon. Haykin, Neural networks: a comprehensive foundation., Prentice
Hall, 1994.
[124] H.W. Hethcote, A thousand and one epidemic models., Lecture Notes in
Biomathematics (1994), 504504.
355

[125]

, The mathematics of infectious diseases., SIAM review (2000),

599653.

[126] H.W. Hethcote, H.W. Stech, and P. van den Driessche, Periodicity
and stability in epidemic models: a survey., Differential Equations and
Applications in Ecology, Epidemics and Population Problems (SN
Busenberg and KL Cooke, eds.) (1981), 6582.
[127] E.C. Hirschman, Innovativeness, novelty seeking, and consumer creativity., Journal of Consumer Research (1980), 283295.
[128] D. Hirshleifer, Informational cascades and social conventions., University
of Michigan Business School Working Paper No. 9705-10 (1997).
[129] P.D. Hoff, A.E. Raftery, and M.S. Handcock, Latent space approaches to
social network analysis., Journal of the American Statistical Association
97 (2002), no. 460, 10901098.
[130] J. Hopcroft and R. Tarjan, Algorithm 447: efficient algorithms for graph
manipulation., Communications of the ACM 16 (1973), no. 6, 372378.
[131] Xia Hu, Jiliang Tang, Huiji Gao, and Huan. Liu, Unsupervised sentiment analysis with emotional signals., Proceedings of the 22nd international conference on World Wide Web, WWW13, ACM, 2013.
[132] Xia Hu, Lei Tang, Jiliang Tang, and Huan. Liu, Exploiting social relations for sentiment analysis in microblogging., Proceedings of the sixth
ACM international conference on Web Search and Data Mining, 2013.
[133] P. Jaccard, Distribution de la Flore Alpine: dans le Bassin des dranses et
dans quelques regions voisines., Rouge, 1901.
[134] M.O. Jackson, Social and economic networks., Princeton University
Press, 2008.
[135] A.K. Jain and R.C. Dubes, Algorithms for clustering data., PrenticeHall, 1988.
[136] A.K. Jain, M.N. Murty, and P.J. Flynn, Data clustering: a review., ACM
Computing Surveys (CSUR) 31 (1999), no. 3, 264323.
356

[137] Anthony Jameson and Barry. Smyth, Recommendation to groups., The

Adaptive Web, Springer, 2007, pp. 596627.
[138] Dietmar Jannach, Markus Zanker, Alexander Felfernig, and Gerhard. Friedrich, Recommender systems: an introduction., Cambridge
University Press, 2010.
[139] T.R. Jensen and B. Toft, Graph coloring problems., vol. 39, WileyInterscience, 1995.
[140] Charles. Kadushin, Understanding Social Networks: theories, concepts,
and findings: theories, concepts, and findings., Oxford University Press,
USA, 2012.
[141] Andreas M. Kaplan and Michael. Haenlein, Users of the world, unite!
the challenges and opportunities of social media., Business horizons 53
(2010), no. 1, 5968.
[142] G. Karypis, E.H. Han, and V. Kumar, Chameleon: hierarchical clustering
using dynamic modeling., Computer 32 (1999), no. 8, 6875.
[143] E. Katz and P.F. Lazarsfeld, Personal influence: the part played by people
in the flow of mass communications., Transaction Pub, 2006.
[144] M.J. Keeling and K.T.D. Eames, Networks and epidemic models., Journal
of the Royal Society Interface 2 (2005), no. 4, 295307.
[145] E. Keller and J. Berry, The influentials: One American in ten tells the
other nine how to vote, where to eat, and what to buy., Free Press, 2003.
Tardos, Maximizing the spread of
[146] D. Kempe, J.M. Kleinberg, and E.
influence through a social network, Proceedings of the ninth ACM
SIGKDD international conference on Knowledge Discovery and
Data Mining, ACM, 2003, pp. 137146.
[147] J. Kennedy, Swarm intelligence., Handbook of Nature-Inspired and
Innovative Computing (2006), 187219.
[148] W.O. Kermack and A.G. McKendrick, Contributions to the mathematical theory of epidemics. ii. the problem of endemicity., Proceedings of the
Royal society of London. Series A 138 (1932), no. 834, 5583.
357

[149] Jan H. Kietzmann, Kristopher Hermkens, Ian P. McCarthy, and

Bruno S. Silvestre, Social media? get serious! understanding the functional building blocks of social media., Business Horizons 54 (2011), no. 3,
241251.
[150] J.M. Kleinberg, Authoritative sources in a hyperlinked environment.,
Journal of the ACM (JACM) 46 (1999), no. 5, 604632.

[151] Jon Kleinberg and Eva.

Tardos, Algorithm design. addison wesley.,
(2006).
[152] Ioannis Konstas, Vassilios Stathopoulos, and Joemon M. Jose, On
social networks and collaborative recommendation., Proceedings of the
32nd international ACM SIGIR conference on Research and Development in Information Retrieval, ACM, 2009, pp. 195202.
[153] R. Kosala and H. Blockeel, Web mining research: a survey., ACM
Sigkdd Explorations Newsletter 2 (2000), no. 1, 115.
[154] G. Kossinets and D.J. Watts, Empirical analysis of an evolving social
network., Science 311 (2006), no. 5757, 88.
[155] P.L. Krapivsky, S. Redner, and F. Leyvraz, Connectivity of growing
random networks., Physical review letters 85 (2000), no. 21, 46294632.
[156] J.B. Kruskal, On the shortest spanning subtree of a graph and the traveling
salesman problem., Proceedings of the American Mathematical Society
7 (1956), no. 1, 4850.
[157] R. Kumar, J. Novak, and A. Tomkins, Structure and evolution of online
social networks., Link Mining: Models, Algorithms, and Applications
(2010), 337357.
[158] R. Kumar, P. Raghavan, S. Rajagopalan, and A. Tomkins, Trawling the
web for emerging cyber-communities., Computer networks 31 (1999),
no. 11-16, 14811493.
[159] S. Kumar, R. Zafarani, and H. Liu, Understanding user migration patterns in social media, 25th AAAI Conference on Artificial Intelligence,
2011.
358

[160] Shamanth Kumar, Fred Morstatter, Reza Zafarani, and Huan. Liu,
Whom should i follow? identifying relevant users during crises., Proceedings of the 24th ACM Conference on Hypertext and Social Media,
2013.
[161] T. La Fond and J. Neville, Randomization tests for distinguishing social
influence and homophily effects, Proceedings of the 19th international
conference on the World Wide Web, ACM, 2010, pp. 601610.
[162] A. Lancichinetti and S. Fortunato, Community detection algorithms: a
comparative analysis., Physical Review E 80 (2009), no. 5, 056117.
[163] P. Langley, Elements of machine learning., Morgan Kaufmann, 1996.
[164] Charles L. Lawson and Richard J. Hanson, Solving least squares problems, vol. 15, SIAM, 1995.
[165] H. Leibenstein, Bandwagon, snob, and veblen effects in the theory of
consumers demand., The Quarterly Journal of Economics 64 (1950),
no. 2, 183207.
[166] E.A. Leicht, P. Holme, and M.E.J. Newman, Vertex similarity in networks., Physical Review E 73 (2006), no. 2, 026120.
[167] J. Leskovec, J.M. Kleinberg, and C. Faloutsos, Graphs over time: densification laws, shrinking diameters and possible explanations, Proceedings
of the 11th ACM SIGKDD international conference on Knowledge
Discovery in Data Mining, ACM, 2005, pp. 177187.
[168] J. Leskovec, K.J. Lang, and M. Mahoney, Empirical comparison of algorithms for network community detection, Proceedings of the 19th international conference on the World Wide Web, ACM, 2010, pp. 631640.
[169] J. Leskovec, M. McGlohon, C. Faloutsos, N. Glance, and M. Hurst,
Cascading behavior in large blog graphs., Arxiv preprint arXiv:0704.2803
(2007).
[170] Jure Leskovec, Lars Backstrom, and Jon. Kleinberg, Meme-tracking
and the dynamics of the news cycle., Proceedings of the 15th ACM
SIGKDD international conference on Knowledge Discovery and
Data Mining, ACM, 2009, pp. 497506.
359

[171] T.G. Lewis, Network Science: theory and Applications., Wiley Publishing, 2009.
[172] D. Liben-Nowell and J.M. Kleinberg, The link-prediction problem for social networks., Journal of the American society for information science
and technology 58 (2007), no. 7, 10191031.
[173] Katri Lietsala and Esa. Sirkkunen, Social media. introduction to the tools
and processes of participatory economy, tampere university., (2008).
[174] B. Liu, Web data mining: exploring hyperlinks, contents, and usage data.,
Springer Verlag, 2007.
[175] Huan Liu and Hiroshi. Motoda, Feature extraction, construction and
selection: a data mining perspective., Springer, 1998.
[176] Huan Liu and Lei. Yu, Toward integrating feature selection algorithms
for classification and clustering., IEEE Transactions on Knowledge and
Data Engineering 17 (2005), no. 4, 491502.
[177] Jiahui Liu, Peter Dolan, and Elin Rnby. Pedersen, Personalized news
recommendation based on click behavior., Proceedings of the 15th international conference on Intelligent User Interfaces, ACM, 2010,
pp. 3140.
[178] F. Lorrain and H.C. White, Structural equivalence of individuals in social
networks., Journal of Mathematical Sociology 1 (1971), no. 1, 4980.
[179] Linyuan Lu and Tao. Zhou, Link prediction in complex networks: a
survey., Physica A: Statistical Mechanics and its Applications 390
(2011), no. 6, 11501170.
[180] Hao Ma, Michael R. Lyu, and Irwin. King, Learning to recommend
with trust and distrust relationships., Proceedings of the third ACM
conference on Recommender Systems, ACM, 2009, pp. 189196.
[181] Hao Ma, Haixuan Yang, Michael R. Lyu, and Irwin. King, Sorec:
social recommendation using probabilistic matrix factorization., Proceedings of the 17th ACM conference on Information and Knowledge
Management, ACM, 2008, pp. 931940.
360

[182] Hao Ma, Dengyong Zhou, Chao Liu, Michael R. Lyu, and Irwin.
King, Recommender systems with social regularization., Proceedings of
the fourth ACM international conference on Web Search and Data
Mining, ACM, 2011, pp. 287296.
[183] M.W. Macy, Chains of cooperation: threshold effects in collective action.,
American Sociological Review (1991), 730747.
[184] M.W. Macy and R. Willer, From factors to actors: computational sociology
and agent-based modeling., Annual Review of Sociology (2002), 143
166.
[185] V. Mahajan, Models for innovation diffusion., no. 48, Sage Publications,
1985.
[186] V. Mahajan and E. Muller, Innovative behavior and repeat purchase diffusion models, Proceedings of the American Marketing Educators Conference, vol. 456, 1982, p. 460.
[187] V. Mahajan and R.A. Peterson, Innovation diffusion in a dynamic potential adopter population., Management Science (1978), 15891597.
[188] E. Mansfield, Technical change and the rate of imitation., Econometrica:
Journal of the Econometric Society (1961), 741766.
[189] J.P. Martino, Technological forecasting for decision making., McGrawHill, 1993.
[190] Paolo Massa and Paolo. Avesani, Trust-aware collaborative filtering for
recommender systems., On the Move to Meaningful Internet Systems
2004: CoopIS, DOA, and ODBASE, Springer, 2004, pp. 492508.
[191] B.D. McKay, Practical graph isomorphism, Practical Graph Isomorphism., vol. 1, Utilitas Mathematica, 1981, p. 45.
[192] M. McPherson, L. Smith-Lovin, and J.M. Cook, Birds of a feather:
homophily in social networks., Annual Review of Sociology (2001),
415444.
[193] D.F. Midgley and G.R. Dowling, Innovativeness: the concept and its
measurement., Journal of Consumer Research (1978), 229242.
361

[194] S. Milgram, Obedience to authority: an experimental view., Harper

Perennial Modern Classics, 2009.
[195] S. Milgram, L. Bickman, and L. Berkowitz, Note on the drawing power of
crowds of different size., Journal of Personality and Social Psychology;
Journal of Personality and Social Psychology 13 (1969), no. 2, 79.
[196] G.W. Milligan and M.C. Cooper, An examination of procedures for determining the number of clusters in a data set., Psychometrika 50 (1985),
no. 2, 159179.
[197] B.G. Mirkin, Clustering for data mining: a data recovery approach., Chapman & Hall/CRC, 2005.
[198] Alan Mislove, Massimiliano Marcon, Krishna P. Gummadi, Peter Druschel, and Bobby. Bhattacharjee, Measurement and analysis of online
social networks., Proceedings of the 7th ACM SIGCOMM conference
on Internet measurement, ACM, 2007, pp. 2942.
[199] T.M. Mitchell, Machine learning. wcb., Mac Graw Hill (1997), 368.
[200] A. Monreale, F. Pinelli, R. Trasarti, and F. Giannotti, Wherenext: a
location predictor on trajectory pattern mining., Proceedings of the 15th
ACM SIGKDD international conference on Knowledge Discovery
and Data Mining, ACM, 2009, pp. 637646.
[201] C. Moore and M.E.J. Newman, Epidemics and percolation in small-world
networks., Physical Review E 61 (2000), no. 5, 5678.
[202] S. Morris, Contagion., The Review of Economic Studies 67 (2000),
no. 1, 5778.
[203] Fred Morstatter, Jurgen Pfeffer, Huan Liu, and Kathleen M. Carley,
Is the sample good enough? comparing data from twitters streaming api
with twitters firehose., Proceedings of ICWSM (2013).
[204] R. Motwani and P. Raghavan, Randomized algorithms., Chapman &
Hall/CRC, 2010.
[205] I.J. Myung, Tutorial on maximum likelihood estimation., Journal of Mathematical Psychology 47 (2003), no. 1, 90100.
362

[206] M.I. Nelson and E.C. Holmes, The evolution of epidemic influenza.,
Nature reviews genetics 8 (2007), no. 3, 196205.
[207] George L. Nemhauser and Laurence A. Wolsey, Integer and combinatorial optimization., vol. 18, Wiley New York, 1988.
[208] John Neter, William Wasserman, Michael H. Kutner, and et al., Applied linear statistical models., vol. 4, Irwin Chicago, 1996.
[209] M.E.J. Newman, Mixing patterns in networks., Physical Review E 67
(2003), no. 2, 026126.
[210]

, Random graphs as models of networks., Handbook of graphs

and networks (2003), 3568.

[211]

, Modularity and community structure in networks., Proceedings

of the National Academy of Sciences 103 (2006), no. 23, 8577.

[212]

, Networks: an introduction., Oxford University Press, 2010.

[213] M.E.J. Newman, A.L. Barabasi, and D.J. Watts, The structure and
dynamics of networks., Princeton University Press, 2006.
[214] M.E.J. Newman, S. Forrest, and J. Balthrop, Email networks and the
spread of computer viruses., Physical Review E 66 (2002), no. 3, 035101.
[215] M.E.J. Newman and M. Girvan, Mixing patterns and community structure in networks., Statistical Mechanics of Complex Networks (2003),
6687.
[216] M.E.J. Newman, S.H. Strogatz, and D.J. Watts, Random graphs with
arbitrary degree distributions and their applications., Physical Review E
64, 026118.
[217] M.E.J. Newman, D.J. Watts, and S.H. Strogatz, Random graph models
of social networks., Proceedings of the National Academy of Sciences
of the United States of America 99 (2002), no. Suppl 1, 2566.
[218] R.T. Ng and J. Han, Efficient and Effective Clustering Methods for Spatial
Data Mining., Proceedings of the 20th International Conference on
Very Large Data Bases (1994), 144155.
363

[219] Jorge Nocedal and S. Wright, Numerical optimization, series in operations research and financial engineering., Springer (2006).
[220] J. Nohl, C.H. Clarke, and et al., The Black Death. A Chronicle of the
Plague. westholme., The Black Death. A Chronicle of the Plague. (1926).
[221] Brendan OConnor, Ramnath Balasubramanyan, Bryan R. Routledge, and Noah A. Smith, From tweets to polls: linking text sentiment
to public opinion time series., Proceedings of the International AAAI
Conference on Weblogs and Social Media, 2010, pp. 122129.
[222] John ODonovan and Barry. Smyth, Trust in recommender systems.,
Proceedings of the 10th international conference on Intelligent User
Interfaces, ACM, 2005, pp. 167174.
[223] J.P. Onnela and F. Reed-Tsochas, Spontaneous emergence of social influence in online systems., Proceedings of the National Academy of
Sciences 107 (2010), no. 43, 1837518380.
[224] L. Page, S. Brin, R. Motwani, and T. Winograd, The pagerank citation
ranking: bringing order to the web, (1999).
[225] G. Palla, I. Derenyi, I. Farkas, and T. Vicsek, Uncovering the overlapping
community structure of complex networks in nature and society., Nature
435 (2005), no. 7043, 814818.
[226] Gergely Palla, Albert-Laszlo Barabasi, and Tamas. Vicsek, Quantifying social group evolution., Nature 446 (2007), no. 7136, 664667.
[227] Bo Pang and Lillian. Lee, Opinion mining and sentiment analysis., Foundations and Trends in Information Retrieval 2 (2008), no. 1-2, 1135.
[228] Christos H. Papadimitriou and Kenneth. Steiglitz, Combinatorial optimization: algorithms and complexity., Courier Dover Publications,
1998.
[229] R. Pastor-Satorras and A. Vespignani, Epidemic spreading in scale-free
networks., Physical review letters 86 (2001), no. 14, 32003203.
[230] K.B. Patterson and T. Runge, Smallpox and the native american., The
American Journal of the Medical Sciences 323 (2002), no. 4, 216.
364

[231] J. Pattillo, N. Youssef, and S. Butenko, Clique relaxation models in social

network analysis., Handbook of Optimization in Complex Networks
(2012), 143162.
[232] D. Peleg, Local majority voting, small coalitions and controlling monopolies in graphs: A review, Proceedings of the third Colloquium
on Structural Information and Communication Complexity, 1997,
pp. 152169.
[233] R.C. Prim, Shortest connection networks and some generalizations., Bell
System Technical Journal 36 (1957), no. 6, 13891401.
[234] J.R. Quinlan, Induction of decision trees., Machine Learning 1 (1986),
no. 1, 81106.
, C4. 5: programs for machine learning., Morgan Kaufmann,

[235]
1993.

[236] W.M. Rand, Objective criteria for the evaluation of clustering methods.,
Journal of the American Statistical Association (1971), 846850.
[237] Paul Resnick and Hal R. Varian, Recommender systems., Communications of the ACM 40 (1997), no. 3, 5658.
[238] T.S. Robertson, The process of innovation and the diffusion of innovation.,
The Journal of Marketing (1967), 1419.
[239] E.M. Rogers, Diffusion of innovations., Free Press, 1995.
[240] J.H. Rohlfs and H.R. Varian, Bandwagon effects in high-technology
industries., The MIT Press, 2003.
[241] P.J. Rousseeuw, Silhouettes: a graphical aid to the interpretation and
validation of cluster analysis., Journal of computational and applied
mathematics 20 (1987), 5365.
[242] B. Ryan and N.C. Gross, The diffusion of hybrid seed corn in two iowa
communities., Rural sociology 8 (1943), no. 1, 1524.
[243] G. Salton, A. Wong, and C.S. Yang, A vector space model for automatic
indexing., Communications of the ACM 18 (1975), no. 11, 613620.
365

[244] Gerard Salton and Michael J. McGill, Introduction to modern information retrieval, mcgraw-hill., (1986).
[245] J. Sander, M. Ester, H.P. Kriegel, and X. Xu, Density-based clustering
in spatial databases: the algorithm GDBSCAN and its applications., Data
Mining and Knowledge Discovery 2 (1998), no. 2, 169194.
[246] Badrul Sarwar, George Karypis, Joseph Konstan, and John. Riedl,
Item-based collaborative filtering recommendation algorithms., Proceedings of the 10th international conference on World Wide Web, ACM,
2001, pp. 285295.
[247] S. Scellato, M. Musolesi, C. Mascolo, V. Latora, and A. Campbell,
Nextplace: a spatio-temporal prediction framework for pervasive systems.,
Pervasive Computing (2011), 152169.
[248] J Ben Schafer, Dan Frankowski, Jon Herlocker, and Shilad. Sen, Collaborative filtering recommender systems., The Adaptive Web, Springer,
2007, pp. 291324.
[249] J Ben Schafer, Joseph Konstan, and John Riedl, Recommender systems
in e-commerce., Proceedings of the first ACM conference on Electronic
Commerce, ACM, 1999, pp. 158166.
[250] D.S. Scharfstein and J.C. Stein, Herd behavior and investment., The
American Economic Review (1990), 465479.
[251] T.C. Schelling, Dynamic models of segregation., Journal of Mathematical
Sociology 1 (1971), no. 2, 143186.
, Micromotives and macrobehavior., WW Norton & Company,

[252]
2006.

[253] John. Scott, Social network analysis., Sociology 22 (1988), no. 1, 109
127.
[254] Shilad Sen, Jesse Vig, and John. Riedl, Tagommenders: connecting users
to items through tags., Proceedings of the 18th international conference
on World Wide Web, ACM, 2009, pp. 671680.
366

[255] C.R. Shalizi and A.C. Thomas, Homophily and contagion are generically
confounded in observational social network studies., Sociological Methods & Research 40 (2011), no. 2, 211239.
[256] R.J. Shiller, Conversation, information, and herd behavior., The American
Economic Review 85 (1995), no. 2, 181185.
[257] Borkur
Sigurbjornsson
and Roelof. Van Zwol, Flickr tag recommen

dation based on collective knowledge., Proceedings of the 17th international conference on World Wide Web, ACM, 2008, pp. 327336.
[258] G. Simmel and E.C. Hughes, The sociology of sociability., American
Journal of Sociology (1949), 254261.
[259] H.A. Simon, Bandwagon and underdog effects and the possibility of election predictions., Public Opinion Quarterly 18 (1954), no. 3, 245253.
[260]

, On a class of skew distribution functions., Biometrika 42 (1955),

no. 3/4, 425440.

[261] T.A.B. Snijders, C.E.G. Steglich, and M. Schweinberger, Modeling the

co-evolution of networks and behavior. in, Longitudinal models in the
behavioral and related sciences (2007), 4171.
[262] R. Solomonoff and A. Rapoport, Connectivity of random nets., Bulletin
of Mathematical Biology 13 (1951), no. 2, 107117.
[263] S. Spaccapietra, C. Parent, M.L. Damiani, J.A. De Macedo, F. Porto,
and C. Vangenot, A conceptual view on trajectories., Data and knowledge engineering 65 (2008), no. 1, 126146.
[264] Martin G. Everett. Stephen P. Borgatti, Two algorithms for computing
regular equivalence., Social Networks 15 (1993), no. 4, 361376.
[265] S.S. Stevens, On the theory of scales of measurement, science, 1946.
[266] D. Strang and S.A. Soule, Diffusion in organizations and social movements: from hybrid corn to poison pills., Annual Review of Sociology
(1998), 265290.
367

[267] A. Strehl, J. Ghosh, and C. Cardie, Cluster ensembles-a knowledge reuse

framework for combining multiple partitions., Journal of Machine Learning Research 3 (2003), no. 3, 583617.
[268] Xiaoyuan Su and Taghi M. Khoshgoftaar, A survey of collaborative
filtering techniques., Advances in Artificial Intelligence 2009 (2009), 4.
[269] J. Sun, C. Faloutsos, S. Papadimitriou, and P.S. Yu, Graphscope:
parameter-free mining of large time-evolving graphs, Proceedings of the
13th ACM SIGKDD international conference on Knowledge Discovery and Data Mining, ACM, 2007, pp. 687696.
[270] P.N. Tan, M. Steinbach, V. Kumar, and et al., Introduction to data
mining., Pearson Addison Wesley Boston, 2006.
[271] Jiliang Tang, Huiji Gao, Xia Hu, and Huan. Liu, Exploiting homophily
effect for trust prediction., WSDM, 2013.
[272] Jiliang Tang, Huiji Gao, and Huan. Liu, mtrust: discerning multifaceted trust in a connected world., WSDM, 2012.
[273] Jiliang Tang, Huiji Gao, Huan Liu, and Atish Das. Sarma, eTrust:
understanding trust evolution in an online world., KDD, 2012.
[274] Jiliang Tang, Xia Hu, Huiji Gao, and Huan. Liu, Exploiting local and
global social context for recommendation., IJCAI, 2013.
[275] Jiliang Tang and Huan. Liu, Feature selection with linked data in social
media., SDM, 2012.
, Unsupervised feature selection for linked social media data., KDD,

[276]
2012.
[277]

, Coselect: Feature selection with instance selection for social media

data., SDM, 2013.

[278] L. Tang and H. Liu, Community detection and mining in social media., Synthesis Lectures on Data Mining and Knowledge Discovery
2 (2010), no. 1, 1137.
368

[279] Lei Tang and Huan. Liu, Relational learning via latent social dimensions.,
Proceedings of the 15th ACM SIGKDD international conference on
Knowledge Discovery and Data Mining, ACM, 2009, pp. 817826.
[280] Lei Tang, Xufei Wang, and Huan. Liu, Community detection via heterogeneous interaction analysis., Data Mining and Knowledge Discovery
(DMKD) 25 (2012), no. 1, 1 33.
[281] G. Tarde, Las leyes de la imitacion: Estudio sociologico., Daniel Jorro,
1907.
[282] N. Thanh and T.M. Phuong, A gaussian mixture model for mobile location prediction., 2007 IEEE International Conference on Research,
Innovation and Vision for the Future, IEEE, 2007, pp. 152157.
[283] W. Trotter, Instincts of the Herd in War and Peace., 1916.
[284] Johan Ugander, Brian Karrer, Lars Backstrom, and Cameron. Marlow,
The anatomy of the facebook social graph, arxiv preprint arxiv:1111.4503.,
Structure 5, 6.
[285] T.W. Valente, Network models of the diffusion of innovations, 1995.
[286]

, Network models of the diffusion of innovations., Computational

& Mathematical Organization Theory 2 (1996), no. 2, 163164.

[287]

, Social network thresholds in the diffusion of innovations., Social

Networks 18 (1996), no. 1, 6989.

[288] T. Veblen, The Theory of the Leisure Class., Houghton Mifflin Boston,
1965.
[289] F. Wang and Q.Y. Huang, The importance of spatial-temporal issues for
case-based reasoning in disaster management., 2010 18th International
Conference on Geoinformatics, IEEE, 2010, pp. 15.
[290] S.S. Wang, S.I. Moon, K.H. Kwon, C.A. Evans, and M.A. Stefanone,
Face off: implications of visual cues on initiating friendship on facebook.,
Computers in Human Behavior 26 (2010), no. 2, 226234.
369

[291] Xufei Wang, Shamanth Kumar, and Huan. Liu, A study of tagging
behavior across social media., In SIGIR Workshop on Social Web Search
and Mining (SWSM), 2011.
[292] Xufei Wang, Lei Tang, Huiji Gao, and Huan. Liu, Discovering overlapping groups in social media., the 10th IEEE International Conference
on Data Mining (ICDM2010) (Sydney, Australia), December 14 - 17
2010.
[293] S. Warshall, A theorem on boolean matrices., Journal of the ACM (JACM)
9 (1962), no. 1, 1112.
[294] S. Wasserman and K. Faust, Social network analysis: Methods and applications., Cambridge University Press (1994).
[295] D.J. Watts, Networks, dynamics, and the small-world phenomenon.,
American Journal of Sociology 105 (1999), no. 2, 493527.
[296]

, A simple model of global cascades on random networks., Proceedings of the National Academy of Sciences 99 (2002), no. 9, 57665771.

[297] D.J. Watts and P.S. Dodds, Influentials, networks, and public opinion
formation., Journal of Consumer Research 34 (2007), no. 4, 441458.
[298] D.J. Watts and S.H. Strogatz, Collective dynamics of small-world networks., nature 393 (1998), no. 6684, 440442.
[299] I. Welch, Sequential sales, learning, and cascades., Journal of Finance
(1992), 695732.
[300] J. Weng, E.P. Lim, J. Jiang, and Q. He, Twitterrank: finding topicsensitive influential twitterers, Proceedings of the third ACM international conference on Web Search and Data Mining, ACM, 2010,
pp. 261270.
[301] D.B. West, Introduction to graph theory., vol. 2, Prentice Hall Upper
Saddle River, NJ.:, 2001.
[302] D.R. White, Structural equivalences in social networks: concepts and
measurement of role structures, Research Methods in Social Network
Analysis Conference, 1980, pp. 193234.
370

[303]

, Regge: a regular graph equivalence algorithm for computing role

distances prior to blockmodeling., Unpublished manuscript, University
of California, Irvine (1984).

[304] I.H. Witten, E. Frank, and M.A. Hall, Data Mining: practical machine
learning tools and techniques., Morgan Kaufmann, 2011.
[305] R. Xu and D. Wunsch, Survey of clustering algorithms., Neural Networks, IEEE Transactions on 16 (2005), no. 3, 645678.
[306] J. Yang and J. Leskovec, Modeling information diffusion in implicit networks, IEEE 10th International Conference on Data Mining, IEEE,
2010, pp. 599608.
[307] H.P. Young, Individual strategy and social structure: an evolutionary
theory of institutions., Princeton University Press, 2001.
[308] G.U. Yule, A mathematical theory of evolution, based on the conclusions
of dr. j.c. willis, frs., Philosophical Transactions of the Royal Society
of London. Series B, Containing Papers of a Biological Character 213
(1925), 2187.
[309] W.W. Zachary, An information flow model for conflict and fission in small
groups., Journal of Anthropological Research (1977), 452473.
[310] Reza Zafarani, William D. Cole, and Huan. Liu, Sentiment propagation in social networks: a case study in livejournal., Advances in Social
Computing, Springer, 2010, pp. 413420.
[311] Reza Zafarani and Huan. Liu, Connecting corresponding identities
across communities., ICWSM, 2009.
[312]

, Connecting users across social media sites: a behavioral-modeling

approach., Proceedings of the 19th ACM SIGKDD international conference on Knowledge Discovery and Data Mining, KDD, 2013.

[313] Zheng Alan Zhao and Huan. Liu, Spectral feature selection for data
mining., Chapman & Hall/CRC, 2011.

371

372

Index

G(n, m) model, 107

G(n, p) model, 106
k-clan, 183
k-clique, 182
k-club, 183
k-means, 156
k-plex, 180

behavior analysis methodology,

320
betweenness centrality, 78
BFS, see breadth-first search
big data, 13
big data paradox, 13
bipartite graph, 44
botnet, 171
breadth-first search, 47
bridge, 45
bridge detection, 60

actor, see node

Adamic and Adar measure, 324
adjacency list, 32
adjacency matrix, 31
adopter types, 229
adoption of hybrid seed corn, 229
affiliation network, 44
Anderson, Lisa R., 218
Asch conformity experiment, 216
Asch, Solomon, 216
assortativity, 255
measuring, 257
nominal attributes, 258
ordinal attributes, 261
significance, 258
attribute, see feature
average path length, 102, 105

cascade maximization, 225

causality testing, 321
centrality, 70
betweenness centrality, 78
computing, 78
closeness centrality, 80
degree centrality, 70
gregariousness, 70
normalized, 71
prestige, 70
eigenvector centrality, 71
group centrality, 81
betweenness, 82
closeness, 82
373

degree, 81
Katz centrality, 74
divergence in computation,
75
PageRank, 76
Christakis, Nicholas A., 245
citizen journalist, 11
class, see class attribute
class attribute, 133
clique identification, 178
clique percolation method, 180
closeness centrality, 80
cluster centroid, 156
clustering, 155, 156
k-means, 156
hierarchical, 191
agglomerative, 191
divisive, 191
partitional, 156
spectral, 187
clustering coefficient, 84, 102
global, 85
local, 86
cohesive subgroup, see community
cold start problem, 286
collaborative filtering, 289, 323
memory-based, 289
item-based, 292
user-based, 290
model-based, 289, 293
collective behavior, 328
collective behavior modeling, 333
common neighbors, 324
community, 171
detection, 175
emit, see explicit
etic, see implicit

evolution, 193
explicit, 173
implicit, 174
community detection, 175
group-based, 175, 184
member-based, 175, 177
node degree, 178
node reachability, 181
node similarity, 183
community evaluation, 200
community membership behavior, 317
commute time, 327
confounding, 255
contact network, 237
content data, 316
content-based recommendation,
288
cosine similarity, 92, 183, 291
covariance, 94, 261
data point, see instance
data preprocessing, 138
aggregation, 138
discretization, 138
feature extraction, 138
feature selection, 138
sampling, 138
random, 139
stratified, 139
with/without replacement,
139
data quality, 137
duplicate data, 137
missing values, 137
noise, 137
outliers, 137
data scientist, 12
374

undirected, 27
visit, 36
edge list, 32
edge-reversal test, 277
eigenvector centrality, 71
entropy, 143
epidemics, 214, 236
infected, 238
recovered, 238
removed, 238
SI model, 239
SIR model, 241
SIRS model, 244
SIS model, 242
susceptible, 238
Erdos, Paul, 107
Euclidean distance, 155
evaluation dilemma, 13
evolutionary clustering, 198
external-influence model, 232

data sparsity, 287

decision tree learning, 141
degree, 25, 28
distribution, 29
degree centrality, 70, 264
gregariousness, 70
prestige, 70
degree distribution, 29, 102
dendrogram, 191
densification, 194
depth-first search, 46
DFS, see depth-first search
diameter, 40
shrinkage, 196
diffusion of innovation, 214
diffusion of innovations, 228
Dijkstra, Edsger, 49
Dijskras algorithm, 49
diminishing returns, 318
directed graph, 28, 33
distinguishing influence and homophily, 276
edge-reversal test, 277
randomization test, 278
shuffle test, 276
early adopters, 229
early majority, 229
Eckart-Young-Mirsky
294
edge, 27
betweenness, 191
directed, 27
incident, 35
loop, 28, 32
self-link, 32
signed, 34
traverse, see visit

F-measure, 202, 307

false negative, 201
false positive, 201
feature, 133
categorical, see nominal
continuous, 133
discrete, 133
interval, 134
nominal, 134
ordinal, 134
ratio, 134
feature selection, 319
Ford-Fulkerson algorithm, 55
Frobenius norm, 293
fully-mixed technique, 237

theorem,

giant component, 108, 194

375

Gilbert, Edgar, 106

Girvan-Newman algorithm, 191
global clustering coefficient, 85
Granger causality, 321
Granger, Clive W. J., 321
graph
k-connected, 188
bipartite graph, 44
bridge, 45
detection, 60
complete graph, 42
component, 38
connected, 38
connectivity, 35
circuit, see tour
cycle, 36
path, 36
tour, 36
trail, 36
walk, 36
cut, 185
densification, 194
density, 190
diameter, 40
directed, 28
directed graph, 33
edge, 25, 27
empty graph, 33
forest, 40
minimum cut, 185
mixed graph, 33
multigraph, 34
node, 25, 26
null graph, 33
partition, 44
planar graph, 43
regular graph, 44
representation

adjacency list, 32
adjacency matrix, 31
edge list, 32
shortest path, 39
Dijskras algorithm, 49
signed graph, 34
simple graph, 34
strongly connected, 38
subgraph, 30
minimum spanning tree, 41
spanning tree, 41
Steiner tree, 42
traversal, 45
breadth-first search, 47
depth-first search, 46
tree, 40
undirected graph, 33
weakly connected, 38
weighted graph, 34
graph Laplacian, 187
graph traversal, 45
breadth-first search, 47
depth-first search, 46
group, see community
group centrality, 81
betweenness, 82
closeness, 82
degree, 81
herd behavior, 214
hitting time, 327
Holt, Charles A., 218
homophily, 255, 274
measuring, 274
modeling, 274
ICM, see independent cascade
model
376

in-degree, 28
independent cascade model, 222
individual behavior, 315
user-community
behavior,
316
user-entity behavior, 316
user-user behavior, 316
individual behavior modeling,
322
influence, 255, 264
measuring, 264
observation-based, 264
prediction-based, 264
modeling, 269
influence flow, 266
influence modeling, 269
explicit networks, 269
implicit networks, 271
information cascade, 214, 221, 323
information pathways, 78
information provenance, 249
innovation characteristics, 228
innovators, 229
instance, 133
labeled, 133
unlabeled, 133
internal-influence model, 232
intervention, 214, 220, 227, 235,
245

Kendalls tau, 308

knowledge
discovery
in
databases, 131
Kronecker delta function, 258
labeled data, 140
laggards, 229
Laplacian, 187
largest connected component, 108
late majority, 229
levels of measurement, 133
LIM, see linear influence model
linear influence model, 269, 334
linear threshold model, 269, 322
link data, 316
link prediction, 299, 323
node
neighborhood-based
methods, 324
path-based methods, 326
local clustering coefficient, 86
logistic function, 276
logistic growth function, 240
logistic regression, 333
logit function, 152
loop, 28, 32
low-rank matrix approximation,
293
LTM, see linear threshold model
MAE, see mean absolute error
Mahalanobis distance, 155
matrix factorization, 301
maximum bipartite matching, 57
mean absolute error, 305
measurement, see feature
measuring homophily, 274
Milgram, Stanley, 105
minimum spanning tree, 41

Jaccard similarity, 92, 183, 324

Katz centrality, 74
divergence in computation, 75
Katz model, 230
Katz, Elihu, 230
KDD process, see knowledge discovery in databases
377

Prims algorithm, 52
mixed graph, 33
mixed-influence model, 232
modeling homophily, 274
modularity, 189, 259
multi-step flow model, 230
mutual information, 203

normalized mean absolute error,

305
normalized mutual information,
203
nuke attack, 287
observation, see instance
obtaining sufficient samples, 13
out-degree, 28
overfitting, 301

naive bayes classifier, 144

nearest neighbor classifier, 146
neighborhood, 28
network flow, 53
augmenting path, 56
capacity, 54
capacity constraint, 54
flow, 54
flow conservation constraint,
54
flow network, 54
Ford-Fulkerson algorithm, 55
residual capacity, 56
residual network, 56
source and sink, 54
network measure, 69
network segmentation, 194
NMI, see normalized mutual information
node, 26
adjacent, 35
connected, 38
degree, 25, 28
in-degree, 28
neighborhood, 28
out-degree, 28
reachable, 38
noise removal fallacy, 13
normalized cut, 185

PageRank, 76, 264

parametric estimation, 273
Pearson correlation coefficient, 94,
263, 290
Perron-Frobenius theorem, 72
Poisson distribution, 112
Power Iteration Method, 98
power-law distribution, 30
powerlaw distribution, 103, 272,
276, 333
precision, 201, 307
preferential attachment measure,
325
preferential attachment model,
333
prestige, 70
prominence, see prestige
purity, 203
push attack, 287
quasi-clique, 191
random graph, 106
average path length, 114
clustering coefficient, 112
degree distribution, 111
evolution, 108
phase transition, 109
378

similarity, 91
cosine similarity, 92
Jaccard similarity, 92, 94
regular equivalence, 94
structural equivalence, 92
SimRank, 328
singleton, 194
singular value decomposition,
152, 293
SIR model, 241
SIRS model, 244
SIS model, 242
six degrees of separation, 105
small world model, 333
small-world, 105
small-world model, 115
average path length, 118
clustering coefficient, 118
degree distribution, 117
Social Atom, 12
social balance, 89
social correlation, 276
social media, 11
social media mining, 12
Social Molecule, 12
social network, 25
social similarity, 255
social status, 35, 90
social tie, see edge
sociomatrix, see adjacency matrix
Solomonoff, Ray, 106
sparse matrix, 32
Spearmans rank correlation coefficient, 268, 308
spectral clustering, 187
star, 194
Stevens, Stanley Smith, 133
Strogatz, Steven, 115

random walk, 36
randomization test, 278, 321
rank correlation, 268
Rapoport, Anatol, 106
ratio cut, 185
raw data, 131
recall, 201, 307
reciprocity, 87
recommendation to groups, 297
least misery, 297
maximizing average satisfaction, 297
most pleasure, 298
recommender systems, 285
regression, 322
regular equivalence, 94, 184
regular graph, 44
ring lattice, 115
regular ring lattice, 115
regularization term, 302
relationships, see edge
relaxing cliques, 180
k-plex, 180
RMSE, see root mean squared error
Rogers, Everett M., 228
root mean squared error, 305
rooted PageRank, 327
Renyi, Alfred, 107
scale-free network, 105
self-link, see loop, 32
sender-centric model, 222
sentiment analysis, 164
shortest path, 39
Dijskras algorithm, 49
shuffle test, 276
SI model, 239
379

see term frequencyinverse document frequency

theory of scales, 133
training data, see labeled data
transitivity, 83
clustering coefficient, 84
global, 85
local, 86
Trotter, Wilfred, 216
true negative, 201
true positive, 201

structural equivalence, 92
submodular function, 225
supervised learning, 140
classification, 141
decision tree learning, 141
naive bayes classifier, 144
nearest neighbor classifier,
146
with network information,
147
evaluation
k-fold cross validation, 154
accuracy, 154
leave-one-out, 154
evaluations, 153
regression, 141, 150
linear, 151
logistic, 152
SVD, see singular value decomposition

TF-IDF,

tabular data, 133

tag clouds, 205
target data, 131
term frequency-inverse document frequency, 135
test data, 140

vector-space model, 135

vectorization, 135
vertex, see node

undirected graph, 33
unsupervised learning, 155
evaluation, 158
cohesiveness, 159
separateness, 159
silhouette index, 160
user migration, 329
user-item matrix, 289

Watts, Duncan J., 115

Zacharys karate club, 173

380