Hui Lin and Ming Li

Introduction to Data
Science
Contents

List of Tables ix

List of Figures xi

Preface xiii

About the Authors xxi

1 Introduction 1
1.1 A brief history of data science . . . . . . . . . . 1
1.2 Data science role and skill tracks . . . . . . . . . 5
1.2.1 Engineering . . . . . . . . . . . . . . . . . 6
1.2.2 Analysis . . . . . . . . . . . . . . . . . . . 8
1.2.3 Modeling . . . . . . . . . . . . . . . . . . 9
1.3 What kind of questions can data science solve? . 13
1.3.1 Prerequisites . . . . . . . . . . . . . . . . 13
1.3.2 Problem type . . . . . . . . . . . . . . . . 15
1.4 Structure data science team . . . . . . . . . . . . 18
1.5 List of potential data science careers . . . . . . . 21

2 Soft Skills for Data Scientists 27

2.1 Comparison between Statistician and Data Scien-
tist . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.2 Beyond Data and Analytics . . . . . . . . . . . . 29
2.3 Three Pillars of Knowledge . . . . . . . . . . . . 31
2.4 Data Science Project Cycle . . . . . . . . . . . . 32
2.4.1 Types of Data Science Projects . . . . . . 32
2.4.2 Problem Formulation and Project Planning
Stage . . . . . . . . . . . . . . . . . . . . 34
2.4.3 Project Modeling Stage . . . . . . . . . . 36

iii
iv Contents

2.4.4 Model Implementation and Post Produc-

tion Stage . . . . . . . . . . . . . . . . . . 37
2.4.5 Project Cycle Summary . . . . . . . . . . 38
2.5 Common Mistakes in Data Science . . . . . . . . 39
2.5.1 Problem Formulation Stage . . . . . . . . 39
2.5.2 Project Planning Stage . . . . . . . . . . . 40
2.5.3 Project Modeling Stage . . . . . . . . . . 41
2.5.4 Model Implementation and Post Produc-
tion Stage . . . . . . . . . . . . . . . . . . 42
2.5.5 Summary of Common Mistakes . . . . . . 43

3 Introduction to The Data 45

3.1 Customer Data for A Clothing Company . . . . 45
3.2 Swine Disease Breakout Data . . . . . . . . . . . 47
3.3 MNIST Dataset . . . . . . . . . . . . . . . . . . 49
3.4 IMDB Dataset . . . . . . . . . . . . . . . . . . . 49

4 Big Data Cloud Platform 53

4.1 Power of Cluster of Computers . . . . . . . . . . 54
4.2 Evolution of Cluster Computing . . . . . . . . . 55
4.2.1 Hadoop . . . . . . . . . . . . . . . . . . . 55
4.2.2 Spark . . . . . . . . . . . . . . . . . . . . 56
4.3 Introduction of Cloud Environment . . . . . . . 56
4.3.1 Open Account and Create a Cluster . . . 57
4.3.2 R Notebook . . . . . . . . . . . . . . . . . 58
4.3.3 Markdown Cells . . . . . . . . . . . . . . 58
4.4 Leverage Spark Using R Notebook . . . . . . . . 60
4.5 Databases and SQL . . . . . . . . . . . . . . . . 67
4.5.1 History . . . . . . . . . . . . . . . . . . . 67
4.5.2 Database, Table and View . . . . . . . . . 69
4.5.3 Basic SQL Statement . . . . . . . . . . . 70
4.5.4 Advanced Topics in Database . . . . . . . 74

5 Data Pre-processing 75
5.1 Data Cleaning . . . . . . . . . . . . . . . . . . . 77
5.2 Missing Values . . . . . . . . . . . . . . . . . . . 80
5.2.1 Impute missing values with median/mode 81
5.2.2 K-nearest neighbors . . . . . . . . . . . . 82
Contents v

5.2.3 Bagging Tree . . . . . . . . . . . . . . . . 83

5.3 Centering and Scaling . . . . . . . . . . . . . . . 84
5.4 Resolve Skewness . . . . . . . . . . . . . . . . . 85
5.5 Resolve Outliers . . . . . . . . . . . . . . . . . . 89
5.6 Collinearity . . . . . . . . . . . . . . . . . . . . . 93
5.7 Sparse Variables . . . . . . . . . . . . . . . . . . 96
5.8 Re-encode Dummy Variables . . . . . . . . . . . 97

6 Data Wrangling 101

6.1 Summarize Data . . . . . . . . . . . . . . . . . . 102
6.1.1 dplyr package . . . . . . . . . . . . . . . . 103
6.1.2 apply(), lapply() and sapply() in base R . . 113
6.2 Tidy and Reshape Data . . . . . . . . . . . . . . 117

7 Model Tuning Strategy 123

7.1 Variance-Bias Trade-Off . . . . . . . . . . . . . . 124
7.2 Data Splitting and Resampling . . . . . . . . . . 132
7.2.1 Data Splitting . . . . . . . . . . . . . . . 133
7.2.2 Resampling . . . . . . . . . . . . . . . . . 143

8 Measuring Performance 149

8.1 Regression Model Performance . . . . . . . . . . 149
8.2 Classification Model Performance . . . . . . . . . 153
8.2.1 Confusion Matrix . . . . . . . . . . . . . . 155
8.2.2 Kappa Statistic . . . . . . . . . . . . . . . 157
8.2.3 ROC . . . . . . . . . . . . . . . . . . . . . 158
8.2.4 Gain and Lift Charts . . . . . . . . . . . . 161

9 Regression Models 165

9.1 Ordinary Least Square . . . . . . . . . . . . . . . 166
9.1.1 The Magic P-value . . . . . . . . . . . . . 171
9.1.2 Diagnostics for Linear Regression . . . . . 174
9.2 Principal Component Regression and Partial Least
Square . . . . . . . . . . . . . . . . . . . . . . . 178

10 Regularization Methods 187

10.1 Ridge Regression . . . . . . . . . . . . . . . . . . 188
10.2 LASSO . . . . . . . . . . . . . . . . . . . . . . . 193
vi Contents

10.3 Variable selection property of the lasso . . . . . . 197

10.4 Elastic Net . . . . . . . . . . . . . . . . . . . . . 199
10.5 Penalized Generalized Linear Model . . . . . . . 200
10.5.1 Introduction to glmnet package . . . . . . 201
10.5.2 Penalized logistic regression . . . . . . . . 206

11 Tree-Based Methods 217

11.1 Tree Basics . . . . . . . . . . . . . . . . . . . . . 217
11.2 Splitting Criteria . . . . . . . . . . . . . . . . . . 220
11.3 Tree Pruning . . . . . . . . . . . . . . . . . . . . 227
11.4 Regression and Decision Tree Basic . . . . . . . . 231
11.4.1 Regression Tree . . . . . . . . . . . . . . . 231
11.4.2 Decision Tree . . . . . . . . . . . . . . . . 236
11.5 Bagging Tree . . . . . . . . . . . . . . . . . . . . 241
11.6 Random Forest . . . . . . . . . . . . . . . . . . . 245
11.7 Gradient Boosted Machine . . . . . . . . . . . . 249
11.7.1 Adaptive Boosting . . . . . . . . . . . . . 249
11.7.2 Stochastic Gradient Boosting . . . . . . . 251
11.7.3 Boosting as Additive Model . . . . . . . . 256

12 Deep Learning 261

12.1 Projection Pursuit Regression . . . . . . . . . . . 265
12.2 Feedforward Neural Network . . . . . . . . . . . 269
12.2.1 Logistic Regression as Neural Network . . 269
12.2.2 Stochastic Gradient Descent . . . . . . . . 271
12.2.3 Deep Neural Network . . . . . . . . . . . 272
12.2.4 Activation Function . . . . . . . . . . . . 276
12.2.5 Optimization . . . . . . . . . . . . . . . . 280
12.2.6 Deal with Overfitting . . . . . . . . . . . . 287
12.2.7 Image Recognition Using FFNN . . . . . . 288
12.3 Convolutional Neural Network . . . . . . . . . . 302
12.3.1 Convolution Layer . . . . . . . . . . . . . 304
12.3.2 Padding Layer . . . . . . . . . . . . . . . 308
12.3.3 Pooling Layer . . . . . . . . . . . . . . . . 309
12.3.4 Convolution Over Volume . . . . . . . . . 313
12.3.5 Image Recognition Using CNN . . . . . . 316
12.4 Recurrent Neural Network . . . . . . . . . . . . 322
Contents vii

12.4.1 RNN Model . . . . . . . . . . . . . . . . . 324

12.4.2 Word Embedding . . . . . . . . . . . . . . 328
12.4.3 Long Short Term Memory . . . . . . . . . 330
12.4.4 Sentiment Analysis Using RNN . . . . . . 333

Appendix 343

A Handling Large Local Data 345

A.1 readr . . . . . . . . . . . . . . . . . . . . . . . . 345
A.2 data.table— enhanced data.frame . . . . . . . . . 351

B R code for data simulation 363

B.1 Customer Data for Clothing Company . . . . . . 363
B.2 Swine Disease Breakout Data . . . . . . . . . . . 368

Bibliography 373

Index 381
List of Tables

ix
List of Figures

1.1 Different roles in data science and the skill require-

ments . . . . . . . . . . . . . . . . . . . . . . . . 23

2.1 Comparison of Statistician and Data Scientist . . 28

2.2 Three Pillars of Knowledge . . . . . . . . . . . . 32

3.1 Sample of MNIST dataset . . . . . . . . . . . . . 50

4.1 Single computer (left) and a cluster of computers

(right) . . . . . . . . . . . . . . . . . . . . . . . . 54
4.2 Example of R Notebook . . . . . . . . . . . . . . 59
4.3 Example of Markdown Cell . . . . . . . . . . . . 59

5.1 Data Pre-processing Outline . . . . . . . . . . . . 75

5.2 Example of skewed distributions . . . . . . . . . . 86

7.1 Types of Model Error . . . . . . . . . . . . . . . . 126

7.2 High bias model . . . . . . . . . . . . . . . . . . . 128
7.3 High variance model . . . . . . . . . . . . . . . . 129
7.4 Parameter Tuning Process . . . . . . . . . . . . . 135
7.5 Compare Maximum Dissimilarity Sampling with
Random Sampling . . . . . . . . . . . . . . . . . 140
7.6 Divide data according to time . . . . . . . . . . . 141
7.7 5-fold cross-validation . . . . . . . . . . . . . . . 144

9.1 Correlation Matrix Plot for Explanatory Variables 169

9.2 Linear Regression Diagnostic Plots: residual plot
(top left), Q-Q plot (top right), standardized resid-
uals plot (lower left), Cook’s distance (lower right) 176

10.1 Test mean squared error for the ridge regression . 191
10.2 Test mean squared error for the lasso regression . 195

xi
xii List of Figures

10.3 Contours of the RSS and constrain functions for the

lasso (left) and ridge regression (right) . . . . . . 198

11.1 This is a classification tree trained from passenger

survival data from the Titanic. Survival probability
is predicted using sex and age. . . . . . . . . . . . 219
11.2 Example of decision regions for different types of
flowers based on petal length and width. Three dif-
ferent types of flowers are classified using decisions
about length and width. . . . . . . . . . . . . . . 221

12.1 Feedforward Neural Network . . . . . . . . . . . . 273

12.2 1-layer Neural Network . . . . . . . . . . . . . . . 273
12.3 Sigmoid Function . . . . . . . . . . . . . . . . . . 276
12.4 Hyperbolic Tangent Function . . . . . . . . . . . 277
12.5 Rectified Linear Unit Function . . . . . . . . . . . 278
12.6 Rectified Linear Unit Function . . . . . . . . . . . 279
12.7 The intuition behind the weighted average gradient 282
12.8 Exponentially weighted averages with and without
corrrection . . . . . . . . . . . . . . . . . . . . . . 284
12.9 Grayscale image is a set of pixels on 2-d space. Each
pixel has a value range from 0 to 255. . . . . . . . 289
12.10Color image is a set of pixels on 3-d space. Each
pixel has a value range from 0 to 255. . . . . . . . 289
12.11There are an input image (left), a filter (middel),
and an output image (right). . . . . . . . . . . . . 304
12.12Convolution step by step . . . . . . . . . . . . . . 305
12.13Edge Detection Example . . . . . . . . . . . . . . 306
12.14Simple Edge Detection Example . . . . . . . . . . 307
12.15Convolution Layer Parameters . . . . . . . . . . . 308
12.16Padding Layer . . . . . . . . . . . . . . . . . . . . 309
12.17Pooling Layer . . . . . . . . . . . . . . . . . . . . 310
12.18Example of max and mean pooling . . . . . . . . 314
12.19Example of convolution over volume . . . . . . . 315
12.20Recurrent Neural Network Unit . . . . . . . . . . 323
12.21An Unrolled Recurrent Neural Network . . . . . . 323
Preface

During the early years of our career as data scientists, we were

bewildered by all kinds of data science hype. There is a lack of
definition of many basic terminologies such as “big data,” “artifi-
cial intelligence,” and “data science.” How big is big data? What is
data science? What is the difference between the sexy title “Data
Scientist” and the traditional “Data Analyst?” The term data sci-
ence stirs so many associations such as machine learning, deep
learning, data mining, pattern recognition. All those struck us as
confusing and vague as real-world data scientists! However, we
always feel something tangible in data science applications, and
it has been developing very fast. After applying data science for
many years, we now have a much better idea about data science
in general. This book is our endeavor to make data science a more
concrete and legitimate field. In addition to “hard” technical as-
pects, the book also covers soft skills and career development in
data science.

Goal of the Book

This is a book on data science with a specific focus on industrial
experience. Data Science is a cross-disciplinary subject involving
hands-on experience and business problem-solving exposures. The
majority of existing introduction books on data science are about
the modeling techniques and the implementation of models using
R or Python. However, many of these books lack the context of the
industrial environment. Moreover, a crucial part, the art of data
science in practice, is often missing. This book intends to fill the
gap.

xiii
xiv Preface

Some key features of this book are as follows:

• It covers both technical and soft skills.
• It has a chapter dedicated to the big data cloud environment.
In the industry, the practice of data science is often in such an
environment.
• It is hands-on. We provide the data and repeatable R and
Python code in notebooks. Readers can repeat the analysis in
the book using the data and code provided. We also suggest
that readers modify the notebook to perform their analyses
with their data and problems whenever possible. The best way
to learn data science is to do it!
• It focuses on the skills needed to solve real-world industrial
problems rather than an academic book.

What This Book Covers

This book provides a comprehensive introduction to various data
science fields, soft and programing skills in data science projects,
and potential career paths for data scientists.
There are many existing data science books, including:
• An Introduction to Data Science by Saltz and Stanton
• A Hands-On Introduction to Data Science by Chirag Shah
• Introduction to Data Science: Data Analysis and Prediction
Algorithms with R by Rafael Irizarry
• Build a Career in Data Science by Robinson and Nolis
• Data Science (The MIT Press Essential Knowledge series) by
Kelleher and Tierney
• The Data Science Handbook by Field Cady
• Data Science from Scratch: First Principles with Python (2nd
Edition) by Joel Grus
As data science fields are fast-growing, we did not find a book
covering all the contents we feel are essential for a data scientist.
Preface xv

The following table compares the data science books mentioned

above with this book for contents in Big Data (Spark), R Code,
Python Code, Data Preprocessing, Deep Learning, ML models,
Career Path, and Project Cycles. We use 0 for not covered, 1 for
minimal coverage, 2 for some coverage, and 3 for extensive cover-
age.

Big
Data R PythonData Deep ML Career Project
Author(Spark) Code Code Preprocessing
LearningModelsPath Cycle
Lin 3 3 3 3 3 3 3 3
and
Li
Saltz 1 3 0 1 0 3 0 0
Shah 0 3 3 1 0 3 1 0
Irizarry0 3 0 3 0 3 0 0
Robinson0 0 0 0 0 0 3 3
Kelleher1 0 0 1 0 2 3 3
Cady 0 0 3 3 0 3 1 1
Grus 1 0 3 2 3 3 1 0

The book is organized as the following.

• Chapters 1-3 discuss various aspects of data science: differ-
ence tracks, career paths, project cycles, soft skills, and common
pitfalls. Chapter 3 is an overview of the data sets used in the
book.
• Chapter 4 introduces typical big data cloud platforms and
uses R library sparklyr as an interface to the big data analytics
engine Spark.
• Chapters 5-6 cover the essential skills to prepare the data
for further analysis and modeling, i.e., data preprocessing and
wrangling.
• Chapter 7 illustrates the practical aspects of model tuning. It
covers different types of model error, sources of model error, hy-
perparameter tuning, how to set up your data, and how to make
sure your model implementation is correct. In practice, apply-
xvi Preface

ing machine learning is a highly iterative process. We discuss

this before introducing the machine learning algorithm because
it applies to nearly all models. You will use cross-validation or
training/developing/testing split to tune the models presented
in later chapters’.
• Chapters 8-14 introduce different types of models. There is a
myriad of learning algorithms to learn the data patterns. This
book doesn’t cover all of them but presents the most common
ones or the foundational methods.

Who This Book Is For

This book is for readers who want to explore potential data sci-
ence career paths and eventually want to become a data scien-
tist. Traditional data-related practitioners such as statisticians,
business analysts, and data analysts will find this book helpful in
expanding their skills for future data science careers. Undergrad-
uate and graduate students from analytics-related areas will find
this book beneficial to learn real-world data science applications.
Non-mathematical readers will appreciate the reproducibility of
the companion R and python codes.

How to Use This Book

What the book assumes
The first two chapters do not have any prerequisite, and the rest
of the chapters do require R or Python programming experience
and undergraduate level statistics. This book does NOT try to
teach the readers to program in the basic sense. It assumes that
readers have experience with R or Python. If you are new to the
programming languages, you may find the code obscure. We pro-
vide some references in the Complementary Reading section
that can help you fill the gap.
Preface xvii

For some chapters (5, 7 - 12), readers need to know elementary

linear algebra (such as matrix manipulations) and understand ba-
sic statistical concepts (such as correlation and simple linear re-
gression). While the book is biased against complex equations, a
mathematical background is good for the deep dive under the hood
mechanism for advanced topics behind applications.

How to run R and Python code

This book uses R in the main text and provides most of the Python
codes on GitHub.
Use R code. You should be able to repeat the R code in your
local R console or RStudio in all the chapters except for Chapter
4. The code in each chapter is self-sufficient, and you don’t need
to run the code in previous chapters first to run the code in the
current chapter. For code within a chapter, you do need to run
from the beginning. At the beginning of each chapter, there is a
code block for installing and loading all required packages. We also
provide the .rmd notebooks that include the code to make it easier
for you to repeat the code. Refer to this page http://bit.ly/3r7cV4s
for a table with the links to the notebooks.
To repeat the code on big data and cloud platforms [chapter 4], you
need to use Databricks, a cloud data platform. We use Databricks
because:
• It provides a user-friendly web-based notebook environment
that can create a Spark cluster on the fly to run R/Python/S-
cala/SQL scripts
• It has a free community edition that is convenient for teaching
purpose
Follow the instructions in section 4.3 on the process of setting up
and using the spark environment.
Use Python code. We provide python notebooks for all the chap-
ters on GitHub. Refer to this page http://bit.ly/3r7cV4s for a ta-
ble with the links to the notebooks. Like R notebooks, you should
be able to repeat all notebooks in your local machine except for
xviii Preface

Chapter 4 with reasons stated above. An easy way to repeat the

notebook is to import and run in Google Colab. To use Colab, you
only need to log in to your Google account in Chrome Browser. To
load the notebook to your colab, you can do any of the following:
• Click the ”Open in Colab” icon on the top of each linked note-
book using the Chrome Brower. It should load the notebook
and open it in your Colab.

• In your Colab, choose File -> Upload notebook -> GitHub.

Copy-paste the notebook’s link in the box, search, and select
the notebook to load it. For example, you can load the python
notebook for data preprocessing like this:

To repeat the code for big data, like running R notebook, you need
to set up Spark in Databricks. Follow the instructions in section
4.3 on the process of setting up and using the spark environment.
Then, run the “Create Spark Data” notebook to create Spark data
Preface xix

frames. After that, you can run the pyspark notebook to learn how
to use pyspark.

Complementary Reading
If you are new to R, we recommend R for Marketing Research
and Analytics by Chris Chapman and Elea McDonnell Feit. The
book is practical and provides repeatable R code. Part I & II of the
book cover basics of R programming and foundational statistics.
It is an excellent book on marketing analytics.
If you are new to Python, we recommend the Python version
of the book mentioned above, Python for Marketing Research and
Analytics by Jason Schwarz, Chris Chapman, and Elea McDonnell
Feit.
If you want to dive deeper into some of the book’s topics, there
are many places to learn more.
• For machine learning, Python Machine Learning 3rd Edition
by Raschka and Mirjalili is a good book on implementing ma-
chine learning in Python. Apply Predictive Modeling by Kuhn
and Johnston is an applied, practitioner-friendly text using R
package caret .
• For statistics models in R, a recommended book is An Intro-
duction to Statistical Learning (ISL) by James, Witten, Hastie,
and Tibshirani. A more advanced treatment of the topics in ISL
is The Elements of Statistical Learning by Friedman, Tibshi-
rani, and Hastie.

Acknowledgements
About the Authors

Hui Lin is currently a Quant Researcher at Google. Before join-

ing Google, Hui held different roles in data science. She was the
head of data science at Netlify where she built and led the data
science team and a Data Scientist at DuPont where she did a
broad range of predictive analytics and market research analysis.
She is the co-founder of Central Iowa R User Group, blogger of
https://scientistcafe.com/, and 2018 Program Chair of ASA Statis-
tics in Marketing Section. She enjoys making analytics accessible
to a broad audience and teaches tutorials and workshops for prac-
titioners on data science. She holds MS and Ph.D. in statistics
from Iowa State University.
Ming Li is currently a Senior Research Scientist at Amazon and
an Adjunct Instructor of University of Washington. He was the
Chair of Quality & Productivity Section of American Statistical
Association for 2017. He was a Data Scientist at Walmart and
a Statistical Leader at General Electric Global Research Center.
He obtained his Ph.D. in Statistics from Iowa State University
at 2010. With deep statistics background and a few years’ experi-
ence in data science, he has trained and mentored numerous junior
data scientist with different background such as statistician, pro-
grammer, software developer and business analyst. He was also
an instructor of Amazon’s internal Machine Learning University
and was one of the key founding members of Walmart’s Analytics
Rotational Program.

xxi
1
Introduction

Data science is a fast-developing area. This chapter discusses var-

ious aspects of data science. What are the different career tracks
and skills required in data science? How is the data science team
structured? We focus on the demand side of data science from
business and industry requirements and application points of view.
We hope this chapter will provide a complementary perspective to
other data sciences books.

1.1 A brief history of data science

Interest in data science-related careers is at an all-time high and
has exploded in popularity in the last few years. Data scientists
today are from various backgrounds. If someone ran into you ask
what data science is all about, what would you tell them? It is not
an easy question to answer. Data science is one of the areas that
everyone is talking about, but no one can define it well.
Media has been hyping about “Data Science,” “Big Data,” and
“Artificial Intelligence” over the past few years. There is an amus-
ing statement from the internet:

“When you’re fundraising, it’s AI. When you’re hiring, it’s ML.
When you’re implementing, it’s logistic regression.”

1
2 1 Introduction

For outsiders, data science is whatever magic that can get useful
information out of data. Everyone should have heard about big
data. Data science trainees now need the skills to cope with such
big data sets. What are those skills? You may hear about: Hadoop,
a system using Map/Reduce to process large data sets distributed
across a cluster of computers, or hear about Spark, a system builds
atop Hadoop for speeding up the same by loading massive datasets
into shared memory (RAM) across clusters with an additional suite
of machine learning functions for big data. The new skills are for
dealing with organizational artifacts of large data sets beyond a
single computer’s memory or hard disk and the large-scale cluster
computing but not for better solving the real problem. A lot of data
means more sophisticated tinkering with computers, especially a
cluster of computers. The computing and programming skills to
handle big data were the biggest hurdle for traditional analysis
practitioners to be a successful data scientist. However, this hurdle
is significantly reduced with the cloud computing revolution, as
described in Chapter 2. After all, it isn’t the size of the data that’s
important. It’s what you do with it. Your first reaction to all of
this might be some combination of skepticism and confusion. We
want to address this upfront that: we had that exact reaction.
To declutter, let’s start with a brief history of data science. If
you hit up the Google Trends website, which shows search key-
word information over time, and check the term “data science,”
you will find the history of data science goes back a little further
than 2004. The way media describes it, you may feel that ma-
chine learning algorithms are new, and there was never “big” data
before Google. That is not true. There are new and exciting devel-
opments in data science. But many of the techniques we are using
are based on decades of work by statisticians, computer scientists,
mathematicians, and scientists of many other fields.
In the early 19th century, when Legendre and Gauss came up
with the least-squares method for linear regression , probably only
physicists would use it to fit linear regression for their data. Now,
nearly anyone can build linear regression using excel with just a lit-
tle bit of self-guided online training. In 1936, Fisher came up with
1.1 A brief history of data science 3

linear discriminant analysis. In the 1940s, we had another widely

used model – logistic regression . In the 1970s, Nelder and Wed-
derburn formulated a “generalized linear mode (GLM) ” which:

“generalized linear regression by allowing the linear model to

be related to the response variable via a link function and by
allowing the magnitude of the variance of each measurement to
be a function of its predicted value.” [from Wikipedia]

By the end of the 1970s, there was a range of analytical models,

and most of them were linear because computers were not powerful
enough to fit non-linear models until the 1980s.
In 1984, Breiman introduced the classification and regression tree
(CART) , one of the oldest and most utilized classification and
regression techniques(Breiman et al., 1984).
After that, Ross Quinlan came up with more tree algorithms such
as ID3, C4.5, and C5.0. In the 1990s, ensemble techniques (meth-
ods that combine many models’ predictions) began to appear.
Bagging is a general approach that uses bootstrapping in con-
junction with regression or classification model to construct an
ensemble. Based on the ensemble idea, Breiman came up with the
random forest model in 2001 (Breiman, 2001a). In the same year,
Leo Breiman published a paper “Statistical Modeling: The Two
Cultures1 ” (Breiman, 2001b) where he pointed out two cultures in
the use of statistical modeling to get information from data:

(1) Data is from a given stochastic data model

(2) Data mechanism is unknown and people approach the

data using algorithmic model
1
http://www2.math.uu.se/~thulin/mm/breiman.pdf
4 1 Introduction

Most of the classical statistical models are the first type of stochas-
tic data model. Black-box models, such as random forest, GMB,
and deep learning , are algorithmic modeling. As Breiman pointed
out, algorithmic models can be used on large complex data as
a more accurate and informative alternative to stochastic data
modeling on smaller data sets. Those algorithms have developed
rapidly with much-expanded applications in fields outside tradi-
tional statistics. That is one of the most important reasons that
statisticians are not the mainstream of today’s data science, both
in theory and practice. We observe that Python is passing R as the
most commonly used language in data science, mainly due to many
data scientists’ background. Since 2000, the approaches to getting
information out of data have shifted from traditional statistical
models to a more diverse toolbox that includes machine learning
and deep learning models. To help readers who are traditional data
practitioners, we provide both R and Python codes.
What is the driving force behind the shifting trend? John Tukey
identified four forces driving data analysis (there was no “data
science” back to 1962):

1. The formal theories of math and statistics

2. Acceleration of developments in computers and display
devices
3. The challenge, in many fields, of more and ever larger
bodies of data
4. The emphasis on quantification in an ever-wider variety
of disciplines

Tukey’s 1962 list is surprisingly modern. Let’s inspect those points

in today’s context. People usually develop theories way before
they find potential applications. In the past 50 years, statisticians,
mathematicians, and computer scientists have laid the theoretical
groundwork for constructing “data science” today. The develop-
ment of computers enables us to apply the algorithmic models
(which can be very computationally expensive) and deliver results
in a friendly and intuitive way. The striking transition to the in-
ternet and the internet of things generates vast amounts of com-
1.2 Data science role and skill tracks 5

mercial data. Industries have also sensed the value of exploiting

that data. Data science seems sure to be a significant preoccupa-
tion of commercial life in the coming decades. All the four forces
John identified exist today and have been driving data science.
The toolbox and application have been expanding fast, benefiting
from the increasing availability of digitized information and the
possibility of distributing it through the internet. Today, people
apply data science in many areas, including business, health, biol-
ogy, social science, politics, etc. Now data science is everywhere.
But what is today’s data science?

1.2 Data science role and skill tracks

There is a widely diffused Chinese parable about a group of blind
men conceptualizing what the elephant is like by touching it. The
first person, whose hand landed on the trunk, said: “This being is
like a thick snake.” For another one whose hand reached its ear,
it seemed like a fan. Another person whose hand was upon its leg
said the elephant is a pillar-like tree trunk. The blind man who
placed his hand upon its side said: “elephant is a wall.” Another
who felt its tail described it as a rope. The last felt its tusk, stating
the elephant is hard, smooth, and spear.
Data science is the elephant. With the data science hype pick-
ing upstream, many professionals changed their titles to be “Data
Scientist” without any necessary qualifications. Today’s data sci-
entists have vastly different backgrounds, yet each conceptualizes
the elephant based on his/her professional training and applica-
tion area. And to make matters worse, most of us are not even
fully aware of our conceptualizations, much less the uniqueness of
the experience from which they are derived.
6 1 Introduction

“We don’t see things as they are, we see them as we are. [by
Anais Nin]”

It is annoying but true. So, the answer to the question “what

is data science?” depends on who you are talking to. Who may
you be talking to then? Data science has three main skill tracks:
engineering, analysis, and modeling (and yes, the order matters!).
There are some representative skills in each track. Different tracks
and combinations of tracks will define different roles in data sci-
ence. 2
When people talk about all the machine learning and AI algo-
rithms, they often overlook the critical data engineering part that
makes everything possible. Data engineering is the unseen iceberg
under the water surface. Does your company need a data scientist?
You are not ready for a data scientist if you don’t have a data en-
gineer yet. You need to have the ability to get data before making
sense of it. If you only deal with small datasets with formatted
data, you may be able to get by with plain text files such as CSV
(i.e., comma-separated values) or even Excel Spreadsheet. As the
data increasing in volume, variety, and velocity, data engineering
becomes a sophisticated discipline in its own right.

1.2.1 Engineering
Data engineering is the foundation that makes everything else pos-
sible. It mainly involves in building the data pipeline infrastruc-
ture. In the (not that) old day, when data is stored on local servers,
computers, or other devices, building the data infrastructure can
be a massive IT project. It involves the software and the hardware
used to store the data and perform data ETL (i.e., extract, trans-
2
This is based on Industry recommendations for academic data science pro-
grams: https://github.com/brohrer/academic_advisory3 with modifications.
It is a collection of thoughts of different data scientist across industries about
what a data scientist does, and what differentiates an exceptional data scien-
tist.
1.2 Data science role and skill tracks 7

form, and load ) process. As cloud service development, it becomes

the new norm to store and compute data on the cloud. Data en-
gineering today, at its core, is software engineering with data flow
as the focus. The fundamental building block for automation is
maintaining the data pipeline through modular, well-commented
code and version control.

(1) Data environment

Designing and setting up the entire environment to support data

science workflow is the prerequisite for data science projects. It
may include setting up storage in the cloud, Kafka platform,
Hadoop and Spark cluster, etc. Each company has a unique data
condition and need. The data environment will be different de-
pending on the size of the data, update frequency, the complexity
of analytics, compatibility with the back-end infrastructure, and
(of course) budget.

(2) Data management

Automated data collection is a common task that includes parsing

the logs (depending on the stage of the company and the type of
industry you are in), web scraping, API queries, and interrogating
data streams. Determine and construct data schema to support
analytical and modeling needs. Use tools, processes, guidelines to
ensure data is correct, standardized, and documented.

(3) Production

If you want to integrate the model or analysis into the production

system, you have to automate all data handling steps. It involves
the whole pipeline from data access, preprocessing, modeling to fi-
nal deployment. It is necessary to make the system work smoothly
with all existing software stacks. So, it requires to monitor the
system through some robust measures, such as rigorous error han-
dling, fault tolerance, and graceful degradation to make sure the
system is running smoothly and the users are happy.
 8 1 Introduction

1.2.2 Analysis
Analysis turns raw information into insights in a fast and often
exploratory way. In general, an analyst needs to have decent do-
main knowledge, do exploratory analysis efficiently, and present
the results using storytelling.

(1) Domain knowledge

Domain knowledge is the understanding of the organization or

industry where you apply data science. You can’t make sense of
data without context. Some questions about the context are:
• What are the critical metrics for this kind of business?
• What are the business questions?
• What type of data do they have, and what does the data repre-
sent?
• How to translate a business need to a data problem?
• What has been tried before, and with what results?
• What are the accuracy-cost-time trade-offs?
• How can things fail?
• What are other factors not accounted for?
• What are the reasonable assumptions, and what are faulty?
In the end, domain knowledge helps you to deliver the results in an
audience-friendly way with the right solution to the right problem.

(2) Exploratory analysis

This type of analysis is about exploration and discovery. Rigor

conclusion is not a concern, which means the goal is to get in-
sights driven by correlation, not causation. The latter one requires
more advanced statistical skills and hence more time and resource
expensive. Instead, this role will help your team look at as much
data as possible so that the decision-makers can get a sense of
what’s worth further pursuing. It often involves different ways to
slice and aggregate data. An important thing to note here is that
you should be careful not to get a conclusion beyond the data. You
 1.2 Data science role and skill tracks 9

don’t need to write production-level robust codes to perform well

in this role.

(3) Storytelling

Storytelling with data is critical to deliver insights and drive better

decision making. It is the art of telling people what the numbers
signify. It usually requires data summarization, aggregation, and
visualization. It is crucial to answering the following questions be-
fore you begin down the path of creating a data story.
• Who is your audience?
• What do you want your audience to know or do?
• How can you use data to help make your point?
A business-friendly report or an interactive dashboard is the typ-
ical outcome of the analysis.

1.2.3 Modeling
Modeling is a process that dives deeper into the data to discover
the pattern we don’t readily see. A fancy machine learning model
is the first thing that comes to people’s minds when the general
public thinks about data science. Unfortunately, fancy models only
occupy a small part of a typical data scientist’s day-to-day time.
Nevertheless, many of those models are powerful tools.

(1) Supervised learning

In supervised learning, each sample corresponds to a response mea-

surement. There are two flavors of supervised learning: regression
and classification. In regression, the response is a real number, such
as the total net sales in 2017 for a company or the yield of corn
next year for a state. The goal for regression is to approximate the
response measurement as much as possible. In classification, the
response is a class label, such as a dichotomous response of yes/no.
The response can also have more than two categories, such as four
segments of customers. A supervised learning model is a function
that maps some input variables (X) with corresponding parame-
 10 1 Introduction

ters (beta) to a response (y). The modeling process is to adjust the

value of parameters to make the mapping fit the given response.
In other words, it is to minimize the discrepancy between given
responses and the model output. When the response y is a real
value number, it is intuitive to define discrepancy as the squared
difference between model output and the response. When y is cate-
gorical, there are other ways to measure the difference, such as the
area under the receiver operating characteristic curve (i.e., AUC)
or information gain.

(2) Unsupervised learning

In unsupervised learning, there is no response variable. For a

long time, the machine learning community overlooked unsuper-
vised learning except for one called clustering. Moreover, many
researchers thought that clustering was the only form of unsuper-
vised learning. One reason is that it is hard to define the goal
of unsupervised learning explicitly. Unsupervised learning can be
used to do the following:
• Identify a good internal representation or pattern of the input
that is useful for subsequent supervised or reinforcement learn-
ing, such as finding clusters;
• It is a dimension reduction tool that provides compact, low di-
mensional representations of the input, such as factor analysis.
• Provide a reduced number of uncorrelated learned features from
original variables, such as principal component regression.

(3) Customized model development

In most cases, after a business problem is fully translated into a

data science problem, a data scientist needs to use out of the box
algorithms to solve the problem with the right data. But in some
situations, there isn’t enough data to use any machine learning
model, or the question doesn’t fit neatly in the specifications of
existing tools, or the model needs to incorporate some prior domain
knowledge. A data scientist may need to develop new models to
 1.2 Data science role and skill tracks 11

accommodate the subtleties of the problem at hand. For example,

people may use Bayesian models to include domain knowledge as
the modeling process’s prior distribution.
Here is a list of questions that can help you decide the type of
technique to use:
• Is your data labeled? It is straightforward since supervised learn-
ing needs labeled data.
• Do you want to deploy your model at scale? There is a fundamen-
tal difference between building and deploying models. It is like
the difference between making bread and making bread machine.
One is a baker who will mix and bake ingredients according to
recipes to make a variety of bread. One is a machine builder who
builds a machine to automate the process and produce bread at
scale.
• Is your data easy to collect? One of the major sources of cost in
deploying machine learning is collecting, preparing, and clean-
ing the data. Because model maintenance includes continuously
collecting data to keep the model updated. If the data collection
process requires too much human labor, the maintenance cost
can be too high.
• Does your problem have a unique context? If so, you may not be
able to find any off-the-shelf method that can directly apply to
your question and need to customize the model.
What others?
There are some common skills to have, regardless of the role people
have in data science.
• Data Preprocessing: the process nobody wants to go
through yet nobody can avoid
No matter what role you hold in the data science team, you will
have to do some data cleaning, which tends to be the least en-
joyable part of anyone’s job. Data preprocessing is the process of
converting raw data into clean data that is proper to use.
 12 1 Introduction

(1) Data preprocessing for data engineer

Getting data from different sources and dumping them into a data
lake, a dumping ground of amorphous data, is far from the data
schema analyst and scientist would use. A data lake is a storage
repository that stores a vast amount of raw data in its native
format, including XML, JSON, CSV, Parquet, etc. It is a data
cesspool rather than a data lake. The data engineer’s job is to get
a clean schema out of the data lake by transforming and formatting
the data. Some common problems to resolve are:
• Enforce new tables’ schema to be the desired one
• Repair broken records in newly inserted data
• Aggregate the data to form the tables with a proper granularity

(2) Data preprocessing for data analyst and scientist

Not just for a data engineer, preprocessing also occupies a large

portion of data analyst and scientist’s working hours. A facility
and a willingness to do these tasks are a prerequisite for a good
data scientist. If you are lucky as a data scientist, you may end up
spending 50% of your time doing this. If you are like most of us,
you will probably spend over 80% of your working hours wrangling
data.
The data a data scientist gets can still be very rough even if it is
from a nice and clean database that a data engineer sets up. For
example, dates and times are notorious for having many represen-
tations and time zone ambiguity. You may also get market survey
responses from your clients in an excel file where the table title
could be multi-line, or the format does not meet the requirements,
such as using 50% to represent the percentage rather than 0.5. In
many cases, you need to set the data to be the right format before
moving on to analysis.
Even the data is in the right format. There are other issues to solve
before or during analysis and modeling. For example, variables can
have missing values. Knowledge about the data collection process
and what it will be used for is necessary to decide a way to handle
 1.3 What kind of questions can data science solve? 13

the missing. Also, different models have different requirements for

the data. For example, some models may require a consistent scale;
some may be susceptible to outliers or collinearity; some may not
be able to handle categorical variables, and so on. The modeler has
to preprocess the data to make it proper for the specific model.
Most of the people in data science today focus on one of the tracks.
A small number of people are experts on two tracks. People who
are proficient in all three? They are unicorns!

1.3 What kind of questions can data science solve?

1.3.1 Prerequisites
Data science is not a panacea, and there are problems data science
can’t help. It is best to make a judgment as early in the analytical
cycle as possible. First and foremost, we need to tell customers,
clients, and stakeholders honestly and clearly when we think data
analytics can’t answer their question after careful evaluation of
the request, data availability, computation resources, and modeling
details. Often, we can tell them what we can do as an alternative.
It is essential to “negotiate” with others what data science can do
specifically; simply answering “we cannot do what you want” will
end the collaboration. Now let’s see what kind of questions data
science can solve:

1. The question needs to be specific enough

Let us look at the two examples below:

• Question 1: How can I increase product sales?
• Question 2: Is the new promotional tool introduced at the be-
ginning of this year boosting the annual sales of P1197 in Iowa
and Wisconsin? (P1197 is an impressive corn seed product from
DuPont Pioneer)
It is easy to see the difference between the two questions. Question
14 1 Introduction

1 is grammatically correct, but it is not proper for data analysis

to answer. Why? It is too general. What is the response variable
here? Product sales? Which product? Is it annual sales or monthly
sales? What are the candidate predictors? We nearly can’t get any
useful information from the questions.
In contrast, question 2 is much more specific. From the analy-
sis point of view, the response variable is clearly “annual sales
of P1197 in Iowa and Wisconsin”. Even we don’t know all the
predictors, but the variable of interest is “the new promotional
tool introduced early this year.” We want to study the impact of
the promotion on sales. We can start there and figure out other
variables that need to be included in the model by further com-
munication.
As a data scientist, we may start with general questions from cus-
tomers, clients, or stakeholders and eventually get to more spe-
cific and data science solvable questions with a series of communi-
cation, evaluation, and negotiation. Effective communication and
in-depth knowledge about the business problem are essential to
converting a general business question into a solvable analytical
problem. Domain knowledge helps data scientists communicate
using the language other people can understand and obtain the
required information.
However, defining the question and variables involved won’t guar-
antee that we can answer it. For example, we could encounter
this situation with a well-defined supply chain problem. The client
may ask us to estimate the stock needed for a product in a partic-
ular area. Why can’t this question be answered? We can try to fit
various models such as a Multivariate Adaptive Regression Spline
(MARS) model and find a reasonable solution from a modeling
perspective. But it can turn out later that the client’s data is an
estimated value, not the actual observation. There is no good way
for data science to solve the problem with the desired accuracy
with inaccurate data.

2. You need to have sound and relevant data

1.3 What kind of questions can data science solve? 15

One cannot make a silk purse out of a sow’s ear. Data scientists
need data, sound, and relevant data. The supply problem men-
tioned above is a case in point. There were relevant data, but not
sound. All the later analytics based on that data was a building
on sand. Of course, data nearly almost have noise, but it has to be
in a certain range. Generally speaking, the accuracy requirement
for the independent variables of interest and response variable is
higher than others. For the above question 2, it is variables related
to the “new promotion” and “sales of P1197”.
The data has to be helpful for the question. If we want to predict
which product consumers are most likely to buy in the next three
months, we need to have historical purchasing data: the last buy-
ing time, the amount of invoice, coupons, etc. Information about
customers’ credit card numbers, ID numbers, and email addresses
will not help much.
Often, the data quality is more important than the quantity, but
you can not overlook the quantity.
Suppose you can guarantee the quality; usually, the more data, the
better. If we have a specific and reasonable question with sound
and relevant data, then congratulations, we can start playing data
science!

1.3.2 Problem type

Many of the data science books classify various models from a tech-
nical point of view. Such as supervised vs. unsupervised models,
linear vs. nonlinear models, parametric models vs. non-parametric
models, and so on. Here we will continue on a “problem-oriented”
track. We first introduce different groups of real-world problems
and then present which models can answer the corresponding cat-
egory of questions.

1. Description

The primary analytic problem is to summarize and explore a data

set with descriptive statistics (mean, standard deviation, and so
forth) and visualization methods. It is the most straightforward
 16 1 Introduction

problem and yet the most crucial and common one. We will need
to describe and explore the dataset before moving on to a more
complex analysis. For problems such as customer segmentation,
after we cluster the sample, the next step is to figure out each
class’s profile by comparing the descriptive statistics of various
variables. Questions of this kind are:
• What is the annual income distribution?
• Are there any outliers?
• What are the mean active days of different accounts?
Data description is often used to check data, find the appropriate
data preprocessing method, and demonstrate the model results.

2. Comparison

The first common modeling problem is to compare different

groups. Is A better in some way than B? Or more comparisons:
Is there any difference among A, B, and C in a particular aspect?
Here are some examples:
• Are males more inclined to buy our products than females?
• Are there any differences in customer satisfaction in different
business districts?
• Do soybean carrying a particular gene have higher oil content?
For those problems, it usually starts with some summary statis-
tics and visualization by groups. After a preliminary visualization,
you can test the differences between the treatment and control
groups statistically. The commonly used statistical tests are chi-
square test, t-test, and ANOVA. There are also methods using
Bayesian methods. In the biology industry, such as new drug de-
velopment, crop breeding, fixed/random/mixed effect models are
standard techniques.

3. Clustering

Clustering is a widespread problem, and it can answer questions

like:
 1.3 What kind of questions can data science solve? 17

• How many reasonable customer segments are there based on his-

torical purchase patterns?
• How are the customer segments different from each other?
Please note that clustering is unsupervised learning; there are no
response variables. The most common clustering algorithms in-
clude K-Means and Hierarchical Clustering.

4. Classification

For classification problems, there are one or more label columns

to define the ground truth of classes. We use other features of
the training dataset as explanatory variables for model training.
We can use the trained classifier to predict the labels of a new
observation. Here are some example questions:
• Will this customer likely to buy our product?
• Is the borrower going to pay us back?
• Is it spam email or not?
There are hundreds of different classifiers. In practice, we do not
need to try all the models but several models that perform well
generally. For example, the random forest algorithm is usually used
as the baseline model to set model performance expectations.

5. Regression

In general, regression deals with a question like “how much is it?”

and return a numerical answer. It is necessary to coerce the model
results to be 0 or round it to the nearest integer in some cases. It
is still the most common problem in the data science world.
• What will be the temperature tomorrow?
• What is the projected net income for the next season?
• How much inventory should we have?

6. Optimization

Optimization is another common type of problems in data science

to find an optimal solution by tuning a few tune-able variables with
 18 1 Introduction

other non-controllable environmental variables. It is an expansion

of comparison problem and can solve problems such as:
• What is the best route to deliver the packages?
• What is the optimal advertisement strategy to promote a new
product?

1.4 Structure data science team

A vast amount of data has become available and readily accessible
for analysis in many companies across different business sectors
during the past decade. The size, complexity, and speed of incre-
ment of data suddenly beyond the traditional scope of statistical
analysis or business intelligence (i.e., BI) reporting. To leverage
the big data collected, do you need an internal data science team
to be a core competency, or can you outsource it? The answer de-
pends on the problems you want to solve using data. If they are
critical to the business, you can’t afford to outsource it. Also, each
company has its business context, and it needs new kinds of data
as the business grows and uses the results in novel ways. Being a
data-driven organization requires cross-organization commitments
to identify what data each department needs to collect, establish
the infrastructure and process for collecting and maintaining that
data, and standardize how to deliver analytical results. Unfortu-
nately, it is unlikely that an off-the-shelf solution will be flexible
enough to adapt to the specific business context. In general, most
of the companies establish their data science team.
Where should the data science team fit? In general, the data sci-
ence team is organized in three ways.

(1) A standalone team

Data science is an autonomous unit parallel to the other orga-

nizations (such as engineering, product, etc.). The head of data
science reports directly to senior leadership, ideally to the CEO or
 1.4 Structure data science team 19

at least someone who understands data strategy and is willing to

invest and give it what it needs. The advantages of this type of
data organization are:
• The data science team has autonomy and is well-positioned to
tackle whatever problems it deems important to the company.
• It is advantageous for people in the data science team to share
knowledge and grow professionally.
• It provides a clear career path for data science professionals and
shows the company treats data as a first-class asset. So, it tends
to attract and retain top talent people.
The biggest concern of this type of organization is the risk of
marginalization. Data science only has value if data drives ac-
tion, which requires collaboration among data scientists, engineers,
product managers, and other business stakeholders across the or-
ganization. Suppose you have a standalone data science team. It
is critical to choose a data science leader who is knowledgeable
about the applications of data science in different areas and has
strong inter-discipline communication skills. The head of data sci-
ence needs to build a strong collaboration with other departments.
As companies grow, each department prefers to be self-sufficient
and tries to hire its data personal under different titles even when
they can get support from the standalone data science team. That
is why it is unlikely for an already mature company to have a
standalone data science team. If you start your data science team
in the early stage as a startup, it is important that the CEO sets
a clear vision from the beginning and sends out a strong message
to the whole company about accessing data support.

(2) An embedded model

There is still a head of data science, but his/her role is mostly

a hiring manager and coach, and he/she may report to a senior
manager in the engineering department. The data science team
brings in talented people and farms them out to the rest of the
company. In other words, it gives up autonomy to ensure utility.
The advantages are:
 20 1 Introduction

• Data science is closer to its applications.

• There is still a data science group, so it is easy to share knowl-
edge.
• It has high flexibility to allocate data science resources across
the company.
However, there are also concerns.
• It brings difficulty to the management since the designated
team’s lead is not responsible for data science professionals’
growth and happiness. In contrast, data science managers are
not directly vested in their work.
• Data scientists are second-class citizens everywhere, and it is
hard to attract and retain top talent.

(3) Integrated team

There is no data science team. Each team hires its data science
people. For example, a marketing analytics group consists of a
data engineer, data analyst, and data scientist. The team leader
is a marketing manager who has an analytical mind and in-depth
business knowledge. The advantages are apparent.
• Data science resource aligns with the organization very well.
• Data science professionals are first-class members and valued in
their team. The manager is responsible for data science profes-
sionals’ growth and happiness.
• The insights from the data are quickly put into action.
It works well in the short term for both the company and the data
science hires. However, there are also many concerns.
• It sacrifices data science hires’ professional growth since they
work in silos and specialize in a specific application. It is also
difficult to share knowledge across different applied areas.
• It is harder to move people around since they are highly associ-
ated with a specific organization’s specific function.
• There is no career path for data science people, and it is difficult
to retain talent.
1.5 List of potential data science careers 21

There is no universal answer to the best way to organize the data

science team. It depends on the answer to many other questions.
How important do you think the data science team is for your
company? What is the stage of your company when you start
to build a data science team? Are you a startup or a relatively
mature company? How valuable it is to use data to tell the truth,
how dangerous it is to use data to affirm existing opinions.
Data science has its skillset, workflow, tooling, integration process,
and culture. If it is critical to your organization, it is best not to
bury it under any part of the organization. Otherwise, data science
will only serve the need for a specific branch. No matter which way
you choose, be aware of both sides of the coin. If you are looking
for a data science position, it is crucial to know where the data
science team fits.

1.5 List of potential data science careers

As companies learn about using data to help with the business,
there is a continuous specialization of different data science roles.
As a result, the old “data scientist” title is fading, and some other
data science job titles are emerging. The misunderstanding of data
science’s fundamental work leads to confusing job postings and
frustrations for both stakeholders and data scientists. Stakeholders
are frustrated that they aren’t getting what they expect, and data
scientists are frustrated that their talent is not appreciated. We are
glad to see that the change is underway. Here is a list of today’s
data science job titles. Some of them are relatively new, and the
others have been around for some time but now are better defined.

Role Skills
Data infrastructure engineer Go, Python, AWS/Google
Cloud/Azure, logstash, Kafka,
and Hadoop
22 1 Introduction

Role Skills
Data engineer spark/scala, python, SQL,
AWS/Google Cloud/Azure,
Data modeling
BI engineer Tableau/looker/Mode, etc.,
data visualization, SQL,
Python
Data analyst SQL, basic statistics, data
visualization
Data scientist R/Python, SQL, basic +
applied statistics, data
visualization, experimental
design
Research scientist R/Python, advanced statistics
+ experimental design, ML,
research background,
publications, conference
contributions, algorithms
Applied scientist ML algorithm design, often
with an expectation of
fundamental software
engineering skills
Machine Learning Engineer More advanced software
engineering skillset,
algorithms, machine learning
algorithm design, system
design

The above table shows some data science roles and common tech-
nical keywords in job descriptions. Those roles are different in the
following key aspects:
• How much business knowledge is required?
• Does it need to deploy code in the production environment?
• How frequently is data updated?
• How much engineering skill is required?
 1.5 List of potential data science careers 23

• How much math/stat knowledge is needed?

• Does the role work with structured or unstructured data?

FIGURE 1.1: Different roles in data science and the skill require-
ments

Data infrastructure engineers work at the beginning of the data

pipeline. They are hardcore engineers who work in the production
system and usually handle high-frequency data. They are respon-
sible for bringing data of different forms and formats and ensur-
ing data coming in smoothly and correctly. They typically don’t
need to know the data’s business context or how data scientists
will use the data. For example, integrate the company’s services
with AWS/GCP/Azure services and set up an Apache Kafka en-
vironment to streaming the events. They work directly with other
engineers (for example, data engineer and backend engineer). Peo-
ple call the pool of data they put together a data lake, a storage
repository that holds a vast amount of raw data in its native for-
mat until needed. As the number of data sources multiplies, having
data scattered all over in various formats prevents the organiza-
tion from using the data to help with business decisions or building
products. That is when data engineers come to help.
Data engineers transform, clean, and organize the data from the
data lake. They commonly design the schemas, store data in query-
able forms, and build and maintain data warehouses. Since the
database is for non-engineers, data engineers need to know a lit-
tle more about the business and how analytical personals use the
data. They use technologies like Hadoop/Spark. Some of them may
have a basic understanding of machine learning to deploy models
developed by data/research scientists.
24 1 Introduction

BI engineers and data Analysts are close to the business, and hence
they need to know the business context well. The critical differ-
ence is that BI engineers build automated dashboards, so they are
engineers. They are usually experts in SQL and have the engi-
neering skill to write production-level code to construct the later
steam data pipeline and automate their work. Data analysts are
technical but not engineers. They analyze ad hoc data and deliver
the results through presentations. The data is, most of the time,
structured. They need to know coding basics (SQL or sometimes
R/Python) but are rarely asked to write production-level code.
This role was mixed with “data scientist” by many companies but
is now much better refined in mature companies.
The most significant difference between a data analyst and a data
scientist is the requirement of mathematics and statistics. Data
analysts usually don’t need to have a quantitative background or
have an advanced degree. The analytics they do are mostly descrip-
tive with visualizations. Most data scientists have a quant back-
ground and do A/B experiments and sometimes machine learning
models. They mainly handle structured and ad hoc data.
Research scientists are experts who have a research background.
They do rigorous analysis and make causal inferences by fram-
ing experiments and developing hypotheses, and proving whether
they are true or not. They are researchers that can create new
models and publish peer-reviewed papers. Most of the small/mid
companies don’t have this role.
Applied scientist is the role that aims to fill the gap between
data/research scientist and data engineers. They have a decent
scientific background, but they are also experts in applying their
knowledge and implementing solutions at scale. They have a dif-
ferent focus than research scientists. Instead of scientific discovery,
they focus on real-life applications. They usually need to pass a
coding bar.
Machine learning engineers have a more advanced software engi-
neering skillset, understand the efficiency of different algorithms
and system design. They focus on deploying the models. They are
1.5 List of potential data science careers 25

in a niche position. They collaborated with data scientists and

applied scientists to read, reinterpret, and rewrite their proof of
concept notebook or scripts into software that can be deployed.
2
Soft Skills for Data Scientists

We want to start the book with soft skills for data scientists. There
are many university courses, online self-learning modules, and ex-
cellent books that teach technical skills. However, there are much
fewer resources discussing the soft skills for data scientists in detail.
However, soft skills are essential for data scientists to succeed in
their career, especially in the early stage. This chapter also intro-
duces the project cycle and some common pitfalls of data science
projects in real life.

2.1 Comparison between Statistician and Data Scientist

Statistics as a scientific area can be traced back to 1749, and statis-
tician as a career has been around for hundreds of years with
well-established theory and application. Data scientist becomes
an attractive career for only a few years, along with the fact that
data size and variety beyond the traditional statistician’s toolbox
and the fast-growing of computation power. Statistician and data
scientist have a lot in common, but there are also significant dif-
ferences, as highlighted in figure 2.1 below:
Both statisticians and data scientists work closely with data.
For typical traditional statisticians, the data set is usually well-
formatted text files with numbers (i.e., numerical variables) and
labels (i.e., categorical variables). The data’s size is typically small
enough to be loaded in a PC’s memory or be saved in a PC’s hard
disk. Comparing to statisticians, data scientists need to deal with
more varieties of data:

27
 28 2 Soft Skills for Data Scientists

FIGURE 2.1: Comparison of Statistician and Data Scientist

• well-formatted data stored in a database system with a size much

larger than a PC’s memory or hard-disk;
• a huge amount of verbatim text, voice, image, and video;
• real-time streaming data and other types of records.
One unique power of statistics is to make statistical inferences
based on a small set of data. Statisticians, especially in academia,
usually spend most of their time developing models and don’t need
to put too much effort into data cleaning. However, data becomes
relatively abundant recently, and modeling is (often small) part of
the overall effort. Due to open source communities’ active develop-
ment, fitting standard models are not too far away from button-
pushing. Data scientists in industry instead spend a lot of time
preprocessing and wrangling the data before feeding them to the
model.
Unlike statisticians, data scientists often focus on delivering ac-
tionable results and sometimes need to deploy the model to the
production system. The data available for model training can be
too large to be processed in a single computer. From the entire
problem-solving cycle, statisticians are usually not well integrated
 2.2 Beyond Data and Analytics 29

with the production system where data is obtained in real-time,

while data scientists are more embedded in the production system
and closer to the data generation procedures. In summary, statis-
ticians focus more on modeling and usually bring data to models,
while data scientists focus more on data and usually bring models
to data.

2.2 Beyond Data and Analytics

Data scientists usually have a good sense of data and analytics,
but data science projects are much more than that. A data science
project may involve people with different roles, especially in a large
company:
• the business owner or leader who identifies business problem and
value;
• the data owner and computation resource/infrastructure owner
from the IT department;
• a dedicated policy owner to make sure the data and model are un-
der model governance, security and privacy guidelines and laws;
• a dedicated engineering team to implement, maintain and refresh
the model;
• a program manager to ensure the data science project fits into
the overall technical program development and to coordinate all
involved parties to set periodical tasks so that the project meets
the preset milestones and results;
The entire team usually will have multiple rounds of discussion
of resource allocation among groups (i.e., who pay for the data
science project) at the beginning of the project and during the
project.
Effective communication and in-depth domain knowledge about
the business problem are essential requirements for a successful
data scientist. A data scientist may interact with people at vari-
ous levels, from senior leaders who set the corporate strategies to
front-line employees who do the daily work. A data scientist needs
30 2 Soft Skills for Data Scientists

to have the capability to view the problem from 10,000 feet above
the ground and down to the detail to the very bottom. To convert
a business question into a data science problem, a data scientist
needs to communicate using the language other people can un-
derstand and obtain the required information through formal and
informal conversations.
In the entire data science project cycle, including defining, plan-
ning, developing, and implementing, every step needs to get a data
scientist involved to ensure the whole team can correctly determine
the business problem and reasonably evaluate the business value
and success. Corporates are investing heavily in data science and
machine learning, and there is a very high expectation of return
for the investment.
However, it is easy to set an unrealistic goal and inflated estimation
for a data science project’s business impact. The team’s data sci-
entist should lead and navigate the discussions to ensure data and
analytics, not wishful thinking, back the goal. Many data science
projects often over-promise in business value and are too optimistic
on the timeline to delivery. These projects eventually fail by not
delivering the pre-set business impact within the promised time-
line. As data scientists, we need to identify these issues early and
communicate with the entire team to ensure the project has a re-
alistic deliverable and timeline. The data scientist team also needs
to work closely with data owners on different things. For example,
identify a relevant internal and external data source, evaluate the
data’s quality and relevancy to the project, and work closely with
the infrastructure team to understand the computation resources
(i.e., hardware and software) availability. It is easy to create scal-
able computation resources through the cloud infrastructure for a
data science project. However, you need to evaluate the dedicated
computation resources’ cost and make sure it fits the budget.
In summary, data science projects are much more than data and
analytics. A successful project requires a data scientist to lead
many aspects of the project.
2.4 Three Pillars of Knowledge 31

2.3 Three Pillars of Knowledge

It is well known there are three pillars of essential knowledge for
a successful data scientist.

(1) Analytics knowledge and toolsets

A successful data scientist needs to have a strong technical back-

ground in data mining, statistics, and machine learning. The in-
depth understanding of modeling with insight about data enables
a data scientist to convert a business problem to a data science
problem. Many chapters of this book are focusing on analytics
knowledge and toolsets.

(2) Domain knowledge and collaboration

A successful data scientist needs in-depth domain knowledge to un-

derstand the business problem well. For any data science project,
the data scientist needs to collaborate with other team members.
Communication and leadership skills are critical for data scientists
during the entire project cycle, especially when there is only one
scientist. The scientist needs to decide the timeline and impact
with uncertainty.

(3) (Big) data management and (new) IT skills

The last pillar is about computation environment and model im-

plementation in a big data platform. It used to be the most difficult
one for a data scientist with a statistics background (i.e., lack com-
puter science knowledge or programming skills). The good news is
that with the rise of the big data platform in the cloud, it is easier
for a statistician to overcome this barrier. The “Big Data Cloud
Platform” chapter of this book will describe this pillar in detail.
32 2 Soft Skills for Data Scientists

FIGURE 2.2: Three Pillars of Knowledge

2.4 Data Science Project Cycle

A data science project has various stages. Many textbooks and
blogs focus on one or two specific stages, and it is rare to see an
end-to-end life cycle of a data science project. To get a good grasp
of the end-to-end process requires years of real-world experience.
Seeing a holistic picture of the whole cycle helps data scientists to
better prepare for real-world applications. We will walk through
the full project cycle in this section.

2.4.1 Types of Data Science Projects

People often use data science projects to describe any project that
uses data to solve a business problem, including traditional busi-
ness analytics, data visualization, or machine learning modeling.
Here we limit our discussion of data science projects that involve
data and some statistical or machine learning models and exclude
basic analytics or visualization. The business problem itself gives
us the flavor of the project. We can view data as the raw ingredi-
ent to start with, and the machine learning model makes the dish.
2.4 Data Science Project Cycle 33

The types of data used and the final model development define the
different kinds of data science projects.
2.4.1.1 Offline and Online Data

There are offline and online data. Offline data are historical data
stored in databases or data warehouses. With the development of
data storage techniques, the cost to store a large amount of data
is low. Offline data are versatile and rich in general (for example,
websites may track and keep each user’s mouse position, click and
typing information while the user is visiting the website). The data
is usually stored in a distributed system, and it can be extracted
in batch to create features used in model training.
Online data are real-time information that flows to models to make
automatic actions. Real-time data can frequently change (for ex-
ample, the keywords a customer is searching for can change at any
given time). Capturing and using real-time online data requires the
integration of a machine learning model to the production infras-
tructure. It used to be a steep learning curve for data scientists not
familiar with computer engineering, but the cloud infrastructure
makes it much more manageable. Based on the offline and online
data and model properties, we can separate data science projects
into three different categories as described below.

2.4.1.2 Offline Training and Offline Application

This type of data science project is for a specific business problem

that needs to be solved once or multiple times. But the dynamic
and disruptive nature of this type of business problem requires
substantial work every time. One example of such a project is
“whether a brand-new business workflow is going to improve effi-
ciency.” In this case, we often use internal/external offline data and
business insight to build models. The final results are delivered as
a report to answer the specific business question. It is similar to
the traditional business intelligence project but with more focus on
data and models. Sometimes the data size and model complexity
are beyond the capacity of a single computer. Then we need to use
distributed storage and computation. Since the model uses histor-
34 2 Soft Skills for Data Scientists

ical data, and the output is a report, there is no need for real-time
execution. Usually, there is no run-time constraint on the machine
learning model unless the model runs beyond a reasonable time
frame, such as a few days. We can call this type of data science
project “offline training, offline application” project.
2.4.1.3 Offline Training and Online Application

Another type of data science project uses offline data for train-
ing and applies the trained model to real-time online data in the
production environment. For example, we can use historical data
to train a personalized advertisement recommendation model that
provides a real-time ad recommendation. The model training uses
historical offline data. The trained model then takes customers’
online real-time data as input features and run the model in real-
time to provide an automatic action. The model training is very
similar to the “offline training, offline application” project. But to
put the trained model into production, there are specific require-
ments. For example, as features used in the offline training have
to be available online in real-time, the model’s online run-time has
to be short enough without impacting user experience. In most
cases, data science projects in this category create continuous and
scalable business value as the model could run millions of times a
day. We will use this type of data science project to describe the
project cycle.
2.4.1.4 Online Training and Online Application

For some business problems, it is so dynamic that even yesterday’s

data is out of date. In this case, we can use online data to train
the model and apply it in real-time. We call this type of data
science project “online training, online application.” This type of
data science project requires high automation and low latency.

2.4.2 Problem Formulation and Project Planning Stage

A data-driven and fact-based planning stage is essential to ensure a
successful data science project. With the recent big data and data
science hype, there is a high demand for data science projects to
 2.4 Data Science Project Cycle 35

create business value across different business sectors. Usually, the

leaders of an organization are those who initiate the data science
project proposals. This top-down style of data science projects
typically has high visibility with some human and computation
resources pre-allocated. However, it is crucial to understand the
business problem first and align the goal across different teams,
including:

(1) the business team, which may include members from the
business operation team, business analytics, insight, and
metrics reporting team;
(2) the technology team, which may include members from
the database and data warehouse team, data engineering
team, infrastructure team, core machine learning team,
and software development team;
(3) the project, program, and product management team de-
pending on the scope of the data science project.

To start the conversation, we can ask everyone in the team the

following questions :
• What are the pain points in the current business operation?
• What data are available, and how is the quality and quantity of
the data?
• What might be the most significant impacts of a data science
project?
• Is there any positive or negative impact on other teams?
• What computation resources are available for model training and
model execution?
• Can we define key metrics to compare and quantify business
value?
• Are there any data security, privacy, and legal concerns?
• What are the desired milestones, checkpoints, and timeline?
• Is the final application online or offline?
• Are the data sources online or offline?
It is likely to have a series of intense meetings and heated discus-
36 2 Soft Skills for Data Scientists

sions to frame the project reasonably. After the planning stage, we

should be able to define a set of key metrics related to the project,
identify some offline and online data sources, request needed com-
putation resources, draft a tentative timeline and milestones, and
form a team of data scientist, data engineer, software developer,
project manager and members from the business operation. Data
scientists should play a significant role in these discussions. If data
scientists do not lead the project formulation and planning, the
project may not catch the desired timeline and milestones.

2.4.3 Project Modeling Stage

Even though we already set some strategies, milestones, and time-
lines at the problem formulation and project planning stage, data
science projects are dynamic. There could be uncertainties along
the road. As a data scientist, communicating any newly encoun-
tered difficulties or opportunities during the modeling stage to the
entire team is essential to keep the data science project progress.
Data cleaning, data wrangling, and exploratory data analysis are
great starting points toward modeling with the available data
source identified at the planning stage. Meanwhile, abstracting
the business problem to be a set of statistical and machine learn-
ing problems is an iterative process. Business problems can rarely
be solved by using just one statistical or machine learning model.
Using a sequence of methods to decompose the business problem
is one of the critical responsibilities for a senior data scientist. The
process requires iterative rounds of discussions with the business
and data engineering team based on each iteration’s new learnings.
Each iteration includes both data-related and model-related parts.

2.4.3.1 Data Related Part

Data cleaning, data preprocessing, and feature engineering are re-

lated procedures that aim to create usable variables or features for
statistical and machine learning models. A critical aspect of data
related procedures is to make sure the data source we are using is a
good representation of the situation where the final trained model
will be applied. The same representation is rarely possible, and it
is ok to start with a reasonable approximation. A data scientist
2.4 Data Science Project Cycle 37

must be clear on the assumptions and communicate the limita-

tions of biased data with the team and quantify its impact on the
application. In the data-related part, sometimes the available data
is not relevant to the business problem we want to solve. We have
to collect more and relevant data before modeling.
2.4.3.2 Model Related Part

There are different types of statistical and machine learning mod-

els, such as supervised learning, unsupervised learning, and causal
inference. For each type, there are various algorithms, libraries, or
packages readily available. To solve a business problem, we some-
times need to piece together a few methods at the model exploring
and developing stage. This stage also includes model training, val-
idation, and testing to ensure the model works well in the produc-
tion environment (i.e., it can be generalized well and not causing
overfitting). The model selection follows Occam’s razor, choosing
the simplest among a set of compatible models. Before we try com-
plicated models, it is good to get some benchmarks by additional
business rules, common-sense decisions, or standard models (such
as random forest for classification and regression problems).

2.4.4 Model Implementation and Post Production Stage

For offline application data science projects, the end product is
often a detailed report with model results and output. However,
for online application projects, a trained model is just halfway from
the finish line. The offline data is stored and processed in a different
environment from the online production environment. Building the
online data pipeline and implementing machine learning models in
a production environment requires lots of additional work. Even
though recent advance in cloud infrastructure lowers the barrier
dramatically, it still takes effort to implement an offline model in
the online production system. Before we promote the model to
production, there are two more steps to go:

1. Shadow mode
2. A/B testing
38 2 Soft Skills for Data Scientists

A shadow mode is like an observation period when the data

pipeline and machine learning models run as fully functional, but
we only record the model output without any actions. Some peo-
ple call it proof of concept (POC). During the shadow mode, peo-
ple frequently check the data pipeline and model and detect bugs
such as a timeout, missing features, version conflict (for example,
Python 2 vs. Python 3), data type mismatch, etc.
Once the online model passes the shadow mode, A/B testing is
the next stage. During A/B testing, all the incoming observations
are randomly separated into two groups: control and treatment.
The control group will skip the machine learning model, while the
treatment group is going through the machine learning model. Af-
ter that, people monitor a list of pre-defined key metrics during a
specific period to compare the control and treatment groups. The
differences in these metrics determine whether the machine learn-
ing model provides business value or not. Real applications can be
complicated. For example, there can be multiple treatment groups,
or hundreds, even thousands of A/B testing running by different
teams at any given time in the same production environment.
Once the A/B testing shows that the model provides significant
business value, we can put it into full production. It is ideal that
the model runs as expected and continues to offer scalable values.
However, the business can change, and a machine learning model
that works now can break tomorrow, and features available now
may not be available tomorrow. We need a monitoring system to
notify us when one or multiple features change. When the model
performance degrades below a pre-defined level, we need to fine-
tune the parameters and thresholds, re-train the model with more
recent data, add or remove features to improve model performance.
Eventually, any model will fail or retire at some time with a pre-
defined model retirement plan.

2.4.5 Project Cycle Summary

Data science end-to-end project cycle is a complicated process that
requires close collaboration among many teams. The data scientist,
maybe the only scientist in the team, has to lead the planning
2.5 Common Mistakes in Data Science 39

discussion and model development based on data available and

communicate key assumptions and uncertainties. A data science
project may fail at any stage, and a clear end-to-end cycle view of
the project helps avoid some mistakes.

2.5 Common Mistakes in Data Science

Data science projects can go wrong at different stages in many
ways. Most textbooks and online blogs focus on technical mistakes
about machine learning models, algorithms, or theories, such as de-
tecting outliers and overfitting. It is important to avoid these tech-
nical mistakes. However, there are common systematic mistakes
across data science projects that are rarely discussed in textbooks.
In this section, we describe these common mistakes in detail so
that readers can proactively identify and avoid these systematic
mistakes in their data science projects.

2.5.1 Problem Formulation Stage

The most challenging part of a data science project is problem
formulation. Data science project stems from pain points of the
business. The draft version of the project’s goal is relatively vague
without much quantification or is the gut feeling of the leadership
team. Often there are multiple teams involved in the initial project
formulation stage, and they have different views. It is easy to have
misalignment across teams, such as resource allocation, milestone
deliverable, and timeline. Data science team members with tech-
nical backgrounds sometimes are not even invited to the initial
discussion at the problem formulation stage. It sounds ridiculous,
but sadly true that a lot of resources are spent on solving the
wrong problem, the number one systematic common mistake in
data science. Formulating a business problem into the right data
science project requires an in-depth understanding of the business
context, data availability and quality, computation infrastructure,
and methodology to leverage the data to quantify business value.
40 2 Soft Skills for Data Scientists

We see people over-promise about business value all the time,

another common mistake that will fail the project at the begin-
ning. With big data and machine learning hype, leaders across
many industries often have unrealistic high expectations on data
science. It is especially true during enterprise transformation when
there is a strong push to adopt new technology to get value out of
the data. The unrealistic expectations are based on assumptions
that are way off the chart without checking the data availability,
data quality, computation resource, and current best practices in
the field. Even when there is some exploratory analysis by the
data science team at the problem formulation stage, project lead-
ers sometimes ignore their data-driven voice.
These two systematic mistakes undermine the organization’s data
science strategy. The higher the expectation, the bigger the dis-
appointment when the project cannot deliver business value. Data
and business context are essential to formulate the business prob-
lem and set reachable business value. It helps avoid mistakes by
having a strong data science leader with a broad technical back-
ground and letting data scientists coordinate and drive the prob-
lem formulation and set realistic goals based on data and business
context.

2.5.2 Project Planning Stage

Now suppose the data science project is formulated correctly with
a reasonable expectation on the business value. The next step is
to plan the project by allocating resources, setting up milestones
and timelines, and defining deliverables. In most cases, project
managers coordinate different teams involved in the project and
use agile project management tools similar to those in software
development. Unfortunately, the project management team may
not have experience with data science projects and hence fail to ac-
count for the uncertainties at the planning stage. The fundamental
difference between data science projects and other projects leads
to another common mistake: too optimistic about the time-
line. For example, data exploratory and data preparation may
2.5 Common Mistakes in Data Science 41

take 60% to 80% of the total time for a given data science project,
but people often don’t realize that.
When there are a lot of data already collected across the orga-
nization, people assume we have enough data for everything. It
leads to the mistake: too optimistic about data availability
and quality. We need not “big data,” but data that can help us
solve the problem. The data available may be of low quality, and
we need to put substantial effort into cleaning the data before we
can use it. There are “unexpected” efforts to bring the right and
relevant data for a specific data science project. To ensure smooth
delivery of data science projects, we need to account for the “unex-
pected” work at the planning stage. Data scientists all know data
preprocessing and feature engineering is usually the most time-
consuming part of a data science project. However, people outside
data science are not aware of it, and we need to educate other
team members and the leadership team.

2.5.3 Project Modeling Stage

Finally, we start to look at the data and fit some models. One
common mistake at this stage is unrepresentative data. The
model trained using historical data may not generalize to the fu-
ture. There is always a problem with biased or unrepresentative
data. As a data scientist, we need to use data that are closer to
the situation where the model will apply and quantify the impact
of model output in production. Another mistake at this stage is
overfitting and obsession with complicated models. Now,
we can easily get hundreds or even thousands of features, and
the machine learning models are getting more complicated. Peo-
ple can use open source libraries to try all kinds of models and are
sometimes obsessed with complicated models instead of using the
simplest among a set of compatible models with similar results.
The data used to build the models is always somewhat biased or
unrepresentative. Simpler models are better to generalize. It has
a higher chance of providing consistent business value once the
model passes the test and is finally implemented in the production
environment. The existing data and methods at hand may be in-
42 2 Soft Skills for Data Scientists

sufficient to solve the business problem. In that case, we can try

to collect more data, do feature engineering, or develop new mod-
els. However, if there is a fundamental gap between data and the
business problem, the data scientist must make the tough decision
to unplug the project.
On the other hand, data science projects usually have high vis-
ibility and may be initiated by senior leadership. Even after the
data science team provided enough evidence that they can’t de-
liver the expected business value, people may not want to stop the
project, which leads to another common mistake at the modeling
stage: take too long to fail. The earlier we can prevent a fail-
ing project, the better because we can put valuable resources into
other promising projects. It damages the data science strategy,
and everyone will be hurt by a long data science project that is
doomed to fail.

2.5.4 Model Implementation and Post Production Stage

Now suppose we have found a model that works great for the
training and testing data. If it is an online application, we are
halfway. The next is to implement the model, which sounds like
alien work for a data scientist without software engineering expe-
rience in the production system. The data engineering team can
help with model production. However, as a data scientist, we need
to know the potential mistakes at this stage. One big mistake is
missing shadow mode and A/B testing and assuming that
the model performance at model training/testing stays the same in
the production environment. Unfortunately, the model trained and
evaluated using historical data nearly never performs the same in
the production environment. The data used in the offline training
may be significantly different from online data, and the business
context may have changed. If possible, machine learning models
in production should always go through shadow mode and A/B
testing to evaluate performance.
In the model training stage, people usually focus on model perfor-
mance, such as accuracy, without paying too much attention to
the model execution time. When a model runs online in real-time,
2.5 Common Mistakes in Data Science 43

each instance’s total run time (i.e., model latency) should not im-
pact the customer’s user experience. Nobody wants to wait for even
one second to see the results after click the “search” button. In the
production stage, feature availability is crucial to run a real-time
model. Engineering resources are essential for model production.
However, in traditional companies, it is common that a data sci-
ence project fails to scale in real-time applications due to
lack of computation capacity, engineering resources, or non-tech
culture and environment.
As the business problem evolves rapidly, the data and model in
the production environment need to change accordingly, or the
model’s performance deteriorates over time. The online production
environment is more complicated than model training and testing.
For example, when we pull online features from different resources,
some may be missing at a specific time; the model may run into a
time-out zone, and various software can cause the version problem.
We need regular checkups during the entire life of the model cycle
from implementation to retirement. Unfortunately, people often
don’t set the monitoring system for data science projects, and it is
another common mistake: missing necessary online checkup.
It is essential to set a monitoring dashboard and automatic alarms,
create model tuning, re-training, and retirement plans.

2.5.5 Summary of Common Mistakes

The data science project is a combination of art, science, and engi-
neering. A data science project may fail in different ways. However,
the data science project can provide significant business value if
we put data and business context at the center of the project, get
familiar with the data science project cycle and proactively iden-
tify and avoid these potential mistakes. Here is the summary of
the mistakes:
• Solving the wrong problem
• Overpromise on business value
• Too optimistic about the timeline
• Too optimistic about data availability and quality
• Unrepresentative data
44 2 Soft Skills for Data Scientists

• Overfitting and obsession with complicated models

• Take too long to fail
• Missing A/B testing
• Fail to scale in real-time applications
• Missing necessary online checkup
3
Introduction to The Data

Before tackling analytics problem, we start by introducing data to

be analyzed in later chapters.

3.1 Customer Data for A Clothing Company

Our first data set represents customers of a clothing company who
sells products in physical stores and online. This data is typical of
what one might get from a company’s marketing data base (the
data base will have more data than the one we show here). This
data includes 1000 customers:

1. Demography
•age: age of the respondent
•gender: male/female
•house: 0/1 variable indicating if the customer owns a
house or not
2. Sales in the past year
•store_exp: expense in store
•online_exp: expense online
•store_trans: times of store purchase
•online_trans: times of online purchase
3. Survey on product preference

It is common for companies to survey their customers and draw

insights to guide future marketing activities. The survey is as be-
low:

45
46 3 Introduction to The Data

How strongly do you agree or disagree with the following state-

ments:

1. Strong disagree
2. Disagree
3. Neither agree nor disagree
4. Agree
5. Strongly agree
• Q1. I like to buy clothes from different brands
• Q2. I buy almost all my clothes from some of my favorite brands
• Q3. I like to buy premium brands
• Q4. Quality is the most important factor in my purchasing de-
cision
• Q5. Style is the most important factor in my purchasing decision
• Q6. I prefer to buy clothes in store
• Q7. I prefer to buy clothes online
• Q8. Price is important
• Q9. I like to try different styles
• Q10. I like to make decision myself and don’t need too much of
others’ suggestions
There are 4 segments of customers:

1. Price
2. Conspicuous
3. Quality
4. Style

Let’s check it:

str(sim.dat,vec.len=3)

## 'data.frame': 1000 obs. of 19 variables:

## $ age : int 57 63 59 60 51 59 57 57 ...
## $ gender : chr "Female" "Female" "Male" ...
## $ income : num 120963 122008 114202 113616 ...
## $ house : chr "Yes" "Yes" "Yes" ...
3.2 Swine Disease Breakout Data 47

## $ store_exp : num 529 478 491 348 ...

## $ online_exp : num 304 110 279 142 ...
## $ store_trans : int 2 4 7 10 4 4 5 11 ...
## $ online_trans: int 2 2 2 2 4 5 3 5 ...
## $ Q1 : int 4 4 5 5 4 4 4 5 ...
## $ Q2 : int 2 1 2 2 1 2 1 2 ...
## $ Q3 : int 1 1 1 1 1 1 1 1 ...
## $ Q4 : int 2 2 2 3 3 2 2 3 ...
## $ Q5 : int 1 1 1 1 1 1 1 1 ...
## $ Q6 : int 4 4 4 4 4 4 4 4 ...
## $ Q7 : int 1 1 1 1 1 1 1 1 ...
## $ Q8 : int 4 4 4 4 4 4 4 4 ...
## $ Q9 : int 2 1 1 2 2 1 1 2 ...
## $ Q10 : int 4 4 4 4 4 4 4 4 ...
## $ segment : chr "Price" "Price" "Price" ...

Refer to Appendix for the simulation code.

3.2 Swine Disease Breakout Data

The swine disease data includes simulated 120 survey questions
from 800 farms. There are three choices for each question. The out-
break status for the 𝑖𝑡ℎ farm is generated from a 𝐵𝑒𝑟𝑛𝑜𝑢𝑙𝑙𝑖(1, 𝑝𝑖 )
distribution with 𝑝𝑖 being a function of the question answers:
𝑝𝑖
𝑙𝑛( ) = 𝛽 0 + Σ𝐺 T
𝑔=1 xi,g �g
1 − 𝑝𝑖

where 𝛽0 is the intercept, xi,g is a three-dimensional indication

vector for question answer and (𝛽𝑔 ) is the parameter vector corre-
sponding to the 𝑔𝑡ℎ predictor. Three types of questions are consid-
ered regarding their effects on the outcome. The first forty survey
questions are important questions such that the coefficients of the
three answers to these questions are all different:

�g = (1, 0, −1) × 𝛾, 𝑔 = 1, … , 40
48 3 Introduction to The Data

The second forty survey questions are also important questions

but only one answer has a coefficient that is different from the
other two answers:

�g = (1, 0, 0) × 𝛾, 𝑔 = 41, … , 80

The last forty survey questions are also unimportant questions

such that all three answers have the same coefficients:

�g = (0, 0, 0) × 𝛾, 𝑔 = 81, … , 120

The baseline coefficient 𝛽0 is set to be − 403 𝛾 so that on average a

farm have 50% of chance to have an outbreak. The parameter 𝛾 in
the above simulation is set to control the strength of the questions’
effect on the outcome. In this simulation study, we consider the
situations where 𝛾 = 0.1, 0.25, 0.5, 1, 2. So the parameter settings
are:

40
�T = ⎛
⎜ , 1, 0, −1, ..., 1, 0, 0 , ..., 0, 0, 0 , ..., 0, 0, 0 ⎞⎟∗𝛾
3 ⏟ ⏟ ⏟ ⏟
⎝ 𝑞𝑢𝑒𝑠𝑡𝑖𝑜𝑛 1 𝑞𝑢𝑒𝑠𝑡𝑖𝑜𝑛 41 𝑞𝑢𝑒𝑠𝑡𝑖𝑜𝑛 81 𝑞𝑢𝑒𝑠𝑡𝑖𝑜𝑛 120⎠

For each value of 𝛾, 20 data sets are simulated. The bigger 𝛾 is,
the larger the corresponding parameter. We provided the data sets
with 𝛾 = 2. Let’s check the data:

disease_dat <- read.csv("http://bit.ly/2KXb1Qi")

# only show the last 7 columns here
head(subset(disease_dat,select=c("Q118.A","Q118.B","Q119.A",
"Q119.B","Q120.A","Q120.B","y")))

## Q118.A Q118.B Q119.A Q119.B Q120.A Q120.B y

## 1 1 0 0 0 0 1 1
## 2 0 1 0 1 0 0 1
## 3 1 0 0 0 1 0 1
## 4 1 0 0 0 0 1 1
## 5 1 0 0 0 1 0 0
3.4 MNIST Dataset 49

## 6 1 0 0 1 1 0 1

Here y indicates the outbreak situation of the farms. y=1 means

there is an outbreak in 5 years after the survey. The rest columns
indicate survey responses. For example Q120.A = 1 means the re-
spondent chose A in Q120. We consider C as the baseline.
Refer to Appendix for the simulation code.

3.3 MNIST Dataset

The MNIST dataset is a popular dataset for image classification
machine learning model tutorials. It is conveniently included in the
Keras library and ready to be loaded with build-in functions for
analysis. The WIKI page of MNIST provides a detailed descrip-
tion of the dataset: https://en.wikipedia.org/wiki/MNIST_database. It
contains 70,000 images of handwritten digits from American Cen-
sus Bureau employees and American high school students. There
are 60,000 training images and 10,000 testing images. Each image
has a resolution of 28x28, and the numerical pixel values are in
greyscale. Each image is represented by a 28x28 matrix with each
element of the matrix an integer between 0 and 255. The label
of each image is the intended digit of the handwritten image be-
tween 0 and 9. We cover the detailed steps to explore the MNIST
dataset in the R and Python notebooks. A sample of the dataset
is illustrated in the figure below 1 :

3.4 IMDB Dataset

The IMDB dataset (http://ai.stanford.edu/~amaas/data/sentiment/)
is a popular dataset for text and language-related machine learning
tutorials. It is also conveniently included in the Keras library, and
1
The image is from https://en.wikipedia.org/wiki/File:MnistExamples.png
50 3 Introduction to The Data

FIGURE 3.1: Sample of MNIST dataset

there are a few build-in functions in Keras for data loading and
pre-processing. It contains 50,000 movie reviews (25,000 in training
and 25,000 in testing) from IMDB, as well as each movie review’s
binary sentiment: positive or negative. The raw data contains the
text of each movie review, and it has to be pre-processed before
being fitted with any machine learning models. By using Keras’s
built-in functions, we can easily get the processed dataset (i.e.,
a numerical data frame) for machine learning algorithms. Keras’
build-in functions perform the following tasks to convert the raw
review text into a data frame:

1. Convert text data into numerical data. Machine learning

models cannot work with raw text data directly, and we
have to convert text into numbers. There are many differ-
ent ways for the conversion and Keras’ build-in function
uses each word’s rank of frequency in the entire train-
ing dataset to replace the raw text in both the training
and testing dataset. For example, the 10th most frequent
word is replaced by integer 10. There are a few additional
setups for this process, including:
a. Skip top frequent words. We usually skip a few top
frequent words as they are mainly stopwords like
“the” “and” or “a,” which usually do not provide
much information. There is a parameter in the
3.4 IMDB Dataset 51

build-in function to specify how many top words to

skip.

b. Set the maximum number of unique words. The en-

tire vocabulary of the unique words in the training
dataset may be large, and many of them have very
low frequencies such as just appearing once in the
entire training dataset. To keep the size of the vocab-
ulary, we can also set up the maximum number of
the unique words using Keras’ built-in function such
that any words with least frequencies will be replaced
with a special index such as “2”.
2. Padding or truncation to keep all the reviews to be
the same length. For most machine learning models, the
algorithms expect to see the same number of features
(i.e., same number of input columns in the data frame).
There is a parameter in the Keras build-in function to
set the maximum number of words in each review (i.e.,
max_length). For reviews that have less than max_legth
words, we pad them with “0”. For reviews that have more
than max_length words, we truncate them.

After the above pre-processing, each review is represented by one

row in the data frame. There is one column for the binary posi-
tive/negative sentiment, and max_length columns input features
converted from the raw review text. In the corresponding R and
Python notebooks, we will go over the details of the data pre-
processing using Keras’ built-in functions.
4
Big Data Cloud Platform

Data has been statisticians and analysts’ friend for hundreds of

years. Tabulated data are the most common format that we use
daily. People used to store data on papers, tapes, diskettes, or hard
drives. Only recently, with the development of computer hardware
and software, the volume, variety, and speed of the data exceed
the capacity of a traditional statistician or analyst. So using data
becomes a science that focuses on the question: how can we store,
access, process, analyze the massive amount of data and provide
actionable insights? In the past few years, by utilizing commod-
ity hardware and open-source software, people created a big data
ecosystem for data storage, data retrieval, and parallel computa-
tion. Hadoop and Spark have become a popular platform that en-
ables data scientists, statisticians, and analysts to access the data
and to build models. Programming skills in the big data platform
have been an obstacle for a traditional statistician or analyst to
become a successful data scientist. However, cloud computing re-
duces the difficulty significantly. The user interface of the data
platform is much more friendly today, and people push mush of
the technical details to the background. Today’s cloud systems also
enable quick implementation of the production environment. Now
data science emphasizes more on the data itself, models and algo-
rithms on top of the data, rather than the platform, infrastructure
and low-level programming such as Java.

53
54 4 Big Data Cloud Platform

4.1 Power of Cluster of Computers

We are familiar with our laptop/desktop computers which have
three main components to do data computation: (1) Hard disk,
(2) Memory, and (3) CPU.
The data and codes stored in the hard disk have specific features
such as slow to read and write, and large capacity of around a
few TB in today’s market. Memory is fast to read and write but
with small capacity in the order of a few dozens of GB in today’s
market. CPU is where all the computation happens.

FIGURE 4.1: Single computer (left) and a cluster of computers

(right)

For statistical software such as R, the amount of data it can process

is limited by the computer’s memory. The memory of computers
before 2000 is less than 1 GB. The memory capacity grows way
slower than the amount of the data. Now it is common that we
need to analyze data far beyond the capacity of a single computer’s
memory, especially in an enterprise environment. Meanwhile, as
the data size increases, to solve the same problem (such as regres-
sions), the computation time is growing faster than linear. Using
a cluster of computers become a common way to solve a big data
problem. In figure 4.1 (right), a cluster of computers can be viewed
as one powerful machine with memory, hard disk and CPU equiv-
4.2 Evolution of Cluster Computing 55

alent to the sum of individual computers. It is common to have

hundreds or even thousands of nodes for a cluster.
In the past, users need to write code (such as MPI) to distribute
data and do parallel computing. Fortunately, with the recent new
development, the cloud environment for big data analysis is more
user-friendly. As data is often beyond the size of the hard disk,
the dataset itself is stored across different nodes (i.e., the Hadoop
system). When doing analysis, the data is distributed across dif-
ferent nodes, and algorithms are parallel to leverage corresponding
nodes’ CPUs to compute (i.e., the Spark system).

4.2 Evolution of Cluster Computing

Using computer clusters to solve general-purpose data and an-
alytics problems needs a lot of effort if we have to specifically
control every element and steps such as data storage, memory
allocation, and parallel computation. Fortunately, high tech com-
panies and open source communities have developed the entire
ecosystem based on Hadoop and Spark. Users need only to know
high-level scripting languages such as Python and R to leverage
computer clusters’ distributed storage, memory and parallel com-
putation power.

4.2.1 Hadoop
The very first problem internet companies face is that a lot of data
has been collected and how to better store these data for future
analysis. Google developed its own file system to provide efficient,
reliable access to data using large clusters of commodity hard-
ware. The open-source version is known as Hadoop Distributed
File System (HDFS). Both systems use Map-Reduce to allocate
computation across computation nodes on top of the file system.
Hadoop is written in Java and writing map-reduce job using Java
is a direct way to interact with Hadoop which is not familiar to
many in the data and analytics community. To help better use
56 4 Big Data Cloud Platform

the Hadoop system, an SQL-like data warehouse system called

Hive, and a scripting language for analytics interface called Pig
were introduced for people with analytics background to inter-
act with Hadoop system. Within Hive, we can create user-defined
functions through R or Python to leverage the distributed and
parallel computing infrastructure. Map-reduce on top of HDFS is
the main concept of the Hadoop ecosystem. Each map-reduce op-
eration requires retrieving data from hard disk, then performing
the computation, and storing the result onto the disk again. So,
jobs on top of Hadoop require a lot of disk operation which may
slow down the entire computation process.

4.2.2 Spark
Spark works on top of a distributed file system including HDFS
with better data and analytics efficiency by leveraging in-memory
operations. Spark is more tailored for data processing and analytics
and the need to interact with Hadoop directly is greatly reduced.
The spark system includes an SQL-like framework called Spark
SQL and a parallel machine learning library called MLib. Fortu-
nately for many in the analytics community, Spark also supports
R and Python. We can interact with data stored in a distributed
file system using parallel computing across nodes easily with R and
Python through the Spark API and do not need to worry about
lower-level details of distributed computing. We will introduce how
to use an R notebook to drive Spark computations.

4.3 Introduction of Cloud Environment

Even though Spark provides a solution for big data analytics, the
maintenance of the computing cluster and Spark system requires a
dedicated team. Historically for each organization the IT depart-
ments own the hardware and the regular maintenance. It usually
takes months for a new environment to be built and the cost is
high. Luckily, the time to deployment and cost are dramatically
4.3 Introduction of Cloud Environment 57

down due to the cloud computation trend. Now we can create a

Spark computing cluster in the cloud in a few minutes with the
desired configuration and the user only pay when the cluster is
up. Cloud computing environments enable smaller organizations
to adopt big data analytics.
There are many cloud computing environments such as Amazon’s
AWS, Google cloud and Microsoft Azure which provide a com-
plete list of functions for heavy-duty enterprise applications. For
example, Netflix runs its business entirely on AWS without own-
ing any data centers. For beginners, however, Databricks provides
an easy to use cloud system for learning purposes. Databricks is a
company founded by the creators of Apache Spark and it provides
a user-friendly web-based notebook environment that can create
a Spark cluster on the fly to run R/Python/Scala/SQL scripts.
We will use Databricks’ free community edition to run demos in
this book. Please note, to help readers to get familiar with the
Databricks cloud system, the content of this section is partially
adopted from the following web pages:
• https://docs.databricks.com/user-guide/faq/sparklyr.html
• http://spark.rstudio.com/index.html

4.3.1 Open Account and Create a Cluster

Anyone can apply for a free Databrick account through https:
//databricks.com/try-databricks and please make sure to choose
the “COMMUNITY EDITION” which does not require credit
card information and will always be free. Once the community
edition account is open and activated. Users can create a cluster
computation environment with Spark. The computing cluster as-
sociated with community edition account is relatively small, but
it is good enough for running all the examples in this book. The
main user interface to the computing environment is notebook:
a collection of cells that contains formatted text or codes. When
a new notebook is created, user will need to choose the default
programming language type (i.e. Python, R, Scala, or SQL) and
every cells in the notebook will assume the default programming
language. However, user can easily override the default selection of
58 4 Big Data Cloud Platform

programming language by adding %sql, %python, %r or %scala at the

first line of each cell to indicate the programming language in that
cell. Allowing running different cells with different programming
language in the same notebook enable user to have the flexibility
to choose the best tools for each task. User can also define a cell
to be markdown cell by adding %md at the first line of the cell. A
markdown cell does not performance computation and it is just
a cell to show formatted text. Well separated cells with compu-
tation, graph and formatted text enable user to create easy to
maintain reproducible reports. The link to a video showing how
to open Databricks account, how to create a cluster, and how to
create notebooks is included in the book’s website.

4.3.2 R Notebook
For this book, we will use R notebook for examples and demos
and the corresponding Python notebook will be available online
too. For an R notebook, it contains multiple cells, and, by default,
the content within each cell are R scripts. Usually, each cell is a
well-managed segment of a few lines of codes that accomplish a
specific task. For example, Figure 4.2 shows the default cell for
an R notebook. We can type in R scripts and comments same as
we are using R console. By default, only the result from the last
line will be shown following the cell. However, you can use print()
function to output results for any lines. If we move the mouse to
the middle of the lower edge of the cell below the results, a “+”
symbol will show up and click on the symbol will insert a new
cell below. When we click any area within a cell, it will make it
editable and you will see a few icons on the top right corn of the
cell where we can run the cell, as well as add a cell below or above,
copy the cell, cut the cell etc. One quick way to run the cell is
Shift+Enter when the cell is chosen. User will become familiar with
the notebook environment quickly.

4.3.3 Markdown Cells

For an R notebook, every cell by default will contain R scripts.
But if we put %md, %sql or %python at the first line of a cell, that
4.4 Introduction of Cloud Environment 59

FIGURE 4.2: Example of R Notebook

cell becomes Markdown cell, SQL script cell, and Python script
cell accordingly. For example, Figure 4.3 shows a markdown cell
with scripts and the actual appearance when exits editing mode.
Markdown cell provides a straightforward way to descript what
each cell is doing as well as what the entire notebook is about. It
is a better way than a simple comment within the code.

FIGURE 4.3: Example of Markdown Cell

60 4 Big Data Cloud Platform

4.4 Leverage Spark Using R Notebook

R is a powerful tool for data analysis given the data can be fit
into memory. Because of the memory bounded dataset limit, R it-
self cannot be used directly for big data analysis where the data is
likely stored in Hadoop and Spark system. By leverage the sparklyr
package created by RStudio, we can use Databricks’ R notebook
to analyze data stored in the Spark system. As the data are stored
across different nodes, Spark enables parallel computation using
the collection of memory and CPU across all nodes. The fundamen-
tal data element in the Spark system is called Spark DataFrames
(SDF). In this section, we will illustrate how to use Databricks’ R
notebook for big data analysis on top of the Spark environment
through sparklyr package.
Install pacakge
First, we need to install sparklyr package which enables the connec-
tion between master or local node to Spark cluster environments.
As it will install more than 10 dependencies, it may take more
than 10 minutes to finish. Be patient while it is installing! Once
the installation finishes, load the sparklyr package as illustrated
by the following code:

# Install sparklyr
if (!require("sparklyr")) {
install.packages("sparklyr")
}
# Load sparklyr package
library(sparklyr)

Create a Spark Connection

Once the library is loaded, we need to create a Spark Connection to
link the computing node (i.e. local node) running the R notebook
to the Spark environment. Here we use the "databricks" option for
parameter method which is specific for databricks’ cloud system. In
4.4 Leverage Spark Using R Notebook 61

other enterprise environments, please consult your administrator

for details. The Spark Connection (i.e. sc) is the pipe to connect R
notebook in the local node with the Spark Cluster. We can think of
the R notebook is running on a local node that has its memory and
CPU; the Spark system has a cluster of connected computation
nodes, and the Spark Connection creates a mechanism to connect
both systems. The Spark Connection can be established with:

# create a sparklyr connection

sc <- spark_connect(method = "databricks")

To simplify the learning process, let us use a very familiar small

dataset: the iris dataset. It is part of the dplyr library and we can
load that library to use the iris data frame. Now the iris dataset
is still on the local node where the R notebook is running on. And
we can check the first a few lines of the iris dataset using the code
below:

library(dplyr)

##
## Attaching package: 'dplyr'

## The following objects are masked from 'package:stats':

##
## filter, lag

## The following objects are masked from 'package:base':

##
## intersect, setdiff, setequal, union

head(iris)

## Sepal.Length Sepal.Width Petal.Length Petal.Width

## 1 5.1 3.5 1.4 0.2
## 2 4.9 3.0 1.4 0.2
## 3 4.7 3.2 1.3 0.2
62 4 Big Data Cloud Platform

## 4 4.6 3.1 1.5 0.2

## 5 5.0 3.6 1.4 0.2
## 6 5.4 3.9 1.7 0.4
## Species
## 1 setosa
## 2 setosa
## 3 setosa
## 4 setosa
## 5 setosa
## 6 setosa

IMPORTANT - Copy Data to Spark Environment

In real applications, the data set may be massive and cannot fit in a
single hard disk and most likely such data are already stored in the
Spark system. If the data is already in Hadoop/Spark ecosystem in
the form of SDF, we can create a local R object to link to the SDF
by the tbl() function where my_sdf is the SDF in the Spark system,
and my_sdf_tbl is the R local object that referring to my_sdf:

my_sdf_tbl <- tbl(sc = sc, my_sdf)

As we just created a brand new Spark computing environment,

there is no SDF in the system yet. We will need to copy a local
dataset to the Spark environment. As we have already created the
Spark Connection sc, it is easy to copy data to spark system using
sdf_copy_to() function as below:

iris_tbl <- sdf_copy_to(sc = sc, x = iris, overwrite = T)

The above one-line code copies iris dataset from the local node to
Spark cluster environment. “sc” is the Spark Connection we just
created; “x” is the data frame that we want to copy; “overwrite”
is the option whether we want to overwrite the target object if
the same name SDF exists in the Spark environment. Finally,
sdf_copy_to() function will return an R object representing the
copied SDF (i.e. creating a “pointer” to the SDF such that we
can refer iris_tbl in the R notebook to operate iris SDF). Now
4.4 Leverage Spark Using R Notebook 63

irir_tblin the local R environment can be used to refer to the iris

SDF in the Spark system.
To check whether the iris data was copied to the Spark environ-
ment successfully or not, we can use src_tbls() function to the
Spark Connection (sc):

## code to return all the data frames associated with sc

src_tbls(sc)

Analyzing the Data

Now we have successfully copied the iris dataset to the Spark
environment as a SDF. This means that iris_tbl is an R object
representing the iris SDF and we can use iris_tbl in R to refer
the iris dataset in the Spark system (i.e. the iris SDF). With
the sparklyr packages, we can nearly all the functions in dplyr
to Spark DataFrame directly through iris_tbl, same as we are
applying dplyr functions to a local R data frame in our laptop.
For example, we can use the %>% operator to pass iris_tbl to the
count() function:

iris_tbl %>% count

or using the head() function to return the first few rows in iris_tbl:

head(iris_tbl)

or applying more advanced data manipulation directly to iris_tbl:

iris_tbl %>%
mutate(Sepal_Width = round(Sepal_Width * 2) / 2) %>%
group_by(Species, Sepal_Width) %>%
summarize(count = n(), Sepal_Length = mean(Sepal_Length),
stdev = sd(Sepal_Length))

Collect Results Back to Master Node

64 4 Big Data Cloud Platform

Even though we can run nearly all of the dplyr functions on SDF,
we cannot apply functions from other packages directly to SDF
(such as ggplot()). For functions that can only work on local R
data frames, we must copy the SDF back to the local node as an
R data frame. To copy SDF back to the local node, we use the
collect() function. The following code using collect() will collect
the results of a few operations and assign the collected data to
iris_summary, a local R data frame:

iris_summary <- iris_tbl %>%

mutate(Sepal_Width = round(Sepal_Width * 2) / 2) %>%
group_by(Species, Sepal_Width) %>%
summarize(count = n(),
Sepal_Length = mean(Sepal_Length),
stdev = sd(Sepal_Length)) %>%
collect()

Now, iris_summary is a local R object to the R notebook and we

can use any R packages and functions to it. In the following code,
we will apply ggplot() to it, exactly the same as a stand along R
console:

library(ggplot2)
ggplot(iris_summary, aes(Sepal_Width, Sepal_Length,
color = Species)) +
geom_line(size = 1.2) +
geom_errorbar(aes(ymin = Sepal_Length - stdev,
ymax = Sepal_Length + stdev),
width = 0.05) +
geom_text(aes(label = count), vjust = -0.2, hjust = 1.2,
color = "black") +
theme(legend.position="top")

In most cases, the heavy-duty data preprocessing and aggregation

is done in Spark using functions in ·dplyr·. Once the data is
aggregated, the size is usually dramatically reduced and such re-
4.4 Leverage Spark Using R Notebook 65

duced data can be collected to an R local object for downstream

analysis.
Fit Regression to SDF
One of the advantages of the Spark system is the parallel machine
learning algorithm. There are many statistical and machine learn-
ing algorithms developed to run in parallel across many CPUs with
data across many memory units for SDF. In this example, we have
already uploaded the iris data to the Spark system, and the data
in the SDF can be referred through iris_tbl as in the last section.
The linear regression algorithm implemented in the Spark system
can be called through ml_linear_regression() function. The syntax
to call the function is to define the local R object that representing
the SDF (i.e. iris_tbl (local R object) for iris (SDF)), response
variable (i.e. the y variable in linear regression in the SDF) and
features (i.e. the x variables in linear regression in the SDF). Now,
we can easily fit a linear regression for large dataset far beyond
the memory limit of one single computer, and it is truly scalable
and only constrained by the resource of the Spark cluster. Below
is an illustration of how to fit a linear regression to SDF using R
notebook:

fit1 <- ml_linear_regression(x = iris_tbl, response = "Sepal_Length",

features = c("Sepal_Width", "Petal_Length", "Petal_Width"))
summary(fit1)

In the above code, x is the R object pointing to the SDF; response

is y-variable, features are the collection of explanatory variables.
For this function, both the data and computation are in the Spark
cluster which leverages multiple CPUs, distributed memories and
parallel computing.
Fit a K-means Cluster
Through the sparklyr package, we can use an R notebook to access
many Spark Machine Learning Library (MLlib) algorithms such
as Linear Regression, Logistic Regression, Survival Regression,
Generalized Linear Regression, Decision Trees, Random Forests,
66 4 Big Data Cloud Platform

Gradient-Boosted Trees, Principal Components Analysis, Naive-

Bayes, K-Means Clustering, and a few other methods. Below codes
fit a k-means cluster algorithm:

## Now fit a k-means clustering using iris_tbl data

## with only two out of four features in iris_tbl
fit2 <- ml_kmeans(x = iris_tbl, centers = 3,
features = c("Petal_Length", "Petal_Width"))
# print our model fit
print(fit2)

After fitting the k-means model, we can apply the model to pre-
dict other datasets through sdf_predict() function. Following code
applies the model to iris_tbl again to predict the cluster and col-
lect the results as a local R object (i.e. prediction) using collect()
function:

prediction = collect(sdf_predict(fit2, iris_tbl))

As prediction is a local R object, we can apply any R functions

from any libraries to it. For example:

prediction %>%
ggplot(aes(Petal_Length, Petal_Width)) +
geom_point(aes(Petal_Width, Petal_Length,
col = factor(prediction + 1)),
size = 2, alpha = 0.5) +
geom_point(data = fit2$centers, aes(Petal_Width, Petal_Length),
col = scales::muted(c("red", "green", "blue")),
pch = 'x', size = 12) +
scale_color_discrete(name = "Predicted Cluster",
labels = paste("Cluster", 1:3)) +
labs(x = "Petal Length",
y = "Petal Width",
title = "K-Means Clustering",
4.5 Databases and SQL 67

subtitle = "Use Spark ML to predict cluster

membership with the iris dataset")

So far, we have illustrated

1. the relationship between a local node (i.e. where R note-

book is running) and Spark Clusters (i..e where data are
stored and computation are done);
2. how to copy a local data frame to a Spark DataFrames
(please note if your data is already in Spark environment,
there is no need to copy and we only need to build the
connection. This is likely to be the case for enterprise
environment);
3. how to manipulate Spark DataFrames for data cleaning
and preprocessing through dplyr functions with the instal-
lation of sparklyr package;
4. how to fit statistical and machine learning models to
Spark DataFrame in a truly parallel manner;
5. how to collect information from Spark DataFrames back
to a local R object (i.e. local R data frame) for future
analysis.

These procedures cover the basics of big data analysis that a data
scientist needs to know as a beginner. We have an R notebook on
the book website that contains the contents of this chapter. We
also have a Python notebook on the book website.

4.5 Databases and SQL

4.5.1 History
Databases have been around for many years to efficiently orga-
nize, store, retrieve, and update data systematically. In the past,
statisticians and analysts usually dealt with small datasets stored
in text or Excel files and often did not interact with database sys-
68 4 Big Data Cloud Platform

tems. Students from the traditional statistics department usually

lack the necessary database knowledge. However, as data grow big-
ger, database knowledge becomes essential and required for statis-
ticians, analysts and data scientists in an enterprise environment
where data are stored in some form of database systems. Databases
often contain a collection of tables and the relationship among
these tables (i.e. schema). The table is the fundamental struc-
ture for databases that contain rows and columns similar to data
frames in R or Python. Database management systems (DBMS)
ensure data integration and security in real time operations. There
are many different DBMS such as Oracle, SQL Server, MySQL,
Teradata, Hive, Redshift and Hana. The majority of database op-
erations are very similar among different DBMS, and Structured
Query Language (SQL) is the standard language to use these sys-
tems.
SQL became a standard of the American National Standards In-
stitute (ANSI) in 1986, and of the International Organization for
Standardization (ISO) in 1987. The most recent version is pub-
lished in December 2016. For typical users, fundamental knowl-
edge is nearly the same across all database systems. In addition
to the standard features, each DBMS providers include their own
specific functions and features. So, for the same query, there may
be slightly different implementations (i.e. SQL script) for different
systems. In this section, we use the Databricks’ SQL implementa-
tion (i.e. all the SQL scripts can run in Databricks SQL notebook).
More recent data is stored in a distributed system such as Hive
for disk storage or Hana for in-memory storage. Most relational
databases are row-based (i.e. data for each row are stored closely),
whereas analytics workflows often favor column-based systems
(i.e. data for each column are stored closely). Fortunately, as a
database user, we only need to learn how to write SQL scripts to
retrieve and manipulate data. Even though there are different im-
plantations of SQL across different DBMS, SQL is nearly universal
across relational databases including Hive and Spark, which means
once we know SQL, our knowledge can be transferred among dif-
ferent database systems. SQL is easy to learn even if you do not
4.5 Databases and SQL 69

have previous experience. In this session, we will go over the key

concepts in the database and SQL.

4.5.2 Database, Table and View

A database is a collection of tables that are related to each other.
A database has its own database name and each table has its
name as well. We can think a database is a “folder” where tables
within a database are “files” within the folder. A table has rows
and columns exactly as an R or Python pandas data frame. Each
row (also called record) represents a unique instance of the sub-
ject and each column (also called field or attribute) represents a
characteristic of the subject on the table. For each table, there is
a special column called the primary key which uniquely identifies
each of its records.
Tables within a specific database contain related information and
the schema of a database illustrates all fields in every table as well
as how these tables and fields relate to each other (i.e. the structure
of a database). Tables can be filtered, joined and aggregated to
return specific information. The view is a virtual table composed
of fields from one or more base tables. The view does not store
data and only store table structure. The view is also referred to
as a saved query. The view is typically used to protect the data
stored in the table and users can only query information from a
view and cannot change or update its contents.
We will use two simple tables to illustrate basic SQL operations.
These two tables are from an R dataset which contains the 50
states’ population and income (https://stat.ethz.ch/R-manual/R-
patched/library/datasets/html/state.html). The first table is called
divisions which has two columns: state and division and the first
few rows are shown in the following table:

state division
Alabama East South Central
Alaska Pacific
Arizona Mountain
 70 4 Big Data Cloud Platform

state division
Arkansas West South Central
California Pacific

The second table is called metrics which contains three columns:

state, population and income and first few rows of the table are
shown below:

state population income

Alabama 3615 3624
Alaska 365 6315
Arizona 2212 4530
Arkansas 2110 3378
California 21198 5114

To illustrate missing information, three more rows are added at the

end of the original division table with state Alberta, Ontario, and
Quebec with their corresponding division NULL. We first creat
these two tables and save them as csv files, and then we upload
these two files as Databricks tables.

4.5.3 Basic SQL Statement

After logging into Databricks and creating two tables, we can now
create a notebook and choose the default language of the notebook
to be SQL. There are a few very easy SQL statements to help us
understand the database and table structure:
• show database:
show current databases in the system
• create database db_name: create a new database with name db_name
• drop database db_name: delete database db_name (be careful when
using it!)
• use db_name: set up the current database to be used
• show tables: show all the tables within the currently used
database
 4.5 Databases and SQL 71

• describe tbl_name:show the structure of table with name tbl_name

(i.e. list of column name and data type)
• drop tbl_name: delete a table with name tbl_name (be careful when
using it!)
• select * from metrics limit 10: show the first 10 rows of a table
If you are familiar with a procedural programming language such
as C and FORTRAN or scripting languages such as R and Python,
you may find SQL code a little bit strange. We should view SQL
code by each specific chunk where it defines a specific task. SQL
codes descript a specific task and DBMS will run and finish the
task. SQL does not follow typical procedure program rules and
we can think SQL is “descriptive” (i.e. we describe what we want
using SQL and DBMS figures out how to do it).
4.5.3.1 SELECT Statement

SELECT is the most used statement in SQL, especially for database

users and business analysts. It is used to extract specific informa-
tion (i.e. column or columns) FROM one or multiple tables. It can be
used to combine multiple tables. WHERE can be used in the SELECT
statement to selected rows with specific conditions (i.e. filters). OR-
DER BY can be used in the SELECT statement to order the results in
descending or ascending order of one or multiple columns. We can
use * after SELECT to represent all columns in the table, or specifi-
cally write the column names separated by a comma. Below is the
basic structure of a SELECT statement:

SELECT Col_Name1, Col_Name2

FROM Table_Name
WHERE Specific_Condition
ORDER BY Col_Name1, Col_Name2;

Here Specific_Condition is the typical logical conditions and only

columns with TRUE for this condition will be chosen. For example, if
we want to choose states and its total income where the population
larger than 10000 and individual income less than 5000 with the
result order by state name, we can use the following query:
72 4 Big Data Cloud Platform

select state, income*population as total_income

from metrics
where population > 10000 and income < 5000
order by state

The SELECT statement is used to slicing and dicing the dataset as

well as create new columns of interest (such as total_income) using
basic computation functions.
4.5.3.2 Aggregation Functions and GROUP BY

We can also use aggregation functions in the SELECT statement to

summarize the data. For example, count(col_name) function will re-
turn the total number of not NULL rows for a specific column. Other
aggregation function on numerical values include min(col_name),
max(col_name), avg(col_name). Let’s use the metrics table again to
illustrate aggregation functions. For aggregation function, it takes
all the rows that match WHERE condition (if any) and return
one number. The following statement will calculate the maximum,
minimum, and average population for all states starts with letter
A to E.

select sum(population) as sum_pop, max(population) as

max_pop, min(population) as min_pop, avg(population)
as avg_pop, count(population) as count_pop
from metrics
where substring(state, 1, 1) in ('A', 'B', 'C', 'D', 'E')

The results from the above query only return one row as expected.
Sometimes we want to find the aggregated value based on groups
that can be defined by one or more columns. Instead of writing
multiple SQL to calculate the aggregated value for each group, we
can easily use the GROUP BY to calculate the aggregated value
for each group in the SELECT statement. For example, if we want to
find how many states in each division, we can use the following:
4.5 Databases and SQL 73

select division, count(state) as number_of_states

from divisions
group by division

Another special aggregation function is to return distinct values

for one column or a combination of multiple columns. Simple use
SELECT DISTINCT col_name1, col_name2 in the first line of the SELECT
statement.
4.5.3.3 Join Multiple Tables

The database system is usually designed such that each table con-
tains a piece of specific information and oftentimes we need to
JOIN multiple tables to achieve a specific task. There are few
types typically JOINs: inner join (keep only rows that match the
join condition from both tables), left outer join (rows from inner
join + unmatched rows from the first table), right outer join (rows
from inner join + unmatched rows from the second table) and
full outer join (rows from inner join + unmatched rows from both
tables). The typical JOIN statement is illustrated below:

SELECT a.col_name1 as var1, b.col_name2 as var2

FROM tbl_one as a
LEFT JOIN tabl_two as b
ON a.col_to_match = b.col_to_match

For example, let us join the division table and metrics table to find
what is the average population and income for each division, and
the results order by division names:

select a.division, avg(b.population) as avg_pop,

avg(b.income) as avg_inc
from divisions as a
inner join metrics as b
on a.state = b.state
74 4 Big Data Cloud Platform

group by division
order by division

4.5.3.4 Add More Content into a Table

We can use the INSERT statement to add additional rows into a

particular table, for example, we can add one more row to the
metrics table by using the following query:

insert into metrics

values ('Alberta', 4146, 7370)

4.5.4 Advanced Topics in Database

There are many advanced topics such as how to efficiently query
data using index; how to take care of data integrity when mul-
tiple users are using the same table; algorithm behind data stor-
age (i.e. column-wise or row-wise data storage); how to design the
database schema. Users can learn these advanced topics gradually.
We hope the basic knowledge covered in this section will kick off
the initial momentum to learn SQL. As you can see, it is easy
to write SQL statement to retrieve, join, slice, dice and aggregate
data. The SQL notebook that contains all the above operations is
included in the book’s website.
5
Data Pre-processing

Many data analysis related books focus on models, algorithms and

statistical inferences. However, in practice, raw data is usually not
directly used for modeling. Data preprocessing is the process of
converting raw data into clean data that is proper for modeling.
A model fails for various reasons. One is that the modeler doesn’t
correctly preprocess data before modeling. Data preprocessing can
significantly impact model results, such as imputing missing value
and handling with outliers. So data preprocessing is a very critical
part.

FIGURE 5.1: Data Pre-processing Outline

In real life, depending on the stage of data cleanup, data has the
following types:

1. Raw data
2. Technically correct data
3. Data that is proper for the model
4. Summarized data
5. Data with fixed format

75
 76 5 Data Pre-processing

The raw data is the first-hand data that analysts pull from the
database, market survey responds from your clients, the experi-
mental results collected by the R & D department, and so on.
These data may be very rough, and R sometimes can’t read them
directly. The table title could be multi-line, or the format does not
meet the requirements:
• Use 50% to represent the percentage rather than 0.5, so R will
read it as a character;
• The missing value of the sales is represented by “-” instead of
space so that R will treat the variable as character or factor type;
• The data is in a slideshow document, or the spreadsheet is not
“.csv” but “.xlsx”
• …
Most of the time, you need to clean the data so that R can import
them. Some data format requires a specific package. Technically
correct data is the data, after preliminary cleaning or format con-
version, that R (or another tool you use) can successfully import
it.
Assume we have loaded the data into R with reasonable column
names, variable format and so on. That does not mean the data is
entirely correct. There may be some observations that do not make
sense, such as age is negative, the discount percentage is greater
than 1, or data is missing. Depending on the situation, there may
be a variety of problems with the data. It is necessary to clean the
data before modeling. Moreover, different models have different
requirements on the data. For example, some model may require
the variables are of consistent scale; some may be susceptible to
outliers or collinearity, some may not be able to handle categorical
variables and so on. The modeler has to preprocess the data to
make it proper for the specific model.
Sometimes we need to aggregate the data. For example, add up
the daily sales to get annual sales of a product at different loca-
tions. In customer segmentation, it is common practice to build
a profile for each segment. It requires calculating some statistics
such as average age, average income, age standard deviation, etc.
5.1 Data Cleaning 77

Data aggregation is also necessary for presentation, or for data

visualization.
The final table results for clients need to be in a nicer format than
what used in the analysis. Usually, data analysts will take the
results from data scientists and adjust the format, such as labels,
cell color, highlight. It is important for a data scientist to make
sure the results look consistent which makes the next step easier
for data analysts.
It is highly recommended to store each step of the data and the R
code, making the whole process as repeatable as possible. The R
markdown reproducible report will be extremely helpful for that. If
the data changes, it is easy to rerun the process. In the remainder
of this chapter, we will show the most common data preprocessing
methods.
Load the R packages first:

# install packages from CRAN

p_needed <- c('imputeMissings','caret','e1071','psych',
'car','corrplot')
packages <- rownames(installed.packages())
p_to_install <- p_needed[!(p_needed %in% packages)]
if (length(p_to_install) > 0) {
install.packages(p_to_install)
}

lapply(p_needed, require, character.only = TRUE)

5.1 Data Cleaning

After you load the data, the first thing is to check how many
variables are there, the type of variables, the distributions, and
data errors. Let’s read and check the data:
78 5 Data Pre-processing

sim.dat <- read.csv("http://bit.ly/2P5gTw4")

summary(sim.dat)

age gender income house store_exp

Min. : 16.0 Female:554 Min. : 41776 No :432 Min. : -500
1st Qu.: 25.0 Male :446 1st Qu.: 85832 Yes:568 1st Qu.: 205
Median : 36.0 Median : 93869 Median : 329
Mean : 38.8 Mean :113543 Mean : 1357
3rd Qu.: 53.0 3rd Qu.:124572 3rd Qu.: 597
Max. :300.0 Max. :319704 Max. :50000
NA's :184
online_exp store_trans online_trans Q1 Q2
Min. : 69 Min. : 1.00 Min. : 1.0 Min. :1.0 Min. :1.00
1st Qu.: 420 1st Qu.: 3.00 1st Qu.: 6.0 1st Qu.:2.0 1st Qu.:1.00
Median :1942 Median : 4.00 Median :14.0 Median :3.0 Median :1.00
Mean :2120 Mean : 5.35 Mean :13.6 Mean :3.1 Mean :1.82
3rd Qu.:2441 3rd Qu.: 7.00 3rd Qu.:20.0 3rd Qu.:4.0 3rd Qu.:2.00
Max. :9479 Max. :20.00 Max. :36.0 Max. :5.0 Max. :5.00

Q3 Q4 Q5 Q6 Q7
Min. :1.00 Min. :1.00 Min. :1.00 Min. :1.00 Min. :1.00
1st Qu.:1.00 1st Qu.:2.00 1st Qu.:1.75 1st Qu.:1.00 1st Qu.:2.50
Median :1.00 Median :3.00 Median :4.00 Median :2.00 Median :4.00
Mean :1.99 Mean :2.76 Mean :2.94 Mean :2.45 Mean :3.43
3rd Qu.:3.00 3rd Qu.:4.00 3rd Qu.:4.00 3rd Qu.:4.00 3rd Qu.:4.00
Max. :5.00 Max. :5.00 Max. :5.00 Max. :5.00 Max. :5.00

Q8 Q9 Q10 segment
Min. :1.0 Min. :1.00 Min. :1.00 Conspicuous:200
1st Qu.:1.0 1st Qu.:2.00 1st Qu.:1.00 Price :250
Median :2.0 Median :4.00 Median :2.00 Quality :200
Mean :2.4 Mean :3.08 Mean :2.32 Style :350
3rd Qu.:3.0 3rd Qu.:4.00 3rd Qu.:3.00
Max. :5.0 Max. :5.00 Max. :5.00

Are there any problems? Questionnaire response Q1-Q10 seem

reasonable, the minimum is 1 and maximum is 5. Recall that
 5.2 Data Cleaning 79

the questionnaire score is 1-5. The number of store transactions

(store_trans) and online transactions (online_trans) make sense
too. Things to pay attention are:
• There are some missing values.
• There are outliers for store expenses (store_exp). The maximum
value is 50000. Who would spend $50000 a year buying clothes?
Is it an imputation error?
• There is a negative value ( -500) in store_exp which is not logical.
• Someone is 300 years old.
How to deal with that? Depending on the situation, if the sample
size is large enough and the missing happens randomly, it does
not hurt to delete those problematic samples. Or we can set these
values as missing and impute them instead of deleting the rows.

# set problematic values as missings

sim.dat$age[which(sim.dat$age > 100)] <- NA
sim.dat$store_exp[which(sim.dat$store_exp < 0)] <- NA
# see the results
summary(subset(sim.dat, select = c("age", "store_exp")))

age store_exp
Min. :16.00 Min. : 155.8
1st Qu.:25.00 1st Qu.: 205.1
Median :36.00 Median : 329.8
Mean :38.58 Mean : 1358.7
3rd Qu.:53.00 3rd Qu.: 597.4
Max. :69.00 Max. :50000.0
NA's :1 NA's :1

Now let’s deal with the missing values in the data.

80 5 Data Pre-processing

5.2 Missing Values

You can write a whole book about missing value. This section
will only show some of the most commonly used methods without
getting too deep into the topic. Chapter 7 of the book by De Waal,
Pannekoek and Scholtus (de Waal et al., 2011) makes a concise
overview of some of the existing imputation methods. The choice
of specific method depends on the actual situation. There is no
best way.
One question to ask before imputation: Is there any auxiliary in-
formation? Being aware of any auxiliary information is critical.
For example, if the system set customer who did not purchase as
missing, then the real purchasing amount should be 0. Is missing
a random occurrence? If so, it may be reasonable to impute with
mean or median. If not, is there a potential mechanism for the
missing data? For example, older people are more reluctant to dis-
close their ages in the questionnaire, so that the absence of age
is not completely random. In this case, the missing values need
to be estimated using the relationship between age and other in-
dependent variables. For example, use variables such as whether
they have children, income, and other survey questions to build a
model to predict age.
Also, the purpose of modeling is important for selecting imputa-
tion methods. If the goal is to interpret the parameter estimate
or statistical inference, then it is important to study the missing
mechanism carefully and to estimate the missing values using non-
missing information as much as possible. If the goal is to predict,
people usually will not study the absence mechanism rigorously
(but sometimes the mechanism is obvious). If the absence mecha-
nism is not clear, treat it as missing at random and use mean, me-
dian, or k-nearest neighbor to impute. Since statistical inference is
sensitive to missing values, researchers from survey statistics have
conducted in-depth studies of various imputation schemes which
focus on valid statistical inference. The problem of missing values
in the prediction model is different from that in the traditional
5.2 Missing Values 81

survey. Therefore, there are not many papers on missing value im-
putation in the prediction model. Those who want to study further
can refer to Saar-Tsechansky and Provost’s comparison of differ-
ent imputation methods (M and F, 007b) and De Waal, Pannekoek
and Scholtus’ book (de Waal et al., 2011).

5.2.1 Impute missing values with median/mode

In the case of missing at random, a common method is to impute
with the mean (continuous variable) or median (categorical vari-
ables). You can use impute() function in imputeMissings package.

# save the result as another object

demo_imp <- impute(sim.dat, method = "median/mode")
# check the first 5 columns
# there is no missing values in other columns
summary(demo_imp[, 1:5])

age gender income house store_exp

Min. :16.00 Female:554 Min. : 41776 No :432 Min. : 155.8
1st Qu.:25.00 Male :446 1st Qu.: 87896 Yes:568 1st Qu.: 205.1
Median :36.00 Median : 93869 Median : 329.8
Mean :38.58 Mean :109923 Mean : 1357.7
3rd Qu.:53.00 3rd Qu.:119456 3rd Qu.: 597.3
Max. :69.00 Max. :319704 Max. :50000.0

After imputation, demo_imp has no missing value. This method is

straightforward and widely used. The disadvantage is that it does
not take into account the relationship between the variables. When
there is a significant proportion of missing, it will distort the data.
In this case, it is better to consider the relationship between vari-
ables and study the missing mechanism. In the example here, the
missing variables are numeric. If the missing variable is a cate-
gorical/factor variable, the impute() function will impute with the
mode.
You can also use preProcess() in package caret, but it is only for
numeric variables, and can not impute categorical variables. Since
82 5 Data Pre-processing

missing values here are numeric, we can use the preProcess() func-
tion. The result is the same as the impute() function. PreProcess() is
a powerful function that can link to a variety of data preprocessing
methods. We will use the function later for other data preprocess-
ing.

imp <- preProcess(sim.dat, method = "medianImpute")

demo_imp2 <- predict(imp, sim.dat)
summary(demo_imp2[, 1:5])

age gender income house store_exp

Min. :16.00 Female:554 Min. : 41776 No :432 Min. : 155.8
1st Qu.:25.00 Male :446 1st Qu.: 87896 Yes:568 1st Qu.: 205.1
Median :36.00 Median : 93869 Median : 329.8
Mean :38.58 Mean :109923 Mean : 1357.7
3rd Qu.:53.00 3rd Qu.:119456 3rd Qu.: 597.3
Max. :69.00 Max. :319704 Max. :50000.0

5.2.2 K-nearest neighbors

K-nearest neighbor (KNN) will find the k closest samples (Euclid-
ian distance) in the training set and impute the mean of those
“neighbors.”
Use preProcess() to conduct KNN:

imp <- preProcess(sim.dat, method = "knnImpute", k = 5)

# need to use predict() to get KNN result
demo_imp <- predict(imp, sim.dat)
# only show the first three elements
lapply(sim.dat, class)[1:3]

age gender income

Min. :-1.5910972 Female:554 Min. :-1.43989
1st Qu.:-0.9568733 Male :446 1st Qu.:-0.53732
Median :-0.1817107 Median :-0.37606
Mean : 0.0000156 Mean : 0.02389
3rd Qu.: 1.0162678 3rd Qu.: 0.21540
5.2 Missing Values 83

Max. : 2.1437770 Max. : 4.13627

The preProcess() in the first line will automatically ignore non-

numeric columns.
Comparing the KNN result with the previous median imputation,
the two are very different. This is because when you tell the prePro-
cess() function to use KNN (the option method =" knnImpute"), it will
automatically standardize the data. Another way is to use Bagging
tree (in the next section). Note that KNN can not impute samples
with the entire row missing. The reason is straightforward. Since
the algorithm uses the average of its neighbors if none of them has
a value, what does it apply to calculate the mean?
Let’s append a new row with all values missing to the original data
frame to get a new object called temp. Then apply KNN to temp
and see what happens:

temp <- rbind(sim.dat, rep(NA, ncol(sim.dat)))

imp <- preProcess(sim.dat, method = "knnImpute", k = 5)
demo_imp <- predict(imp, temp)

Error in FUN(newX[, i], ...) :

cannot impute when all predictors are missing in the new data point

There is an error saying “cannot impute when all predictors are

missing in the new data point”. It is easy to fix by finding and
removing the problematic row(s):

idx <- apply(temp, 1, function(x) sum(is.na(x)))

as.vector(which(idx == ncol(temp)))

It shows that row 1001 is problematic. You can go ahead to delete

it.

5.2.3 Bagging Tree

Bagging (Bootstrap aggregating) was originally proposed by Leo
Breiman. It is one of the earliest ensemble methods (L, 966a).
84 5 Data Pre-processing

When used in missing value imputation, it will use the remaining

variables as predictors to train a bagging tree and then use the tree
to predict the missing values. Although theoretically, the method
is powerful, the computation is much more intense than KNN.
In practice, there is a trade-off between computation time and
the effect. If a median or mean meet the modeling needs, even
bagging tree may improve the accuracy a little, but the upgrade is
so marginal that it does not deserve the extra time. The bagging
tree itself is a model for regression and classification. Here we use
preProcess() to impute sim.dat:

imp <- preProcess(sim.dat, method = "bagImpute")

demo_imp <- predict(imp, sim.dat)
summary(demo_imp[, 1:5])

age gender income house store_exp

Min. :16.00 Female:554 Min. : 41776 No :432 Min. : 155.8
1st Qu.:25.00 Male :446 1st Qu.: 86762 Yes:568 1st Qu.: 205.1
Median :36.00 Median : 94739 Median : 329.0
Mean :38.58 Mean :114665 Mean : 1357.7
3rd Qu.:53.00 3rd Qu.:123726 3rd Qu.: 597.3
Max. :69.00 Max. :319704 Max. :50000.0

5.3 Centering and Scaling

It is the most straightforward data transformation. It centers and
scales a variable to mean 0 and standard deviation 1. It ensures
that the criterion for finding linear combinations of the predic-
tors is based on how much variation they explain and therefore
improves the numerical stability. Models involving finding linear
combinations of the predictors to explain response/predictors vari-
ation need data centering and scaling, such as PCA (Jolliffe, 2002),
PLS (Geladi P, 1986) and EFA (Mulaik, 2009). You can quickly
write code yourself to conduct this transformation.
5.4 Resolve Skewness 85

Let’s standardize the variable income from sim.dat:

income <- sim.dat$income

# calculate the mean of income
mux <- mean(income, na.rm = T)
# calculate the standard deviation of income
sdx <- sd(income, na.rm = T)
# centering
tr1 <- income - mux
# scaling
tr2 <- tr1/sdx

Or the function preProcess() can apply this transformation to a set

of predictors.

sdat <- subset(sim.dat, select = c("age", "income"))

# set the 'method' option
trans <- preProcess(sdat, method = c("center", "scale"))
# use predict() function to get the final result
transformed <- predict(trans, sdat)

Now the two variables are in the same scale. You can check the
result using summary(transformed). Note that there are missing val-
ues.

5.4 Resolve Skewness

Skewness1 is defined to be the third standardized central moment.
The formula for the sample skewness statistics is:
∑(𝑥𝑖 − 𝑥)̄ 3
𝑠𝑘𝑒𝑤𝑛𝑒𝑠𝑠 =
(𝑛 − 1)𝑣3/2
∑(𝑥𝑖 − 𝑥)̄ 2
𝑣=
(𝑛 − 1)
1
https://en.wikipedia.org/wiki/Skewness
86 5 Data Pre-processing

Skewness=0 means that the destribution is symmetric, i.e. the

probability of falling on either side of the distribution’s mean is
equal.

left skew, skewnwss = -1.88 right skew, skewness = 1.88

0.30

0.30
0.20

0.20
Density

Density
0.10

0.10
0.00

0.00

0 5 10 15 0 5 10 15

X2 X1

FIGURE 5.2: Example of skewed distributions

There are different ways may help to remove skewness such as log,
square root or inverse. However, it is often difficult to determine
from plots which transformation is most appropriate for correct-
ing skewness. The Box-Cox procedure automatically identified a
transformation from the family of power transformations that are
indexed by a parameter 𝜆(Box G, 1964).

𝑥𝜆 −1
𝑖𝑓 𝜆 ≠ 0
𝑥∗ = { 𝜆
𝑙𝑜𝑔(𝑥) 𝑖𝑓 𝜆 = 0

It is easy to see that this family includes log transformation

(𝜆 = 0), square transformation (𝜆 = 2), square root (𝜆 = 0.5),
inverse (𝜆 = −1) and others in-between. We can still use func-
tion preProcess() in package caret to apply this transformation by
chaning the method argument.
5.4 Resolve Skewness 87

describe(sim.dat)

vars n mean sd median trimmed mad ...

age 1 1000 38.84 16.42 36 37.69 16.31
gender* 2 1000 1.45 0.50 1 1.43 0.00
income 3 816 113543.07 49842.29 93869 104841.94 28989.47
house* 4 1000 1.57 0.50 2 1.58 0.00
store_exp 5 1000 1356.85 2774.40 329 839.92 196.45
online_exp 6 1000 2120.18 1731.22 1942 1874.51 1015.21
store_trans 7 1000 5.35 3.70 4 4.89 2.97
online_trans 8 1000 13.55 7.96 14 13.42 10.38
...

It is easy to see the skewed variables. If mean and trimmed differ a

lot, there is very likely outliers. By default, trimmed reports mean
by dropping the top and bottom 10%. It can be adjusted by setting
argument trim=. It is clear that store_exp has outliers.
As an example, we will apply Box-Cox transformation on
store_trans and online_trans:

# select the two columns and save them as dat_bc

dat_bc <- subset(sim.dat, select = c("store_trans", "online_trans"))
(trans <- preProcess(dat_bc, method = c("BoxCox")))

## Created from 1000 samples and 2 variables

##
## Pre-processing:
## - Box-Cox transformation (2)
## - ignored (0)
##
## Lambda estimates for Box-Cox transformation:
## 0.1, 0.7

The last line of the output shows the estimates of 𝜆 for each vari-
able. As before, use predict() to get the transformed result:
88 5 Data Pre-processing

transformed <- predict(trans, dat_bc)

par(mfrow = c(1, 2), oma = c(2, 2, 2, 2))
hist(dat_bc$store_trans, main = "Before Transformation",
xlab = "store_trans")
hist(transformed$store_trans, main = "After Transformation",
xlab = "store_trans")

Before Transformation After Transformation

150
250
Frequency

Frequency

100
150

50
0 50

0 5 10 15 20 0.0 1.0 2.0 3.0

store_trans store_trans

Before the transformation, the stroe_trans is skewed right. Box-

CoxTrans () can also conduct Box-Cox transform. But note that
BoxCoxTrans () can only be applied to a single variable, and it is
not possible to transform difference columns in a data frame at
the same time.

(trans <- BoxCoxTrans(dat_bc$store_trans))

## Box-Cox Transformation
##
## 1000 data points used to estimate Lambda
##
## Input data summary:
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 1.00 3.00 4.00 5.35 7.00 20.00
5.5 Resolve Outliers 89

##
## Largest/Smallest: 20
## Sample Skewness: 1.11
##
## Estimated Lambda: 0.1
## With fudge factor, Lambda = 0 will be used for transformations

transformed <- predict(trans, dat_bc$store_trans)

skewness(transformed)

## [1] -0.2155

The estimate of 𝜆 is the same as before (0.1). The skewness of the

original observation is 1.1, and -0.2 after transformation. Although
it is not strictly 0, it is greatly improved.

5.5 Resolve Outliers

Even under certain assumptions we can statistically define outliers,
it can be hard to define in some situations. Box plot, histogram
and some other basic visualizations can be used to initially check
whether there are outliers. For example, we can visualize numerical
non-survey variables in sim.dat:

# select numerical non-survey data

sdat <- subset(sim.dat, select = c("age", "income", "store_exp",
"online_exp", "store_trans", "online_trans"))
# use scatterplotMatrix() function from car package
par(oma = c(2, 2, 1, 2))
car::scatterplotMatrix(sdat, diagonal = TRUE)
90 5 Data Pre-processing
50000 250000 0 4000 0 10 25

age

250
50
income
50000

40000
store_exp

0
online_exp
0 6000

store_trans

15
5
online_trans
20
0

50 150 300 0 20000 50000 5 10 20

It is also easy to observe the pair relationship from the plot. age
is negatively correlated with online_trans but positively correlated
with store_trans. It seems that older people tend to purchase from
the local store. The amount of expense is positively correlated with
income. Scatterplot matrix like this can reveal lots of information
before modeling.
In addition to visualization, there are some statistical methods to
define outliers, such as the commonly used Z-score. The Z-score
for variable Y is defined as:

𝑌𝑖 − 𝑌 ̄
𝑍𝑖 =
𝑠

where 𝑌 ̄ and 𝑠 are mean and standard deviation for 𝑌 . Z-score is

a measurement of the distance between each observation and the
mean. This method may be misleading, especially when the sample
size is small. Iglewicz and Hoaglin proposed to use the modified
Z-score to determine the outlier(Iglewicz and Hoaglin, 1993)�

0.6745(𝑌𝑖 − 𝑌 ̄ )
𝑀𝑖 =
𝑀 𝐴𝐷
5.5 Resolve Outliers 91

Where MAD is the median of a series of |𝑌𝑖 − 𝑌 ̄ |, called the me-

dian of the absolute dispersion. Iglewicz and Hoaglin suggest that
the points with the Z-score greater than 3.5 corrected above are
possible outliers. Let’s apply it to income:

# calculate median of the absolute dispersion for income

ymad <- mad(na.omit(sdat$income))
# calculate z-score
zs <- (sdat$income - mean(na.omit(sdat$income)))/ymad
# count the number of outliers
sum(na.omit(zs > 3.5))

## [1] 59

According to modified Z-score, variable income has 59 outliers.

Refer to (Iglewicz and Hoaglin, 1993) for other ways of detecting
outliers.
The impact of outliers depends on the model. Some models are
sensitive to outliers, such as linear regression, logistic regression.
Some are pretty robust to outliers, such as tree models, support
vector machine. Also, the outlier is not wrong data. It is real ob-
servation so cannot be deleted at will. If a model is sensitive to
outliers, we can use spatial sign transformation (Serneels S, 2006)
to minimize the problem. It projects the original sample points to
the surface of a sphere by:
𝑥𝑖𝑗
𝑥∗𝑖𝑗 = 𝑝
√∑𝑗=1 𝑥2𝑖𝑗

where 𝑥𝑖𝑗 represents the 𝑖𝑡ℎ observation and 𝑗𝑡ℎ variable. As shown
in the equation, every observation for sample 𝑖 is divided by its
square mode. The denominator is the Euclidean distance to the
center of the p-dimensional predictor space. Three things to pay
attention here:

1. It is important to center and scale the predictor data be-

fore using this transformation
92 5 Data Pre-processing

2. Unlike centering or scaling, this manipulation of the pre-

dictors transforms them as a group
3. If there are some variables to remove (for example, highly
correlated variables), do it before the transformation

Function spatialSign() caret package can conduct the transforma-

tion. Take income and age as an example:

# KNN imputation
sdat <- sim.dat[, c("income", "age")]
imp <- preProcess(sdat, method = c("knnImpute"), k = 5)
sdat <- predict(imp, sdat)
transformed <- spatialSign(sdat)
transformed <- as.data.frame(transformed)
par(mfrow = c(1, 2), oma = c(2, 2, 2, 2))
plot(income ~ age, data = sdat, col = "blue", main = "Before")
plot(income ~ age, data = transformed, col = "blue", main = "After")

Before After
1.0
4
3

0.5
income

income
2

0.0
1

-1.0 -0.5
0
-1

0 5 10 15 -1.0 0.0 0.5 1.0

age age

Some readers may have found that the above code does not seem
to standardize the data before transformation. Recall the intro-
duction of KNN, preProcess() with method="knnImpute" by default
will standardize data.
5.6 Collinearity 93

5.6 Collinearity
It is probably the technical term known by the most un-technical
people. When two predictors are very strongly correlated, includ-
ing both in a model may lead to confusion or problem with a
singular matrix. There is an excellent function in corrplot package
with the same name corrplot() that can visualize correlation struc-
ture of a set of predictors. The function has the option to reorder
the variables in a way that reveals clusters of highly correlated
ones.

# select non-survey numerical variables

sdat <- subset(sim.dat, select = c("age", "income", "store_exp",
"online_exp", "store_trans", "online_trans"))
# use bagging imputation here
imp <- preProcess(sdat, method = "bagImpute")
sdat <- predict(imp, sdat)
# get the correlation matrix
correlation <- cor(sdat)
# plot
par(oma = c(2, 2, 2, 2))
corrplot.mixed(correlation, order = "hclust", tl.pos = "lt",
upper = "ellipse")
 online_trans

store_trans
online_exp

store_exp
94 5 Data Pre-processing

income
age
1
online_trans 0.8

0.6
age -0.74
0.4

online_exp 0.23 -0.26 0.2

0
store_exp -0.14 0.07 0.53 -0.2

-0.4
income -0.36 0.29 0.52 0.6
-0.6

store_trans -0.44 0.34 0.44 0.54 0.71 -0.8

-1

The closer the correlation is to 0, the lighter the color is and the
closer the shape is to a circle. The elliptical means the correlation
is not equal to 0 (because we set the upper = "ellipse"), the greater
the correlation, the narrower the ellipse. Blue represents a positive
correlation; red represents a negative correlation. The direction
of the ellipse also changes with the correlation. The correlation
coefficient is shown in the lower triangle of the matrix.
The variables relationship from previous scatter matrix are clear
here: the negative correlation between age and online shopping,
the positive correlation between income and amount of purchasing.
Some correlation is very strong ( such as the correlation between
online_trans andage is -0.7414) which means the two variables con-
tain duplicate information.
Section 3.5 of “Applied Predictive Modeling” (Kuhn and Johnston,
2013) presents a heuristic algorithm to remove a minimum number
of predictors to ensure all pairwise correlations are below a certain
threshold:
5.7 Collinearity 95

(1) Calculate the correlation matrix of the predictors.

(2) Determine the two predictors associated with the
largest absolute pairwise correlation (call them pre-
dictors A and B).
(3) Determine the average correlation between A and the
other variables. Do the same for predictor B.
(4) If A has a larger average correlation, remove it; oth-
erwise, remove predictor B.
(5) Repeat Step 2-4 until no absolute correlations are
above the threshold.

The findCorrelation() function in package caret will apply the

above algorithm.

(highCorr <- findCorrelation(cor(sdat), cutoff = 0.7))

## [1] 2 6

It returns the index of columns need to be deleted. It tells us that

we need to remove the first column to make sure the correlations
are all below 0.7.

# delete highly correlated columns

sdat <- sdat[-highCorr]
# check the new correlation matrix
(cor(sdat))

The absolute value of the elements in the correlation matrix af-

ter removal are all below 0.7. How strong does a correlation have
to get, before you should start worrying about multicollinearity?
There is no easy answer to that question. You can treat the thresh-
old as a tuning parameter and pick one that gives you best pre-
diction accuracy.
 96 5 Data Pre-processing

5.7 Sparse Variables

Other than the highly related predictors, predictors with degen-
erate distributions can cause the problem too. Removing those
variables can significantly improve some models’ performance and
stability (such as linear regression and logistic regression but the
tree based model is impervious to this type of predictors). One
extreme example is a variable with a single value which is called
zero-variance variable. Variables with very low frequency of unique
values are near-zero variance predictors. In general, detecting those
variables follows two rules:
• The fraction of unique values over the sample size
• The ratio of the frequency of the most prevalent value to the
frequency of the second most prevalent value.
nearZeroVar() function in the caret package can filter near-zero vari-
ance predictors according to the above rules. In order to show the
useage of the function, let’s arbitaryly add some problematic vari-
ables to the origional data sim.dat:

# make a copy
zero_demo <- sim.dat
# add two sparse variable zero1 only has one unique value zero2 is a
# vector with the first element 1 and the rest are 0s
zero_demo$zero1 <- rep(1, nrow(zero_demo))
zero_demo$zero2 <- c(1, rep(0, nrow(zero_demo) - 1))

The function will return a vector of integers indicating which

columns to remove:

nearZeroVar(zero_demo,freqCut = 95/5, uniqueCut = 10)

## [1] 20 21

As expected, it returns the two columns we generated. You can

go ahead to remove them. Note the two arguments in the function
 5.8 Re-encode Dummy Variables 97

freqCut = and uniqueCut = are corresponding to the previous two

rules.
• freqCut:
the cutoff for the ratio of the most common value to the
second most common value
• uniqueCut: the cutoff for the percentage of distinct values out of
the number of total samples

5.8 Re-encode Dummy Variables

A dummy variable is a binary variable (0/1) to represent sub-
groups of the sample. Sometimes we need to recode categories to
smaller bits of information named “dummy variables.” For exam-
ple, some questionnaires have five options for each question, A,
B, C, D, and E. After you get the data, you will usually convert
the corresponding categorical variables for each question into five
nominal variables, and then use one of the options as the baseline.
Let’s encode gender and house from sim.dat to dummy variables.
There are two ways to implement this. The first is to use
class.ind() from nnet package. However, it only works on one vari-
able at a time.

dumVar <- nnet::class.ind(sim.dat$gender)

head(dumVar)

## Female Male
## [1,] 1 0
## [2,] 1 0
## [3,] 0 1
## [4,] 0 1
## [5,] 0 1
## [6,] 0 1

Since it is redundant to keep both, we need to remove one of

98 5 Data Pre-processing

them when modeling. Another more powerful function is dummy-

Vars() from caret:

# use "origional variable name + level" as new name

dumMod <- dummyVars(~gender + house + income,
data = sim.dat,
levelsOnly = F)
head(predict(dumMod, sim.dat))

## genderFemale genderMale houseNo houseYes income

## 1 1 0 0 1 120963
## 2 1 0 0 1 122008
## 3 0 1 0 1 114202
## 4 0 1 0 1 113616
## 5 0 1 0 1 124253
## 6 0 1 0 1 107661

dummyVars() can also use formula format. The variable on the right-
hand side can be both categorical and numeric. For a numerical
variable, the function will keep the variable unchanged. The ad-
vantage is that you can apply the function to a data frame without
removing numerical variables. Other than that, the function can
create interaction term:

dumMod <- dummyVars(~gender + house + income + income:gender,

data = sim.dat,
levelsOnly = F)
head(predict(dumMod, sim.dat))

## genderFemale genderMale houseNo houseYes income

## 1 1 0 0 1 120963
## 2 1 0 0 1 122008
## 3 0 1 0 1 114202
## 4 0 1 0 1 113616
## 5 0 1 0 1 124253
## 6 0 1 0 1 107661
## genderFemale:income genderMale:income
5.8 Re-encode Dummy Variables 99

## 1 120963 0
## 2 122008 0
## 3 0 114202
## 4 0 113616
## 5 0 124253
## 6 0 107661

If you think the impact income levels on purchasing behavior is

different for male and female, then you may add the interaction
term between income and gender. You can do this by adding income:
gender in the formula.
6
Data Wrangling

This chapter focuses on some of the most frequently used data ma-
nipulations and shows how to implement them in R and Python.
It is critical to explore the data with descriptive statistics (mean,
standard deviation, etc.) and data visualization before analysis.
Transform data so that the data structure is in line with the re-
quirements of the model. You also need to summarize the results
after analysis.
When the data is too large to fit in a computer’s memory, we can
use some big data analytics engine like Spark on a cloud platform
(see Chapter 4). Even the user interface of many data platforms is
much more friendly now; it is still easier to manipulate the data as
a local data frame. Spark’s R and Python interfaces aim to keep
the data manipulation syntax consistent with popular packages
for local data frames. As shown in Section 4.4, we can run nearly
all of the dplyr functions on a spark data frame once setting up
the Spark environment. And the Python interface pyspark uses a
similar syntax as pandas. This chapter focuses on data manipula-
tions on standard data frames, which is also the foundation of big
data manipulation.
Even when the data can fit in the memory, there may be a situ-
ation where it is slow to read and manipulate due to a relatively
large size. Some R packages can make the process faster with the
cost of familiarity, especially for data wrangling. But it avoids the
hurdle of setting up Spark cluster and working in an unfamiliar en-
vironment. It is not a topic in this chapter but Appendix A briefly
introduces some of the alternative R packages to read, write and
wrangle a data set that is relatively large but not too big to fit in
the memory.

101
 102 6 Data Wrangling

There are many fundamental data processing functions in R. They

lack consistent coding and can’t flow together easily. Learning all
of them is a daunting task and unnecessary. R Studio developed a
collection of packages and bundled them in tidyverse to systemize
data wrangling and analysis tasks. You can see the package list
in tidyverse on the website1 . This chapter focuses on some of the
tidyverse packages to do data wrangling for the following reasons:

• Those packages are widely used among R users in data science.

• The code is more efficient.
• The code syntax is consistent, which makes it easier to remember
and read.
Section 6.1.2 introduces some base R functions outside the tidy-
verse universe, such as apply(), lapply() and sapply(). They are
complementary functions when you are working with a data frame.
Load the R packages first:

# install packages from CRAN

p_needed <- c('dplyr','tidyr')
packages <- rownames(installed.packages())
p_to_install <- p_needed[!(p_needed %in% packages)]
if (length(p_to_install) > 0) {
install.packages(p_to_install)
}

lapply(p_needed, require, character.only = TRUE)

6.1 Summarize Data

1
https://www.tidyverse.org/packages/
 6.1 Summarize Data 103

6.1.1 dplyr package

dplyr provides a flexible grammar of data manipulation focusing
on tools for working with data frames (hence the d in the name).
It is faster and more friendly:
• It identifies the most important data manipulations and make
they easy to use from R
• It performs faster for in-memory data by writing key pieces in
C++ using Rcpp
• The interface is the same for data frame, data table or database.
We will illustrate the following functions in order using the clothing
company data:

1. Display
2. Subset
3. Summarize
4. Create new variable
5. Merge

# Read data
sim.dat <- read.csv("http://bit.ly/2P5gTw4")

Display
• tbl_df():Convert the data to tibble which offers better checking
and printing capabilities than traditional data frames. It will
adjust output width according to fit the current window.

tbl_df(sim.dat)

• glimpse(): This is like a transposed version of tbl_df()

glimpse(sim.dat)

Subset
Get rows with income more than 300000:
104 6 Data Wrangling

filter(sim.dat, income >300000) %>%

tbl_df()

Here we use the operator %>%. It is called a “pipe operator” which

pipes a value forward into an expression or function call. What
you get in the left operation will be the first argument or the only
argument in the right operation.

x %>% f(y) = f(x, y)

y %>% f(x, ., z) = f(x, y, z )

It is an operator from magrittr which can be really beneficial. Look

at the following code. Can you tell me what it does?

ave_exp <- filter(

summarise(
group_by(
filter(
sim.dat,
!is.na(income)
),
segment
),
ave_online_exp = mean(online_exp),
n = n()
),
n > 200
)

Now look at the identical code using “%>%”:

ave_exp <- sim.dat %>%

filter(!is.na(income)) %>%
group_by(segment) %>%
summarise(
6.1 Summarize Data 105

ave_online_exp = mean(online_exp),
n = n() ) %>%
filter(n > 200)

Isn’t it much more straightforward now? Let’s read it:

1. Delete observations from sim.dat with missing income val-

ues
2. Group the data from step 1 by variable segment
3. Calculate mean of online expense for each segment and
save the result as a new variable named ave_online_exp
4. Calculate the size of each segment and saved it as a new
variable named n
5. Get segments with size larger than 200

You can use distinct() to delete duplicated rows.

dplyr::distinct(sim.dat)

sample_frac() will randomly select some rows with a specified per-

centage. sample_n() can randomly select rows with a specified num-
ber.

dplyr::sample_frac(sim.dat, 0.5, replace = TRUE)

dplyr::sample_n(sim.dat, 10, replace = TRUE)

slice() will select rows by position:

dplyr::slice(sim.dat, 10:15)

It is equivalent to sim.dat[10:15,].

top_n() will select the order top n entries:

106 6 Data Wrangling

dplyr::top_n(sim.dat,2,income)

If you want to select columns instead of rows, you can use select().
The following are some sample codes:

# select by column name

dplyr::select(sim.dat,income,age,store_exp)

# select columns whose name contains a character string

dplyr::select(sim.dat, contains("_"))

# select columns whose name ends with a character string

# similar there is "starts_with"
dplyr::select(sim.dat, ends_with("e"))

# select columns Q1,Q2,Q3,Q4 and Q5

select(sim.dat, num_range("Q", 1:5))

# select columns whose names are in a group of names

dplyr::select(sim.dat, one_of(c("age", "income")))

# select columns between age and online_exp

dplyr::select(sim.dat, age:online_exp)

# select all columns except for age

dplyr::select(sim.dat, -age)

Summarize
A standard marketing problem is customer segmentation. It usu-
ally starts with designing survey and collecting data. Then run a
cluster analysis on the data to get customer segments. Once we
have different segments, the next is to understand how each group
of customer look like by summarizing some key metrics. For exam-
ple, we can do the following data aggregation for different segments
of clothes customers.
6.1 Summarize Data 107

dat_summary <- sim.dat %>%

dplyr::group_by(segment) %>%
dplyr::summarise(Age = round(mean(na.omit(age)), 0),
FemalePct = round(mean(gender == "Female"), 2),
HouseYes = round(mean(house == "Yes"), 2),
store_exp = round(mean(na.omit(store_exp),
trim = 0.1), 0),
online_exp = round(mean(online_exp), 0),
store_trans = round(mean(store_trans), 1),
online_trans = round(mean(online_trans), 1))

# transpose the data frame for showing purpose

# due to the limit of output width
cnames <- dat_summary$segment
dat_summary <- dplyr::select(dat_summary, - segment)
tdat_summary <- t(dat_summary) %>% data.frame()
names(tdat_summary) <- cnames
tdat_summary

## Conspicuous Price Quality Style

## Age 42.00 60.00 35.00 24.00
## FemalePct 0.32 0.45 0.47 0.81
## HouseYes 0.86 0.94 0.34 0.27
## store_exp 4990.00 501.00 301.00 200.00
## online_exp 4898.00 205.00 2013.00 1962.00
## store_trans 10.90 6.10 2.90 3.00
## online_trans 11.10 3.00 16.00 21.10

Now, let’s peel the onion.

The first line sim.dat is easy. It is the data you want to work on. The
second line group_by(segment) tells R that in the following steps you
want to summarise by variable segment. Here we only summarize
data by one categorical variable, but you can group by multiple
variables, such as group_by(segment, house). The third argument sum-
marise tells R the manipulation(s) to do. Then list the exact actions
 108 6 Data Wrangling

inside summarise() . For example, Age=round(mean(na.omit(age)),0)

tell R the following things:

1. Calculate the mean of column age ignoring missing value

for each customer segment
2. Round the result to the specified number of decimal places
3. Store the result in a new variable named Age

The rest of the command above is similar. In the end, we calculate

the following for each segment:

1. Age:average age for each segment

2. FemalePct: percentage for each segment
3. HouseYes: percentage of people who own a house
4. stroe_exp: average expense in store
5. online_exp: average expense online
6. store_trans: average times of transactions in the store
7. online_trans: average times of online transactions

There is a lot of information you can extract from those simple

averages.
• Conspicuous: average age is about 40. It is a group of middle-age
wealthy people. 1/3 of them are female, and 2/3 are male. They
buy regardless the price. Almost all of them own house (0.86).
It makes us wonder what is wrong with the rest 14%?
• Price: They are older people with average age 60. Nearly all of
them own a house(0.94). They are less likely to purchase online
(store_trans = 6 while online_trans = 3). It is the only group that
is less likely to buy online.
• Quality: The average age is 35. They are not way different with
Conspicuous regarding age. But they spend much less. The per-
centages of male and female are similar. They prefer online shop-
ping. More than half of them don’t own a house (0.66).
• Style: They are young people with average age 24. The major-
ity of them are female (0.81). Most of them don’t own a house
6.1 Summarize Data 109

(0.73). They are very likely to be digital natives and prefer online
shopping.
You may notice that Style group purchase more frequently online
(online_trans) but the expense (online_exp) is not higher. It makes
us wonder what is the average expense each time, so you have a
better idea about the price range of the group.
The analytical process is aggregated instead of independent steps.
The current step will shed new light on what to do next. Sometimes
you need to go back to fix something in the previous steps. Let’s
check average one-time online and instore purchase amounts:

sim.dat %>%
group_by(segment) %>%
summarise(avg_online = round(sum(online_exp)/sum(online_trans), 2),
avg_store = round(sum(store_exp)/sum(store_trans), 2))

## `summarise()` ungrouping output (override with `.groups` argument)

## # A tibble: 4 x 3
## segment avg_online avg_store
## <chr> <dbl> <dbl>
## 1 Conspicuous 442. 479.
## 2 Price 69.3 81.3
## 3 Quality 126. 105.
## 4 Style 92.8 121.

Price group has the lowest averaged one-time purchase. The Con-
spicuous group will pay the highest price. When we build customer
profile in real life, we will also need to look at the survey summa-
rization. You may be surprised how much information simple data
manipulations can provide.
Another comman task is to check which column has missing values.
It requires the program to look at each column in the data. In this
case you can use summarise_all:
110 6 Data Wrangling

# apply function anyNA() to each column

# you can also assign a function vector
# such as: c("anyNA","is.factor")
dplyr::summarise_all(sim.dat, funs_(c("anyNA")))

## Warning: `funs_()` is deprecated as of dplyr 0.7.0.

## Please use `funs()` instead.
## See vignette('programming') for more help
## This warning is displayed once every 8 hours.
## Call `lifecycle::last_warnings()` to see where this warning was generated.

## Warning: `funs()` is deprecated as of dplyr 0.8.0.

## Please use a list of either functions or lambdas:
##
## # Simple named list:
## list(mean = mean, median = median)
##
## # Auto named with `tibble::lst()`:
## tibble::lst(mean, median)
##
## # Using lambdas
## list(~ mean(., trim = .2), ~ median(., na.rm = TRUE))
## This warning is displayed once every 8 hours.
## Call `lifecycle::last_warnings()` to see where this warning was generated.

## age gender income house store_exp online_exp

## 1 FALSE FALSE TRUE FALSE FALSE FALSE
## store_trans online_trans Q1 Q2 Q3 Q4
## 1 FALSE FALSE FALSE FALSE FALSE FALSE
## Q5 Q6 Q7 Q8 Q9 Q10 segment
## 1 FALSE FALSE FALSE FALSE FALSE FALSE FALSE

The above code returns a vector indicating if there is any value

missing in each column.
Create new variable
There are often situations where you need to create new variables.
For example, adding online and store expense to get total expense.
6.1 Summarize Data 111

In this case, you will apply window function to the columns and
return a column with the same length. mutate() can do it for you
and append one or more new columns:

dplyr::mutate(sim.dat, total_exp = store_exp + online_exp)

The above code sums up two columns and appends the result
(total_exp) to sim.dat. Another similar function is transmute(). The
difference is that transmute() will delete the original columns and
only keep the new ones.

dplyr::transmute(sim.dat, total_exp = store_exp + online_exp)

Merge
Similar to SQL, there are different joins in dplyr. We create two
baby data sets to show how the functions work.

(x <- data.frame(cbind(ID = c("A", "B", "C"), x1 = c(1, 2, 3))))

## ID x1
## 1 A 1
## 2 B 2
## 3 C 3

(y <- data.frame(cbind(ID = c("B", "C", "D"), y1 = c(T, T, F))))

## ID y1
## 1 B TRUE
## 2 C TRUE
## 3 D FALSE

# join to the left

# keep all rows in x
left_join(x, y, by = "ID")
112 6 Data Wrangling

## ID x1 y1
## 1 A 1 <NA>
## 2 B 2 TRUE
## 3 C 3 TRUE

# get rows matched in both data sets

inner_join(x, y, by = "ID")

## ID x1 y1
## 1 B 2 TRUE
## 2 C 3 TRUE

# get rows in either data set

full_join(x, y, by = "ID")

## ID x1 y1
## 1 A 1 <NA>
## 2 B 2 TRUE
## 3 C 3 TRUE
## 4 D <NA> FALSE

# filter out rows in x that can be matched in y

# it doesn't bring in any values from y
semi_join(x, y, by = "ID")

# the opposite of semi_join()

# it gets rows in x that cannot be matched in y
# it doesn't bring in any values from y
anti_join(x, y, by = "ID")

There are other functions(intersect(), union() and setdiff()). Also

the data frame version of rbind and cbind which are bind_rows()
and bind_col(). We are not going to go through them all. You can
try them yourself. If you understand the functions we introduced
so far. It should be easy for you to figure out the rest.
6.1 Summarize Data 113

6.1.2 apply(), lapply() and sapply() in base R

There are some powerful functions to summarize data in base R,
such as apply(), lapply() and sapply(). They do the same basic
things and are all from “apply” family: apply functions over parts
of data. They differ in two important respects:

1. the type of object they apply to

2. the type of result they will return

When do we use apply()? When we want to apply a function to

margins of an array or matrix. That means our data need to be
structured. The operations can be very flexible. It returns a vector
or array or list of values obtained by applying a function to margins
of an array or matrix.
For example you can compute row and column sums for a matrix:

## simulate a matrix
x <- cbind(x1 =1:8, x2 = c(4:1, 2:5))
dimnames(x)[[1]] <- letters[1:8]
apply(x, 2, mean)

## x1 x2
## 4.5 3.0

col.sums <- apply(x, 2, sum)

row.sums <- apply(x, 1, sum)

You can also apply other functions:

ma <- matrix(c(1:4, 1, 6:8), nrow = 2)

ma

## [,1] [,2] [,3] [,4]

## [1,] 1 3 1 7
## [2,] 2 4 6 8
 114 6 Data Wrangling

apply(ma, 1, table) #--> a list of length 2

## [[1]]
##
## 1 3 7
## 2 1 1
##
## [[2]]
##
## 2 4 6 8
## 1 1 1 1

apply(ma, 1, stats::quantile) # 5 x n matrix with rownames

## [,1] [,2]
## 0% 1 2.0
## 25% 1 3.5
## 50% 2 5.0
## 75% 4 6.5
## 100% 7 8.0

Results can have different lengths for each call. This is a trickier
example. What will you get?

## Example with different lengths for each call

z <- array(1:24, dim = 2:4)
zseq <- apply(z, 1:2, function(x) seq_len(max(x)))
zseq ## a 2 x 3 matrix
typeof(zseq) ## list
dim(zseq) ## 2 3
zseq[1,]
apply(z, 3, function(x) seq_len(max(x)))

• applies a function over a list, data.frame or vector and

lapply()
returns a list of the same length.
• sapply() is a user-friendly version and wrapper of lapply(). By
 6.1 Summarize Data 115

default it returns a vector, matrix or if simplify = "array", an

array if appropriate. apply(x, f, simplify = FALSE, USE.NAMES =
FALSE) is the same as lapply(x, f). If simplify=TRUE, then it will
return a data.frame instead of list.
Let’s use some data with context to help you better understand
the functions.
• Get the mean and standard deviation of all numerical variables
in the dataset.

# Get numerical variables

sdat <- sim.dat[, lapply(sim.dat, class) %in% c("integer", "numeric")]
## Try the following code with apply() function apply(sim.dat,2,class)
## What is the problem?

The data frame sdat only includes numeric columns. Now we can
go head and use apply() to get mean and standard deviation for
each column:

apply(sdat, MARGIN = 2, function(x) mean(na.omit(x)))

## age income store_exp online_exp

## 3.884e+01 1.135e+05 1.357e+03 2.120e+03
## store_trans online_trans Q1 Q2
## 5.350e+00 1.355e+01 3.101e+00 1.823e+00
## Q3 Q4 Q5 Q6
## 1.992e+00 2.763e+00 2.945e+00 2.448e+00
## Q7 Q8 Q9 Q10
## 3.434e+00 2.396e+00 3.085e+00 2.320e+00

Here we defined a function using function(x) mean(na.omit(x)). It is

a very simple function. It tells R to ignore the missing value when
calculating the mean. MARGIN = 2 tells R to apply the function to
each column. It is not hard to guess what MARGIN = 1 mean. The
result show that the average online expense is much higher than
store expense. You can also compare the average scores across
116 6 Data Wrangling

different questions. The command to calculate standard deviation

is very similar. The only difference is to change mean() to sd():

apply(sdat, MARGIN = 2, function(x) sd(na.omit(x)))

## age income store_exp online_exp

## 16.417 49842.287 2774.400 1731.224
## store_trans online_trans Q1 Q2
## 3.696 7.957 1.450 1.168
## Q3 Q4 Q5 Q6
## 1.402 1.155 1.284 1.439
## Q7 Q8 Q9 Q10
## 1.456 1.154 1.118 1.136

Even the average online expense is higher than store expense, the
standard deviation for store expense is much higher than online ex-
pense which indicates there is very likely some big/small purchase
in store. We can check it quickly:

summary(sdat$store_exp)

## Min. 1st Qu. Median Mean 3rd Qu. Max.

## -500 205 329 1357 597 50000

summary(sdat$online_exp)

## Min. 1st Qu. Median Mean 3rd Qu. Max.

## 69 420 1942 2120 2441 9479

There are some odd values in store expense. The minimum value
is -500 which indicates that you should preprocess data before an-
alyzing it. Checking those simple statistics will help you better
understand your data. It then gives you some idea how to prepro-
cess and analyze them. How about using lapply() and sapply()?
Run the following code and compare the results:
6.2 Tidy and Reshape Data 117

lapply(sdat, function(x) sd(na.omit(x)))

sapply(sdat, function(x) sd(na.omit(x)))
sapply(sdat, function(x) sd(na.omit(x)), simplify = FALSE)

6.2 Tidy and Reshape Data

“Tidy data” represent the information from a dataset as data
frames where each row is an observation, and each column con-
tains the values of a variable (i.e., an attribute of what we are
observing). Depending on the situation, the requirements on what
to present as rows and columns may change. To make data easy
to work with for the problem at hand, in practice, we often need
to convert data between the “wide” and the “long” format. The
process feels like kneading the dough.
In this section, we will show how to tidy and reshape data using
tidyr packages. It is built to simplify the process of creating tidy
data. We will go through four fundamental functions: - gather():
reshape data from wide to long - spread(): reshape data from long
to wide - separate(): split a column into multiple columns - unite():
combine multiple columns to one column
Take a baby subset of our exemplary clothes consumers data to
illustrate:

sdat<-sim.dat[1:5,1:6]
sdat

## age gender income house store_exp online_exp

## 1 57 Female 120963 Yes 529.1 303.5
## 2 63 Female 122008 Yes 478.0 109.5
## 3 59 Male 114202 Yes 490.8 279.2
## 4 60 Male 113616 Yes 347.8 141.7
## 5 51 Male 124253 Yes 379.6 112.2
118 6 Data Wrangling

For the above data sdat, what if we want to reshape the data
to have a column indicating the purchasing channel (i.e. from
store_exp or online_exp) and a second column with the correspond-
ing expense amount? Assume we want to keep the rest of the
columns the same. It is a task to change data from “wide” to
“long”.

dat_long <- tidyr::gather(sdat, "Channel","Expense",

store_exp, online_exp)
dat_long

## age gender income house Channel Expense

## 1 57 Female 120963 Yes store_exp 529.1
## 2 63 Female 122008 Yes store_exp 478.0
## 3 59 Male 114202 Yes store_exp 490.8
## 4 60 Male 113616 Yes store_exp 347.8
## 5 51 Male 124253 Yes store_exp 379.6
## 6 57 Female 120963 Yes online_exp 303.5
## 7 63 Female 122008 Yes online_exp 109.5
## 8 59 Male 114202 Yes online_exp 279.2
## 9 60 Male 113616 Yes online_exp 141.7
## 10 51 Male 124253 Yes online_exp 112.2

The above code gathers two variables (store_exp and online_exp),

and collapses them into key-value pairs (Channel and Expense), du-
plicating all other columns as needed.
You can run a regression to study the effect of purchasing channel
as follows:

# Here we use all observations from sim.dat

# Don't show result here

msdat <- tidyr::gather(sim.dat[, 1:6], "Channel","Expense",

store_exp, online_exp)
fit <- lm(Expense ~ gender + house + income + Channel + age,
6.2 Tidy and Reshape Data 119

data = msdat)
summary(fit)

Sometimes we want to reshape the data from “long” to “wide”.

For example, you want to compare the online and in-store
expense between male and female based on house owner-
ship.
We need to reshape the wide data frame dat_wide to a long format
by spreading the key-value pairs across multiple columns. And
then summarize the long data frame dat_long, grouping byhouse
and gender.

dat_wide = tidyr::spread(dat_long, Channel, Expense)

# you can check what dat_long is like
dat_wide %>%
dplyr::group_by(house, gender) %>%
dplyr::summarise(total_online_exp = sum(online_exp),
total_store_exp = sum(store_exp))

## # A tibble: 2 x 4
## # Groups: house [1]
## house gender total_online_exp total_store_exp
## <chr> <chr> <dbl> <dbl>
## 1 Yes Female 413. 1007.
## 2 Yes Male 533. 1218.

The above code also uses the functions in the dplyr package in-
troduced in the previous section. Here we use package::function to
make clear the package name. It is not necessary if the package is
already loaded.
Another pair of functions that do opposite manipulations are sep-
arate() and unite().

sepdat<- dat_long %>%

120 6 Data Wrangling

separate(Channel, c("Source", "Type"))

sepdat

## age gender income house Source Type Expense

## 1 57 Female 120963 Yes store exp 529.1
## 2 63 Female 122008 Yes store exp 478.0
## 3 59 Male 114202 Yes store exp 490.8
## 4 60 Male 113616 Yes store exp 347.8
## 5 51 Male 124253 Yes store exp 379.6
## 6 57 Female 120963 Yes online exp 303.5
## 7 63 Female 122008 Yes online exp 109.5
## 8 59 Male 114202 Yes online exp 279.2
## 9 60 Male 113616 Yes online exp 141.7
## 10 51 Male 124253 Yes online exp 112.2

You can see that the function separates the original column
“Channel” to two new columns “Source” and “Type”. You can use
sep = to set the string or regular expression to separate the col-
umn. By default, it is “_”.
The unite() function will do the opposite: combining two columns.
It is the generalization of paste() to a data frame.

sepdat %>%
unite("Channel", Source, Type, sep = "_")

## age gender income house Channel Expense

## 1 57 Female 120963 Yes store_exp 529.1
## 2 63 Female 122008 Yes store_exp 478.0
## 3 59 Male 114202 Yes store_exp 490.8
## 4 60 Male 113616 Yes store_exp 347.8
## 5 51 Male 124253 Yes store_exp 379.6
## 6 57 Female 120963 Yes online_exp 303.5
## 7 63 Female 122008 Yes online_exp 109.5
## 8 59 Male 114202 Yes online_exp 279.2
## 9 60 Male 113616 Yes online_exp 141.7
## 10 51 Male 124253 Yes online_exp 112.2
6.2 Tidy and Reshape Data 121

The reshaping manipulations may be the trickiest part. You have

to practice a lot to get familiar with those functions. Unfortunately,
there is no shortcut.
7
Model Tuning Strategy

When training a machine learning model, there are many decisions

to make. For example, when training a random forest, you need
to decide the number of trees and the number of variables at each
node. For lasso method, you need to determine the penalty param-
eter. There may be standard settings for some of the parameters,
but it’s unlikely to guess the right values for all of these correctly.
Other than that, making good choices on how you split the data
into training and testing sets can make a huge difference in helping
you find a high-performance model efficiently.
This chapter will illustrate the practical aspects of model tuning.
We will talk about different types of model error, sources of model
error, hyperparameter tuning, how to set up your data and how
to make sure your model implementation is correct. In practice
applying machine learning is a highly iterative process.
Load the R packages first:

# install packages from CRAN

p_needed <- c('ggplot2','tidyr', 'caret', 'dplyr',
'lattice', 'proxy', 'caret')
packages <- rownames(installed.packages())
p_to_install <- p_needed[!(p_needed %in% packages)]
if (length(p_to_install) > 0) {
install.packages(p_to_install)
}

lapply(p_needed, require, character.only = TRUE)

123
124 7 Model Tuning Strategy

7.1 Variance-Bias Trade-Off

Assume X is 𝑛 × 𝑝 observation matrix and y is response variable,
we have:

y = 𝑓(X) + 𝝐 (7.1)

where 𝝐 is the random error with a mean of zero. The function

𝑓(⋅) is our modeling target, which represents the information in
the response variable that predictors can explain. The main goal
of estimating 𝑓(⋅) is inference or prediction, or sometimes both. In
general, there is a trade-off between flexibility and interpretability
of the model. So data scientists need to comprehend the delicate
balance between these two.
Depending on the modeling purposes, the requirement for inter-
pretability varies. If the prediction is the only goal, then as long as
the prediction is accurate enough, the interpretability is not under
consideration. In this case, people can use “black box” model, such
as random forest, boosting tree, neural network and so on. These
models are very flexible but nearly impossible to explain. Their
accuracy is usually higher on the training set, but not necessary
when it predicts. It is not surprising since those models have a huge
number of parameters and high flexibility that they can “memo-
rize” the entire training data. A paper by Chiyuan Zhang et al. in
2017 pointed out that “Deep neural networks (even just two-layer
net) easily fit random labels” (Zhang et al., 2017). The traditional
forms of regularization, such as weight decay, dropout, and data
augmentation, fail to control generalization error. It poses a con-
ceptual challenge to statistical theory and also calls our attention
when we use such black-box models.
There are two kinds of application problems: complete informa-
tion problem and incomplete information problem. The complete
information problem has all the information you need to know the
correct response. Take the famous cat recognition, for example, all
the information you need to identify a cat is in the picture. In this
7.1 Variance-Bias Trade-Off 125

situation, the algorithm that penetrates the data the most wins.
There are some other similar problems such as the self-driving car,
chess game, facial recognition and speech recognition. But in most
of the data science applications, the information is incomplete. If
you want to know whether a customer is going to purchase again
or not, it is unlikely to have 360-degree of the customer’s informa-
tion. You may have their historical purchasing record, discounts
and service received. But you don’t know if the customer sees your
advertisement, or has a friend recommends competitor’s product,
or encounters some unhappy purchasing experience somewhere.
There could be a myriad of factors that will influence the cus-
tomer’s purchase decision while what you have as data is only a
small part. To make things worse, in many cases, you don’t even
know what you don’t know. Deep learning doesn’t have any advan-
tage in solving those problems. Instead, some parametric models
often work better in this situation. You will comprehend this more
after learning the different types of model error. Assume we have
𝑓 ̂ which is an estimator of 𝑓. Then we can further get 𝐲̂ = 𝑓(X).
̂
The predicted error is divided into two parts, systematic error, and
random error:

𝐸(y − y)̂ 2 = ̂
𝐸[𝑓(X) + 𝝐 − 𝑓(X)] 2

= 𝐸[𝑓(X) ̂
− 𝑓(X)]2 + 𝑉 𝑎𝑟(𝝐) (7.2)
⏟⏟⏟⏟⏟⏟⏟ ⏟
(1) (2)

It is also called Mean Square Error (MSE) where (1) is the sys-
tematic error. It exists because 𝑓 ̂ usually does not entirely describe
the “systematic relation” between X and y which refers to the sta-
ble relationship that exists across different samples or time. Model
improvement can help reduce this kind of error; (2) is the random
error which represents the part of y that cannot be explained by
X. A more complex model does not reduce the error. There are
three reasons for random error:

1. the current sample is not representative, so the pattern

in one sample set does not generalize to a broader scale.
126 7 Model Tuning Strategy

2. The information is incomplete. In other words, you don’t

have all variables needed to explain the response.
3. Measurement error in the variables.

Deep learning has significant success solving problems with com-

plete information and usually low measurement error. As men-
tioned before, in a task like image recognition, all you need are the
pixels in the pictures. So in deep learning applications, increasing
the sample size can improve the model performance significantly.
But it may not perform well in problems with incomplete informa-
tion. The biggest problem with the black-box model is that it fits
random error, i.e., over-fitting. The notable feature of random er-
ror is that it varies over different samples. So one way to determine
whether overfitting happens is to reserve a part of the data as the
test set and then check the performance of the trained model on
the test data. Note that overfitting is a general problem from which
any model could suffer. However, since black-box models usually
have a large number of parameters, it is much more suspectable
to over-fitting.

FIGURE 7.1: Types of Model Error

̂
The systematic error 𝐸[𝑓(X) − 𝑓(X)]2
can be further decomposed
as:

̂
(𝑓(X) − 𝐸[𝑓(X)] ̂
+ 𝐸[𝑓(X)] ̂
− 𝑓(X))
2 2
̂
= 𝐸 (𝐸[𝑓(X)] − 𝑓(X)) + 𝐸 (𝑓(X) ̂ ̂
− 𝐸[𝑓(X)]) (7.3)
= ̂
[𝐵𝑖𝑎𝑠(𝑓(X))] 2 ̂
+ 𝑉 𝑎𝑟(𝑓(X))
7.1 Variance-Bias Trade-Off 127

The systematic error consists of two parts, 𝐵𝑖𝑎𝑠(𝑓(X)) ̂ and

̂
𝑉 𝑎𝑟(𝑓(X)). To minimize the systematic error, we need to min-
imize both. The bias represents the error caused by the model’s
approximation of the reality, i.e., systematic relation, which may
be very complex. For example, linear regression assumes a linear
relationship between the predictors and the response, but rarely is
there a perfect linear relationship in real life. So linear regression
is more likely to have a high bias. In general, the more flexible the
model is, the higher the variance. However, this does not guar-
antee complex models will outperform a much simpler one, such
as linear regression. If the real relationship 𝑓 is linear, then linear
regression is unbiased. It is difficult for a more flexible model to
compete. Good learning method requires low variance and bias.
However, it is easy to find a model with extremely low bias but
high variance (by fitting a tree) or a method with low variance but
high bias (by fitting a straight line). That why we call it trade-off.
To explore bias and variance, let’s begin with a simple simulation.
We will simulate a data with a non-linear relationship and fit dif-
ferent models on it. An intuitive way to show these is to compare
the plots of various models.
The code below simulates one predictor (x) and one response
variable (fx). The relationship between x and fx is non-
linear. You need to load the multiplot function by running
source('http://bit.ly/2KeEIg9'). The function assembles multiple
plots on a canvas.

source('http://bit.ly/2KeEIg9')
# randomly simulate some non-linear samples
x = seq(1, 10, 0.01) * pi
e = rnorm(length(x), mean = 0, sd = 0.2)
fx <- sin(x) + e + sqrt(x)
dat = data.frame(x, fx)

Then fit a simple linear regression on these data:

128 7 Model Tuning Strategy

# plot fitting result

library(ggplot2)
ggplot(dat, aes(x, fx)) +
geom_point() +
geom_smooth(method = "lm", se = FALSE)

## `geom_smooth()` using formula 'y ~ x'

4
fx

10 20 30
x

FIGURE 7.2: High bias model

Despite a large sample size, trained linear regression cannot de-

scribe the relationship very well. In other words, in this case, the
model has a high bias (Fig. 7.2). People also call it underfitting.
Since the estimated parameters will be somewhat different for dif-
ferent samples, there is the variance of estimates. Intuitively, it
gives you some sense that if we fit the same model with different
samples (presumably, they are from the same population), how
much will the estimates change. Ideally, the change is trivial. For
high variance models, small changes in the training data result in
very different estimates. In general, a model with high flexibility
also has high variance., such as the CART tree, and the initial
7.1 Variance-Bias Trade-Off 129

boosting method. To overcome that problem, the Random For-

est and Gradient Boosting Model aim to reduce the variance by
summarizing the results obtained from different samples.
Let’s fit the above data using a smoothing method which is highly
flexible and can fit the current data tightly:

ggplot(dat, aes(x, fx)) + geom_smooth(span = 0.03)

4
fx

10 20 30
x

FIGURE 7.3: High variance model

The resulting plot (Fig. 7.3) indicates the smoothing method fit
the data much better and it has a much smaller bias. However,
this method has a high variance. If we simulate different subsets
of the sample, the result curve will change significantly:

# set random seed

set.seed(2016)
# sample part of the data to fit model sample 1
idx1 = sample(1:length(x), 100)
dat1 = data.frame(x1 = x[idx1], fx1 = fx[idx1])
p1 = ggplot(dat1, aes(x1, fx1)) +
130 7 Model Tuning Strategy

geom_smooth(span = 0.03) +
geom_point()

# sample 2
idx2 = sample(1:length(x), 100)
dat2 = data.frame(x2 = x[idx2], fx2 = fx[idx2])
p2 = ggplot(dat2, aes(x2, fx2)) +
geom_smooth(span = 0.03) +
geom_point()

# sample 3
idx3 = sample(1:length(x), 100)
dat3 = data.frame(x3 = x[idx3], fx3 = fx[idx3])
p3 = ggplot(dat3, aes(x3, fx3)) +
geom_smooth(span = 0.03) +
geom_point()

# sample 4
idx4 = sample(1:length(x), 100)
dat4 = data.frame(x4 = x[idx4], fx4 = fx[idx4])
p4 = ggplot(dat4, aes(x4, fx4)) +
geom_smooth(span = 0.03) +
geom_point()

multiplot(p1, p2, p3, p4, cols = 2)

7.1 Variance-Bias Trade-Off 131

6 6

5
4
fx1

fx3
4

3
2
2

1
10 20 30 10 20 30
x1 x3

6
6

4 4
fx2

fx4
2 2

10 20 30 10 20 30
x2 x4

The fitted lines (blue) change over different samples which means it
has high variance. People also call it overfitting. Fitting the linear
model using the same four subsets, the result barely changes:

p1 = ggplot(dat1, aes(x1, fx1)) +

geom_smooth(method = "lm", se = FALSE) +
geom_point()

p2 = ggplot(dat2, aes(x2, fx2)) +

geom_smooth(method = "lm", se = FALSE) +
geom_point()

p3 = ggplot(dat3, aes(x3, fx3)) +

geom_smooth(method = "lm", se = FALSE) +
geom_point()

p4 = ggplot(dat4, aes(x4, fx4)) +

geom_smooth(method = "lm", se = FALSE) +
geom_point()
132 7 Model Tuning Strategy

multiplot(p1, p2, p3, p4, cols = 2)

6 6

5
4
fx1

fx3
4

3
2
2

10 20 30 10 20 30
x1 x3

6
6

4
4
fx2

fx4

2 2

10 20 30 10 20 30
x2 x4

In general, the variance (𝑉 𝑎𝑟(𝑓(X)))̂ increases and the bias

̂
(𝐵𝑖𝑎𝑠(𝑓(X))) decreases as the model flexibility increases. Vari-
ance and bias together determine the systematic error. As we
increase the flexibility of the model, at first the rate at which
̂
𝐵𝑖𝑎𝑠(𝑓(X)) ̂
decreases is faster than 𝑉 𝑎𝑟(𝑓(X)), so the MSE de-
creases. However, to some degree, higher flexibility has little effect
̂
on 𝐵𝑖𝑎𝑠(𝑓(X)) ̂
but 𝑉 𝑎𝑟(𝑓(X)) increases significantly, so the MSE
increases.

7.2 Data Splitting and Resampling

Those highly adaptable models can model complex relationships.
However, they tend to overfit which leads to the poor prediction
by learning too much from the data. It means that the model is
susceptible to the specific sample used to fit it. When future data
 7.2 Data Splitting and Resampling 133

is not exactly like the past data, the model prediction may have
big mistakes. A simple model like ordinary linear regression tends
instead to underfit which leads to a bad prediction by learning
too little from the data. It systematically over-predicts or under-
predicts the data regardless of how well future data resemble past
data. Without evaluating models, the modeler will not know about
the problem before the future samples. Data splitting and resam-
pling are fundamental techniques to build sound models for pre-
diction.

7.2.1 Data Splitting

Data splitting is to put part of the data aside as testing set (or
Hold-outs, out of bag samples) and use the rest for model training.
Training samples are also called in-sample. Model performance
metrics evaluated using in-sample are retrodictive, not predictive.
The traditional business intelligence usually handles data descrip-
tion. Answer simple questions by querying and summarizing the
data, such as:
• What is the monthly sales of a product in 2020?
• What is the number of visits to our site in the past month?
• What is the sales difference in 2021 for two different product
designs?
There is no need to go through the tedious process of splitting the
data, tuning and testing model to answer questions of this kind.
Instead, people usually use as complete data as possible and then
sum or average the parts of interest.
Many models have parameters which cannot be directly estimated
from the data, such as 𝜆 in the lasso (penalty parameter), the
number of trees in the random forest. This type of model param-
eter is called tuning parameter, and there is no analytical formula
available to calculate the optimized value. Tuning parameters of-
ten control the complexity of the model. A poor choice can result
in over-fitting or under-fitting. A standard approach to estimate
tuning parameters is through cross-validation which is a data re-
sampling approach.
134 7 Model Tuning Strategy

To get a reasonable precision of the performance based on a single

test set, the size of the test set may need to be large. So a con-
ventional approach is to use a subset of samples to fit the model
and use the rest to evaluate model performance. This process will
repeat multiple times to get a performance profile. In that sense,
resampling is based on splitting. The general steps are:

Algorithm 1 General resampling steps

1: Define a set of candidate values for tuning parameter(s)
2: for Each candidate value in the set do
3: Resample data
4: Fit model
5: Predict hold-out
6: Calculate performance
7: end for
8: Aggregate the results
9: Determine the final tuning parameter
10: Refit the model with the entire data set

The above is an outline of the general procedure to tune parame-

ters. Now let’s focus on the critical part of the process: data split-
ting. Ideally, we should evaluate model using samples that were
not used to build or fine-tune the model. So it provides an unbi-
ased sense of model effectiveness. When the sample size is large, it
is a good practice to set aside part of the samples to evaluate the
final model. People use “training” data to indicate samples used
to fit or fine-tune the model and “test” or “validation” data set is
used to validate performance.
The first decision to make for data splitting is to decide the pro-
portion of data in the test set. There are two factors to consider
here: (1) sample size; (2) computation intensity. If the sample size
is large enough which is the most common situation according to
my experience, you can try to use 20%, 30% and 40% of the data
as the test set, and see which one works the best. If the model is
computationally intense, then you may consider starting from a
smaller sample of data to train the model hence will have a higher
portion of data in the test set. Depending on how it performs, you
7.2 Data Splitting and Resampling 135

FIGURE 7.4: Parameter Tuning Process

may need to increase the training set. If the sample size is small,
you can use cross-validation or bootstrap which is the topic in the
next section.
The next decision is to decide which samples are in the test set.
There is a desire to make the training and test sets as similar
as possible. A simple way is to split data by random sampling
which, however, does not control for any of the data attributes,
such as the percentage of the retained customer in the data. So
it is possible that the distribution of outcomes is substantially
different between the training and test sets. There are three main
ways to split the data that account for the similarity of resulted
data sets. We will describe the three approaches using the clothing
company customer data as examples.
136 7 Model Tuning Strategy

(1) Split data according to the outcome variable

Assume the outcome variable is customer segment (column seg-

ment) and we decide to use 80% as training and 20% test. The
goal is to make the proportions of the categories in the two sets
as similar as possible. The createDataPartition() function in caret
will return a balanced splitting based on assigned variable.

# load data
sim.dat <- read.csv("http://bit.ly/2P5gTw4")
library(caret)
# set random seed to make sure reproducibility
set.seed(3456)
trainIndex <- createDataPartition(sim.dat$segment,
p = 0.8,
list = FALSE,
times = 1)
head(trainIndex)

## Resample1
## [1,] 1
## [2,] 2
## [3,] 3
## [4,] 4
## [5,] 6
## [6,] 7

The list = FALSE in the call to createDataPartition is to return

a data frame. The times = 1 tells R how many times you want
to split the data. Here we only do it once, but you can repeat
the splitting multiple times. In that case, the function will return
multiple vectors indicating the rows to training/test. You can set
times 2 and rerun the above code to see the result. Then we can
use the returned indicator vector trainIndex to get training and
test sets:
7.2 Data Splitting and Resampling 137

# get training set

datTrain <- sim.dat[trainIndex, ]
# get test set
datTest <- sim.dat[-trainIndex, ]

According to the setting, there are 800 samples in the training set
and 200 in test set. Let’s check the distribution of the two sets:

datTrain %>%
dplyr::group_by(segment) %>%
dplyr::summarise(count = n(),
percentage = round(length(segment)/nrow(datTrain), 2))

## # A tibble: 4 x 3
## segment count percentage
## <chr> <int> <dbl>
## 1 Conspicuous 160 0.2
## 2 Price 200 0.25
## 3 Quality 160 0.2
## 4 Style 280 0.35

datTest %>%
dplyr::group_by(segment) %>%
dplyr::summarise(count = n(),
percentage = round(length(segment)/nrow(datTrain), 2))

## # A tibble: 4 x 3
## segment count percentage
## <chr> <int> <dbl>
## 1 Conspicuous 40 0.05
## 2 Price 50 0.06
## 3 Quality 40 0.05
## 4 Style 70 0.09

The percentages are the same for these two sets. In practice, it is
 138 7 Model Tuning Strategy

possible that the distributions are not exactly identical but should
be close.

(2) Divide data according to predictors

An alternative way is to split data based on the predictors. The

goal is to get a diverse subset from a dataset so that the sample
is representative. In other words, we need an algorithm to identify
the 𝑛 most diverse samples from a dataset with size 𝑁 . However,
the task is generally infeasible for non-trivial values of 𝑛 and 𝑁
(Willett, 2004). And hence practicable approaches to dissimilarity-
based selection involve approximate methods that are sub-optimal.
A major class of algorithms split the data on maximum dissimi-
larity sampling. The process starts from:
• Initialize a single sample as starting test set
• Calculate the dissimilarity between this initial sample and each
remaining samples in the dataset
• Add the most dissimilar unallocated sample to the test set
To move forward, we need to define the dissimilarity between
groups. Each definition results in a different version of the al-
gorithm and hence a different subset. It is the same problem as
in hierarchical clustering where you need to define a way to mea-
sure the distance between clusters. The possible approaches are
to use minimum, maximum, sum of all distances, the average of
all distances, etc. Unfortunately, there is not a single best choice,
and you may have to try multiple methods and check the resulted
sample sets. R users can implement the algorithm using maxDissim()
function from caret package. The obj argument is to set the defi-
nition of dissimilarity. Refer to the help documentation for more
details (?maxDissim).
Let’s use two variables (age and income) from the customer data as
an example to illustrate how it works in R and compare maximum
dissimilarity sampling with random sampling.
7.2 Data Splitting and Resampling 139

library(lattice)
# select variables
testing <- subset(sim.dat, select = c("age", "income"))

Random select 5 samples as initial subset (start) , the rest will be

in samplePool:

set.seed(5)
# select 5 random samples
startSet <- sample(1:dim(testing)[1], 5)
start <- testing[startSet, ]
# save the rest in data frame 'samplePool'
samplePool <- testing[-startSet, ]

Use maxDissim() to select another 5 samples from samplePool that

are as different as possible with the initical set start:

selectId <- maxDissim(start, samplePool, obj = minDiss, n = 5)

minDissSet <- samplePool[selectId, ]

The obj = minDiss in the above code tells R to use minimum dissim-
ilarity to define the distance between groups. Next, random select
5 samples from samplePool in data frame RandomSet:

selectId <- sample(1:dim(samplePool)[1], 5)

RandomSet <- samplePool[selectId, ]

Plot the resulted set to compare different sampling methods:

start$group <- rep("Initial Set", nrow(start))

minDissSet$group <- rep("Maximum Dissimilarity Sampling",
nrow(minDissSet))
RandomSet$group <- rep("Random Sampling",
nrow(RandomSet))
xyplot(age ~ income,
 140 7 Model Tuning Strategy

data = rbind(start, minDissSet, RandomSet),

grid = TRUE,
group = group,
auto.key = TRUE)

Initial Set
Maximum Dissimilarity Sampling
Random Sampling

60

50
age

40

30

20

50000 100000 150000 200000 250000 300000

income

FIGURE 7.5: Compare Maximum Dissimilarity Sampling with

Random Sampling

The points from maximum dissimilarity sampling are far away

from the initial samples ( Fig. 7.5, while the random samples are
much closer to the initial ones. Why do we need a diverse subset?
Because we hope the test set to be representative. If all test set
samples are from respondents younger than 30, model performance
on the test set has a high risk to fail to tell you how the model
will perform on more general population.
• Divide data according to time
For time series data, random sampling is usually not the best way.
There is an approach to divide data according to time-series. Since
time series is beyond the scope of this book, there is not much dis-
cussion here. For more detail of this method, see (Hyndman and
7.2 Data Splitting and Resampling 141

Athanasopoulos, 2013). We will use a simulated first-order autore-

gressive model [AR (1)] time-series data with 100 observations to
show how to implement using the function createTimeSlices () in
the caret package.

# simulte AR(1) time series samples

timedata = arima.sim(list(order=c(1,0,0), ar=-.9), n=100)
# plot time series
plot(timedata, main=(expression(AR(1)~~~phi==-.9)))

AR(1) φ = − 0.9
6
4
2
timedata

0
-2
-4
-6

0 20 40 60 80 100

Time

FIGURE 7.6: Divide data according to time

Fig. 7.6 shows 100 simulated time series observation. The goal is
to make sure both training and test set to cover the whole period.

timeSlices <- createTimeSlices(1:length(timedata),

initialWindow = 36,
horizon = 12,
fixedWindow = T)
str(timeSlices,max.level = 1)

## List of 2
 142 7 Model Tuning Strategy

## $ train:List of 53
## $ test :List of 53

There are three arguments in the above createTimeSlices().

• initialWindow:The initial number of consecutive values in each

training set sample
• horizon: the number of consecutive values in test set sample
• fixedWindow: if FALSE, all training samples start at 1
The function returns two lists, one for the training set, the other
for the test set. Let’s look at the first training sample:

# get result for the 1st training set

trainSlices <- timeSlices[[1]]
# get result for the 1st test set
testSlices <- timeSlices[[2]]
# check the index for the 1st training and test set
trainSlices[[1]]

## [1] 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
## [18] 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34
## [35] 35 36

testSlices[[1]]

## [1] 37 38 39 40 41 42 43 44 45 46 47 48

The first training set is consist of sample 1-36 in the dataset

(initialWindow = 36). Then sample 37-48 are in the first test set
( horizon = 12). Type head(trainSlices) or head(testSlices) to check
the later samples. If you are not clear about the argument fixed-
Window, try to change the setting to be F and check the change in
trainSlices and testSlices.

Understand and implement data splitting is not difficult. But there

are two things to note:
7.2 Data Splitting and Resampling 143

1. The randomness in the splitting process will lead to un-

certainty in performance measurement.
2. When the dataset is small, it can be too expensive to leave
out test set. In this situation, if collecting more data is
just not possible, the best shot is to use leave-one-out
cross-validation which is in the next section.

7.2.2 Resampling
You can consider resampling as repeated splitting. The basic idea
is: use part of the data to fit model and then use the rest of data
to calculate model performance. Repeat the process multiple times
and aggregate the results. The differences in resampling techniques
usually center around the ways to choose subsamples. There are
two main reasons that we may need resampling:

1. Estimate tuning parameters through resampling. Some

examples of models with such parameters are Support
Vector Machine (SVM), models including the penalty
(LASSO) and random forest.
2. For models without tuning parameter, such as ordinary
linear regression and partial least square regression, the
model fitting doesn’t require resampling. But you can
study the model stability through resampling.

We will introduce three most common resampling techniques: k-

fold cross-validation, repeated training/test splitting, and boot-
strap.
7.2.2.1 k-fold cross-validation

k-fold cross-validation is to partition the original sample into 𝑘

equal size subsamples (folds). Use one of the 𝑘 folds to validate
the model and the rest 𝑘 − 1 to train model. Then repeat the
process 𝑘 times with each of the 𝑘 folds as the test set. Aggregate
the results into a performance profile.
̂ (𝑋) the fitted function, computed with the 𝜅𝑡ℎ fold
Denote by 𝑓 −𝜅
144 7 Model Tuning Strategy

removed and 𝑥𝜅𝑖 the predictors for samples in left-out fold. The
process of k-fold cross-validation is as follows:

Algorithm 2 k-fold cross-validation

1: Partition the original sample into 𝑘 equal size folds
2: for 𝜅 = 1…𝑘 do
3: Use data other than fold 𝜅 to train the model 𝑓 −𝜅 ̂ (𝑋)
4: ̂
−𝜅 ̂
−𝜅
Apply 𝑓 (𝑋) to predict fold 𝜅 to get 𝑓 (𝑥𝑖 ) 𝜅

5: end for
6: Aggregate the results

̂ = 1 Σ𝑘𝜅=1 Σ𝑥𝜅 𝐿(𝑦𝑖𝜅 , 𝑓 −𝜅

𝐸𝑟𝑟𝑜𝑟 ̂ (𝑥𝜅 ))
𝑖
𝑁 𝑖

It is a standard way to find the value of tuning parameter that gives

you the best performance. It is also a way to study the variability
of model performance.
The following figure represents a 5-fold cross-validation example.

FIGURE 7.7: 5-fold cross-validation

A special case of k-fold cross-validation is Leave One Out Cross

Validation (LOOCV) where 𝑘 = 1. When sample size is small, it
is desired to use as many data to train the model. Most of the
functions have default setting 𝑘 = 10. The choice is usually 5-10
in practice, but there is no standard rule. The more folds to use,
the more samples are used to fit model, and then the performance
estimate is closer to the theoretical performance. Meanwhile, the
7.2 Data Splitting and Resampling 145

variance of the performance is larger since the samples to fit model

in different iterations are more similar. However, LOOCV has high
computational cost since the number of interactions is the same
as the sample size and each model fit uses a subset that is nearly
the same size of the training set. On the other hand, when k is
small (such as 2 or 3), the computation is more efficient, but the
bias will increase. When the sample size is large, the impact of 𝑘
becomes marginal.
Chapter 7 of (Hastie T, 2008) presents a more in-depth and more
detailed discussion about the bias-variance trade-off in k-fold cross-
validation.
You can implement k-fold cross-validation using createFolds() in
caret:

library(caret)
class <- sim.dat$segment
# creat k-folds
set.seed(1)
cv <- createFolds(class, k = 10, returnTrain = T)
str(cv)

## List of 10
## $ Fold01: int [1:900] 1 2 3 4 5 6 7 8 9 10 ...
## $ Fold02: int [1:900] 1 2 3 4 5 6 7 9 10 11 ...
## $ Fold03: int [1:900] 1 2 3 4 5 6 7 8 10 11 ...
## $ Fold04: int [1:900] 1 2 3 4 5 6 7 8 9 11 ...
## $ Fold05: int [1:900] 1 3 4 6 7 8 9 10 11 12 ...
## $ Fold06: int [1:900] 1 2 3 4 5 6 7 8 9 10 ...
## $ Fold07: int [1:900] 2 3 4 5 6 7 8 9 10 11 ...
## $ Fold08: int [1:900] 1 2 3 4 5 8 9 10 11 12 ...
## $ Fold09: int [1:900] 1 2 4 5 6 7 8 9 10 11 ...
## $ Fold10: int [1:900] 1 2 3 5 6 7 8 9 10 11 ...

The above code creates ten folds (k=10) according to the customer
segments (we set class to be the categorical variable segment). The
function returns a list of 10 with the index of rows in training set.
146 7 Model Tuning Strategy

7.2.2.2 Repeated Training/Test Splits

In fact, this method is nothing but repeating the training/test

set division on the original data. Fit the model with the train-
ing set, and evaluate the model with the test set. Unlike k-fold
cross-validation, the test set generated by this procedure may have
duplicate samples. A sample usually shows up in more than one
test sets. There is no standard rule for split ratio and number of
repetitions. The most common choice in practice is to use 75% to
80% of the total sample for training. The remaining samples are
for validation. The more sample in the training set, the less biased
the model performance estimate is. Increasing the repetitions can
reduce the uncertainty in the performance estimates. Of course, it
is at the cost of computational time when the model is complex.
The number of repetitions is also related to the sample size of
the test set. If the size is small, the performance estimate is more
volatile. In this case, the number of repetitions needs to be higher
to deal with the uncertainty of the evaluation results.
We can use the same function (createDataPartition ()) as before.
If you look back, you will see times = 1. The only thing to change
is to set it to the number of repetitions.

trainIndex <- createDataPartition(sim.dat$segment,

p = .8,
list = FALSE,
times = 5)
dplyr::glimpse(trainIndex)

## int [1:800, 1:5] 1 3 4 5 6 7 8 9 10 11 ...

## - attr(*, "dimnames")=List of 2
## ..$ : NULL
## ..$ : chr [1:5] "Resample1" "Resample2" "Resample3" "Resample4" ...

Once know how to split the data, the repetition comes naturally.
7.2 Data Splitting and Resampling 147

7.2.2.3 Bootstrap Methods

Bootstrap is a powerful statistical tool (a little magic too). It can

be used to analyze the uncertainty of parameter estimates (Efron
and Tibshirani, 1986) quantitatively. For example, estimate the
standard deviation of linear regression coefficients. The power of
this method is that the concept is so simple that it can be easily
applied to any model as long as the computation allows. However,
you can hardly obtain the standard deviation for some models by
using the traditional statistical inference.
Since it is with replacement, a sample can be selected multiple
times, and the bootstrap sample size is the same as the original
data. So for every bootstrap set, there are some left-out samples,
which is also called “out-of-bag samples.” The out-of-bag sample
is used to evaluate the model. Efron points out that under normal
circumstances (Efron, 1983), bootstrap estimates the error rate of
the model with more certainty.The probability of an observation 𝑖
in bootstrap sample B is:
𝑁
𝑃 𝑟𝑖 ∈ 𝐵 = 1 − (1 − 𝑁1 )
≈ 1 − 𝑒−1
= 0.632
On average, 63.2% of the observations appeared at least once in a
bootstrap sample, so the estimation bias is similar to 2-fold cross-
validation. As mentioned earlier, the smaller the number of folds,
the larger the bias. Increasing the sample size will ease the prob-
lem. In general, bootstrap has larger bias and smaller uncertainty
than cross-validation. Efron came up the following “.632 estimator”
to alleviate this bias:

(0.632×𝑜𝑟𝑖𝑔𝑖𝑛𝑎𝑙 𝑏𝑜𝑜𝑡𝑠𝑡𝑟𝑎𝑝 𝑒𝑠𝑡𝑖𝑚𝑎𝑡𝑒)+(0.368×𝑎𝑝𝑝𝑎𝑟𝑒𝑛𝑡 𝑒𝑟𝑟𝑜𝑟 𝑟𝑎𝑡𝑒)

The apparent error rate is the error rate when the data is used
twice, both to fit the model and to check its accuracy and it is ap-
parently over-optimistic. The modified bootstrap estimate reduces
the bias but can be unstable with small samples size. This esti-
mate can also be unduly optimistic when the model severely over-
148 7 Model Tuning Strategy

fits since the apparent error rate will be close to zero. Efron and
Tibshirani (Efron and Tibshirani, 1997) discuss another technique,
called the “632+ method,” for adjusting the bootstrap estimates.
8
Measuring Performance

To compare different models, we need a way to measure model

performance. There are various metrics to use. To better under-
stand the strengths and weaknesses of a model, you need to look
at it through multiple metrics. In this chapter, we will introduce
some of the most common performance measurement metrics.
Load the R packages first:

# install packages from CRAN

p_needed <- c('caret', 'dplyr', 'randomForest',
'readr', 'car', 'pROC', 'fmsb', 'caret')

packages <- rownames(installed.packages())

p_to_install <- p_needed[!(p_needed %in% packages)]

if (length(p_to_install) > 0) {
install.packages(p_to_install)
}

lapply(p_needed, require, character.only = TRUE)

8.1 Regression Model Performance

MSE and RMSE
Mean Squared Error (MSE) measures the average of the squares
of the errors—that is, the average squared difference between the

149
150 8 Measuring Performance

estimated values and the actual value. The Root Mean Squared
Error (RMSE) is the root square of the MSE.

1 𝑛
𝑀 𝑆𝐸 = ∑(𝑦𝑖 − 𝑦𝑖̂ )2
𝑛 𝑖=1
1 𝑛
𝑅𝑀 𝑆𝐸 = √ ∑(𝑦𝑖 − 𝑦𝑖̂ )2
𝑛 𝑖=1

Both are the common measurements for the regression model per-
formance. Let’s use the previous income prediction as an example.
Fit a simple linear model:

sim.dat <- read.csv("http://bit.ly/2P5gTw4")

fit<- lm(formula = income ~ store_exp + online_exp + store_trans +
online_trans, data = sim.dat)
summary(fit)

##
## Call:
## lm(formula = income ~ store_exp + online_exp + store_trans +
## online_trans, data = sim.dat)
##
## Residuals:
## Min 1Q Median 3Q Max
## -128768 -15804 441 13375 150945
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 85711.680 3651.599 23.47 < 2e-16 ***
## store_exp 3.198 0.475 6.73 3.3e-11 ***
## online_exp 8.995 0.894 10.06 < 2e-16 ***
## store_trans 4631.751 436.478 10.61 < 2e-16 ***
## online_trans -1451.162 178.835 -8.11 1.8e-15 ***
## ---
## Signif. codes:
## 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
8.1 Regression Model Performance 151

##
## Residual standard error: 31500 on 811 degrees of freedom
## (184 observations deleted due to missingness)
## Multiple R-squared: 0.602,Adjusted R-squared: 0.6
## F-statistic: 306 on 4 and 811 DF, p-value: <2e-16

The fitted results fit shows the RMSE is 31530 (at the bottom of
the output after Residual standard error:).
Another common performance measure for the regression model
is R-Squared, often denoted as 𝑅2 . It is the square of the corre-
lation between the fitted value and the observed value. It is often
explained as the percentage of the information in the data that
can be explained by the model. The above model returns a R-
squared�0.6, which indicates the model can explain 60% of the
variance in variable income. While 𝑅2 is easy to explain, it is not
a direct measure of model accuracy but correlation. Here the 𝑅2
value is not low but the RMSE is 0.6 which means the average dif-
ference between model fitting and the observation is 3.153×104 . It
is a big discrepancy from an application point of view. When the
response variable has a large scale and high variance, a high 𝑅2
doesn’t mean the model has enough accuracy. It is also important
to remember that 𝑅2 is dependent on the variation of the outcome
variable. If the data has a response variable with a higher variance,
the model based on it tends to have a higher 𝑅2 .
We used 𝑅2 to show the impact of the error from independent
and response variables in Chapter 7 where we didn’t consider the
impact of the number of parameters (because the number of pa-
rameters is very small compared to the number of observations).
However, 𝑅2 increases as the number of parameters increases. So
people usually use Adjusted R-squared which is designed to miti-
gate the issue. The original 𝑅2 is defined as:

𝑅𝑆𝑆
𝑅2 = 1 −
𝑇 𝑆𝑆
𝑛 𝑛
where 𝑅𝑆𝑆 = ∑𝑖=1 (𝑦𝑖 − 𝑦𝑖̂ )2 and 𝑇 𝑆𝑆 = ∑𝑖=1 (𝑦𝑖 − 𝑦)̄ 2 .
Since RSS is always decreasing as the number of parameters in-
152 8 Measuring Performance

creases, 𝑅2 increases as a result. For a model with 𝑝 parameters,

the adjusted 𝑅2 is defined as:

𝑅𝑆𝑆/(𝑛 − 𝑝 − 1)
𝐴𝑑𝑗𝑢𝑠𝑡𝑒𝑑 𝑅2 = 1 −
𝑇 𝑆𝑆/(𝑛 − 1)

To maximize the adjusted 𝑅2 is identical to minimize 𝑅𝑆𝑆/(𝑛−𝑝−

1). Since the number of parameters 𝑝 is reflected in the equation,
𝑅𝑆𝑆/(𝑛 − 𝑝 − 1) can increase or decrease as 𝑝 increases. The
idea behind this is that the adjusted R-squared increases if the
new variable improves the model more than would be expected by
chance. It decreases when a predictor improves the model by less
than expected by chance. While values are usually positive, they
can be negative as well.
Another measurement is 𝐶𝑝 . For a least squared model with 𝑝
parameters:

1
𝐶𝑝 = (𝑅𝑆𝑆 + 2𝑝𝜎̂ 2 )
𝑛

where 𝜎̂ 2 is the estimator of the model random effect 𝜖. 𝐶𝑝 is to

add penalty 2𝑝𝜎̂ 2 to the training set 𝑅𝑆𝑆. The goal is to adjust the
over-optimistic measurement based on training data. As the num-
ber of parameters increases, the penalty increases. It counteracts
the decrease of 𝑅𝑆𝑆 due to increasing the number of parameters.
We choose the model with a smaller 𝐶𝑝 .
Both AIC and BIC are based on the maximum likelihood. In linear
regression, the maximum likelihood estimate is the least squared
estimate. The definitions of the two are:

𝐴𝐼𝐶 = 𝑛 + 𝑛𝑙𝑜𝑔(2𝜋) + 𝑛𝑙𝑜𝑔(𝑅𝑆𝑆/𝑛) + 2(𝑝 + 1)

𝐵𝐼𝐶 = 𝑛 + 𝑛𝑙𝑜𝑔(2𝜋) + 𝑛𝑙𝑜𝑔(𝑅𝑆𝑆/𝑛) + 𝑙𝑜𝑔(𝑛)(𝑝 + 1)

R function AIC() and BIC() will calculate the AIC and BIC value
according to the above equations. Many textbooks ignore content
8.2 Classification Model Performance 153

item 𝑛 + 𝑛𝑙𝑜𝑔(2𝜋), and use 𝑝 instead of 𝑝 + 1. Those slightly differ-

ent versions give the same results since we are only interested in
the relative value. Comparing to AIC, BIC puts a heavier penalty
on the number of parameters.

8.2 Classification Model Performance

This section focuses on performance measurement for models with
a categorical response. The metrics in the previous section are for
models with a continuous response and they are not appropriate
in the context of classification. Most of the classification problems
are dichotomous, such as an outbreak of disease, spam email, etc.
There are also cases with more than two categories as the seg-
ments in the clothing company data. We use swine disease data to
illustrate different metrics. Let’s train a random forest model as
an example. We will discuss the model in Chapter 11.

disease_dat <- read.csv("http://bit.ly/2KXb1Qi")

# you can check the data using glimpse()
# glimpse(disease_dat)

Separate the data to be training and testing sets. Fit model using
training data (xTrain and yTrain) and applied the trained model on
testing data (xTest and yTest) to evaluate model performance. We
use 70% of the sample as training and the rest 30% as testing.

set.seed(100)
# separate the data to be training and testing
trainIndex <- createDataPartition(disease_dat$y, p = 0.8,
list = F, times = 1)
xTrain <- disease_dat[trainIndex, ] %>% dplyr::select(-y)
xTest <- disease_dat[-trainIndex, ] %>% dplyr::select(-y)
# the response variable need to be factor
 154 8 Measuring Performance

yTrain <- disease_dat$y[trainIndex] %>% as.factor()

yTest <- disease_dat$y[-trainIndex] %>% as.factor()

Train a random forest model:

train_rf <- randomForest(yTrain ~ .,

data = xTrain,
mtry = trunc(sqrt(ncol(xTrain) - 1)),
ntree = 1000,
importance = T)

Apply the trained random forest model to the testing data to get
two types of predictions:
• probability (a value between 0 to 1)

yhatprob <- predict(train_rf, xTest, "prob")

set.seed(100)
car::some(yhatprob)

## 0 1
## 47 0.831 0.169
## 101 0.177 0.823
## 196 0.543 0.457
## 258 0.858 0.142
## 274 0.534 0.466
## 369 0.827 0.173
## 389 0.852 0.148
## 416 0.183 0.817
## 440 0.523 0.477
## 642 0.836 0.164

• category prediction (0 or 1)

yhat <- predict(train_rf, xTest)

car::some(yhat)
8.2 Classification Model Performance 155

## 146 232 269 302 500 520 521 575 738 781
## 0 0 1 0 0 0 1 0 0 0
## Levels: 0 1

We will use the above two types of predictions to show different

performance metrics.

8.2.1 Confusion Matrix

Confusion Matrix is a counting table to describe the perfor-
mance of a classification model. For the true response yTest and
prediction yhat, the confusion matrix is:

yhat = as.factor(yhat) %>% relevel("1")

yTest = as.factor(yTest) %>% relevel("1")
table(yhat, yTest)

## yTest
## yhat 1 0
## 1 56 1
## 0 15 88

The top-left and bottom-right are the numbers of correctly clas-

sified samples. The top-right and bottom-left are the numbers of
wrongly classified samples. A general confusion matrix for a binary
classifier is following:

Predicted Yes Predicted No

Actual Yes TP FN
Actual No FP TN

where TP is true positive, FP is false positive, TN is true negative,

FN is false negative. The cells along the diagonal line from top-
left to bottom-right contain the counts of correctly classified sam-
ples. The cells along the other diagonal line contain the counts of
wrongly classified samples. The most straightforward performance
156 8 Measuring Performance

measure is the total accuracy which is the percentage of correctly

classified samples:

𝑇𝑃 + 𝑇𝑁
𝑇 𝑜𝑡𝑎𝑙 𝑎𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁

You can calculate the total accuracy when there are more than
two categories. This statistic is straightforward but has some dis-
advantages. First, it doesn’t differentiate different error types. In
a real application, different types of error may have different im-
pacts. For example, it is much worse to tag an important email as
spam and miss it than failing to filter out a spam email. Provost
et al. (Provost F, 1998) discussed in detail the problem of using
total accuracy on different classifiers. There are some other met-
rics based on the confusion matrix that measure different types of
error.
Precision is a metric to measure how accurate positive predic-
tions are (i.e. among those emails predicted as spam, how many
percentages of them are spam emails?):

𝑇𝑃
𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =
𝑇𝑃 + 𝐹𝑃

Sensitivity is to measure the coverage of actual positive samples

(i.e. among those spam emails, how many percentages of them are
predicted as spam) :

𝑇𝑃
𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 =
𝑇𝑃 + 𝐹𝑁

Specificity is to measure the coverage of actual negative samples

(i.e. among those non-spam emails, how many percentages of them
pass the filter):

𝑇𝑁
𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 =
𝑇𝑁 + 𝐹𝑃

Since wrongly tagging an important email as spam has a bigger

8.2 Classification Model Performance 157

impact, in the spam email case, we want to make sure the model
specificity is high enough.
Second, total accuracy doesn’t reflect the natural frequencies of
each class. For example, the percentage of fraud cases for insurance
may be very low, like 0.1%. A model can achieve nearly perfect
accuracy (99.9%) by predicting all samples to be negative. The
percentage of the largest class in the training set is also called the
no-information rate. In this example, the no-information rate is
99.9%. You need to get a model that at least beats this rate.

8.2.2 Kappa Statistic

Another metric is the Kappa statistic. It measures the agreement
between the observed and predicted classes. It was originally come
up by Cohen etc. (J, 1960). Kappa takes into account the accuracy
generated simply by chance. It is defined as:

𝑃0 − 𝑃𝑒
𝐾𝑎𝑝𝑝𝑎 =
1 − 𝑃𝑒

Let 𝑛 = 𝑇 𝑃 + 𝑇 𝑁 + 𝐹 𝑃 + 𝐹 𝑁 be the total number of

samples, where 𝑃0 = 𝑇 𝑃 +𝑇 𝑛
𝑁
is the observed accuracy, 𝑃𝑒 =
(𝑇 𝑃 +𝐹 𝑃 )(𝑇 𝑃 +𝐹 𝑁)+(𝐹 𝑁+𝑇 𝑁)(𝐹 𝑃 +𝑇 𝑁)
𝑛2 is the expected accuracy based
on the marginal totals of the confusion matrix. Kappa can take on
a value from -1 to 1. The higher the value, the higher the agree-
ment. A value of 0 means there is no agreement between the ob-
served and predicted classes, while a value of 1 indicates perfect
agreement. A negative value indicates that the prediction is in the
opposite direction of the observed value. The following table may
help you “visualize” the interpretation of kappa (Landis JR, 1977):

Kappa Agreement
<0 Less than chance agreement
0.01–0.20 Slight agreement
0.21– 0.40 Fair agreement
0.41–0.60 Moderate agreement
0.61–0.80 Substantial agreement
158 8 Measuring Performance

Kappa Agreement
0.81–0.99 Almost perfect agreement

In general, a value between 0.3 to 0.5 indicates a reasonable agree-

ment. If a model has a high accuracy of 90%, while the expected
accuracy is also high, say 85%. The Kappa statistics is 31 . It means
the prediction and the observation have a fair agreement. You can
calculate Kappa when the number of categories is larger than 2.
The package fmsb has a function Kappa.test() to calculate Cohen’s
Kappa statistics. The function can also return the hypothesis test
result and a confidence interval. Use the above observation vector
yTest and prediction vector yhat as an example, you can calculate
the statistics:

# install.packages("fmsb")
kt<-fmsb::Kappa.test(table(yhat,yTest))
kt$Result

##
## Estimate Cohen's kappa statistics and test the
## null hypothesis that the extent of agreement is
## same as random (kappa=0)
##
## data: table(yhat, yTest)
## Z = 9.7, p-value <2e-16
## 95 percent confidence interval:
## 0.6972 0.8894
## sample estimates:
## [1] 0.7933

The output of the above function contains an object named Judge-

ment:

kt$Judgement

## [1] "Substantial agreement"

8.2 Classification Model Performance 159

8.2.3 ROC
Receiver Operating Characteristic (ROC) curve uses the predicted
class probabilities and determines an effective threshold such that
values above the threshold are indicative of a specific event. We
have shown the definitions of sensitivity and specificity above. The
sensitivity is the true positive rate and specificity is true negative
rate. “1 - specificity” is the false negative rate. ROC is a graph
of pairs of true positive rate (sensitivity) and false positive rate
(1-specificity) values that result as the test’s cutoff value is varied.
The Area Under the Curve(AUC) is a common measure for two-
class problem. There is usually a trade-off between sensitivity and
specificity. If the threshold is set lower, then there are more sam-
ples predicted as positive and hence the sensitivity is higher. Let’s
look at the predicted probability yhatprob in the swine disease ex-
ample. The predicted probability object yhatprob has two columns,
one is the predicted probability that a farm will have an outbreak,
the other is the probability that farm will NOT have an outbreak.
So the two add up to have value 1. We use the probability of out-
break (the 2nd column) for further illustration. You can use roc()
function to get an ROC object (rocCurve) and then apply different
functions on that object to get needed plot or ROC statistics. For
example, the following code produces the ROC curve:

rocCurve <- pROC::roc(response = yTest,

predictor = yhatprob[,2])

## Setting levels: control = 1, case = 0

## Setting direction: controls > cases

plot(1-rocCurve$specificities,
rocCurve$sensitivities,
type = 'l',
xlab = '1 - Specificities',
ylab = 'Sensitivities')
160 8 Measuring Performance

1.0
0.8
Sensitivities

0.6
0.4
0.2
0.0

0.0 0.2 0.4 0.6 0.8 1.0

1 - Specificities

The first argument of the roc() is, response, the observation vec-
tor. The second argument is predictor is the continuous prediction
(probability or link function value). The x-axis of ROC curve is
“1 - specificity” and the y-axis is “sensitivity”. ROC curve starts
from (0, 0) and ends with (1, 1). A perfect model that correctly
identifies all the samples will have 100% sensitivity and specificity
which corresponds to the curve that also goes through (0, 1). The
area under the perfect curve is 1. A model that is totally useless
corresponds to a curve that is close to the diagonal line and an
area under the curve about 0.5.
You can visually compare different models by putting their ROC
curves on one plot. Or use the AUC to compare them. DeLong et
al. came up a statistic test to compare AUC based on U-statistics
(E.R. DeLong, 1988) which can give a p-value and confidence in-
terval. You can also use bootstrap to get a confidence interval for
AUC (Hall P, 2004).
We can use the following code in R to get an estimate of AUC and
its confidence interval:

# get the estimate of AUC

auc(rocCurve)
8.2 Classification Model Performance 161

## Area under the curve: 0.989

# get a confidence interval based on DeLong et al.

ci.auc(rocCurve)

## 95% CI: 0.979-0.999 (DeLong)

AUC is robust to class imbalance(Provost F, 1998; T, 2006) hence

a popular measurement. But it still boils a lot of information down
to one number so there is inevitably a loss of information. It is bet-
ter to double check by comparing the curve at the same time. If
you care more about getting a model that will have high speci-
ficity, which is the lower part of the curve, as in the spam filtering
case, you can use the area of the lower part of the curve as the
performance measurement (D, 1989). ROC is only for two-class
case. Some researchers generalized it to situations with more than
two categories (Hand D, 2001; Lachiche N, 2003; Li J, 2008).

8.2.4 Gain and Lift Charts

Gain and lift chart is a visual tool for evaluating the performance
of a classification model. In the previous swine disease example,
there are 160 samples in the testing data and 89 of them have a
positive outcome.

table(yTest)

## yTest
## 1 0
## 71 89

If we order the testing samples by the predicted probability, one

would hope that the positive samples are ranked higher than the
negative ones. That is what the lift charts do: rank the samples
by their scores and calculate the cumulative positive rate as more
samples are evaluated. In the perfect scenario, the highest-ranked
71 samples would contain all 71 positive samples. When the model
is totally random, the highest-ranked x% of the data would contain
162 8 Measuring Performance

about x% of the positive sample. The gain/lift charts compare the

ratio between the results obtained with and without a model.
Let’s plot the lift charts to compare the predicted outbreak prob-
ability (modelscore <- yhatprob[ ,2]) from random forest model we
fit before with some random scores generated from a uniform dis-
tribution (randomscore <- runif(length(yTest))).

# predicted outbreak probability

modelscore <- yhatprob[ ,2]
# randomly sample from a uniform distribution
randomscore <- runif(length(yTest))
labs <- c(modelscore="Random Forest Prediction",
randomscore="Random Number")
liftCurve <- caret::lift(yTest ~ modelscore + randomscore,
class = "1",
labels = labs)
xyplot(liftCurve, auto.key = list(columns = 2, lines = T, points = F))

Random Forest Prediction Random Number

100

80
% Samples Found

60

40

20

0 20 40 60 80 100

% Samples Tested

The x-axis is the percentage of samples tested and the y-axis is

the percentage of positive samples that are detected by the model.
8.2 Classification Model Performance 163

For example, the point on the curve of random forest prediction,

�8.125,18.31�, indicates that if you order the testing samples by
the predicted probability from high to low, the top 8.125% of the
samples contain 0.1831% of the total positive outcomes.
Similar to the ROC curve, we can choose the model by comparing
their lift charts. Some parts of the lift curves may be more inter-
esting than the rest. For example, if you only have a budget to
clean 50% of the farms, then you should pick the model that gives
the highest point when the x-axis is 50%.
9
Regression Models

In this chapter, we will cover ordinary linear regression and a few

more advanced regression methods. The linear combination of vari-
ables seems simple compared to many of today’s machine learning
models. However, many advanced models use linear combinations
of variables as one of its major components or steps. For example,
for each neuron in the deep neural network, all the input signals
are first linearly combined before feeding to a non-linear activation
function. To understand many of today’s machine learning models,
it is helpful to understand the key ideas across different modeling
frameworks.
First, we will introduce multivariate linear regression (i.e. the typ-
ical least square regression) which is one of the simplest supervised
learning methods. Even though it is simple, the general ideas and
procedures of fitting a regression model are applied to a boarder
scope. Having a solid understanding of the basic linear regression
model enables us to learn more advanced models easily. For exam-
ple, we will introduce two “shrinkage” versions of linear regression:
ridge regression and LASSO regression. While the parameters are
fitted by the least square method, the extra penalty can effectively
shrink model parameters towards zero. It mediates overfitting and
maintains the robustness of the model when data size is small
compared to the number of explanatory variables. We first intro-
duce basic knowledge of each model and then provide R codes to
show how to fit the model. We only cover the major properties of
these models and the listed reference will provide more in-depth
discussion.
We will use the clothing company data as an example. We want to
answer business questions such as “which variables are the driving

165
166 9 Regression Models

factors of total revenue (both online and in-store purchase)?” The

answer to this question can help the company to decide where to
invest (such as design, quality, etc). Note that the driving factor
here does not guarantee a causal relationship. Linear regression
models reveal correlation rather than causation. For example, if a
survey on car purchase shows a positive correlation between price
and customer satisfaction, does it suggest the car dealer should in-
crease the price? Probably not! It is more likely that the customer
satisfaction is impacted by quality. And a higher quality car tends
to be more expensive. Causal inference is much more difficult to
establish and we have to be very careful when interpreting regres-
sion model results.
Load the R packages first:

# install packages from CRAN

p_needed <- c('caret', 'dplyr', 'lattice',
'elasticnet', 'lars', 'corrplot',
'pls')

packages <- rownames(installed.packages())

p_to_install <- p_needed[!(p_needed %in% packages)]
if (length(p_to_install) > 0) {
install.packages(p_to_install)
}

lapply(p_needed, require, character.only = TRUE)

9.1 Ordinary Least Square

For a typical linear regression with 𝑝 explanatory variables, we
have a linear combinations of these variables:
9.1 Ordinary Least Square 167

𝑝
𝑓(X) = X𝜷 = 𝛽0 + ∑ x.j 𝛽𝑗
𝑗=1

where 𝜷 is the parameter vector with length 𝑝 + 1. Here we use

x.j for column vector and xi. for row vector. Least square is the
method to find a set of value for 𝜷𝐓 = (𝛽0 , 𝛽1 , ..., 𝛽𝑝 ) such that it
minimizes the residual sum of square (RSS):

𝑁 𝑁 𝑝
𝑅𝑆𝑆(𝛽) = ∑(𝑦𝑖 − 𝑓(xi. )) = ∑(𝑦𝑖 − 𝛽0 − ∑ 𝑥𝑖𝑗 𝛽𝑗 )2
2

𝑖=1 𝑖=1 𝑗=1

The process of finding a set of values has been implemented in R.

Now let’s load the data:

dat <- read.csv("http://bit.ly/2P5gTw4")

Before fitting the model, we need to clean the data, such as re-
moving bad data points that are not logical (negative expense).

dat <- subset(dat, store_exp > 0 & online_exp > 0)

Use 10 survey question variables as our explanatory variables.

modeldat <- dat[, grep("Q", names(dat))]

The response variable is the sum of in-store spending and online

spending.

# total expense = in store expense + online expense

modeldat$total_exp <- dat$store_exp + dat$online_exp

To fit a linear regression model, let us first check if there are any
missing values or outliers:
168 9 Regression Models

par(mfrow = c(1, 2))

hist(modeldat$total_exp, main = "", xlab = "total_exp")
boxplot(modeldat$total_exp)
800

50000
600
Frequency

30000
400
200

10000
0

0 20000 40000

total_exp

There is no missing value in the response variable, but there are

outliers. Outliers are usually best described by the problem to
solve itself such that we know from domain knowledge that it is
not possible to have such values. We can also use a statistical
threshold to remove extremely large or small outlier values from
the data. We use the Z-score to find and remove outliers described
in section 5.5. Readers can refer to the section for more detail.

y <- modeldat$total_exp
# Find data points with Z-score larger than 3.5
zs <- (y - mean(y))/mad(y)
modeldat <- modeldat[-which(zs > 3.5), ]

We will not perform log-transformation for the response variable

at this stage. Let us first check the correlation among explanatory
variables:
9.1 Ordinary Least Square 169

correlation <- cor(modeldat[, grep("Q", names(modeldat))])

corrplot::corrplot.mixed(correlation, order = "hclust", tl.pos = "lt",
upper = "ellipse")

Q1
0.8
Q8 0.24
0.6
Q6 0.53 0.68
0.4
Q10 0.72 0.6 0.85
0.2
Q7 -0.61 -0.74 -0.93 -0.88
0
Q5 -0.79 -0.56 -0.86 -0.91 0.9
-0.2
Q9 -0.78 -0.48 -0.8 -0.86 0.83 0.92
-0.4
Q2 0.2 -0.06 0.4 0.27 -0.25 -0.28 -0.3
-0.6
Q3 0.21 -0.64 -0.1 -0.07 0.24 0.01 -0.07 0.49
-0.8
Q4 0.57 -0.34 0.2 0.32 -0.14 -0.42 -0.46 0.44 0.75
-1

FIGURE 9.1: Correlation Matrix Plot for Explanatory Variables

As shown in figure 9.1, there are some highly correlated variables.

Let us use the method described in section 5.6 to find potential
highly correlated explanatory variables to remove with a threshold
of 0.75:

highcor <- findCorrelation(correlation, cutoff = 0.75)

modeldat <- modeldat[, -highcor]

The dataset is now ready to fit a linear regression model. The

standard format to define a regression in R is:

(1) response variable is at the left side of ~

170 9 Regression Models

(2) the explanatory variables are at the right side of ~

(3) if all the variables in the dataset except the response vari-
able are included in the model, we can use . at the right
side of ~
(4) if we want to consider the interaction between two vari-
ables such as Q1 and Q2, we can add an interaction term
Q1*Q2

(5) transformation of variables can be added directly to vari-

able names such as log(total_exp).

lmfit <- lm(log(total_exp) ~ ., data = modeldat)

summary(lmfit)

##
## Call:
## lm(formula = log(total_exp) ~ ., data = modeldat)
##
## Residuals:
## Min 1Q Median 3Q Max
## -1.1749 -0.1372 0.0128 0.1416 0.5623
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 8.09831 0.05429 149.18 < 2e-16 ***
## Q1 -0.14534 0.00882 -16.47 < 2e-16 ***
## Q2 0.10228 0.01949 5.25 2.0e-07 ***
## Q3 0.25445 0.01835 13.87 < 2e-16 ***
## Q6 -0.22768 0.01152 -19.76 < 2e-16 ***
## Q8 -0.09071 0.01650 -5.50 5.2e-08 ***
## ---
## Signif. codes:
## 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.226 on 805 degrees of freedom
## Multiple R-squared: 0.854,Adjusted R-squared: 0.853
9.1 Ordinary Least Square 171

## F-statistic: 943 on 5 and 805 DF, p-value: <2e-16

The summary(lmfit) presents a summary of the model fit. It shows

the point estimate of each explanatory variable (the Estimate col-
umn), their corresponding standard error (the Std. Error column),
t values (t value), and p values (Pr(>|t|)).

9.1.1 The Magic P-value

Let us pause a little to have a short discussion about p-value.
Misuse of p-value is common in many research fields. There were
heated discussions about P-value in the past. Siegfried commented
in his 2010 Science News article:

“It’s science’s dirtiest secret: The scientific method of testing

hypotheses by statistical analysis stands on a flimsy foundation.”

American Statistical Association (i.e., ASA) released an official

statement on p-value in 2016 (Ronald L. Wassersteina, 2016). It
was the first time to have an organization level announcement
about p-value. ASA stated that the goal to release this guidance
was to

“improve the conduct and interpretation of quantitative science

and inform the growing emphasis on reproducibility of science
research.”

The statement also noted that

172 9 Regression Models

“the increased quantification of scientific research and a pro-

liferation of large, complex data sets has expanded the scope
for statistics and the importance of appropriately chosen tech-
niques, properly conducted analyses, and correct interpretation.”

The statement’s six principles, many of which address misconcep-

tions and misuse of the P-value, are the following:

1. P-values can indicate how incompatible the data are with

a specified statistical model.
2. P-values do not measure the probability that the studied
hypothesis is true or the probability that the data were
produced by random chance alone.
3. Scientific conclusions and business or policy decisions
should not be based only on whether a p-value passes
a specific threshold.
4. Proper inference requires full reporting and transparency.
5. A p-value, or statistical significance, does not measure the
size of an effect or the importance of a result.
6. By itself, a p-value does not provide a good measure of
evidence regarding a model or hypothesis.

The 𝑝 = 0.05 threshold is not based on any scientific calculation

but is an arbitrary number. It means that practitioners can use a
different threshold if they think it better fits the problem to solve.
We do not promote p-value in this book. However, the p-value is
hard to avoid in classical statistical inference. In practice, when
making classic statistical inferences, we recommend reporting con-
fidence interval whenever possible instead of P-value.
The Bayesian paradigm is an alternative to the classical paradigm.
A Bayesian can state probabilities about the parameters, which
are considered random variables. However, it is not possible in the
classical paradigm. In our work, We use hierarchical (generalize)
linear models in practice instead of classical linear regression. Hier-
archical models pool information across clusters (for example, you
9.1 Ordinary Least Square 173

can treat each customer segment as a cluster). This pooling tends

to improve estimates of each cluster, especially when sampling is
imbalanced. Because the models automatically cope with differing
uncertainty introduced by sampling imbalance (bigger cluster has
smaller variance), it prevents over-sampled clusters from unfairly
dominating inference.
This book does not cover the Bayesian framework. The best ap-
plied Bayesian book is Statistical Rethinking1 by Richard McEl-
reath (McElreath, 2020). The book provides outstanding concep-
tual explanations and a wide range of models from simple to ad-
vanced with detailed, repeatable code. The text uses R, but there
are code examples for Python and Julia on the book website.
Now let us come back to our example. We will not spend too much
time on p-values, while we will focus on the confidence interval for
the parameter estimate for each explanatory variable. In R, the
function confint() can produce the confidence interval for each
parameter:

confint(lmfit,level=0.9)

## 5 % 95 %
## (Intercept) 8.00892 8.18771
## Q1 -0.15987 -0.13081
## Q2 0.07018 0.13437
## Q3 0.22424 0.28466
## Q6 -0.24665 -0.20871
## Q8 -0.11787 -0.06354

The above output is for a 90% confidence level as level=0.9 indi-

cated in the function call. We can change the confidence level by
adjusting the level setting.
Fitting a linear regression is so easy using R that many analysts
directly write reports without thinking about whether the model
is meaningful. On the other hand, we can easily use R to check
1
https://xcelab.net/rm/statistical-rethinking/
 174 9 Regression Models

model assumptions. In the following sections, we will introduce

a few commonly used diagnostic methods for linear regression to
check whether the model assumptions are reasonable.

9.1.2 Diagnostics for Linear Regression

In linear regression , we would like the Ordinary Least Square
(OLS) estimate to be the Best Linear Unbiased Estimate (BLUE)
. In other words, we hope the expected value of the estimate is the
actual parameter value (i.e., unbiased) and achieving minimized
residual (i.e., best). Based on the Gauss-Markov theorem, the OLS
estimate is BLUE under the following conditions:

1. Explanatory variables (x.j ) and random error (𝝐) are in-

dependent: 𝑐𝑜𝑣(x.j , �) = 0 for ∀𝑗 = 𝑗 ∈ 1...𝑝.
2. The expected value of random error is zero: 𝐸(𝝐|𝐗) = 0
3. Random errors are independent with each other, and the
variance of random error is consistent: 𝑉 𝑎𝑟(𝝐) = 𝜎2 𝐼,
where 𝜎 is positive and 𝐼 is a 𝑛 × 𝑛 identical matrix.

We will introduce four graphic diagnostics for the above assump-

tions.

(1) Residual plot

It is a scatter plot with residual on the Y-axis and fitted value on

the X-axis. We can also put any of the explanatory variables on the
X-axis. Under the assumption, residuals are randomly distributed,
and we need to check the following:
• Are residuals centered around zero?
• Are there any patterns in the residual plots (such as residuals
with x-values farther from 𝑥̄ have greater variance than residuals
with x-values closer to 𝑥)?
̄
• Are the variances of the residual consistent across a range of
fitted values?
Please note that even if the variance is not consistent, the re-
9.1 Ordinary Least Square 175

gression parameter’s point estimate is still unbiased. However, the

variance estimate is not unbiased. Because the significant test for
regression parameters is based on the random error distribution,
these tests are no longer valid if the variance is not constant.

(2) Normal quantile-quantile Plot (Q-Q Plot)

Q-Q Plot is used to check the normality assumption for the resid-
ual. For normally distributed residuals, the data points should fol-
low a straight line along the Q-Q plot. The more departure from
a straight line, the more departure from a normal distribution for
the residual.

(3) Standardized residuals plot

Standardized residual is the residual normalized by an estimate of

its standard deviation. Like the residual plot, the X-axis is still the
fitted value, but the y-axis is now standardized residuals. Because
of the normalization, the y-axis shows the number of standard de-
viations from zero. A value greater than 2 or less than -2 indicates
observations with large standardized residuals. The plot is useful
because when the variance is not consistent, it can be difficult to
detect the outliers using the raw residuals plot.

(4) Cook’s distance

Cook’s distance can check influential points in OLS based linear

regression models. In general, we need to pay attention to data
points with Cook’s distance > 0.5.
In R, these diagnostic graphs are built in theplot() function.

par(mfrow = c(2, 2))

plot(lmfit, which = 1)
plot(lmfit, which = 2)
plot(lmfit, which = 3)
plot(lmfit, which = 4)
 176 9 Regression Models

Standardized residuals
Residuals vs Fitted Normal Q-Q

0 2
Residuals

0.0
678

-1.0
155 678
155

-4
960 960

6.5 7.0 7.5 8.0 8.5 -3 -2 -1 0 1 2 3

Fitted values Theoretical Quantiles

Standardized residuals

Scale-Location Cook's distance

Cook's distance
960 257
254
310
2.0

155 678

0.03
1.0

0.00
0.0

6.5 7.0 7.5 8.0 8.5 0 200 400 600 800

Fitted values Obs. number

FIGURE 9.2: Linear Regression Diagnostic Plots: residual plot

(top left), Q-Q plot (top right), standardized residuals plot (lower
left), Cook’s distance (lower right)

The above diagnostic plot examples show:

• Residual plot: residuals are generally distributed around 𝑦 = 0
horizontal line. There are no significant trends or patterns in
this residual plot (there are two bumps but does not seem too
severe). So the linear relationship assumption between the re-
sponse variable and explanatory variables is reasonable.
• Q-Q plot: data points are pretty much along the diagonal line
of Y=X, indicating no significant normality assumption depar-
ture for the residuals. Because we simulate the data, we know
the response variable before log transformation follows a normal
distribution. The shape of the distribution does not deviate from
a normal distribution too much after log transformation.
• Standardized residual plot: if the constant variance assumption is
valid, then the plot’s data points should be randomly distributed
around the horizontal line. We can see there are three outliers
on the plot. Let us check those points:
 9.2 Ordinary Least Square 177

modeldat[which(row.names(modeldat) %in% c(960, 678, 155)), ]

## Q1 Q2 Q3 Q6 Q8 total_exp
## 155 4 2 1 4 4 351.9
## 678 2 1 1 1 2 1087.3
## 960 2 1 1 1 3 658.3

It is not easy to see why those records are outliers from the above
output. It will be clear conditional on the independent variables
(Q1, Q2, Q3, Q6, and Q8). Let us examine the value of total_exp for
samples with the same Q1, Q2, Q3, Q6, and Q8 answers as the
3rd row above.

datcheck = modeldat %>% filter(Q1 ==2 & Q2 == 1 & Q3 == 1 & Q6 == 1 & Q8 == 3)

nrow(datcheck)

## [1] 87

There are 87 samples with the same values of independent vari-

ables. The response variable’s (total_exp) distribution is:

summary(datcheck$total_exp)

## Min. 1st Qu. Median Mean 3rd Qu. Max.

## 658 1884 2215 2204 2554 3197

Now it is easy to see why row 960 with total_exp = 658.3 is an

outlier. All the other 86 records with the same survey responses
have a much higher total expense!
• Cook’s distance: the maximum of Cook’s distance is around 0.05.
Even though the graph does not have any point with Cook’s
distance of more than 0.5, we could spot some outliers.
The graphs suggest some outliers, but it is our decision what to
do. We can either remove it or investigate it further. If the values
are not due to any data error, we should consider them in our
analysis.
178 9 Regression Models

9.2 Principal Component Regression and Partial Least

Square
In real-life applications, explanatory variables are usually related
to each other, containing similar information. For example, in
the previous chapter, we used expenditure variables to predict
consumer income. In that model, store expenditure (store_exp),
online expenditure (online_exp), number of store transactions
(store_trans), and number of online transactions (online_trans) are
correlated to a certain extent, especially the number of transac-
tions and expenditure. If there is a strong correlation among ex-
planatory variables, then the least square-based linear regression
model may not be robust. If the number of observations is less than
the number of explanatory variables, the standard least square
method cannot provide a unique set of coefficient estimates. We
can perform data preprocessing, such as remove highly correlated
variables with a preset threshold for such cases. However, this
approach cannot guarantee a low correlation of the linear com-
bination of the variables with other variables. In that case, the
standard least square method will still not be robust. We need to
be aware that removing highly correlated variables is not always
guarantee a robust solution. We can also apply feature engineering
procedures to explanatory variables such as principal component
analysis (PCA). By using principal components, we can ensure
they are uncorrelated with each other. However, the drawback of
using PCA is that these components are linear combinations of
original variables, and it is difficult to explain the fitted model.
Principal component regression (PCR) is described in more detail
in (W, 1965). It can be used when there are strong correlations
among variables or when the number of observations is less than
the number of variables.
In theory, we can use PCR to reduce the number of variables used
in a linear model, but the results are not good. Because the first
a few principal components may not be good candidates to model
the response variable. PCA is unsupervised learning such that the
9.2 Principal Component Regression and Partial Least Square 179

entire process does not consider the response variable. In PCA,

it only focuses on the variability of explanatory variables. When
the independent variables and response variables are related, PCR
can well identify the systematic relationship between them. How-
ever, when there exist independent variables not associated with
response variable, it will undermine PCR’s performance. We need
to be aware that PCR does not make feature selections, and each
principal component is a combination of original variables.
Partial least square regression (PLS) is the supervised version
of PCR. Similar to PCR, PLS can also reduce the number of
variables in the regression model. As PLS is also related to
the variables’ variance, we usually standardize or normalize vari-
ables before PLS. Suppose we have a list of explanatory variables
X = [𝑋1 , 𝑋2 , ..., 𝑋𝑝 ]𝑇 , and their variance-covariance matrix is Σ.
PLS also transforms the original variables using linear combina-
tion to new uncorrelated variables (𝑍1 , 𝑍2 , … , 𝑍𝑚 ). When 𝑚 = 𝑝,
the result of PLS is the same as OLS. The main difference be-
tween PCR and PLS is the process of creating new variables. PLS
considers the response variable.
PLS is from Herman Wold’s Nonlinear Iterative Partial Least
Squares (NIPALS) algorithm (Wold, 1973; Wold and Jöreskog,
1982) . Later NIPALS was applied to regression problems, which
was then called PLS . PLS is a method of linearizing nonlinear re-
lationships through hidden layers. It is similar to the PCR, except
that PCR does not take into account the information of the depen-
dent variable when selecting the components, and its purpose is to
find the linear combinations (i.e., unsupervised) that capture the
most variance of the independent variables. And PLS maximizes
the linear combination of dependencies with the response variable.
In the current case, the more complicated PLS does not perform
better than simple linear regression. We will not discuss the PLS
algorithm’s detail, and the reference mentioned above provides a
more detailed description of the algorithm.
We focus on using R library caret to fit PCR and PLS models.
Let us use the 10 survey questions (Q1-Q10) as the explanatory
180 9 Regression Models

variables and income as the response variable. First load the data
and preprocessing the data:

library(lattice)
library(caret)
library(dplyr)
library(elasticnet)
library(lars)

# Load Data
sim.dat <- read.csv("http://bit.ly/2P5gTw4")
ymad <- mad(na.omit(sim.dat$income))

# Calculate Z values
zs <- (sim.dat$income - mean(na.omit(sim.dat$income)))/ymad
# which(na.omit(zs>3.5)) find outlier
# which(is.na(zs)) find missing values
idex <- c(which(na.omit(zs > 3.5)), which(is.na(zs)))
# Remove rows with outlier and missing values
sim.dat <- sim.dat[-idex, ]

Now let’s define explanatory variable matrix (xtrain) by selecting

these 10 survey questions columns, and define response variable
(ytrain):

xtrain = dplyr::select(sim.dat, Q1:Q10)

ytrain = sim.dat$income

We also set up random seed and 10-folder cross-validation:

set.seed(100)
ctrl <- trainControl(method = "cv", number = 10)

Fit PLS model using number of explanatory variables as the hyper-

parameter to tune. As there are at most 10 explanatory variables
9.2 Principal Component Regression and Partial Least Square 181

in the model, we set up the hyper-parameter tuning range to be 1

to 10:

plsTune <- train(xtrain, ytrain,

method = "pls",
# set hyper-parameter tuning range
tuneGrid = expand.grid(.ncomp = 1:10),
trControl = ctrl)
plsTune

## Partial Least Squares

##
## 772 samples
## 10 predictor
##
## No pre-processing
## Resampling: Cross-Validated (10 fold)
## Summary of sample sizes: 696, 693, 694, 694, 696, 695, ...
## Resampling results across tuning parameters:
##
## ncomp RMSE Rsquared MAE
## 1 27777 0.6534 19845
## 2 24420 0.7320 15976
## 3 23175 0.7590 14384
## 4 23011 0.7625 13808
## 5 22977 0.7631 13737
## 6 22978 0.7631 13729
## 7 22976 0.7631 13726
## 8 22976 0.7631 13726
## 9 22976 0.7631 13726
## 10 22976 0.7631 13726
##
## RMSE was used to select the optimal model using
## the smallest value.
## The final value used for the model was ncomp = 7.

From the result, we can see that the optimal number of variables is
182 9 Regression Models

7. However, if we pay attention to the RMSE improvement, we will

find only minimum improvement in RMSE after three variables.
In practice, we could choose to use the model with three variables
if the improvement does not make a practical difference, and we
would rather have a simpler model.
We can also find the relative importance of each variable during
PLS model tuning process, as described using the following code:

plsImp <- varImp(plsTune, scale = FALSE)

plot(plsImp, top = 10, scales = list(y = list(cex = 0.95)))

Q6

Q2

Q3

Q1

Q4

Q7

Q8

Q9

Q5

Q10

2000 4000 6000 8000

Importance

The above plot shows that Q1, Q2, Q3, and Q6, are more impor-
tant than other variables.Now let’s fit a PCR model with number
of principal components as the hyper-parameter:

# Set random seed

set.seed(100)
pcrTune <- train(x = xtrain, y = ytrain,
method = "pcr",
# set hyper-parameter tuning range
9.2 Principal Component Regression and Partial Least Square 183

tuneGrid = expand.grid(.ncomp = 1:10),

trControl = ctrl)
pcrTune

## Principal Component Analysis

##
## 772 samples
## 10 predictor
##
## No pre-processing
## Resampling: Cross-Validated (10 fold)
## Summary of sample sizes: 696, 693, 694, 694, 696, 695, ...
## Resampling results across tuning parameters:
##
## ncomp RMSE Rsquared MAE
## 1 45958 0.03243 36599
## 2 32460 0.52200 24041
## 3 23235 0.75774 14516
## 4 23262 0.75735 14545
## 5 23152 0.75957 14232
## 6 23133 0.76004 14130
## 7 23114 0.76049 14129
## 8 23115 0.76045 14130
## 9 22991 0.76283 13801
## 10 22976 0.76308 13726
##
## RMSE was used to select the optimal model using
## the smallest value.
## The final value used for the model was ncomp = 10.

From the output, the default recommendation is ten components.

However, if we pay attention to RMSE improvement with more
components, we will find little difference after the model with three
components. Again, in practice, we can keep models with three
components.
184 9 Regression Models

Now let’s compare the hyper-parameter tuning process for PLS

and PCR:

# Save PLS model tuning information to plsResamples

plsResamples <- plsTune$results
plsResamples$Model <- "PLS"
# Save PCR model tuning information to plsResamples
pcrResamples <- pcrTune$results
pcrResamples$Model <- "PCR"
# Combine both output for plotting
plsPlotData <- rbind(plsResamples, pcrResamples)
# Leverage xyplot() function from lattice library
xyplot(RMSE ~ ncomp,
data = plsPlotData,
xlab = "# Components",
ylab = "RMSE (Cross-Validation)",
auto.key = list(columns = 2),
groups = Model,
type = c("o", "g"))

PCR PLS

45000
RMSE (Cross-Validation)

40000

35000

30000

25000

2 4 6 8 10

# Components
9.2 Principal Component Regression and Partial Least Square 185

The plot confirms our choice of using a model with three compo-
nents for both PLS and PCR.
10
Regularization Methods

The regularization method is also known as the shrinkage method.

It is a technique that constrains or regularizes the coefficient esti-
mates. By imposing a penalty on the size of coefficients, it shrinks
the coefficient estimates towards zero. It also intrinsically conduct
feature selection and is naturally resistant to non-informative pre-
dictors. It may not be obvious why this technique improves model
performance, but it turns out to be a very effective modeling tech-
nique. In this chapter, we will introduce two best-known regular-
ization methods: ridge regression and lasso. The elastic net is a
combination of ridge and lasso, or it is a general representation of
the two.
We talked about the variance bias trade-off in section 7.1. The
variance of a learning model is the amount by which 𝑓 ̂ would
change if we estimated it using a different training data set. In
general, model variance increases as flexibility increases. The reg-
ularization technique decreases the model flexibility by shrinking
the coefficient and hence significantly reduce the model variance.
Load the R packages:

# install packages from CRAN

p_needed <- c('caret', 'elasticnet', 'glmnet', 'devtools',
'MASS', 'grplasso')

packages <- rownames(installed.packages())

p_to_install <- p_needed[!(p_needed %in% packages)]
if (length(p_to_install) > 0) {
install.packages(p_to_install)
}

187
188 10 Regularization Methods

lapply(p_needed, require, character.only = TRUE)

# install packages from GitHub

p_needed_gh <- c('NetlifyDS')

if (! p_needed_gh %in% packages) {

devtools::install_github("netlify/NetlifyDS")
}

library(NetlifyDS)

10.1 Ridge Regression

Recall that the least square estimates minimize RSS:

𝑅𝑆𝑆 = Σ𝑛𝑖=1 (𝑦𝑖 − 𝛽0 − Σ𝑝𝑗=1 𝛽𝑗 𝑥𝑖𝑗 )2

Ridge regression (Hoerl and Kennard, 1970) is similar but it finds

̂ that optimizes a slightly different function:
𝛽𝑅

Σ𝑛𝑖=1 (𝑦𝑖 − 𝛽0 − Σ𝑝𝑗=1 𝛽𝑗 𝑥𝑖𝑗 )2 + 𝜆Σ𝑝𝑗=1 𝛽𝑗2 = 𝑅𝑆𝑆 + 𝜆Σ𝑝𝑗=1 𝛽𝑗2 (10.1)

where 𝜆 > 0 is a tuning parameter. As with the least squares, ridge

regression considers minimizing RSS. However, it adds a shrinkage
penalty 𝜆Σ𝑝𝑗=1 𝛽𝑗2 that takes account of the number of parameters
in the model. When 𝜆 = 0, it is identical to least squares. As 𝜆 gets
larger, the coefficients start shrinking towards 0. When 𝜆 → ∞, the
rest of the coefficients 𝛽1 , ..., 𝛽𝑝 are close to 0. Here, the penalty
is not applied to 𝛽0 . The tuning parameter 𝜆 is used to adjust
the impact of the two parts in equation (10.1). Every value of 𝜆
corresponds to a set of parameter estimates.
There are many R packages for ridge regression, such as lm.ridge()
10.1 Ridge Regression 189

function from MASS, function enet() from elasticnet. If you know the
value of 𝜆, you can use either of the function to fit ridge regression.
A more convenient way is to use train() function from caret. Let’s
use the 10 survey questions to predict the total purchase amount
(sum of online and store purchase).

dat <- read.csv("http://bit.ly/2P5gTw4")

# data cleaning: delete wrong observations
# expense can't be negative
dat <- subset(dat, store_exp > 0 & online_exp > 0)
# get predictors
trainx <- dat[ , grep("Q", names(dat))]
# get response
trainy <- dat$store_exp + dat$online_exp

Use train() function to tune parameter. Since ridge regression adds

the penalty parameter 𝜆 in front of the sum of squares of the
parameters, the scale of the parameters matters. So here it is better
to center and scale the predictors. This preprocessing is generally
recommended for all techniques that puts penalty to parameter
estimates. In this example, the 10 survey questions are already
with the same scale so data preprocessing doesn’t make too much
different. It is a good idea to set the preprocessing as a standard.

# set cross validation

ctrl <- trainControl(method = "cv", number = 10)
# set the parameter range
ridgeGrid <- data.frame(.lambda = seq(0, .1, length = 20))
set.seed(100)
ridgeRegTune <- train(trainx, trainy,
method = "ridge",
tuneGrid = ridgeGrid,
trControl = ctrl,
## center and scale predictors
preProc = c("center", "scale"))
ridgeRegTune
190 10 Regularization Methods

## Ridge Regression
##
## 999 samples
## 10 predictor
##
## Pre-processing: centered (10), scaled (10)
## Resampling: Cross-Validated (10 fold)
## Summary of sample sizes: 899, 899, 899, 899, 899, 900, ...
## Resampling results across tuning parameters:
##
## lambda RMSE Rsquared MAE
## 0.000000 1744 0.7952 754.0
## 0.005263 1744 0.7954 754.9
## 0.010526 1744 0.7955 755.9
## 0.015789 1744 0.7955 757.3
## 0.021053 1745 0.7956 758.8
## 0.026316 1746 0.7956 760.6
## 0.031579 1747 0.7956 762.4
## 0.036842 1748 0.7956 764.3
## 0.042105 1750 0.7956 766.4
## 0.047368 1751 0.7956 768.5
## 0.052632 1753 0.7956 770.6
## 0.057895 1755 0.7956 772.7
## 0.063158 1757 0.7956 774.9
## 0.068421 1759 0.7956 777.2
## 0.073684 1762 0.7956 779.6
## 0.078947 1764 0.7955 782.1
## 0.084211 1767 0.7955 784.8
## 0.089474 1769 0.7955 787.6
## 0.094737 1772 0.7955 790.4
## 0.100000 1775 0.7954 793.3
##
## RMSE was used to select the optimal model using
## the smallest value.
## The final value used for the model was lambda
## = 0.005263.

The results show that the best value of 𝜆 is 0.005 and the RMSE
10.1 Ridge Regression 191

and 𝑅2 are 1744 and 0.7954 correspondingly. You can see from the
figure 10.1, as the 𝜆 increase, the RMSE first slightly decreases and
then increases.

plot(ridgeRegTune)

1770
RMSE (Cross-Validation)

1760

1750

0.00 0.02 0.04 0.06 0.08 0.10

Weight Decay

FIGURE 10.1: Test mean squared error for the ridge regression

Once you have the tuning parameter value, there are different func-
tions to fit a ridge regression. Let’s look at how to use enet() in
elasticnet package.

ridgefit = enet(x = as.matrix(trainx), y = trainy, lambda = 0.01,

# center and scale predictors
normalize = TRUE)

Note here ridgefit only assigns the value of the tuning parameter
for ridge regression. Since the elastic net model include both ridge
and lasso penalty, we need to use predict() function to get the
model fit. You can get the fitted results by setting s = 1 and mode
= "fraction". Here s = 1 means we only use the ridge parameter.
We will come back to this when we get to lasso regression.
192 10 Regularization Methods

ridgePred <- predict(ridgefit, newx = as.matrix(trainx),

s = 1, mode = "fraction", type = "fit")

By setting type = "fit", the above returns a list object. The fit
item has the predictions:

names(ridgePred)

## [1] "s" "fraction" "mode" "fit"

head(ridgePred$fit)

## 1 2 3 4 5 6
## 1290.5 224.2 591.4 1220.6 853.4 908.2

If you want to check the estimated coefficients, you can set

type="coefficients":

ridgeCoef<-predict(ridgefit,newx = as.matrix(trainx),
s=1, mode="fraction", type="coefficients")

It also returns a list and the estimates are in the coefficients item:

# didn't show the results

RidgeCoef = ridgeCoef$coefficients

Comparing to the least square regression, ridge regression per-

forms better because of the bias-variance-trade-off we mentioned
in section 7.1. As the penalty parameter 𝜆 increases, the flexibility
of the ridge regression decreases. It decreases the variance of the
model but increases the bias at the same time.
10.2 LASSO 193

10.2 LASSO
Even though the ridge regression shrinks the parameter estimates
towards 0, it won’t shink any estimates to be exactly 0 which
means it includes all predictors in the final model. So it can’t
select variables. It may not be a problem for prediction but it is
a huge disadvantage if you want to interpret the model especially
when the number of variables is large. A popular alternative to the
ridge penalty is the Least Absolute Shrinkage and Selection
Operator (LASSO) (R, 1996).
Similar to ridge regression, lasso adds a penalty. The lasso coeffi-
cients 𝛽𝜆𝐿̂ minimize the following:

Σ𝑛𝑖=1 (𝑦𝑖 −𝛽0 −Σ𝑝𝑗=1 𝛽𝑗 𝑥𝑖𝑗 )2 +𝜆Σ𝑝𝑗=1 |𝛽𝑗 | = 𝑅𝑆𝑆 +𝜆Σ𝑝𝑗=1 |𝛽𝑗 | (10.2)

The only difference between lasso and ridge is the penalty. In sta-
tistical parlance, ridge uses 𝐿2 penalty (𝛽𝑗2 ) and lasso uses 𝐿1
penalty (|𝛽𝑗 |). 𝐿1 penalty can shrink the estimates to 0 when 𝜆 is
big enough. So lasso can be used as a feature selection tool. It is
a huge advantage because it leads to a more explainable model.
Similar to other models with tuning parameters, lasso regres-
sion requires cross-validation to tune the parameter. You can use
train() in a similar way as we showed in the ridge regression sec-
tion. To tune parameter, we need to set cross-validation and pa-
rameter range. Also, it is advised to standardize the predictors:

ctrl <- trainControl(method = "cv", number = 10)

lassoGrid <- data.frame(fraction = seq(.8, 1, length = 20))
set.seed(100)
lassoTune <- train(trainx, trainy,
## set the method to be lasso
method = "lars",
tuneGrid = lassoGrid,
trControl = ctrl,
194 10 Regularization Methods

## standardize the predictors

preProc = c("center", "scale"))
lassoTune

## Least Angle Regression

##
## 999 samples
## 10 predictor
##
## Pre-processing: centered (10), scaled (10)
## Resampling: Cross-Validated (10 fold)
## Summary of sample sizes: 899, 899, 899, 899, 899, 900, ...
## Resampling results across tuning parameters:
##
## fraction RMSE Rsquared MAE
## 0.8000 1763 0.7921 787.5
## 0.8105 1760 0.7924 784.1
## 0.8211 1758 0.7927 780.8
## 0.8316 1756 0.7930 777.7
## 0.8421 1754 0.7933 774.6
## 0.8526 1753 0.7936 771.8
## 0.8632 1751 0.7939 769.1
## 0.8737 1749 0.7942 766.6
## 0.8842 1748 0.7944 764.3
## 0.8947 1746 0.7947 762.2
## 0.9053 1745 0.7949 760.1
## 0.9158 1744 0.7951 758.3
## 0.9263 1743 0.7952 756.7
## 0.9368 1743 0.7953 755.5
## 0.9474 1742 0.7954 754.5
## 0.9579 1742 0.7954 754.0
## 0.9684 1742 0.7954 753.6
## 0.9789 1743 0.7953 753.4
## 0.9895 1743 0.7953 753.5
## 1.0000 1744 0.7952 754.0
##
10.2 LASSO 195

## RMSE was used to select the optimal model using

## the smallest value.
## The final value used for the model was fraction
## = 0.9579.

The results show that the best value of the tuning parameter
(fraction from the output) is 0.957 and the RMSE and 𝑅2 are
1742 and 0.7954 correspondingly. The performance is nearly the
same with ridge regression. You can see from the figure 10.2, as
the 𝜆 increase, the RMSE first decreases and then increases.

plot(lassoTune)

1760
RMSE (Cross-Validation)

1755

1750

1745

0.80 0.85 0.90 0.95 1.00

Fraction

FIGURE 10.2: Test mean squared error for the lasso regression

Once you select a value for tuning parameter, there are different
functions to fit lasso regression, such as lars() in lars, enet() in
elasticnet, glmnet() in glmnet. They all have very similar syntax.

Here we continue using enet(). The syntax is similar to ridge re-

gression. The only difference is that you need to set lambda = 0
because the argument lambda here is to control the ridge penalty.
When it is 0, the function will return the lasso model object.
196 10 Regularization Methods

lassoModel <- enet(x = as.matrix(trainx), y = trainy,

lambda = 0, normalize = TRUE)

Set the fraction value to be 0.957 (the value we got above):

lassoFit <- predict(lassoModel, newx = as.matrix(trainx),

s = 0.957, mode = "fraction", type = "fit")

Again by setting type = "fit", the above returns a list object. The
fit item has the predictions:

head(lassoFit$fit)

## 1 2 3 4 5 6
## 1357.3 300.5 690.2 1228.2 838.4 1010.1

You need to set type = "coefficients" to get parameter estimates:

lassoCoef <- predict(lassoModel,

newx = as.matrix(trainx),
s = 0.95,
mode = "fraction",
type = "coefficients")

It also returns a list and the estimates are in the coefficients item:

# didn't show the results

LassoCoef = lassoCoef$coefficients

Many researchers applied lasso to other learning methods, such as

linear discriminant analysis (Line Clemmensen and Ersbøll, 2011),
partial least squares regression(Chun and Keleş, 2010). However,
since the 𝐿1 norm is not differentiable, optimization for lasso re-
gression is more complicated. People come up with different al-
gorithms to solve the computation problem. The biggest break-
through is Least Angle Regression (LARS) from Bradley Efron
10.3 Variable selection property of the lasso 197

etc. This algorithm works well for lasso regression especially when
the dimension is high.

10.3 Variable selection property of the lasso

You may ask why lasso (𝐿1 penalty) has the feature selection prop-
erty but not the ridge (𝐿2 penalty). To answer that question, let’s
look at the alternative representations of the optimization problem
for lasso and ridge. For lasso regression, it is identical to optimize
the following two functions:

Σ𝑛𝑖=1 (𝑦𝑖 −𝛽0 −Σ𝑝𝑗=1 𝛽𝑗 𝑥𝑖𝑗 )2 +𝜆Σ𝑝𝑗=1 |𝛽𝑗 | = 𝑅𝑆𝑆 +𝜆Σ𝑝𝑗=1 |𝛽𝑗 | (10.3)

2
𝑚𝑖𝑛 {Σ𝑛𝑖=1 (𝑦𝑖 − 𝛽0 − Σ𝑝𝑗=1 𝛽𝑗 𝑥𝑖𝑗 ) } , Σ𝑝𝑗=1 |𝛽𝑗 | ≤ 𝑠 (10.4)
𝛽

For any value of tuning parameter 𝜆, there exists a 𝑠 such that the
coefficient estimates optimize equation (10.3) also optimize equa-
tion (10.4). Similarly, for ridge regression, the two representations
are identical:

Σ𝑛𝑖=1 (𝑦𝑖 − 𝛽0 − Σ𝑝𝑗=1 𝛽𝑗 𝑥𝑖𝑗 )2 + 𝜆Σ𝑝𝑗=1 𝛽𝑗2 = 𝑅𝑆𝑆 + 𝜆Σ𝑝𝑗=1 𝛽𝑗2 (10.5)

2
𝑚𝑖𝑛 {Σ𝑛𝑖=1 (𝑦𝑖 − 𝛽0 − Σ𝑝𝑗=1 𝛽𝑗 𝑥𝑖𝑗 ) } , Σ𝑝𝑗=1 𝛽𝑗2 ≤ 𝑠 (10.6)
𝛽

̂
When 𝑝 2, lasso estimates (𝛽𝑙𝑎𝑠𝑠𝑜 in figure 10.3) have the smallest
RSS among all points that satisfy |𝛽 ∗ 1| + |𝛽2 | ≤ 𝑠 (i.e. within the
diamond in figure 10.3). Ridge estimates have the smallest RSS
among all points that satisfy 𝛽12 + 𝛽22 ≤ 𝑠 (i.e. within the circle
in figure 10.3). As s increases, the diamond and circle regions ex-
pand and get less restrictive. If s is large enough, the restrictive
region will cover the least squares estimate (𝛽𝐿𝑆𝐸 ̂ ). Then, equa-
tions (10.4) and (10.6) will simply yield the least squares estimate.
198 10 Regularization Methods

In contrast, if s is small, the grey region in figure 10.3 will be small

and hence restrict the magnitude of 𝛽. With the alternative formu-
lations, we can answer the question of feature selection property.

FIGURE 10.3: Contours of the RSS and constrain functions for

the lasso (left) and ridge regression (right)

Figure 10.3 illustrates the situations for lasso (left) and ridge
(right). The least square estimate is marked as 𝛽.̂ The restric-
tive regions are in grey, the diamond region is for the lasso, the
circle region is for the ridge. The least square estimates lie out-
side the grey region, so they are different from the lasso and ridge
estimates. The ellipses centered around 𝛽 ̂ represent contours of
RSS. All the points on a given ellipse share an RSS. As the el-
lipses expand, the RSS increases. The lasso and ridge estimates
are the first points at which an ellipse contracts the grey region.
Since ridge regression has a circular restrictive region that doesn’t
have a sharp point, the intersecting point can’t drop on the axis.
But it is possible for lasso since it has corners at each of the axes.
When the intersecting point is on an axis, one of the parameter
estimates is 0. If p > 2, the restrictive regions become sphere or
hypersphere. In that case, when the intersecting point drops on an
axis, multiple coefficient estimates can equal 0 simultaneously.
10.4 Elastic Net 199

10.4 Elastic Net

Elastic Net is a generalization of lasso and ridge regression(Zou
and Hastie, 2005). It combines the two penalties. The estimates of
coefficients optimize the following function:

Σ𝑛𝑖=1 (𝑦𝑖 − 𝑦𝑖̂ )2 + 𝜆1 Σ𝑝𝑗=1 𝛽𝑗2 + 𝜆2 Σ𝑝𝑗=1 |𝛽𝑗 | (10.7)

Ridge penalty shrinks the coefficients of correlated predictors to-

wards each other while the lasso tends to pick one and discard
the others. So lasso estimates have a higher variance. However,
ridge regression doesn’t have a variable selection property. The
advantage of the elastic net is that it keeps the feature selection
quality from the lasso penalty as well as the effectiveness of the
ridge penalty. (Zou and Hastie, 2005) suggest that it deals with
highly correlated variables more effectively.
We can still use train() function to tune the parameters in the
elastic net. As before, set the cross-validation and parameter range.
Standardize the predictors:

enetGrid <- expand.grid(.lambda = seq(0,0.2,length=20),

.fraction = seq(.8, 1, length = 20))
set.seed(100)
enetTune <- train(trainx, trainy,
method = "enet",
tuneGrid = enetGrid,
trControl = ctrl,
preProc = c("center", "scale"))
enetTune

Elasticnet

999 samples
10 predictor
200 10 Regularization Methods

Pre-processing: centered (10), scaled (10)

Resampling: Cross-Validated (10 fold)
Summary of sample sizes: 899, 899, 899, 899, 899, 900, ...
Resampling results across tuning parameters:

lambda fraction RMSE Rsquared MAE

0.00000 0.8000 1763 0.7921 787.5
0.00000 0.8105 1760 0.7924 784.1
.
.
.
0.09474 0.9158 1760 0.7945 782.5
0.09474 0.9263 1761 0.7947 782.5
0.09474 0.9368 1761 0.7949 782.7
0.09474 0.9474 1763 0.7950 783.3
0.09474 0.9579 1764 0.7951 784.3
0.09474 0.9684 1766 0.7953 785.7
0.09474 0.9789 1768 0.7954 787.1
0.09474 0.9895 1770 0.7954 788.8
0.09474 1.0000 1772 0.7955 790.4
[ reached getOption("max.print") -- omitted 200 rows ]

RMSE was used to select the optimal model using the smallest value.
The final values used for the model were fraction = 0.9579 and lambda = 0.

The results show that the best values of the tuning parameters are
fraction = 0.9579 and lambda = 0 It also indicates that the final
model is lasso only (the ridge penalty parameter lambda is 0). The
RMSE and 𝑅2 are 1742.2843 and 0.7954 correspondingly.

10.5 Penalized Generalized Linear Model

Adding penalties is a general technique that can be applied to
many methods other than linear regression. In this section, we
will introduce the penalized generalized linear model. It is to fit
the generalized linear model by minimizing a penalized maximum
10.5 Penalized Generalized Linear Model 201

likelihood. The penalty can be 𝐿1 , 𝐿2 or a combination of the two.

The estimates of coefficients minimize the following:

1 𝑁
𝑚𝑖𝑛 Σ𝑖=1 𝑤𝑖 𝑙(𝑦𝑖 , 𝛽0 + 𝜷𝐓 𝐱𝐢 ) + 𝜆[(1 − 𝛼) ∥ 𝜷 ∥22 /2 + 𝛼 ∥ 𝜷 ∥1 ]
𝛽0 ,𝜷 𝑁

where

𝑙(𝑦𝑖 , 𝛽0 + 𝜷𝐓 𝐱𝐢 ) = −𝑙𝑜𝑔[ℒ(𝑦𝑖 , 𝛽0 + 𝜷𝐓 𝐱𝐢 )]

It is the negative logarithm of the likelihood, ℒ(𝑦𝑖 , 𝛽0 + 𝜷𝐓 𝐱𝐢 ).

Maximize likelihood is to minimize 𝑙(𝑦𝑖 , 𝛽0 + 𝜷𝐓 𝐱𝐢 ).
Parameter 𝛼 decides the penalty, i.e, between 𝐿2 (𝛼 = 0) and
𝐿1 (𝛼 = 1). 𝜆 controls the weight of the whole penalty item.
The higher 𝜆 is, the more weight the penalty carries comparing
to likelihood. As discussed above, the ridge penalty shrinks the
coefficients towards 0 but can’t be exactly 0. The lasso penalty
can set 0 estimates so it has the property of feature selection. The
elastic net combines both. Here we have two tuning parameters, 𝛼
and 𝜆.

10.5.1 Introduction to glmnet package

glmnet is a package that fits a penalized generalized linear model
using cyclical coordinate descent. It successively optimizes the ob-
jective function over each parameter with others fixed, and cycles
repeatedly until convergence. Since the linear model is a special
case of the generalized linear model, glmnet can also fit a penal-
ized linear model. Other than that, it can also fit penalized logistic
regression, multinomial, Poisson, and Cox regression models.
The default family option in the function glmnet() is gaussian. It
is the linear regression we discussed so far in this chapter. But
the parameterization is a little different in the generalized linear
model framework (we have 𝛼 and 𝜆). Let’s start from our previous
example, using the same training data but glmnet() to fit model:
202 10 Regularization Methods

dat <- read.csv("http://bit.ly/2P5gTw4")

# data cleaning: delete wrong observations with expense < 0
dat <- subset(dat, store_exp > 0 & online_exp > 0)
# get predictors
trainx <- dat[, grep("Q", names(dat))]
# get response
trainy <- dat$store_exp + dat$online_exp
glmfit = glmnet::glmnet(as.matrix(trainx), trainy)

The object glmfit returned by glmnet() has the information of the

fitted model for the later operations. An easy way to extract the
components is through various functions on glmfit, such as plot(),
print(), coef() and predict() . For example, the following code
visualizes the path of coefficients as penalty increases:

plot(glmfit, label = T)

0 3 3 7 9
1000

3
6
2
5
500
Coefficients

4
0

1
10
-500

0 1000 2000 3000 4000

L1 Norm

Each curve in the plot represents one predictor. The default setting
is 𝛼 = 1 which means there is only lasso penalty. From left to right,
𝐿𝐼 norm is increasing which means 𝜆 is decreasing. The bottom
x-axis is 𝐿1 norm (i.e. ∥ 𝜷 ∥1 ). The upper x-axis is the effective
10.5 Penalized Generalized Linear Model 203

degrees of freedom (df) for the lasso. You can check the detail for
every step by:

print(glmfit)

Call: glmnet(x = as.matrix(trainx), y = trainy)

Df %Dev Lambda
1 0 0.000 3040
2 2 0.104 2770
3 2 0.192 2530
4 2 0.265 2300
5 3 0.326 2100
6 3 0.389 1910
7 3 0.442 1740
8 3 0.485 1590
9 3 0.521 1450
...

The first column Df is the degree of freedom (i.e. the number of non-
zero coefficients), %Dev is the percentage of deviance explained and
Lambda is the value of tuning parameter 𝜆. By default, the function
will try 100 different values of 𝜆. However, if as 𝜆 changes, the %Dev
doesn’t change sufficiently, the algorithm will stop before it goes
through all the values of 𝜆. We didn’t show the full output above.
But it only uses 68 different values of 𝜆. You can also set the value
of 𝜆 using s= :

coef(glmfit, s = 1200)

## 11 x 1 sparse Matrix of class "dgCMatrix"

## 1
## (Intercept) 2255.2
## Q1 -390.9
## Q2 653.6
## Q3 624.4
## Q4 .
204 10 Regularization Methods

## Q5 .
## Q6 .
## Q7 .
## Q8 .
## Q9 .
## Q10 .

When 𝜆 = 1200, there are three coefficients with non-zero esti-

mates(Q1, Q2 and Q3). You can apply models with different values
of tuning parameter to new data using predict():

newdat = matrix(sample(1:9, 30, replace = T), nrow = 3)

predict(glmfit, newdat, s = c(1741, 2000))

## 1 2
## [1,] 6004 5968
## [2,] 7101 6674
## [3,] 9158 8411

Each column corresponds to a value of 𝜆. To tune the value of

𝜆, we can easily use cv.glmnet() function to do cross-validation.
cv.glmnet() returns the cross-validation results as a list object. We
store the object in cvfit and use it for further operations.

cvfit = cv.glmnet(as.matrix(trainx), trainy)

We can plot the object using plot(). The red dotted line is the
cross-validation curve. Each red point is the cross-validation mean
squared error for a value of 𝜆. The grey bars around the red points
indicate the upper and lower standard deviation. The two gray dot-
ted vertical lines represent the two selected values of 𝜆, one gives
the minimum mean cross-validated error (lambda.min), the other
gives the error that is within one standard error of the minimum
(lambda.1se).

plot(cvfit)
10.5 Penalized Generalized Linear Model 205
9 9 9 9 9 9 9 7 7 6 6 4 3 3 3 3 2

1.5e+07
Mean-Squared Error

1.0e+07
5.0e+06

2 3 4 5 6 7 8

Log(λ)

You can check the two selected 𝜆 values by:

# lambda with minimum mean cross-validated error

cvfit$lambda.min

## [1] 12.57

# lambda with one standard error of the minimum

cvfit$lambda.1se

## [1] 1200

You can look at the coefficient estimates for different 𝜆 by:

# coefficient estimates for model with the error

# that is within one standard error of the minimum
coef(cvfit, s = "lambda.1se")

## 11 x 1 sparse Matrix of class "dgCMatrix"

## 1
## (Intercept) 2255.3
## Q1 -391.1
206 10 Regularization Methods

## Q2 653.7
## Q3 624.5
## Q4 .
## Q5 .
## Q6 .
## Q7 .
## Q8 .
## Q9 .
## Q10 .

10.5.2 Penalized logistic regression

10.5.2.1 Multivariate logistic regression model

Logistic regression is a traditional statistical method for a two-

category classification problem. It is simple yet useful. Here we
use the swine disease breakout data as an example to illustrate
the learning method and code implementation. Refer to section
3.2 for more details about the dataset. The goal is to predict if a
farm will have a swine disease outbreak (i.e build a risk scoring
system).
Consider risk scoring system construction using a sample of 𝑛 ob-
servations, with information collected for 𝐺 categorical predictors
and one binary response variable for each observation. The pre-
dictors are 120 survey questions (i.e. G=120). There were three
possible answers for each question (A, B and C). So each pre-
dictor is encoded to two dummy variables (we consider C as the
baseline.). Let xi,g be the vector of dummy variables associated
with the 𝑔𝑡ℎ categorical predictor for the 𝑖𝑡ℎ observation, where
𝑖 = 1, ⋯ , 𝑛, 𝑔 = 1, ⋯ , 𝐺. For example, if the first farm chooses B
for question 2, then the corresponding observation is x12 = (0, 1)𝑇 .
Each question has a degree of freedom of 2.
We denote the degrees of freedom of the 𝑔𝑡ℎ predictor by 𝑑𝑓𝑔 ,
which is also the length of vector xi,g . Let 𝑦𝑖 (= 1, diseased; or
0, not diseased) be the binary response for the 𝑖th observation.
Denote the probability of disease for 𝑖th subject by 𝜃𝑖 , the model
can be formulated as:
10.5 Penalized Generalized Linear Model 207

𝑦𝑖 ∼ 𝐵𝑜𝑢𝑛𝑜𝑢𝑙𝑙𝑖(𝜃𝑖 )

𝐺
𝜃𝑖
𝑙𝑜𝑔 ( ) = 𝜂𝜷 (𝑥𝑖 ) = 𝛽0 + ∑ xi,g 𝑇 𝜷𝐠
1 − 𝜃𝑖 𝑔=1

where 𝛽0 is the intercept and 𝜷𝐠 is the parameter vector cor-

responding to the 𝑔𝑡ℎ predictor. As we mentioned, here 𝜷𝐠 has
length 2.
Traditional estimation of logistic parameters 𝜷 =
(𝛽0𝑇 , 𝜷𝟏𝑇 , 𝜷𝟐𝑇 , ..., 𝜷𝐆
𝑇 𝑇
) is done through maximizing the log-
likelihood
𝑛
𝑦
𝑙(𝜷) = 𝑙𝑜𝑔[∏ 𝜃𝑖 𝑖 (1 − 𝜃𝑖 )1−𝑦𝑖 ]
𝑖=1
𝑛
= ∑{𝑦𝑖 𝑙𝑜𝑔(𝜃𝑖 ) + (1 − 𝑦𝑖 )𝑙𝑜𝑔(1 − 𝜃𝑖 )}
𝑖=1
𝑛
= ∑{ 𝑦𝑖 𝜂𝜷 (xi ) − 𝑙𝑜𝑔[1 + 𝑒𝑥𝑝(𝜂𝜷 (xi ))] }
𝑖=1

For logistic regression analysis with a large number of explana-

tory variables, complete- or quasi-complete-separation may result
which makes the maximum likelihood estimation unstable (Wed-
derburn, 1976), (A and J, 1984). For example:

library(MASS)
dat <- read.csv("http://bit.ly/2KXb1Qi")
fit <- glm(y~., dat, family = "binomial")

## Warning: glm.fit: algorithm did not converge

## Warning: glm.fit: fitted probabilities numerically 0 or

## 1 occurred

There is an error saying “algorithm did not converge”. It is because

208 10 Regularization Methods

there is complete separation. It happens when there are a large

number of explanatory variables which makes the estimation of
the coefficients unstable. To stabilize the estimation of parameter
coefficients, one popular approach is the lasso algorithm with 𝐿1
norm penalty proposed by Tibshirani(R, 1996). Because the lasso
algorithm can estimate some variable coefficients to be 0, it can
also be used as a variable selection tool.
10.5.2.2 Penalized logistic regression

Penalized logistic regression adds penalty to the likelihood func-

tion:
𝑛
∥ 𝜷 ∥22
∑{ 𝑦𝑖 𝜂𝜷 (xi )−𝑙𝑜𝑔[1+𝑒𝑥𝑝(𝜂𝜷 (xi ))] }+𝜆[(1−𝛼) ]+𝛼 ∥ 𝜷 ∥1 ]
𝑖=1
2

dat <- read.csv("http://bit.ly/2KXb1Qi")

trainx = dplyr::select(dat, -y)
trainy = dat$y
fit <- glmnet(as.matrix(trainx), trainy, family = "binomial")

The error message is gone when we use penalized regression. We

can visualize the shrinking path of coefficients as penalty increases.
The use of predict() function is a little different. For the general-
ized linear model, you can return different results by setting the
type argument. The choices are:

• link: return the link function value

• response: return the probability
• class: return the category (0/1)
• coefficients: return the coefficient estimates
• nonzero: return an indicator for non-zero estimates (i.e. which
variables are selected)
The default setting is to predict the probability of the second level
of the response variable. For example, the second level of the re-
sponse variable for trainy here is:
10.5 Penalized Generalized Linear Model 209

levels(as.factor(trainy))

## [1] "0" "1"

The first column of the above output is the predicted link function
value when 𝜆 = 0.02833. The second column of the output is the
predicted link function when 𝜆 = 0.0311. Similarly, you can change
the setting for type to produce different outputs. You can use the
cv.glmnet() function to tune parameters. The parameter setting
is nearly the same as before, the only difference is the setting
of type.measure. Since the response is categorical, not continuous,
we have different performance measurements. The most common
settings of type.measure for classification are:
• class:error rate
• auc:it is the area under the ROC for the dichotomous problem
So the model is to predict the probability of outcome “1”. Take
a baby example of 3 observations and 2 values of 𝜆 to show the
usage of predict() function:

newdat = as.matrix(trainx[1:3, ])
predict(fit, newdat, type = "link", s = c(2.833e-02, 3.110e-02))

## 1 2
## 1 0.1943 0.1443
## 2 -0.9913 -1.0077
## 3 -0.5841 -0.5496

For example:

cvfit = cv.glmnet(as.matrix(trainx), trainy,

family = "binomial", type.measure = "class")
plot(cvfit)
210 10 Regularization Methods
213 211 212 209 206 204 197 181 146 91 47 8

0.5
Misclassification Error

0.4
0.3
0.2

-10 -8 -6 -4

Log(λ)

The above uses error rate as performance criteria and use 10-fold
cross-validation. Similarly, you can get the 𝜆 value for the mini-
mum error rate and the error rate that is 1 standard error from
the minimum:

cvfit$lambda.min

## [1] 2.643e-05

cvfit$lambda.1se

## [1] 0.003334

You can use the same way to get the parameter estimates and
make prediction.

10.5.2.3 Group lasso logistic regression

For models with categorical survey questions (explanatory vari-

ables), however, the original lasso algorithm only selects individual
dummy variables instead of sets of the dummy variables grouped
by the question in the survey. Another disadvantage of applying
lasso to grouped variables is that the estimates are affected by the
10.5 Penalized Generalized Linear Model 211

way dummy variables are encoded. Thus the group lasso (Yuan
and Lin, 2007) method has been proposed to enable variable se-
lection in linear regression models on groups of variables, instead
of on single variables. For logistic regression models, the group
lasso algorithm was first studied by Kim et al. (Y. Kim and Kim,
2006). They proposed a gradient descent algorithm to solve the
corresponding constrained problem, which does, however, depend
on unknown constants. Meier et al. (L Meier and Buhlmann, 2008)
proposed a new algorithm that could work directly on the penal-
ized problem and its convergence property does not depend on
unknown constants. The algorithm is especially suitable for high-
dimensional problems. It can also be applied to solve the corre-
sponding convex optimization problem in generalized linear mod-
els. The group lasso estimator proposed by Meier et al. (L Meier
and Buhlmann, 2008) for logistic regression has been shown to be
statistically consistent, even with a large number of categorical
predictors.
In this section, we illustrate how to use the logistic group lasso
algorithm to construct risk scoring systems for predicting disease.
Instead of maximizing the log-likelihood in the maximum likeli-
hood method, the logistic group lasso estimates are calculated by
minimizing the convex function:

𝐺
𝑆𝜆 (𝜷) = −𝑙(𝜷) + 𝜆 ∑ 𝑠(𝑑𝑓𝑔 ) ∥ 𝜷𝐠 ∥2
𝑔=1

where 𝜆 is a tuning parameter for the penalty and 𝑠(⋅) is a function

to rescale the penalty. In lasso algorithms, the selection of 𝜆 is
usually determined by cross-validation using data. For 𝑠(⋅), we use
the square root function 𝑠(𝑑𝑓𝑔 ) = 𝑑𝑓𝑔0.5 as suggested in Meier et
al.(L Meier and Buhlmann, 2008). It ensures the penalty is of the
order of the number of parameters 𝑑𝑓𝑔 as used in (Yuan and Lin,
2007).
Here we consider selection of the tuning param-
eter 𝜆 from a multiplicative grid of 100 values
212 10 Regularization Methods

{0.96𝜆𝑚𝑎𝑥 , 0.962 𝜆𝑚𝑎𝑥 , 0.963 𝜆𝑚𝑎𝑥 , ..., 0.96100 𝜆𝑚𝑎𝑥 }. Here 𝜆𝑚𝑎𝑥 is
defined as

1
𝜆𝑚𝑎𝑥 = 𝑚𝑎𝑥 { ∥ xg 𝑇 (y − y)̄ ∥2 } , (10.8)
𝑔∈{1,...,𝐺} 𝑠(𝑑𝑓𝑔 )

such that when 𝜆 = 𝜆𝑚𝑎𝑥 , only the intercept is in the model. When
𝜆 goes to 0, the model is equivalent to ordinary logistic regression.
Three criteria may be used to select the optimal value of 𝜆. One
is AUC which you should have seem many times in this book
by now. The log-likelihood score used in Meier et al. (L Meier
and Buhlmann, 2008) is taken as the average of log-likelihood of
the validation data overall cross-validation sets. Another one is
the maximum correlation coefficient in Yeo and Burge (Yeo and
Burge, 2004) that is defined as:

𝜌𝑚𝑎𝑥 = 𝑚𝑎𝑥{𝜌𝜏 |𝜏 ∈ (0, 1)},

where 𝜏 ∈ (0, 1) is a threshold to classify the predicted probability

into binary disease status and 𝜌𝜏 is the Pearson correlation coef-
ficient between the true binary disease status and the predictive
disease status with threshold 𝜏 .
You can use the following package to implement the model. Install
the package using:

devtools::install_github("netlify/NetlifyDS")

Load the package:

library("NetlifyDS")

The package includes the swine disease breakout data and you can
load the data by:
10.5 Penalized Generalized Linear Model 213

data("sim1_da1")

You can use cv_glasso() function to tune the parameters:

# the last column of sim1_da1 response variable y

# trainx is the explanatory variable matrix
trainx = dplyr::select(sim1_da1, -y)
# save response variable as as trainy
trainy = sim1_da1$y
# get the group indicator
index <- gsub("\\..*", "", names(trainx))

Dummy variables from the same question are in the same group:

index[1:50]

## [1] "Q1" "Q1" "Q2" "Q2" "Q3" "Q3" "Q4" "Q4"

## [9] "Q5" "Q5" "Q6" "Q6" "Q7" "Q7" "Q8" "Q8"
## [17] "Q9" "Q9" "Q10" "Q10" "Q11" "Q11" "Q12" "Q12"
## [25] "Q13" "Q13" "Q14" "Q14" "Q15" "Q15" "Q16" "Q16"
## [33] "Q17" "Q17" "Q18" "Q18" "Q19" "Q19" "Q20" "Q20"
## [41] "Q21" "Q21" "Q22" "Q22" "Q23" "Q23" "Q24" "Q24"
## [49] "Q25" "Q25"

Set a series of tuning parameter values. nlam is the num-

ber of values we want to tune. It is the parameter 𝑚
in {0.96𝜆𝑚𝑎𝑥 , 0.962 𝜆𝑚𝑎𝑥 , 0.963 𝜆𝑚𝑎𝑥 , ..., 0.96𝑚 𝜆𝑚𝑎𝑥 }. The tuning
process returns a long output and we will not report all:

# Tune over 100 values

nlam <- 100
# set the type of prediction
# - `link`: return the predicted link function
# - `response`: return the predicted probability
# number of cross-validation folds
kfold <- 10
214 10 Regularization Methods

cv_fit <- cv_glasso(trainx, trainy,

nlam = nlam, kfold = kfold, type = "link")
# only show part of the results
str(cv_fit)

Here we only show part of the output:

...
$ auc : num [1:100] 0.573 0.567 0.535 ...
$ log_likelihood : num [1:100] -554 -554 -553 ...
$ maxrho : num [1:100] -0.0519 0.00666 ...
$ lambda.max.auc : Named num [1:2] 0.922 0.94
..- attr(*, "names")= chr [1:2] "lambda" "auc"
$ lambda.1se.auc : Named num [1:2] 16.74 0.81
..- attr(*, "names")= chr [1:2] "" "se.auc"
$ lambda.max.loglike: Named num [1:2] 1.77 -248.86
..- attr(*, "names")= chr [1:2] "lambda" "loglike"
$ lambda.1se.loglike: Named num [1:2] 9.45 -360.13
..- attr(*, "names")= chr [1:2] "lambda" "se.loglike"
$ lambda.max.maxco : Named num [1:2] 0.922 0.708
..- attr(*, "names")= chr [1:2] "lambda" "maxco"
$ lambda.1se.maxco : Named num [1:2] 14.216 0.504
..- attr(*, "names")= chr [1:2] "lambda" "se.maxco"

In the returned results:

• $ auc: the AUC values
• $ log_likelihood: log-likelihood
• $ maxrho: maximum correlation coefficient
• $ lambda.max.auc: the max AUC and the corresponding value of
𝜆
• $ lambda.1se.auc: one standard error to the max AUC and the
corresponding 𝜆
• $ lambda.max.loglike: max log-likelihood and the corresponding 𝜆
• $ lambda.1se.loglike: one standard error to the max log-likelihood
and the corresponding 𝜆
10.5 Penalized Generalized Linear Model 215

• $ lambda.max.maxco: maximum correlation coefficient and the cor-

responding 𝜆
• $ lambda.1se.maxco:one standard error to the maximum correla-
tion coefficient and the corresponding 𝜆
The most common criterion is AUC. You can compare the selec-
tions from different criteria. If they all point to the same value
of the tuning parameter, you can have more confidence about the
choice. If they suggest very different values, then you need to con-
cern if the tuning process is stable. You can visualize the cross
validation result:

plot(cv_fit)
1.0
0.8
auc

0.6
0.4

0 10 20 30 40 50

Lambda

The x-axis is the value of the tuning parameter, the y-axis is AUC.
The two dash lines are the value of 𝜆 for max AUC and the value
for the one standard deviation to the max AUC. Once you choose
the value of the tuning parameter, you can use fitglasso() to fit
the model. For example, we can fit the model using the parameter
value that gives the max AUC, which is 𝜆 = 0.922:

fitgl <- fitglasso(trainx, trainy,

lambda = 0.922, na_action = na.pass)
216 10 Regularization Methods

Lambda: 0.922 nr.var: 229

You can use coef() to get the estimates of coefficients:

coef(fitgl)

0.922
Intercept -5.318e+01
Q1.A 1.757e+00
Q1.B 1.719e+00
Q2.A 2.170e+00
Q2.B 6.939e-01
Q3.A 2.102e+00
Q3.B 1.359e+00
...

Use predict_glasso() to predict new samples:

prey <- predict_glasso(fitgl, trainx)

11
Tree-Based Methods

Tree-based models such as random forest and gradient boosted

trees are frequent winners in data challenges and competitions
which use standard numerical and categorical datasets. These
methods, in general, provide a good baseline for model perfor-
mance. This chapter describes the fundamentals of tree-based
models and provides a set of standard modeling procedures.
Load R packages:

# install packages from CRAN

p_needed <- c('rpart', 'caret', 'partykit',
'pROC', 'dplyr', 'ipred',
'e1071', 'randomForest', 'gbm')

packages <- rownames(installed.packages())

p_to_install <- p_needed[!(p_needed %in% packages)]
if (length(p_to_install) > 0) {
install.packages(p_to_install)
}

lapply(p_needed, require, character.only = TRUE)

11.1 Tree Basics

The tree-based models can be used for regression and classification.
The goal is to stratify or segment the predictor space into a number
of sub-regions. For a given observation, use the mean (regression)

217
218 11 Tree-Based Methods

or the mode (classification) of the training observations in the

sub-region as the prediction. Tree-based methods are conceptually
simple yet powerful. This type of model is often referred to as
Classification And Regression Trees (CART). They are popular
tools for many reasons:

1. Do not require user to specify the form of the relationship

between predictors and response
2. Do not require (or if they do, very limited) data pre-
processing and can handle different types of predictors
(sparse, skewed, continuous, categorical, etc.)
3. Robust to co-linearity
4. Can handle missing data
5. Many pre-built packages make implementation as easy as
a button push

CART can refer to the tree model in general, but most of the time,
it represents the algorithm initially proposed by Breiman (Breiman
et al., 1984). After Breiman, there are many new algorithms, such
as ID3, C4.5, and C5.0. C5.0 is an improved version of C4.5, but
since C5.0 is not open source, the C4.5 algorithm is more popular.
C4.5 was a major competitor of CART. But now, all those seem
outdated. The most popular tree models are Random Forest (RF)
and Gradient Boosting Machine (GBM). Despite being out of favor
in application, it is important to understand the mechanism of the
basic tree algorithm. Because the later models are based on the
same foundation.
The original CART algorithm targets binary classification, and the
later algorithms can handle multi-category classification. A single
tree is easy to explain but has poor accuracy. More complicated
tree models, such as RF and GBM, can provide much better pre-
diction at the cost of explainability. As the model becoming more
complicated, it is more like a black-box which makes it very diffi-
cult to explain the relationship among predictors. There is always
a trade-off between explainability and predictability.
The reason why it is called “tree” is of course because the structure
 11.1 Tree Basics 219

FIGURE 11.1: This is a classification tree trained from passenger

survival data from the Titanic. Survival probability is predicted
using sex and age.

has similarities. But the direction of the decision tree is opposite

of a real tree, the root is on the top, and the leaf is on the bot-
tom (figure 11.1). From the root node, a decision tree divides to
different branches and generates more nodes. The new nodes are
child nodes, and the previous node is the parent node. At each
child node, the algorithm will decide whether to continue divid-
ing. If it stops, the node is called a leaf node. If it continues, then
the node becomes the new parent node and splits to produce the
next layer of child nodes. At each non-leaf node, the algorithm
needs to decide if it will split into branches. A leaf node contains
the final “decision” on the sample’s value. Here are the important
definitions in the tree model:
• Classification tree: the outcome is discrete
• Regression tree: the outcome is continuous (e.g. the price of a
house)
• Non-leaf node (or split node): the algorithm needs to decide
a split at each non-leaf node (eg: age >= 9.5)
• Root node�the beginning node where the tree starts
 220 11 Tree-Based Methods

• Leaf node (or Terminal node): the node stops splitting. It

has the final decision of the model
• Degree of the node: the number of subtrees of a node
• Degree of the tree: the maximum degree of a node in the tree
• Pruning: remove parts of the tree that do not provide power to
classify instances
• Branch (or Subtree): the whole part under a non-leaf node
• Child: the node directly after and connected to another node
• Parent: the converse notion of a child
A single tree is easy to explain, but it can be very non-robust,
which means a slight change in the data can significantly change
the fitted tree. The predictive accuracy is not as good as other re-
gression and classification approaches in this book since a series of
rectangular decision regions (figure 11.2) defined by a single tree
is often too naive to represent the relationship between the depen-
dent variable and the predictors. To overcome these shortcomings,
researchers have proposed ensemble methods which combine many
trees. Ensemble tree models typically have much better predictive
performance than a single tree. We will introduce those models in
later sections.

11.2 Splitting Criteria

The splitting criteria used by the regression tree and the classifi-
cation tree are different. Like the regression tree, the goal of the
classification tree is to divide the data into smaller, more homo-
geneous groups. Homogeneity means that most of the samples at
each node are from one class. The original CART algorithm uses
Gini impurity as the splitting criterion; The later ID3, C4.5, and
C5.0 use entropy. We will look at three most common splitting
criteria.
Gini impurity
Gini impurity (Breiman et al., 1984) is a measure of non-
homogeneity. It is widely used in classification tree. It is defined
11.2 Splitting Criteria 221

FIGURE 11.2: Example of decision regions for different types of

flowers based on petal length and width. Three different types of
flowers are classified using decisions about length and width.

as:
𝐺𝑖𝑛𝑖 𝐼𝑛𝑑𝑒𝑥 = Σ𝑖 𝑝𝑖 (1 − 𝑝𝑖 )

where 𝑝𝑖 is the probability of class 𝑖 and the interval of Gini is

[0, 0.5]. For a two-class problem, the Gini impurity for a given
node is:

𝑝1 (1 − 𝑝1 ) + 𝑝2 (1 − 𝑝2 )

It is easy to see that when the sample set is pure, one of the
probability is 0 and the Gini score is the smallest. Conversely, when
𝑝1 = 𝑝2 = 0.5, the Gini score is the largest, in which case the purity
of the node is the smallest. Let’s look at an example. Suppose
we want to determine which students are computer science (CS)
majors. Here is the simple hypothetical classification tree result
obtained with the gender variable.
222 11 Tree-Based Methods

Let’s calculate the Gini impurity for splitting node “Gender”:

1. Gini impurity for “Female” = 16 × 65 + 56 × 16 = 5

18
2. Gini impurity for “Male” = 0 × 1 + 1 × 0 = 0

The Gini impurity for the node “Gender” is the following weighted
average of the above two scores:

3 5 2 1
× + ×0=
5 18 5 6

The Gini impurity for the 50 samples in the parent node is 12 . It is

easy to calculate the Gini impurity drop from 21 to 16 after splitting.
The split using “gender” causes a Gini impurity decrease of 13 . The
algorithm will use different variables to split the data and choose
the one that causes the most substantial Gini impurity decrease.
Information Gain (IG)
Looking at the samples in the following three nodes, which one is
the easiest to describe? It is obviously C. Because all the samples in
C are of the same type, so the description requires the least amount
of information. On the contrary, B needs more information, and
A needs the most information. In other words, C has the highest
11.2 Splitting Criteria 223

purity, B is the second, and A has the lowest purity. We need less
information to describe nodes with higher purity.

A measure of the degree of disorder is entropy which is defined as:

𝐸𝑛𝑡𝑟𝑜𝑝𝑦 = −Σ𝑖 𝑝𝑖 𝑙𝑜𝑔2 (𝑝𝑖 )

where 𝑝𝑖 is the probability of class 𝑖 and the interval of entropy is
[0, 1]. For a two-class problem:

𝐸𝑛𝑡𝑟𝑜𝑝𝑦 = −𝑝𝑙𝑜𝑔2 𝑝 − (1 − 𝑝)𝑙𝑜𝑔2 (1 − 𝑝)

where p is the percentage of one type of samples. If all the samples

in one node are of one type (such as C), the entropy is 0. If the
proportion of each type in a node is 50% - 50%, the entropy is 1. We
can use entropy as splitting criteria. The goal is to decrease entropy
as the tree grows. As an analogy, entropy in physics quantifies the
level of disorder and the goal here is to have the least disorder.
Similarly, the entropy of a splitting node is the weighted average
of the entropy of each child. In the above tree for the students,
the entropy of the root node with all 50 students is − 25 25
50 𝑙𝑜𝑔2 50 −
25 25
50 𝑙𝑜𝑔2 50 = 1. Here an entropy of 1 indicates that the purity of the
node is the lowest, that is, each type takes up half of the samples.
The entropy of the split using variable “gender” can be calculated
in three steps:
5 5
1. Entropy for “Female” = − 30 𝑙𝑜𝑔2 30 − 25 25
30 𝑙𝑜𝑔2 30 = 0.65
2. Entropy for “Male” = 0 × 1 + 1 × 0 = 0
224 11 Tree-Based Methods

3. Entropy for the node “Gender” is the weighted average of

the above two entropy numbers: 53 × 0.65 + 25 × 0 = 0.39

So entropy decreases from 1 to 0.39 after the split and the IG for
“Gender” is 0.61.
Information Gain Ratio (IGR)
ID3 uses information gain as the splitting criterion to train the
classification tree. A drawback of information gain is that it is
biased towards choosing attributes with many values, resulting in
overfitting (selecting a feature that is non-optimal for prediction)
(HSSINA et al., 2014).
To understand why let’s look at another hypothetical scenario. As-
sume that the training set has students’ birth month as a feature.
You might say that the birth month should not be considered in
this case because it intuitively doesn’t help tell the student’s ma-
jor. Yes, you’re right. However, practically, we may have a much
more complicated dataset, and we may not have such intuition
for all the features. So, we may not always be able to determine
whether a feature makes sense or not. If we use the birth month
to split the data, the corresponding entropy of the mode “Birth
Month” is 0.24 (the sum of column “Weighted Entropy” in the ta-
ble), and the information gain is 0.76, which is larger than the IG
of “Gender” (0.61). So between the two features, IG would choose
“Birth Month” to split the data.
11.2 Splitting Criteria 225

To overcome this problem, C4.5 uses the “information gain ratio”

instead of “information gain.” The gain ratio is defined as:

𝐼𝑛𝑓𝑜𝑟𝑚𝑎𝑡𝑖𝑜𝑛 𝐺𝑎𝑖𝑛
𝐺𝑎𝑖𝑛 𝑅𝑎𝑡𝑖𝑜 =
𝑆𝑝𝑙𝑖𝑡 𝐼𝑛𝑓𝑜𝑟𝑚𝑎𝑡𝑖𝑜𝑛
where split information is:

𝑆𝑝𝑙𝑖𝑡 𝐼𝑛𝑓𝑜𝑟𝑚𝑎𝑡𝑖𝑜𝑛 = −Σ𝐶

𝑐=1 𝑝𝑐 𝑙𝑜𝑔(𝑝𝑐 )

𝑝𝑐 is the proportion of samples in category 𝑐. For example, there

are three students with the birth month in Jan, 6% of the total
50 students. So the 𝑝𝑐 for “Birth Month = Jan” is 0.06. The split
information measures the intrinsic information that is independent
of the sample distribution inside different categories. The gain ratio
corrects the IG by taking the intrinsic information of a split into
account.

The split information for the birth month is 3.4, and the gain ratio
is 0.22, which is smaller than that of gender (0.63). The gain ratio
refers to use gender as the splitting feature rather than the birth
month. Gain ratio favors attributes with fewer categories and leads
to better generalization (less overfitting).
226 11 Tree-Based Methods

Sum of Squared Error (SSE)

The previous two metrics are for classification tree. The SSE is
the most widely used splitting metric for regression. Suppose you
want to divide the data set 𝑆 into two groups of 𝑆1 and 𝑆2 , where
the selection of 𝑆1 and 𝑆2 needs to minimize the sum of squared
errors:

𝑆𝑆𝐸 = Σ𝑖∈𝑆1 (𝑦𝑖 − 𝑦1̄ )2 + Σ𝑖∈𝑆2 (𝑦𝑖 − 𝑦2̄ )2 (11.1)

In equation (11.1), 𝑦1̄ and 𝑦2̄ are the average of the sample in 𝑆1
and 𝑆2 . The way regression tree grows is to automatically decide
on the splitting variables and split points that can maximize SSE
reduction. Since this process is essentially a recursive segmenta-
tion, this approach is also called recursive partitioning.
Take a look at this simple regression tree for the height of 10
students:

You can calculate the SSE using the following code:

1. SSE for “Female” is 136

2. SSE for “Male” is 32
3. SSE for splitting node “Gender” is the sum of the above
two numbers which is 168

SSE for the 10 students in root node is 522.9. After the split, SSE
decreases from 522.9 to 168.
 11.3 Tree Pruning 227

If there is another possible way of splitting, divide it by major, as

follows:

In this situation:

1. SSE for “Math” is 184

2. SSE for “English” is 302.8
3. SSE for splitting node “Major” is the sum of the above
two numbers which is 486.8

Splitting data using variable “gender” reduced SSE from 522.9

to 168; using variable “major” reduced SSE from 522.9 to 486.8.
Based on SSE reduction, you should use gender to split the data.
The three splitting criteria mentioned above are the basis for build-
ing a tree model.

11.3 Tree Pruning

Pruning is the process that reduces the size of decision trees. It
reduces the risk of overfitting by limiting the size of the tree or
removing sections of the tree that provide little power.
Limit the size
You can limit the tree size by setting some parameters.
• Minimum sample size at each node: Defining the minimum sam-
 228 11 Tree-Based Methods

ple size at the node helps to prevent the leaf nodes having only
one sample. The sample size can be a tuning parameter. If it
is too large, the model tends to under-fit. If it is too small, the
model tends to over-fit. In the case of severe class imbalance,
the minimum sample size may need to be smaller because the
number of samples in a particular class is small.
• Maximum depth of the tree: If the tree grows too deep, the model
tends to over-fit. It can be a tuning parameter.
• Maximum number of terminal nodes: Limit on the terminal
nodes works the same as the limit on the depth of the tree.
They are proportional.
• The number of variables considered for each split: the algorithm
randomly selects variables used in finding the optimal split point
at each level. In general, the square root of the number of all
variables works best, which is also the default setting for many
functions. However, people often treat it as a tuning parameter.
Remove branches
Another way is to first let the tree grow as much as possible and
then go back to remove insignificant branches. The process reduces
the depth of the tree. The idea is to overfit the training set and then
correct using cross-validation. There are different implementations.
• cost/complexity penalty
The idea is that the pruning minimizes the penalized error 𝑆𝑆𝐸𝜆
with a certain value of tuning parameter 𝜆.

𝑆𝑆𝐸𝜆 = 𝑆𝑆𝐸 + 𝜆 × (𝑐𝑜𝑚𝑝𝑙𝑒𝑥𝑖𝑡𝑦)

Here complexity is a function of the number of leaves. For every

given 𝜆, we want to find the tree that minimizes this penalized
error. Breiman presents the algorithm to solve the optimization
(Breiman et al., 1984).
To find the optimal pruning tree, you need to iterate through a
series of values of 𝜆 and calculate the corresponding SSE. For the
 11.3 Tree Pruning 229

same 𝜆, SSE changes over different samples. Breiman et al. sug-

gested using cross-validation (Breiman et al., 1984) to study the
variation of SSE under each 𝜆 value. They also proposed a standard
deviation criterion to give the simplest tree: within one standard
deviation, find the simplest tree that minimizes the absolute er-
ror. Another method is to choose the tree size that minimizes the
numerical error (Hastie T, 2008).
• Error-based pruning
This method was first proposed by Quinlan (Quinlan, 1999). The
idea behind is intuitive. All split nodes of the tree are included
in the initial candidate pool. Pruning a split node means remov-
ing the entire subtree under the node and setting the node as a
terminal node. The data is divided into 3 subsets for:

(1) training a complete tree;

(2) pruning;
(3) testing the final model.

You train a complete tree using the subset (1) and apply the tree
on the subset (2) to calculate the accuracy. Then prune the tree
based on a node and apply that on the subset (2) to calculate
another accuracy. If the accuracy after pruning is higher or equal
to that from the complete tree, then we set the node as a terminal
node. Otherwise, keep the subtree under the node. The advantage
of this method is that it is easy to compute. However, when the
size of the subset (2) is much smaller than that of the subset (1),
there is a risk of over-pruning. Some researchers found that this
method results in more accurate trees than pruning process based
on tree size (F. Espoito and Semeraro, 1997).
• Error-complexity pruning
This method is to search for a trade-off between error and com-
plexity. Assume we have a splitting node 𝑡, and the corresponding
subtree 𝑇 . The error cost of the node is defined as:
 230 11 Tree-Based Methods

𝑚𝑖𝑠𝑐𝑙𝑎𝑠𝑠𝑖𝑓𝑖𝑒𝑑 𝑠𝑎𝑚𝑝𝑙𝑒 𝑠𝑖𝑧𝑒 𝑜𝑓 𝑡ℎ𝑒 𝑛𝑜𝑑𝑒

𝑅(𝑡) = 𝑟(𝑡) × 𝑝(𝑡) =
𝑡𝑜𝑡𝑎𝑙 𝑠𝑎𝑚𝑝𝑙𝑒 𝑠𝑖𝑧𝑒

where 𝑟(𝑡) is the error rate associate with the node as if it is a

terminal node:

𝑚𝑖𝑠𝑐𝑙𝑎𝑠𝑠𝑖𝑓𝑖𝑒𝑑 𝑠𝑎𝑚𝑝𝑙𝑒 𝑠𝑖𝑧𝑒 𝑜𝑓 𝑡ℎ𝑒 𝑛𝑜𝑑𝑒

𝑟(𝑡) =
𝑠𝑎𝑚𝑝𝑙𝑒 𝑠𝑖𝑧𝑒 𝑜𝑓 𝑡ℎ𝑒 𝑛𝑜𝑑𝑒

𝑝(𝑡) is the ratio of the sample of the node to the total sample�

𝑠𝑎𝑚𝑝𝑙𝑒 𝑠𝑖𝑧𝑒 𝑜𝑓 𝑡ℎ𝑒 𝑛𝑜𝑑𝑒

𝑝(𝑡) =
𝑡𝑜𝑡𝑎𝑙 𝑠𝑎𝑚𝑝𝑙𝑒 𝑠𝑖𝑧𝑒

The multiplication 𝑟(𝑡) × 𝑝(𝑡) cancels out the sample size of the
node. If we keep node 𝑡, the error cost of the subtree 𝑇 is:

𝑅(𝑇 ) = Σ𝑖=𝑛𝑜. 𝑜𝑓 𝑙𝑒𝑎𝑣𝑒𝑠 𝑖𝑛 𝑠𝑢𝑏𝑡𝑟𝑒𝑒 𝑇 𝑅(𝑖)

The error-complexity measure of the node is:

𝑅(𝑡) − 𝑅(𝑇 )𝑡
𝑎(𝑡) =
𝑛𝑜. 𝑜𝑓 𝑙𝑒𝑎𝑣𝑒𝑠 − 1

Based on the metrics above, the pruning process is (Patel and

Upadhyay, 2012)�

1. Calculate 𝑎 for every node 𝑡

2. Prune the node with the lowest value
3. Repeat 1 and 2. It produces a pruned tree each time and
they form a forest.
4. Select the tree with the best overall accuracy
• Minium error pruning
Niblett and Brotko came it up in 1991 (Cestnik and Bratko, 1991).
The process is bottom-up which seeks a single tree that minimizes
the expected error rate on new samples. If we prune a splitting
 11.4 Regression and Decision Tree Basic 231

point 𝑡, all the samples under 𝑡 will be classified as from one cat-
egory, say category 𝑐. If we prune the subtree, the expected error
rate is:

𝑛𝑡 − 𝑛𝑡,𝑐 + 𝑘 − 1
𝐸(𝑡) =
𝑛𝑡 + 𝑘

where:

𝑘 = 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑐𝑎𝑡𝑒𝑔𝑜𝑟𝑖𝑒𝑠
𝑛𝑡 = 𝑠𝑎𝑚𝑝𝑙𝑒 𝑠𝑖𝑧𝑒 𝑢𝑛𝑑𝑒𝑟 𝑛𝑜𝑑𝑒 𝑡
𝑛𝑡,𝑐 = 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑠𝑎𝑚𝑝𝑙𝑒 𝑢𝑛𝑑𝑒𝑟 𝑡 𝑡ℎ𝑎𝑡 𝑏𝑒𝑙𝑜𝑛𝑔 𝑡𝑜 𝑐𝑎𝑡𝑒𝑔𝑜𝑟𝑦 𝑐

Based on the above definition, the pruning process is (F. Espoito

and Semeraro, 1997):
• Calculate the expected error rate for each non-leave node if that
subtree is pruned
• Calculate the expected error rate for each non-leave node if that
subtree is not pruned
• If pruning the node leads to higher expected rate, then keep the
subtree; otherwise, prune it.

11.4 Regression and Decision Tree Basic

11.4.1 Regression Tree
Let’s look at the process of building a regression tree
(Gareth James and Tibshirani, 2015). There are two steps:

1. Divide predictors space — that is a set of possible values

of 𝑋1 , 𝑋2 , … , 𝑋𝑝 — into 𝐽 distinct and non-overlapping
regions: 𝑅1 , 𝑅2 , … , 𝑅𝐽
2. For every observation that falls into the region 𝑅𝑗 , the
prediction is the mean of the response values for the train-
ing observations in 𝑅𝑗
232 11 Tree-Based Methods

Let’s go back to the previous baby example. If we use the variable

“Gender” to divide the observations, we obtain two regions 𝑅1
(female) and 𝑅2 (male).

y1 <- c(156, 167, 165, 163, 160, 170, 160)

y2 <- c(172, 180, 176)

The sample average for region 𝑅1 is 163, for region 𝑅2 is 176. For
a new observation, if it is female, the model predicts the height to
be 163, if it is male, the predicted height is 176. Calculating the
mean is easy. Let’s look at the first step in more detail which is to
divide the space into 𝑅1 , 𝑅2 , … , 𝑅𝐽 .
In theory, the region can be any shape. However, to simplify the
problem, we divide the predictor space into high-dimensional rect-
angles. The goal is to divide the space in a way that minimize RSS.
Practically, it is nearly impossible to consider all possible parti-
tions of the feature space. So we use an approach named recursive
binary splitting, a top-down , greedy algorithm. The process starts
from the top of the tree (root node) and then successively splits
the predictor space. Each split produces two branches (hence bi-
nary). At each step of the process, it chooses the best split at that
particular step, rather than looking ahead and picking a split that
leads to a better tree in general (hence greedy).

𝑅1 (𝑗, 𝑠) = {𝑋|𝑋𝑗 < 𝑠} 𝑎𝑛𝑑 𝑅2 (𝑗, 𝑠) = {𝑋|𝑋𝑗 ≥ 𝑠}

11.4 Regression and Decision Tree Basic 233

Calculate the RSS decrease after the split. For different (𝑗, 𝑠),
search for the combination that minimizes the RSS, that is to
minimize the following:

̂ 1 )2 + Σ𝑖∶𝑥𝑖 ∈𝑅2 (𝑗,𝑠) (𝑦𝑖 − 𝑦𝑅

Σ𝑖∶𝑥𝑖 ∈𝑅1 (𝑗,𝑠) (𝑦𝑖 − 𝑦𝑅 ̂ 2 )2

where 𝑦𝑅̂ 1 is the mean of all samples in 𝑅1 , 𝑦𝑅

̂ 2 is the mean of
samples in 𝑅2 . It can be quick to optimize the equation above.
Especially when 𝑝 is not too large.
Next, we continue to search for the split that optimize the RSS.
Note that the optimization is limited in the sub-region. The pro-
cess keeps going until a stopping criterion is reaches. For example,
continue until no region contains more than 5 samples or the RSS
decreases less than 1%. The process is like a tree growing.

There are multiple R packages for building regression tree, such as

ctree, rpart and tree. rpart is widely used for building a single tree.
The split is based on CART algorithm, using rpart() function from
the package. There are some parameters that controls the model
fitting, such as the minimum number of observations that must
exist in a node in order for a split to be attempted, the minimum
number of observations in any leaf node etc. You can can set those
parameter using rpart.control.
A more convenient way is to use train() function in caret package.
The package can call rpart() function and train the model through
234 11 Tree-Based Methods

cross-validation. In this case, the most common parameters are cp

(complexity parameter) and maxdepth (the maximum depth of any
node of the final tree). To tune the complexity parameter, set
method = "rpart". To tune the maximum tree depth, set method =
"rpart2" :

dat <- read.csv("http://bit.ly/2P5gTw4")

# data cleaning: delete wrong observations
dat <- subset(dat, store_exp > 0 & online_exp > 0)
# use the 10 survey questions as predictors
trainx <- dat[, grep("Q", names(dat))]
# use the sum of store and online expenditure as response variable
# total expenditure = store expenditure + online expenditure
trainy <- dat$store_exp + dat$online_exp
set.seed(100)
rpartTune <- train(trainx, trainy,
method = "rpart2",
tuneLength = 10,
trControl = trainControl(method = "cv"))
plot(rpartTune)

1800
RMSE (Cross-Validation)

1750

1700

1650

2 4 6 8 10 12

Max Tree Depth

11.4 Regression and Decision Tree Basic 235

RMSE doesn’t change much when the maximum is larger than 2.

So we set the maximum depth to be 2 and refit the model:

rpartTree <- rpart(trainy ~ ., data = trainx, maxdepth = 2)

You can check the result using print():

print(rpartTree)

## n= 999
##
## node), split, n, deviance, yval
## * denotes terminal node
##
## 1) root 999 1.581e+10 3479.0
## 2) Q3< 3.5 799 2.374e+09 1819.0
## 4) Q5< 1.5 250 3.534e+06 705.2 *
## 5) Q5>=1.5 549 1.919e+09 2326.0 *
## 3) Q3>=3.5 200 2.436e+09 10110.0 *

You can see that the final model picks Q3 and Q5 to predict total
expenditure. To visualize the tree, you can convert rpart object to
party object using partykit then use plot() function:

rpartTree2 <- as.party(rpartTree)

plot(rpartTree2)
236 11 Tree-Based Methods

1
Q3

< 3.5 ≥ 3.5

2
Q5

< 1.5 ≥ 1.5

Node 3 (n = 250) Node 4 (n = 549) Node 5 (n = 200)

50000 50000 50000
40000 40000 40000
30000 30000 30000
20000 20000 20000
10000 10000 10000
0 0 0

11.4.2 Decision Tree

Similar to a regression tree, the goal of a classification tree is to
stratifying the predictor space into a number of sub-regions that
are more homogeneous. The difference is that a classification tree
is used to predict a categorical response rather than a continuous
one. For a classification tree, the prediction is the most commonly
occurring class of training observations in the region to which an
observation belongs. The splitting criteria for a classification tree
are different. The most common criteria are entropy and Gini im-
purity. CART uses Gini impurity and C4.5 uses entropy.
When the predictor is continuous, the splitting process is straight-
forward. When the predictor is categorical, the process can take
different approaches:

1. Keep the variable as categorical and group some cate-

gories on either side of the split. In this way, the model
can make more dynamic splits but must treat the cate-
gorical predictor as an ordered set of bits.
2. Encode the categorical variable to be a set of dummy
variables. In this way, the information in the categorical
11.4 Regression and Decision Tree Basic 237

variable is decomposed to independent bits of informa-

tion. The model considers these dummy variables sepa-
rately and evaluates for each of these on one split point
(because there are only two possible values: 0/1).

When fitting tree models, people need to choose the way to treat
categorical predictors. If you know some of the categories have
higher predictability, then the first approach may be better. In
the rest of this section, we will build tree models using the above
two approaches and compare them.
Let’s build a classification model to identify the gender of the
customer. The population is relatively balanced (55.4% male and
44.6% female).

dat <- read.csv("http://bit.ly/2P5gTw4")

# use the 10 survey questions as predictors
trainx1 <- dat[, grep("Q", names(dat))]
# add a categorical predictor
# use two ways to treat categorical predictor
# trainx1: use approach 1, without encoding
trainx1$segment <- dat$segment

# trainx2: use approach 2, encode it to a set of dummy variables

dumMod <- dummyVars(
~.,
data = trainx1,
# Combine the previous variable and the level name
# as the new dummy variable name
levelsOnly = F
)
trainx2 <- predict(dumMod, trainx1)
# the response variable is gender
trainy <- dat$gender

Here we use train() function in caret package to call rpart to build

238 11 Tree-Based Methods

the model. We can compare the model results from the two ap-
proaches:
CART

1000 samples
11 predictor
2 classes: 'Female', 'Male'

No pre-processing
Resampling: Cross-Validated (10 fold)
Summary of sample sizes: 901, 899, 900, 900, 901, 900, ...
Resampling results across tuning parameters:

cp ROC Sens Spec

0.00000 0.6937 0.6517 0.6884
0.00835 0.7026 0.6119 0.7355
0.01670 0.6852 0.5324 0.8205
0.02505 0.6803 0.5107 0.8498
0.03340 0.6803 0.5107 0.8498

......

0.23380 0.6341 0.5936 0.6745

0.24215 0.5556 0.7873 0.3240

ROC was used to select the optimal model using the largest value.
The final value used for the model was cp = 0.00835.

The above keeps the variable as categorical without encoding. Here

cp is the complexity parameter. It is used to decide when to stop
growing the tree. cp = 0.01 means the algorithm only keeps the
split that improves the corresponding metric by more than 0.01.
Next, let’s encode the categorical variable to be a set of dummy
variables and fit the model again:
11.4 Regression and Decision Tree Basic 239

rpartTune2 <- caret::train(

trainx2, trainy, method = "rpart",
tuneLength = 30,
metric = "ROC",
trControl = trainControl(method = "cv",
summaryFunction = twoClassSummary,
classProbs = TRUE,
savePredictions = TRUE)
)

Compare the results of the two approaches.

rpartRoc <- pROC::roc(response = rpartTune1$pred$obs,

predictor = rpartTune1$pred$Female,
levels = rev(levels(rpartTune1$pred$obs)))

rpartFactorRoc <- pROC::roc(response = rpartTune2$pred$obs,

predictor = rpartTune2$pred$Female,
levels = rev(levels(rpartTune1$pred$obs)))

plot.roc(rpartRoc,
type = "s",
print.thres = c(.5),
print.thres.pch = 3,
print.thres.pattern = "",
print.thres.cex = 1.2,
col = "red", legacy.axes = TRUE,
print.thres.col = "red")

plot.roc(rpartFactorRoc,
type = "s",
add = TRUE,
print.thres = c(.5),
print.thres.pch = 16, legacy.axes = TRUE,
print.thres.pattern = "",
print.thres.cex = 1.2)
240 11 Tree-Based Methods

legend(.75, .2,
c("Grouped Categories", "Independent Categories"),
lwd = c(1, 1),
col = c("black", "red"),
pch = c(16, 3))
1.0
0.8
0.6
Sensitivity
0.4
0.2

Grouped Categories
Independent Categories
0.0

0.0 0.5 1.0

1 - Specificity

In this case, the two approaches lead to similar model performance.

Single tree is straightforward and easy to interpret but it has prob-
lems:

1. Low accuracy
2. Unstable: little change in the training data leads to very
different trees.

One way to overcome those is to use an ensemble of trees. In the

rest of this chapter, we will introduce three ensemble methods
(combine many models’ predictions): bagging tree, random forest,
and gradient boosted machine. Those ensemble approaches have
11.5 Bagging Tree 241

significant higher accuracy and stability. However, it comes with

the cost of interpretability.

11.5 Bagging Tree

As mentioned before, a single tree is unstable. If you randomly
separate the sample to be two parts and fit tree model on each,
you can get two very different trees. A stable model should give a
similar result on different random samples. Some traditional statis-
tical models have high stability, such as linear regression. Ensem-
ble methods appeared in the 1990s which can effectively stabilize
the model. Bootstrapping is a type of process where you repeated
draw samples of the same size from a single original sample with
replacement (B and R, 1986). Bootstrap aggregation (Bagged) is
an ensemble technique proposed by Leo Breiman (Breiman 1996a).
It uses bootstrapping in conjunction with any model to construct
an ensemble. The process is very straightforward:

Algorithm 3 Bagging tree

1: Build a model on different bootstrap samples to form an en-
semble, say 𝐵 samples
2: For a new sample, each model will give a prediction:
𝑓 1̂ (𝑥), 𝑓 2̂ (𝑥) … , 𝑓 𝐵̂ (𝑥)
3: The bagged model’s prediction is the average of all the predic-
tions:
̂ (𝑥) = 1 Σ𝐵 𝑓 𝑏̂ (𝑥)
𝑓𝑎𝑣𝑔
𝐵 𝑏=1

Assume that there are 𝑛 independent random variables 𝑍1 , … , 𝑍𝑛

2
with variance 𝜎2 . Then the variance of the mean 𝑍 ̄ is 𝜎𝑛 . It is easy
to see why bagged models have less variance. Since bootstrapping
is to sample with replacement, it means some samples are selected
multiple times and some not at all. Those left out samples are
called out-of-bag. You can use the out-of-bag sample to access
the model performance. For regression, the prediction is a simple
 242 11 Tree-Based Methods

average. For classification, the prediction is the category with the

most “votes”. Here, the number of trees, 𝐵 is a parameter you need
to decide, i.e. tuning parameter. Bagging is a general approach
that can be applied to different learners. Here we only
discuss in the context of decision trees.
The advantages of bagging tree are:
• Bagging stabilizes the model predictions by averaging the results.
If we have 10 bootstrap samples and fit a single tree on each of
those, we may get 10 trees with very different structures and
leading to different predictions for a new sample. But if we use
the average of the 10 predictions as the final prediction, then
the result is much more stable. It means if we have another 10
samples and do it all-over again, we will get very similar averaged
prediction.
• Bagging provides more accurate predictions. If the goal is to pre-
dict rather than interpret, then the ensemble approach definitely
has an advantage, especially for unstable models. However, for
stable models (such as regression, MARS), bagging may bring
marginal improvement for the model performance.
• Bagging can use out-of-bag samples to evaluate model perfor-
mance. For each model in the ensemble, we can calculate the
value of the model performance metric (you can decide what
metric to use). You can use the average of all the out-of-bag
performance values to gauge the predictive performance of the
entire ensemble. This correlates well with either cross-validation
estimates or test set estimates. On average, each tree uses about
2/3 of the samples, and the rest 1/3 is used as out-of-bag. When
the number of bootstrap samples is large enough, the out-of-
bag performance estimate approximates that from leave one out
cross-validation.
You need to choose the number of bootstrap samples. The au-
thor of “Applied Predictive Modeling” (Kuhn and Johnston, 2013)
points out that often people see an exponential decrease in pre-
dictive improvement as the number of iterations increases. Most
of the predictive power is from a small portion of the trees. Based
 11.5 Bagging Tree 243

on their experience, model performance can have small improve-

ments up to 50 bagging iterations. If it is still not satisfying, they
suggest trying other more powerfully predictive ensemble methods
such as random forests and boosting which will be described in the
following sections.
The disadvantages of bagging tree are:
• As the number of bootstrap samples increases, the computation
and memory requirements increase as well. You can mitigate
this disadvantage by parallel computing. Since each bootstrap
sample and modeling is independent of any other sample and
model, you can easily parallelize the bagging process by building
those models separately and bring back the results in the ed to
generate the prediction.
• The bagged model is difficult to explain which is common for all
ensemble approaches. However, you can still get variable impor-
tance by combining measures of importance across the ensemble.
For example, we can calculate the RSS decrease for each variable
across all trees and use the average as the measurement of the
importance.
• Since the bagging tree uses all of the original predictors as ev-
erey split of every tree, those trees are related with each other.
The tree correlation prevents bagging from optimally reducing
the variance of the predicted values. See (Hastie T, 2008) for a
mathematical illustration of the tree correlation phenomenon.
Let’s look at how to use R to build bagging tree. Get the predictors
and response variable first:

dat <- read.csv("http://bit.ly/2P5gTw4")

# use the 10 survey questions as predictors
trainx <- dat[, grep("Q", names(dat))]
# add segment as a predictor
# don't need to encode it to dummy variables
trainx$segment <- as.factor(dat$segment)
244 11 Tree-Based Methods

# use gender as the response variable

trainy <- as.factor(dat$gender)

Then fit the model using train function in caret package. Here
we just set the number of trees to be 1000. You can tune that
parameter.

set.seed(100)
bagTune <- caret::train(trainx, trainy,
method = "treebag",
nbagg = 1000,
metric = "ROC",
trControl = trainControl(method = "cv",
summaryFunction = twoClassSummary,
classProbs = TRUE,
savePredictions = TRUE))

The model results are:

bagTune

## Bagged CART
##
## 1000 samples
## 11 predictor
## 2 classes: 'Female', 'Male'
##
## No pre-processing
## Resampling: Cross-Validated (10 fold)
## Summary of sample sizes: 901, 899, 900, 900, 901, 900, ...
## Resampling results:
##
## ROC Sens Spec
## 0.7093 0.6533 0.6774

Since we only have a handful of variables in this example, the

11.6 Random Forest 245

maximum AUC doesn’t improve by using bagging tree. But it

makes a difference when we have more predictors.

11.6 Random Forest

Since the tree correlation prevents bagging from optimally reduc-
ing the variance of the predicted values, a natural way to improve
the model performance is to reduce the correlation among trees.
That is what random forest aims to do: improve the performance
of bagging by de-correlating trees.
From a statistical perspective, you can de-correlate trees by in-
troducing randomness when you build each tree. One approach
(T, 1998; Y and D, 1997) is to randomly choose 𝑚 variables to
use each time you build a tree. Dietterich(T, 2000) came up with
the idea of random split selection which is to randomly choose 𝑚
variables to use at each splitting node. Based on the different gen-
eralizations to the original bagging algorithm, Breiman (Breiman,
2001a) came up with a unified algorithm called random forest.
When building a tree, the algorithm randomly chooses 𝑚 vari-
ables to use at each splitting node. Then choose the best one out
√
of the 𝑚 to use at that node. In general, people use 𝑚 = 𝑝.
For example, if we use 10 questions from the questionnaire as pre-
dictors, then at each node, the algorithm will randomly choose 4
candidate variables. Since those trees in the forest don’t always use
the same variables, tree correlation is less than that in bagging.
It tends to work better when there are more predictors. Since we
only have 10 predictors here, the improvement from the random
forest is marginal. The number of randomly selected predictors is
a tuning parameter in the random forest. Since random forest is
computationally intensive, we suggest starting with 5 value around
√
𝑚 = 𝑝. Another tuning parameter is the number of trees in the
forest. You can start with 1000 trees and then increase the number
until performance levels off. The basic random forest is shown in
Algorithm 4.
246 11 Tree-Based Methods

Algorithm 4 Random forest

1: Select the number of trees, B
2: for i=1 to B do
3: Generate a bootstrap sample of the original data
4: Train a tree on this sample:
5: for each split do
6: Randomly select m (< p) predictors
7: Choose the best one out of the 𝑚 and partition the
data
8: end for
9: Use typical tree model stopping criteria to determine when
a tree is complete without pruning
10: end for

When 𝑚 = 𝑝, random forest is equal to the bagging tree. When

the predictors are highly correlated, then smaller 𝑚 tends to work
better. Let’s use the caret package to train a random forest:

# tune across a list of numbers of predictors

mtryValues <- c(1:5)
set.seed(100)
rfTune <- train(x = trainx,
y = trainy,
# set the model to be random forest
method = "rf",
ntree = 1000,
tuneGrid = data.frame(.mtry = mtryValues),
importance = TRUE,
metric = "ROC",
trControl = trainControl(method = "cv",
summaryFunction = twoClassSummary,
classProbs = TRUE,
savePredictions = TRUE))
11.6 Random Forest 247

rfTune

## Random Forest
##
## 1000 samples
## 11 predictor
## 2 classes: 'Female', 'Male'
##
## No pre-processing
## Resampling: Cross-Validated (10 fold)
## Summary of sample sizes: 899, 900, 900, 899, 899, 901, ...
## Resampling results across tuning parameters:
##
## mtry ROC Sens Spec
## 1 0.7169 0.5341 0.8205
## 2 0.7137 0.6334 0.7175
## 3 0.7150 0.6478 0.6995
## 4 0.7114 0.6550 0.6950
## 5 0.7092 0.6514 0.6882
##
## ROC was used to select the optimal model using
## the largest value.
## The final value used for the model was mtry = 1.

In this example, since the number of predictors is small, the result

of the model indicates that the optimal number of candidate vari-
ables at each node is 1. The optimal AUC is not too much higher
than that from bagging tree:
If you have selected the values of tuning parameters, you can also
use the randomForest package to fit a random forest.

rfit = randomForest(trainy ~ ., trainx, mtry = 1, ntree = 1000)

Since bagging tree is a special case of random forest, you can fit
the bagging tree by setting 𝑚𝑡𝑟𝑦 = 𝑝. Function importance() can
return the importance of each predictor:
248 11 Tree-Based Methods

importance(rfit)

## MeanDecreaseGini
## Q1 9.056
## Q2 7.582
## Q3 7.611
## Q4 12.308
## Q5 5.628
## Q6 9.740
## Q7 6.638
## Q8 7.829
## Q9 5.955
## Q10 4.781
## segment 11.185

You can use varImpPlot() function to visualize the predictor im-

portance:

varImpPlot(rfit)

rfit

Q4
segment
Q6
Q1
Q8
Q3
Q2
Q7
Q9
Q5
Q10

0 2 4 6 8 10 12

MeanDecreaseGini
11.7 Gradient Boosted Machine 249

It is easy to see from the plot that segment and Q4 are the top two
variables to classify gender.

11.7 Gradient Boosted Machine

Boosting models were developed in the 1980s (L, 1984; M and L,
1989) and were originally for classification problems. Due to the
excellent model performance, they were widely used for a variety
of applications, such as gene expression (Dudoit S and T, 2002;
et. al, 2000), chemical substructure classification (Varmuza K and
K, 2003), music classification (et. al, 2006), etc. The first effective
implementation of boosting is Adaptive Boosting (AdaBoost) al-
gorithm came up by Yoav Freund and Robert Schapire in 1996
(YFR, 1999). After that, some researchers (Friedman et al., 2000)
started to connect the boosting algorithm with some statistical
concepts, such as loss function, additive model, logistic regression.
Friedman pointed out that boosting can be considered as a for-
ward stagewise additive model that minimizes exponential loss.
The new view of boosting in a statistical framework enabled the
method to be extended to regression problems.
The idea is to combine a group of weak learners (a classifier that is
marginally better than random guess) to produce a strong learner.
Like bagging, boosting is a general approach that can be applied
to different learners. Here we focus on the decision tree. Recall
that both bagging and random forest create multiple copies of
the original training data using the bootstrap, fitting a separate
decision tree to each copy and combining all the results to create
a single prediction. Boosting also creates different trees but the
trees are grown sequentially and each tree is a weak learner. Any
modeling technique with tuning parameters can produce a range of
learners, from weak to strong. You can easily make a weak learner
by restricting the depth of the tree. There are different types of
boosting. Here we introduce two main types: adaptive boosting
and stochastic gradient boosting.
250 11 Tree-Based Methods

11.7.1 Adaptive Boosting

Yoav Freund and Robert Schapire (Freund and Schapire, 1997)
came up the AdaBoost.M1 algorithm. Consider a binary classi-
fication problem where the response variable has two categories
𝑌 ∈ {−1, 1}. Given predictor matrix, 𝑋, construct a classifier
𝐺(𝑋) that predicts 1 or −1. The corresponding error rate in the
training set is:

1 𝑁
𝑒𝑟𝑟̄ = Σ 𝐼(𝑦 ≠ 𝐺(𝑥𝑖 ))
𝑁 𝑖=1 𝑖
The algorithm produces a series of classifiers 𝐺𝑚 (𝑥), 𝑚 =
1, 2, ..., 𝑀 from different iterations. In each iteration, it finds the
best classifier based on the current weights. The misclassified sam-
ples in the 𝑚𝑡ℎ iteration will have higher weights in the (𝑚+1)𝑡ℎ it-
eration and the correctly classified samples will have lower weights.
As it moves on, the algorithm will put more effort into the “diffi-
cult” samples until it can correctly classify them. So it requires the
algorithm to change focus at each iteration. At each iteration, the
algorithm will calculate a stage weight based on the error rate. The
final prediction is a weighted average of all those weak classifiers
using stage weights from all the iterations:

𝐺(𝑥) = 𝑠𝑖𝑔𝑛(Σ𝑀
𝑚=1 𝛼𝑚 𝐺𝑚 (𝑥))

where 𝛼1 , 𝛼2 , ..., 𝛼𝑀 are the weights from different iterations.

Since the classifier 𝐺𝑚 (𝑥) returns discrete value, the AdaBoost.M1
algorithm is known as “Discrete AdaBoost” (Friedman et al.,
2000). You can revise the above algorithm if it returns continuousf
value, for example, a probability (Friedman et al., 2000). As men-
tioned before, boosting is a general approach that can be applied to
different learners. Since you can easily create weak learners by lim-
iting the depth of the tree, the boosting tree is a common method.
Since the classification tree is a low bias/high variance technique,
ensemble decreases model variance and lead to low bias/low vari-
ance model. See Breinman (Breiman, 1998) for more explanation
about why the boosting tree performs well in general. However,
11.7 Gradient Boosted Machine 251

Algorithm 5 AdaBoost.M1
1: Response variables have two values: +1 and -1
2: Initialize the observation to have the same weights: 𝑤𝑖 =
1
𝑁 , 𝑖 = 1, ..., 𝑁
3: for m = 1 to M do
4: Fit a classifier 𝐺𝑚 (𝑥) to the training data using weights
𝑤𝑖
5: Compute the error rate for the 𝑚𝑡 ℎ classifier: 𝑒𝑟𝑟𝑚 =
Σ𝑁
𝑖=1 𝑤𝑖 𝐼(𝑦𝑖 ≠𝐺𝑚 (𝑥𝑖 ))
Σ𝑁𝑖=1 𝑤𝑖
6: Compute the stage weight for 𝑚𝑡ℎ iteration: 𝛼𝑚 =
𝑙𝑜𝑔 1−𝑒𝑟𝑟
𝑒𝑟𝑟𝑚
𝑚

7: Update 𝑤𝑖 = 𝑤𝑖 ⋅ 𝑒𝑥𝑝[𝛼𝑚 ⋅ 𝐼(𝑦𝑖 ≠ 𝐺𝑚 (𝑥𝑖 ))], 𝑖 = 1, 2, … , 𝑁

8: end for
9: Calculate the prediction�𝐺(𝑥) = 𝑠𝑖𝑔𝑛[Σ𝑀 𝑚=1 𝛼𝑚 𝐺𝑚 (𝑥)], where
𝑠𝑖𝑔𝑛(⋅) means if ⋅ is positive, then the sample is classified as
+1, -1 otherwise.

boosting can not significantly improve the low variance model.

So applying boosting to Latent Dirichlet Allocation (LDA) or K-
Nearest Neighbor (KNN) doesn’t lead to as good improvement as
applying boosting to statistical learning methods like naive Bayes
(Bauer and Kohavi, 1999).

11.7.2 Stochastic Gradient Boosting

As mentioned before, Friedman (Friedman et al., 2000) provided
a statistical framework for the AdaBoost algorithm and pointed
out that boosting can be considered as a forward stagewise addi-
tive model that minimizes exponential loss. The framework led to
some generalized algorithms such as Real AdaBoost, Gentle Ad-
aBoost, and LogitBoost. Those algorithms later were unified under
a framework called gradient boosting machine. The last section of
the chapter illustrates how boosting can be considered as an ad-
ditive model.
Consider a 2-class classification problem. You have the response
𝑦 ∈ {−1, 1} and the sample proportion of class 1 from the training
252 11 Tree-Based Methods

set is 𝑝. 𝑓(𝑥) is the model prediction in the range of [−∞, +∞] and
1
the predicted event probability is 𝑝̂ = 1+𝑒𝑥𝑝[−𝑓(𝑥)] . The gradient
boosting for this problem is as follows:

Algorithm 6 Stochastic gradient boosting for 2-class classifica-

tion
1: Response variables have two values: +1 and -1
𝑝̂
2: Initialize all predictions to the sample log-odds: 𝑓𝑖 = 𝑙𝑜𝑔 1−𝑝̂
3: for j=1 ... M do
1
4: Compute predicted event probability: 𝑝𝑖̂ = 1+𝑒𝑥𝑝[−𝑓
𝑖 (𝑥)]
5: Compute the residual (i.e. gradient): 𝑧𝑖 = 𝑦𝑖 − 𝑝𝑖̂
6: Randomly sample the training data
7: Train a tree model on the random subset using the residuals
as the outcome
8: Compute the terminal node estimates of the Pearson resid-
1/𝑛Σ𝑛 (𝑦𝑖 −𝑝̂𝑖 )
uals: 𝑟𝑖 = 1/𝑛Σ𝑛𝑖𝑝̂ (1− 𝑝̂ )
𝑖 𝑖 𝑖
(𝑗)
9: Update f�𝑓𝑖 = 𝑓𝑖 + 𝜆𝑓𝑖
10: end for

When using the tree as the base learner, basic gradient boosting
has two tuning parameters: tree depth and the number of iter-
ations. You can further customize the algorithm by selecting a
different loss function and gradient(Hastie T, 2008). The final line
of the loop includes a regularization strategy. Instead of adding
(𝑗)
𝑓𝑖 to the previous iteration’s 𝑓𝑖 , only a fraction of the value is
added. This fraction is called learning rate which is 𝜆 in the algo-
rithm. It can take values between 0 and 1 which is another tuning
parameter of the model.
The way to calculate variable importance in boosting is similar
to a bagging model. You get variable importance by combining
measures of importance across the ensemble. For example, we can
calculate the Gini index improvement for each variable across all
trees and use the average as the measurement of the importance.
Boosting is a very popular method for classification. It is one of the
methods that can be directly applied to the data without requir-
ing a great deal of time-consuming data preprocessing. Applying
11.7 Gradient Boosted Machine 253

boosting on tree models significantly improves predictive accuracy.

Some advantages of trees that are sacrificed by boosting are speed
and interpretability.
Let’s look at the R implementation.

gbmGrid <- expand.grid(interaction.depth = c(1, 3, 5, 7, 9),

n.trees = 1:5,
shrinkage = c(.01, .1),
n.minobsinnode = c(1:10))

set.seed(100)
gbmTune <- caret::train(x = trainx,
y = trainy,
method = "gbm",
tuneGrid = gbmGrid,
metric = "ROC",
verbose = FALSE,
trControl = trainControl(method = "cv",
classProbs = TRUE,
savePredictions = TRUE))

# only show part of the output

gbmTune

Stochastic Gradient Boosting

1000 samples
11 predictor
2 classes: 'Female', 'Male'

No pre-processing
Resampling: Cross-Validated (10 fold)
Summary of sample sizes: 899, 900, 900, 899, 899, 901, ...
Resampling results across tuning parameters:
254 11 Tree-Based Methods

shrinkage interaction.depth n.minobsinnode n.trees ROC Sens Spec

0.01 1 1 1 0.6821 1.00 0.00
0.01 1 1 2 0.6882 1.00 0.00
.
.
.
0.01 5 8 4 0.7096 1.00 0.00
0.01 5 8 5 0.7100 1.00 0.00
0.01 5 9 1 0.7006 1.00 0.00
0.01 5 9 2 0.7055 1.00 0.00
[ reached getOption("max.print") -- omitted 358 rows ]

ROC was used to select the optimal model using the largest value.
The final values used for the model were n.trees = 4,
interaction.depth = 3, shrinkage = 0.01 and n.minobsinnode = 6.

The results show that the tuning parameter settings that lead to
the best ROC are n.trees = 4 (number of trees), interaction.depth =
3 (depth of tree), shrinkage = 0.01 (learning rate) and n.minobsinnode
= 6 (minimum number of observations in each node).

Now, let’s compare the results from the three tree models.

treebagRoc <- pROC::roc(response = bagTune$pred$obs,

predictor = bagTune$pred$Female,
levels = rev(levels(bagTune$pred$obs)))

rfRoc <- pROC::roc(response = rfTune$pred$obs,

predictor = rfTune$pred$Female,
levels = rev(levels(rfTune$pred$obs)))

gbmRoc <- pROC::roc(response = gbmTune$pred$obs,

predictor = gbmTune$pred$Female,
levels = rev(levels(gbmTune$pred$obs)))
11.7 Gradient Boosted Machine 255

plot.roc(rpartRoc,
type = "s",
print.thres = c(.5), print.thres.pch = 16,
print.thres.pattern = "", print.thres.cex = 1.2,
col = "black", legacy.axes = TRUE,
print.thres.col = "black")

plot.roc(treebagRoc,
type = "s",
add = TRUE,
print.thres = c(.5), print.thres.pch = 3,
legacy.axes = TRUE, print.thres.pattern = "",
print.thres.cex = 1.2,
col = "red", print.thres.col = "red")

plot.roc(rfRoc,
type = "s",
add = TRUE,
print.thres = c(.5), print.thres.pch = 1,
legacy.axes = TRUE, print.thres.pattern = "",
print.thres.cex = 1.2,
col = "green", print.thres.col = "green")

plot.roc(gbmRoc,
type = "s",
add = TRUE,
print.thres = c(.5), print.thres.pch = 10,
legacy.axes = TRUE, print.thres.pattern = "",
print.thres.cex = 1.2,
col = "blue", print.thres.col = "blue")

legend(0.2, 0.5, cex = 0.8,

c("Single Tree", "Bagged Tree",
"Random Forest", "Boosted Tree"),
lwd = c(1, 1, 1, 1),
256 11 Tree-Based Methods

col = c("black", "red", "green", "blue"),

pch = c(16, 3, 1, 10))
1.0
0.8
0.6
Sensitivity

Single Tree
0.4

Bagged Tree
Random Forest
Boosted Tree
0.2
0.0

0.0 0.5 1.0

1 - Specificity

Since the data here doesn’t have many variables, we don’t see
a significant difference among the models. But you can still see
those ensemble methods are better than a single tree. In most
of the real applications, ensemble methods perform much better.
Random forest and boosting trees can be a baseline model. Before
exploring different models, you can quickly run a random forest to
see the performance and then try to improve that performance. If
the performance you got from the random forest is not too much
better than guessing, you should consider collecting more data or
reviewing the problem to frame it a different way instead of trying
other models. Because it usually means the current data is not
enough to solve the problem.

11.7.3 Boosting as Additive Model

This section illustrates how boosting is a forward stagewise ad-
ditive model that minimizes exponential loss (Friedman et al.,
11.7 Gradient Boosted Machine 257

2000). Many seemingly different models can be represented as a

basis expansion model. Recall the classifier obtained by the
AdaBoost.M1 algorithm:

𝐺(𝑥) = 𝑠𝑖𝑔𝑛(Σ𝑀
𝑚=1 𝛼𝑚 𝐺𝑚 (𝑥))

The above expression fits in the framework of basis expansion

which is as following:

𝑓(𝑥) = Σ𝑀
𝑚=1 𝛽𝑚 𝑏(𝑥, 𝛾𝑚 ) (11.2)

where 𝛽𝑚 , 𝑚 = 1, … , 𝑀 is expansion coefficient and 𝑏(𝑥, 𝛾) ∈ ℝ is

basis function. Many of the learning methods fit into this additive
framework. Hastie et al discuss basis expansion in detail in Chapter
5 of Hastie T (2008) and cover the additive expansion for different
learning techniques, such as single-hidden-layer neural networks
(Chapter 11) and MARS (Section 9.4). These models are fit by
minimizing a designated loss function averaged over the training
data:

𝑚𝑖𝑛 Σ𝑁 𝑀
𝑖=1 𝐿 (𝑦𝑖 , Σ𝑚=1 𝛽𝑚 𝑏(𝑥𝑖 ; 𝛾𝑚 )) (11.3)
{𝛽𝑚 ,𝛾𝑚 }𝑀
𝑖

Different models have different basis function 𝑏(𝑥𝑖 ; 𝛾𝑚 ). You can

also customize loss function 𝐿(⋅), such as squared-error loss, or
likelihood-based loss. Optimizing the loss function across the whole
training set is usually computation costly no matter the choice of
basis and loss. The good news is that the problem can be simplified
as fitting a single basis function.

𝑚𝑖𝑛 = Σ𝑁
𝑖=1 𝐿(𝑦𝑖 , 𝛽𝑏(𝑥𝑖 ; 𝛾))
𝛽,𝛾

The forward stagewise algorithm approximates the optimal solu-

tion of equation (11.3). It adds new basis functions to the expan-
sion without adjusting the previous ones. The forward stagewise
additive algorithm is as following (Section 10.2 of Hastie T (2008)):
At iteration m, it will search for the optimal 𝑏(𝑥; 𝛾𝑚 ) and 𝛽𝑚 based
258 11 Tree-Based Methods

Algorithm 7 Forward stagewise algorithm

1: Initialize 𝑓0 (𝑥) = 0
2: for 𝑚 = 1, … , 𝑀 do
3: Compute

(𝛽𝑚 , 𝛾𝑚 ) = 𝑎𝑟𝑔𝑚𝑖𝑛Σ𝑁
𝑖=1 𝐿(𝑦𝑖 , 𝑓𝑚−1 (𝑥𝑖 ) + 𝛽𝑏(𝑥𝑖 ; 𝛾))
𝛽,𝛾

4: Set 𝑓𝑚 (𝑥) = 𝑓𝑚−1 (𝑥) + 𝛽𝑚 𝑏(𝑥; 𝛾𝑚 )

5: end for

on the previous basis function 𝑓𝑚−1 (𝑥). And then add the new
basis 𝑏(𝑥; 𝛾𝑚 )𝛽𝑚 to the previous basis function to get a new basis
function 𝑓𝑚 (𝑥) without changing any parameters from previous
steps. Assume we use squared-error loss:

𝐿(𝑦, 𝑓(𝑥)) = (𝑦 − 𝑓(𝑥))2

Then we have:

𝐿(𝑦𝑖 , 𝑓𝑚−1 (𝑥𝑖 ) + 𝛽𝑏(𝑥𝑖 ; 𝛾)) = (𝑦𝑖 − 𝑓𝑚−1 (𝑥𝑖 ) − 𝛽𝑏(𝑥𝑖 ; 𝛾))2

where 𝑦𝑖 − 𝑓𝑚−1 (𝑥𝑖 ) is the residual for sample i based on the pre-
vious model. That is to say, it is fitting the new basis using the
residual of the previous step.
However, the squared-error loss is generally not a good choice. For
a regression problem, it is susceptible to outliers. It doesn’t fit the
classification problem since the response is categorical. Hence we
often consider other loss functions.
Now let is go back to the AdaBoost.M1 algorithm. It is actually
a special case of the above forward stagewise model when the loss
function is:

𝐿(𝑦, 𝑓(𝑥)) = 𝑒𝑥𝑝(−𝑦𝑓(𝑥))

In AdaBoost.M1, the basis function is the classifier from each it-

11.7 Gradient Boosted Machine 259

eration 𝐺𝑚 (𝑥) ∈ {−1, 1}. If we use the exponential loss, the opti-
mization problem is

(𝛽𝑚 , 𝐺𝑚 ) = 𝑎𝑟𝑔𝑚𝑖𝑛Σ𝑁
𝑖=1 𝑒𝑥𝑝[−𝑦𝑖 (𝑓𝑚−1 (𝑥𝑖 ) + 𝛽𝐺(𝑥𝑖 ))]
𝛽,𝐺
= 𝑎𝑟𝑔𝑚𝑖𝑛Σ𝑁
𝑖=1 𝑒𝑥𝑝[−𝑦𝑖 𝛽𝐺(𝑥𝑖 )] ⋅ 𝑒𝑥𝑝[−𝑦𝑖 𝑓𝑚−1 (𝑥𝑖 )]
𝛽,𝐺
= 𝑎𝑟𝑔𝑚𝑖𝑛Σ𝑁 𝑚
𝑖=1 𝑤𝑖 𝑒𝑥𝑝[−𝑦𝑖 𝛽𝐺(𝑥𝑖 )]
𝛽,𝐺
(11.4)
where 𝑤𝑖𝑚 = 𝑒𝑥𝑝[−𝑦𝑖 𝑓𝑚−1 (𝑥𝑖 )]. It does not depend on 𝛽 and 𝐺(𝑥).
So we can consider it as the weight for each sample. Since the
weight is related to 𝑓𝑚−1 (𝑥𝑖 ), it changes each iteration. We can
further decompose equation (11.4):

𝑎𝑟𝑔𝑚𝑖𝑛Σ𝑁 𝑚
𝑖=1 𝑤𝑖 𝑒𝑥𝑝[−𝑦𝑖 𝛽𝐺(𝑥𝑖 )]
𝛽,𝐺
= 𝑎𝑟𝑔𝑚𝑖𝑛Σ𝑁 𝑚 −𝛽 𝑚 𝛽
𝑖=1 {𝑤𝑖 𝑒 𝐼(𝑦𝑖 = 𝐺(𝑥)) + 𝑤𝑖 𝑒 𝐼(𝑦𝑖 ≠ 𝐺(𝑥))}
𝛽,𝐺
= 𝑎𝑟𝑔𝑚𝑖𝑛Σ𝑁 𝑚 −𝛽 𝑚 𝛽
𝑖=1 {𝑤𝑖 𝑒 [1 − 𝐼(𝑦𝑖 ≠ 𝐺(𝑥))] + 𝑤𝑖 𝑒 𝐼(𝑦𝑖 ≠ 𝐺(𝑥))}
𝛽,𝐺
= 𝑎𝑟𝑔𝑚𝑖𝑛 {(𝑒𝛽 − 𝑒−𝛽 ) ⋅ Σ𝑁 𝑚
𝑖=1 𝑤𝑖 𝐼(𝑦𝑖 ≠ 𝐺(𝑥𝑖 )) + 𝑒
−𝛽
⋅ Σ𝑁 𝑚
𝑖=1 𝑤𝑖 }
𝛽,𝐺
(11.5)
when 𝛽 > 0, the solution for equation (11.5) is:

𝐺𝑚 = 𝑎𝑟𝑔𝑚𝑖𝑛Σ𝑁 𝑚
𝑖=1 𝑤𝑖 𝐼(𝑦𝑖 ≠ 𝐺(𝑥))
𝐺

It is the classifier that minimizes the weighted error. Plug the

above 𝐺𝑚 into equation (11.5). Take the derivative with respect
to 𝛽 and set it to be 0 to solve the optimal 𝛽𝑚 :

1 1 − 𝑒𝑟𝑟𝑚
𝛽𝑚 = 𝑙𝑛
2 𝑒𝑟𝑟𝑚

where

Σ𝑁 𝑚
𝑖=1 𝑤𝑖 𝐼(𝑦𝑖 ≠ 𝐺𝑚 (𝑥𝑖 ))
𝑒𝑟𝑟𝑚 = 𝑚
Σ𝑁
𝑖=1 𝑤𝑖
260 11 Tree-Based Methods

According to the forward stagewise algorithm, the result is up-

dated as:

𝑓𝑚 (𝑥) = 𝑓𝑚−1 (𝑥) + 𝛽𝑚 𝐺𝑚 (𝑥)

We can go ahead and get the weight for the next iteration:

𝑤𝑖𝑚+1 = 𝑒𝑥𝑝[−𝑦𝑖 𝑓𝑚 (𝑥𝑖 )]

= 𝑒𝑥𝑝[−𝑦𝑖 𝑓𝑚−1 (𝑥) − 𝑦𝑖 𝛽𝑚 𝐺𝑚 (𝑥)] (11.6)
= 𝑤𝑖𝑚 ⋅ 𝑒𝑥𝑝[−𝛽𝑚 𝑦𝑖 𝐺𝑚 (𝑥𝑖 )]

Since −𝑦𝑖 𝐺𝑚 (𝑥𝑖 ) = 2 ⋅ 𝐼(𝑦𝑖 ≠ 𝐺𝑚 (𝑥𝑖 )) − 1, equation (11.5) can be

written as:

𝑤𝑖𝑚+1 = 𝑤𝑖𝑚 ⋅ 𝑒𝑥𝑝[𝛼𝑚 𝐼(𝑦𝑖 ≠ 𝐺𝑚 (𝑥𝑖 ))] ⋅ 𝑒𝑥𝑝[−𝛽𝑚 ]

where 𝛼𝑚 = 2𝛽𝑚 = 𝑙𝑜𝑔 1−𝑒𝑟𝑟 𝑒𝑟𝑟𝑚 is the same with the 𝛼𝑚 in Ad-
𝑚

aBoost.M1 algorithm we showed before. So AdaBoost.M1 is a spe-

cial case of a forward stagewise additive model using exponential
loss. For comparing and selecting different loss functions, you can
refer to section 10.5 and 10.6 in (Hastie T, 2008).
12
Deep Learning

With everyday applications in language, voice, image, and auto-

matic driving cars, deep learning has become a popular concept to
the general public in the past few years. However, many of the con-
cepts of deep learning started as early as the 1940s. For example,
the binary perceptron classifier, invented in the late 1950s, uses a
linear combination of input signals and a step activation function.
This is the same as a single neuron in a modern deep learning
network that uses the same linear combination of input signals
from neurons at the previous layer and a more efficient nonlinear
activation function. The perceptron model was further defined by
minimizing the classification error and trained by using one data
point at a time to update the model parameters during the op-
timization process. Modern neural networks are trained similarly
by minimizing a loss function but with more modern optimization
algorithms such as stochastic gradient descent and its variations.
Even though the theoretical foundation of deep learning has been
continually developed in the past few decades, real-world appli-
cations of deep learning are fairly recent due to some real world
constraints: data, network structure, algorithm, and computation
power.
Data
We are all familiar with all sorts of data today: structured tabu-
lated data in database tables or CSV files, free form text, images,
and other unstructured datasets. However, historical datasets are
relatively small in size, especially for data with accurately labeled
ground truth. Statisticians have been working on datasets that
only have a few thousand rows and a few dozen columns for decades
to solve business problems. Even with modern computers, the size

261
262 12 Deep Learning

of the data is usually limited to the memory of a computer. Now

we know, to enable deep learning applications, the algorithm needs
a much larger dataset than traditional machine learning methods.
It is usually at the order of millions of samples with high-quality
ground truth labels for supervised deep learning models.
The first widely used large dataset with accurate labels was the
ImageNet dataset which was first created in 2010. It now con-
tains more than 14 million images with more than 20k synsets
(i.e. meaningful categories). Every image in the dataset was
human-annotated with quality control to ensure the ground truth
labels are accurate. One of the direct results of ImageNet was the
Large Scale Visual Recognition Challenge (ILSVRC) which eval-
uated different algorithms in image-related tasks. The ILSVRC
competition provided a perfect stage for deep learning applica-
tions to debut to the general public. For 2010 and 2011, the best
record of error from traditional image classifications methods was
around 26%. In 2012, a method based on the convolutional neural
network became the state of the art with an error rate of around
16%, a dramatic improvement from the traditional methods.
With the prevalence of the modern internet, the amount of text,
voice, image, and video data increased exponentially. The quality
and quantity of data for deep learning applications enabled the use
of deep learning for applications such as image classification, voice
recognition, and natural language understanding. Data is the fuel
for deep learning engines. With more and more varieties of data
created, captured, and saved, there will be more applications of
deep learning discovered every day.
Network Structure
Lacking high-quality high-volume data was not the only constraint
for early deep learning years. For perceptron with just one single
neuron, it is just a linear classifier. Real applications are nearly
always non-linear. To solve this problem, we have to grow one
neuron to multiple layers with multiple neurons per layer. This
multi layer perceptron (MLP) is also referred to as a feedforward
neural network. In the 1990s, the universal approximation theorem
12.0 263

was proven and it assured us that a feedforward network with a

single hidden layer containing a finite number of neurons can ap-
proximate continuous functions. Even though the one-layer neural
network theoretically can solve a general non-linear problem, the
reality is that we have grown the neural network to many layers of
neurons. The number of layers in the network is the “depth” of a
network. Loosely speaking, deep learning is a neural network with
the many layers (i.e. the depth is deep).
The MLP is the basic structure for the modern deep learning appli-
cations. MLP can be used for classification problems or regression
problems with response variables as the output and a collection of
explanatory variables as the input (i.e. the traditionally structured
datasets). Many of the problems that can be solved using classi-
cal classification methods such as random forest can be solved
by MLP. However, MLP is not the best option for image and
language-related tasks. For image-related tasks, pixels for a local
neighbor region collectively provide useful information to solve a
task. To take advantage of the 2D spatial relationship among pix-
els, the convolutional neural network (CNN) structure is a better
choice. For language-related tasks, the sequence of the text pro-
vides additional information than just a collection of single words.
The recurrent neural network (RNN) is a better structure for such
sequence-related data. There are other more complicated neural
network structures and it is still a fast-developing area. MLP,
CNN, and RNN are just the starting point of deep learning meth-
ods.
Algorithm
In addition to data and neural network structure, there were a few
key algorithm breakthroughs to enable the widespread adoption of
deep learning. For an entry-level neural network structure, there
are hundreds of thousands of parameters to be estimated from
the data. With a large amount of training data, stochastic gradi-
ence decent and mini-batch gradience decent are efficient ways to
utilize a subset of training data to update the model parameters.
Within the process, one of the key steps is back-propagation which
was introduced in the 1980s for efficient weight update. There is a
264 12 Deep Learning

non-linear activation for each neuron in deep learning models, and

sigmoid or hyperbolic tangent functions were often used. How-
ever, it has the problem of gradient vanishing when the number of
layers of the network grows large (i.e. deeper network). To solve
this problem, the rectified linear unit (ReLu) was introduced
to deep learning in the 2000s and it increases the convergence
speed dramatically. ReLu is so simple (i.e. y = x when x >= 0
and y = 0 otherwise), but it indeed cleverly solved one of the big
headaches in deep learning. We will talk more about activation
functions in section 12.2.4.
With hundreds of thousands of parameters in the model, deep
learning is easy to overfit. In order to mitigate this, dropout, a
form of regularization, was introduced in 2012. It randomly drops
out a certain percentage of neurons in the network during the op-
timization process to achieve more robust model performance. It is
similar to the concept of random forest where features and train-
ing data are randomly chosen. There are many other algorithm
improvements to get better models such as batch normalization
and using residuals from previous layers. With backpropagation
in stochastic gradience decent, ReLu activation function, dropout,
and other techniques, modern deep learning methods begin to out-
perform traditional machine learning methods.
Computation Power
With data, network structure and algorithms ready, modern deep
learning still requires a certain amount of computation power for
training. The entire framework involves heavy linear algebra oper-
ations with large matrices and tensors. These types of operations
are much faster on modern graphical processing units (GPUs) than
the computer’s central processor (CPU).
With the vast potential application of deep learning, major tech
companies contribute heavily to open-source deep learning frame-
works. For example, Google has open-sourced its TensorFlow
framework; Facebook has open-sourced its PyTorch framework,
and Amazon has significantly contributed to the MXNet open-
source framework. With thousands of software developers and sci-
12.1 Projection Pursuit Regression 265

entists behind these deep learning frameworks, users can confi-

dently pick one framework and start training their deep learning
models right away in popular cloud environments. Much of the
heavy lifting to train a deep learning model has been embedded in
these open-source frameworks and there are also many pre-trained
models available for users to adopt. Users can now enjoy the rela-
tively easy access to software and hardware to develop their own
deep learning applications. In this book, we will demonstrate deep
learning examples using Keras, a high-level abstraction of Tensor-
Flow, using the Databricks Community Edition platform.
In summary, deep learning has not just developed in the past
few years but have in fact been ongoing research for the past few
decades. The accumulation of data, the advancement of new opti-
mization algorithms and the improvement of computation power
has finally enabled every day deep learning applications. In the
foreseeable future, deep learning will continue to revolutionize ma-
chine learning methods across many more areas.

12.1 Projection Pursuit Regression

Before moving onto neural networks, let us start with a broader
framework, Projection Pursuit Regression (PPR). It has a form
of additive model of the derived features rather than the inputs
themselves. Another widely used algorithm, AdaBoost, also fits an
additive model in a base learner.
Assume XT = (𝑋1 , 𝑋2 , … , 𝑋𝑝 ) is a vector with 𝑝 variables. 𝑌 is
the corresponding response variable. �m , 𝑚 = 1, 2, … , 𝑀 is param-
eter vector with 𝑝 elements.

𝑀
𝑓(X) = ∑ 𝑔𝑚 (�T
m X)
𝑚=1

The new feature Vm = �Tm X is a linear combination of input vari-

ables X. The additive model is based on the new features. Here �m
266 12 Deep Learning

is a unit vector, and the new feature vm is actually the projection

of X on �m . It projects the p-dimensional independent variable
space onto the new M-dimensional feature space. This is similar
to the principal component analysis except that the principal com-
ponent is orthogonal projection but it is not necessarily orthogonal
here.
I know it is very abstract. Let’s look at some examples. Assume √𝑝 =
2, i.e. there are two variables 𝑥1 and√
𝑥2 . If 𝑀 = 1, �T = ( 12 , 23 ),
then the corresponding 𝑣 = 21 𝑥1 + 23 𝑥2 . Let’s try different setings
and compare the results:
√ √
1. �T = ( 12 , 3
2 ), 𝑣 = 21 𝑥1 + 3
2 𝑥2 , 𝑔(𝑣) = 1
1+𝑒−𝑣
1
2. �T = (1, 0), 𝑣 = 𝑥1 , 𝑔(𝑣) = (𝑣 + 5)𝑠𝑖𝑛( 𝑣 +0.1 )
3

𝑣2
3. �T = (0, 1), 𝑣 = 𝑥2 , 𝑔(𝑣) = 𝑒 5

1
4. �T = (1, 0), 𝑣 = 𝑥1 , 𝑔(𝑣) = (𝑣 + 0.1)𝑠𝑖𝑛( 𝑣 +0.1 )
3

Here is how you can simulate the data and plot it using R:

# use plot3D package to generate 3D plot

library(plot3D)

## Warning in system2("/usr/bin/otool", c("-L",

## shQuote(DSO)), stdout = TRUE): running command ''/
## usr/bin/otool' -L '/Library/Frameworks/R.framework/
## Resources/library/tcltk/libs//tcltk.so'' had status 1

# get x1 and x2 note here x1 and x2 need to be matrix if

# you check the two objects, you will find: columns in
# x1 are identical rows in x2 are identical mesh() is
# funciton from plot3D package you may need to think a
# little here
M <- mesh(seq(-13.2, 13.2, length.out = 50), seq(-37.4,
37.4, length.out = 50))
12.1 Projection Pursuit Regression 267

x1 <- M$x
x2 <- M$y
## setting 1 map X using w to get v
v <- (1/2) * x1 + (sqrt(3)/2) * x2
# apply g() on v
g1 <- 1/(1 + exp(-v))
par(mfrow = c(2, 2), mar = c(0, 0, 1, 0))
surf3D(x1, x2, g1, colvar = g1, border = "black", colkey = FALSE,
box = FALSE, main = "Setting 1")
## setting 2
v <- x1
g2 <- (v + 5) * sin(1/(v/3 + 0.1))
surf3D(x1, x2, g2, colvar = g2, border = "black", colkey = FALSE,
box = FALSE, main = "Setting 2")
## setting 3
v <- x2
g3 <- exp(v^2/5)
surf3D(x1, x2, g3, colvar = g3, border = "black", colkey = FALSE,
box = FALSE, main = "Setting 3")
## setting 4
v <- x1
g4 <- (v + 0.1) * sin(1/(v/3 + 0.1))
surf3D(x1, x2, g4, colvar = g4, border = "black", colkey = FALSE,
box = FALSE, main = "Setting 4")
268 12 Deep Learning

Setting 1 Setting 2

Setting 3 Setting 4

You can see that this framework is very flexible. In essence, it is to

do a non-linear transformation of the linear combination. You can
use this way to capture varies of relationships. For example,𝑥1 𝑥2
2 2
can be written as (𝑥1 +𝑥2 ) −(𝑥
4
1 −𝑥2 )
, where 𝑀 = 2. All the higher
order factors of 𝑥1 and 𝑥2 can be represented similarly. If 𝑀 is large
enough, this framework can approximate any continuous function
on ℝ𝑝 . So the model family covers a board area, but with a price.
That is the interpretability. Because the number of parameters
increases with M and the mode is nested.
PPR in 1981 was a new idea then which lead to the debut of
the neural network model. The basic technical idea behind deep
learning has been around for decades. However, why did they take
off in recent years? Here are some main drivers behind the rise.
First, thanks to the digitalization where lots of human activities
are now in the digital realm and captured as data. So in the last 10
year, for many problems, we went from having a relatively small
12.2 Feedforward Neural Network 269

amount of data to accumulating a large amount of data. The tradi-

tional learning algorithms, like Logistic Regression, Support Vec-
tor Machine, Random Forest cannot effectively take advantage of
such big data. Second, the increasing computation power enables
us to train a large neural network either on a CPU or GPU us-
ing big data. The scale of data and computation ability lead to
much progress, but tremendous algorithmic innovation is also an
important driver. Many of the innovations are about speeding up
the optimization of neural network. One of the examples is to use
ReLU as intermediate layer activation function instead of the pre-
vious sigmoid function. The change has made the optimization
process much faster because the previous sigmoid function suf-
fers from vanishing gradient. We will talk more about that in the
following sections. Here we just want to show an example of the
impact of algorithmic innovation.

12.2 Feedforward Neural Network

12.2.1 Logistic Regression as Neural Network
Let’s look at logistic regression from the lens of the neural network.
For a binary classification problem, for example spam classifier,
given 𝑚 samples {(𝑥(1) , 𝑦(1) ), (𝑥(2) , 𝑦(2) ), ..., (𝑥(𝑚) , 𝑦(𝑚) )}, we need
to use the input feature 𝑥(𝑖) (they may be the frequency of various
words such as “money”, special characters like dollar signs, and
the use of capital letters in the message etc.) to predict the output
𝑦(𝑖) (if it is a spam email). Assume that for each sample 𝑖, there
are 𝑛𝑥 input features. Then we have:

(1) (2) (𝑚)

𝑥 𝑥 ⋯ 𝑥1
⎡ 1(1) 1(2) (𝑚) ⎤
𝑥2 𝑥2 ⋯ 𝑥2
𝑋=⎢
⎢ ⋮
⎥ ∈ ℝ𝑛𝑥 ×𝑚
⎥ (12.1)
⎢ ⋮ ⋮ ⋮ ⎥
(1) (2) (𝑚)
⎣ 𝑥𝑛𝑥 𝑥𝑛𝑥 … 𝑥𝑛𝑥 ⎦

𝑦 = [𝑦(1) , 𝑦(2) , … , 𝑦(𝑚) ] ∈ ℝ1×𝑚

270 12 Deep Learning

To predict if sample 𝑖 is a spam email, we first get the inactivated

neuro 𝑧 (𝑖) by a linear transformation of the input 𝑥(𝑖) , which is
𝑧 (𝑖) = 𝑤𝑇 𝑥(𝑖) +𝑏. Then we apply a function to “activate” the neuro
𝑧 (𝑖) and we call it “activation function”. In logistic regression, the
activation function is sigmoid function and the “activated” 𝑧 (𝑖) is
the prediction:

𝑦(𝑖)
̂ = 𝜎(𝑤𝑇 𝑥(𝑖) + 𝑏)
1
where 𝜎(𝑧) = 1+𝑒−𝑧 . The following figure summarizes the process:

There are two types of layers. The last layer connects directly to
the output. All the rest are intermediate layers. Depending on your
definition, we call it “0-layer neural network” where the layer count
only considers intermediate layers. To train the model, you need
a cost function which is defined as equation (12.2).

1 𝑚
𝐽 (𝑤, 𝑏) = Σ 𝐿(𝑦(𝑖)
̂ , 𝑦(𝑖) ) (12.2)
𝑚 𝑖=1
where

𝐿(𝑦(𝑖)
̂ , 𝑦(𝑖) ) = −𝑦(𝑖) 𝑙𝑜𝑔(𝑦(𝑖)
̂ ) − (1 − 𝑦(𝑖) )𝑙𝑜𝑔(1 − 𝑦(𝑖)
̂ )

To fit the model is to minimize the cost function.

12.2 Feedforward Neural Network 271

12.2.2 Stochastic Gradient Descent

The general approach to minimize 𝐽 (𝑤, 𝑏) is by gradient descent,
also known as back-propagation. The optimization process is a for-
ward and backward sweep over the network.

The forward propagation takes the current weights, calculates the

prediction and cost. The backward propagation computes the gra-
dient descent for the parameters by the chain rule for differentia-
tion. In logistic regression, it is easy to calculate the gradient w.r.t
the parameters (𝑤, 𝑏).
Let’s look at the Stochastic Gradient Descent (SGD) for logistic
regression across 𝑚 samples. SGD updates one sample each time.
The detailed process is as follows.
272 12 Deep Learning

First initialize 𝑤1 , 𝑤2 , … , 𝑤𝑛𝑥 , and 𝑏. Then plug in the initial-

ized value to the forward and backward propagation. The forward
propagation takes the current weights and calculates the predic-
tion ℎ̂ (𝑖) and cost 𝐽 (𝑖) . The backward propagation calculates the
gradient descent for the parameters. After iterating through all 𝑚
samples, you can calculate gradient descent for the parameters.
Then update the parameter by:
𝜕𝐽
𝑤 ∶= 𝑤 − 𝛾
𝜕𝑤
𝜕𝐽
𝑏 ∶= 𝑏 − 𝛾
𝜕𝑏
Repeat the forward and backward process using the updated pa-
rameter until the cost 𝐽 stabilizes.

12.2.3 Deep Neural Network

Before people coined the term deep learning, a neural network
refers to single hidden layer network. Neural networks with more
than one layers are called deep learning. Network with the struc-
12.2 Feedforward Neural Network 273

ture in figure 12.1 is the multiple layer perceptron (MLP) or

feedforward neural network (FFNN).

FIGURE 12.1: Feedforward Neural Network

Let’s look at a simple one-hidden-layer neural network (figure

12.2). First only consider one sample. From left to right, there
is an input layer with 3 features (𝑥1 , 𝑥2 , 𝑥3 ), a hidden layer with
four neurons and an output later to produce a prediction 𝑦.̂

FIGURE 12.2: 1-layer Neural Network

274 12 Deep Learning

From input to the first hidden layer

Each inactivated neuron on the first layer is a linear transforma-
[1] [1]𝑇 [1]
tion of the input vector 𝑥. For example, 𝑧1 = 𝑤1 𝑥(𝑖) + 𝑏1 is
the first inactivated neuron for hidden layer one. We use super-
script [l] to denote a quantity associated with the 𝑙𝑡ℎ layer
and the subscript i to denote the 𝑖𝑡ℎ entry of a vector (a
[1]
neuron or feature). Here 𝑤[1] and 𝑏1 are the weight and bias
[1]𝑇
parameters for layer 1. 𝑤[1] is a 4 × 1 vector and hence 𝑤1 𝑥(𝑖) us
a linear combination of the four input features. Then use a sigmoid
[1] [1]
function 𝜎(⋅) to activate the neuron 𝑧1 to get 𝑎1 .
From the first hidden layer to the output
Next, do a linear combination of the activated neurons from the
[2]
first layer to get inactivated output, 𝑧1 . And then activate the
neuron to get the predicted output 𝑦.̂ The parameters to estimate
[2]
in this step are 𝑤[2] and 𝑏1 .
If you fully write out the process, it is the bottom right of figure
12.2. When you implement a neural network, you need to do sim-
ilar calculation four times to get the activated neurons in the first
hidden layer. Doing this with a for loop is inefficient. So people
vectorize the four equations. Take an input and compute the cor-
responding 𝑧 and 𝑎 as a vector. You can vectorize each step and
get the following representation:

𝑧 [1] = 𝑊 [1] 𝑥 + 𝑏[1] 𝜎[1] (𝑧 [1] ) = 𝑎[1]

𝑧 [2] = 𝑊 [2] 𝑎[1] + 𝑏[2] 𝜎[2] (𝑧 [2] ) = 𝑎[2] = 𝑦 ̂

𝑏[1] is the column vector of the four bias parameters shown above.
𝑧 [1] is a column vector of the four non-active neurons. When you
apply an activation function to a matrix or vector, you apply it
element-wise. 𝑊 [1] is the matrix by stacking the four row-vectors:
12.2 Feedforward Neural Network 275

[1]𝑇
𝑤
⎡ 1[1]𝑇 ⎤
𝑤2
𝑊 [1]
=⎢
⎢ [1]𝑇
⎥
⎥
⎢ 𝑤3 ⎥
[1]𝑇
⎣ 𝑤4 ⎦

So if you have one sample, you can go through the above forward
propagation process to calculate the output 𝑦 ̂ for that sample.
If you have 𝑚 training samples, you need to repeat this process
each of the 𝑚 samples. We use superscript (i) to denote a
quantity associated with 𝑖𝑡ℎ sample. You need to do the same
calculation for all 𝑚 samples.
For i = 1 to m, do:

𝑧 [1](𝑖) = 𝑊 [1] 𝑥(𝑖) + 𝑏[1] 𝜎[1] (𝑧 [1](𝑖) ) = 𝑎[1](𝑖)

𝑧 [2](𝑖) = 𝑊 [2] 𝑎[1](𝑖) + 𝑏[2] 𝜎[2] (𝑧 [2](𝑖) ) = 𝑎[2](𝑖) = 𝑦(𝑖)
̂

Recall that we defined the matrix X to be equal to our training

samples stacked up as column vectors in equation (12.1). We do
a similar thing here to stack vectors with the superscript (i) to-
gether across 𝑚 samples. This way, the neural network computes
the outputs on all the samples on at the same time:

𝑍 [1] = 𝑊 [1] 𝑋 + 𝑏[1] 𝜎[1] (𝑍 [1] ) = 𝐴[1]

𝑍 [2] = 𝑊 [2] 𝐴[1] + 𝑏[2] 𝜎[2] (𝑍 [2] ) = 𝐴[2] = 𝑌 ̂

where
| | |
⎡
𝑋=⎢ 𝑥 (1) (1) (𝑚) ⎤
𝑥 ⋯ 𝑥 ⎥,
⎣ | | | ⎦
| | |
⎡
[𝑙]
𝐴 =⎢ 𝑎 [𝑙](1)
𝑎 [𝑙](1)
⋯ 𝑎 [𝑙](𝑚) ⎤
,
⎥
⎣ | | | ⎦𝑙=1 𝑜𝑟 2

| | |
𝑍 [𝑙] = ⎡
⎢ 𝑧 [𝑙](1)
𝑧 [𝑙](1)
⋯ 𝑧 [𝑙](𝑚) ⎤
⎥
⎣ | | | ⎦
𝑙=1 𝑜𝑟 2
 276 12 Deep Learning

You can add layers like this to get a deeper neural network as
shown in the bottom right of figure 12.1.

12.2.4 Activation Function

• Sigmoid and Softmax Function
We have used the sigmoid (or logistic) activation function. The
function is S-shape with an output value between 0 to 1. Therefore
it is used as the output layer activation function to predict the
probability when the response 𝑦 is binary. However, it is rarely
used as an intermediate layer activation function. One of the main
reasons is that when 𝑧 is away from 0, then the derivative of the
function drops fast which slows down the optimization process
through gradient descent. Even the fact that it is differentiable
provides some convenience, the decreasing slope can cause a neural
network to get stuck at the training time.
1.0

1
0.8

σ(z) =
1 + e−z
0.6
σ(z)

0.4
0.2
0.0

-5 0 5

FIGURE 12.3: Sigmoid Function

When the output has more than 2 categories, people use softmax
function as the output layer activation function.

𝑒𝑧𝑖
𝑓𝑖 (z) = (12.3)
Σ𝐽𝑗=1 𝑒𝑧𝑗

where z is a vector.
 12.2 Feedforward Neural Network 277

• Hyperbolic Tangent Function (tanh)

Another activation function with a similar S-shape is the hyper-
bolic tangent function. It often works better than the sigmoid func-
tion as the intermediate layer.

𝑒𝑧 − 𝑒−𝑧
𝑡𝑎𝑛ℎ(𝑧) = (12.4)
𝑒𝑧 + 𝑒−𝑧
1.0

ez − e−z
0.5

tanh(z) =
ez + e−z
0.0
-0.5
-1.0

-5 0 5

FIGURE 12.4: Hyperbolic Tangent Function

The tanh function crosses point (0, 0) and the value of the function
is between 1 and -1 which makes the mean of the activated neurons
closer to 0. The sigmoid function doesn’t have that property. When
you preprocess the training input data, you sometimes center the
data so that the mean is 0. The tanh function is doing that data
processing in some way which makes learning for the next layer a
little easier. This activation function is used a lot in the recurrent
neural networks where you want to polarize the results.
• Rectified Linear Unit (ReLU) Function
The most popular activation function is the Rectified Linear Unit
(ReLU) function. It is a piecewise function, or a half rectified func-
tion:
278 12 Deep Learning

𝑅(𝑧) = 𝑚𝑎𝑥(0, 𝑧) (12.5)

The derivative is 1 when z is positive and 0 when z is negative.

You can define the derivative as either 0 or 1 when z is 0. When
you implement this, it is unlikely that z equals to exactly 0 even
it can be very close to 0.
8

R(z) = max(0, z)
6
4
2
0

-5 0 5

FIGURE 12.5: Rectified Linear Unit Function

The advantage of the ReLU is that when z is positive, the deriva-

tive doesn’t vanish as z getting larger. So it leads to faster compu-
tation than sigmoid or tanh. It is non-linear with an unconstrained
response. However, the disadvantage is that when z is negative, the
derivative is 0. It may not map the negative values appropriately.
In practice, this doesn’t cause too much trouble but there is an-
other version of ReLu called leaky ReLu that attempts to solve
the dying ReLU problem. The leaky ReLu is

𝑧 𝑧≥0
𝑅(𝑧)𝐿𝑒𝑎𝑘𝑦 = {
𝑎𝑧 𝑧<0

Instead of being 0 when z is negative, it adds a slight slope such

as 𝑎 = 0.01 as shown in figure 12.6 (can you see the leaky part
there? : ).
 12.2 Feedforward Neural Network 279

8
6
4
2
0

-5 0 5

FIGURE 12.6: Rectified Linear Unit Function

You may notice that all activation functions are non-linear. Since
the composition of two linear functions is still linear, using a lin-
ear activation function doesn’t help to capture more information.
That is why you don’t see people use a linear activation function
in the intermediate layer. One exception is when the output 𝑦 is
continuous, you may use linear activation function at the output
layer. To sum up, for intermediate layers:
• ReLU is usually a good choice. If you don’t know what to choose,
then start with ReLU. Leaky ReLu usually works better than the
ReLU but it is not used as much in practice. Either one works
fine. Also, people usually use a = 0.01 as the slope for leaky
ReLU. You can try different parameters but most of the people
a = 0.01.
• tanh is used sometimes especially in recurrent neural network.
But you nearly never see people use sigmoid function as inter-
mediate layer activation function.
For the output layer:
• When it is binary classification, use sigmoid with binary cross-
entropy as loss function
 280 12 Deep Learning

• When there are multiple classes, use softmax function with cat-
egorical cross-entropy as loss function
• When the response is continuous, use identity function (i.e. y =
x)

12.2.5 Optimization
So far, we have introduced the core components of deep learning
architecture, layer, weight, activation function, and loss function.
With the architecture in place, we need to determine how to update
the network based on a loss function (a.k.a. objective function).
In this section, we will look at variants of optimization algorithms
that will improve the training speed.
12.2.5.1 Batch, Mini-batch, Stochastic Gradient Descent

Stochastic Gradient Descent (SGD) updates one sample each

time. We showed the SGD for logistic regression across 𝑚 samples
in section 12.2.1. If you process the entire training set each time to
calculate the gradient, it is Batch Gradient Descent (BGD).
The vector representation section 12.2.3 using all 𝑚 is an example
of BGD.
In deep learning applications, the training set is often huge, with
hundreds of thousands or millions of samples. If processing all the
samples can only lead to one step of gradient descent, it could be
slow. An intuitive way to improve the algorithm is to make some
progress before finishing the entire data set. In particular, split
up the training set into smaller subsets and fit the model using
one subset each time. The subsets are mini-batches. Mini-batch
Gradient Descent (MGD) is to split the training set to be
smaller mini-batches. For example, if the mini-batch size is 1000,
the algorithm will process 1000 samples each time, calculate the
gradient and update the parameters. Then it moves on to the next
mini-batch set until it goes through the whole training set. We call
it one pass through training set using mini-batch gradient descent
or one epoch. Epoch is a common keyword in deep learning, which
means a single pass through the training set. If the training set has
60,000 samples, one epoch leads to 60 gradient descent steps. Then
 12.2 Feedforward Neural Network 281

it will start over and take another pass through the training set. It
means one more decision to make, the optimal number of epochs.
It is decided by looking at the trends of performance metrics on a
holdout set of training data. We discussed the data splitting and
sampling in section 7.2. People often use a single holdout set to
tune the model in deep learning. But it is important to use a big
enough holdout set to give high confidence in your model’s overall
performance. Then you can evaluate the chosen model on your test
set that is not used in the training process. MGD is what everyone
in deep learning uses when training on a large data set.

𝑥= [𝑥 (1) , 𝑥(2) , ⋯ , 𝑥(1000) / ⋯ / ⋯ 𝑥(𝑚) ]

⏟⏟ ⏟⏟⏟⏟⏟⏟⏟
(𝑛𝑥 , 𝑚) 𝑚𝑖𝑛𝑖 − 𝑏𝑎𝑡𝑐ℎ 1
(1) , 𝑦 (2) , ⋯ , 𝑦 (1000) / ⋯ / ⋯ 𝑦 (𝑚) ]
𝑦 = [𝑦⏟⏟ ⏟⏟⏟⏟⏟⏟⏟
(1, 𝑚) 𝑚𝑖𝑛𝑖 − 𝑏𝑎𝑡𝑐ℎ 1
• Mini-batch size = m: batch gradient descent, too long per itera-
tion
• Mini-batch size = 1: stochastic gradient descent, lose speed from
vectorization
• Mini-batch size in between: mini-batch gradient descent, make
progress without processing all training set, typical batch sizes
are 26 = 64, 27 = 128, 27 = 256, 28 = 512
12.2.5.2 Optimization Algorithms

In the history of deep learning, researchers proposed different opti-

mization algorithms and showed that they worked well in a specific
scenario. But the optimization algorithms didn’t generalize well to
a wide range of neural networks. So you will need to try different
optimizers in your application. We will introduce three commonly
used optimizers here, and they are all based on something called
exponentially weighted averages. To understand the intuition be-
hind it, let’s look at a hypothetical example shown in figure 12.7.
We have two parameters, b and w. The blue dot represents the
current parameter value, and the red point is the optimum value
we want to reach. The current value is close to the target verti-
282 12 Deep Learning

FIGURE 12.7: The intuition behind the weighted average gra-

dient

cally but far away horizontally. In this situation, we hope to have

slower learning on b and faster learning on w. A way to adjust is to
use the average of the gradients from different iterations instead of
the current iteration to update the parameter. Vertically, since the
current value is close to the target value of parameter b, the gra-
dient of b is likely to jump between positive and negative values.
The average tends to cancel out the positive and negative deriva-
tives along the vertical direction and slow down the oscillations.
Horizontally, since it is still further away from the target value of
parameter w, all the gradients are likely pointing in the same di-
rection. Using an average won’t have too much impact there. That
is the fundamental idea behind many different optimizers: adjust
the learning rate using a weighted average of various iterations’
gradients.
Exponentially Weighted Averages
Before we get to more sophisticated optimization algorithms that
implement this idea, let’s first introduce the basic weighted moving
average framework.
Suppose we have the following 100 days’ temperature data:

𝜃1 = 49𝐹 , 𝜃2 = 53𝐹 , … , 𝜃99 = 70𝐹 , 𝜃100 = 69𝐹

12.2 Feedforward Neural Network 283

The weighted average is defined as:

𝑉𝑡 = 𝛽𝑉𝑡−1 + (1 − 𝛽)𝜃𝑡
And we have:

𝑉0 = 0
𝑉1 = 𝛽𝑉1 + (1 − 𝛽)𝜃1
𝑉2 = 𝛽𝑉1 + (1 − 𝛽)𝜃2
⋮
𝑉100 = 𝛽𝑉99 + (1 − 𝛽)𝜃100

For example, for 𝛽 = 0.95:

𝑉0 = 0
𝑉1 = 0.05𝜃1 ......
𝑉2 = 0.0475𝜃1 + 0.05𝜃2

The black line in the right plot of figure 12.8 is the exponentially
weighted averages of simulated temperature data with 𝛽 = 0.95.
1
𝑉𝑡 is approximately average over the previous 1−𝛽 days. So 𝛽 =
0.95 approximates a 20 days’ average. The red line corresponds to
𝛽 = 0.8, which approximates 5 days’ average. As 𝛽 increases, it
averages over a larger window of the previous values, and hence the
curve gets smoother. A larger 𝛽 also means that it gives the current
value 𝜃𝑡 less weight (1 − 𝛽), and the average adapts more slowly.
It is easy to see from the plot that the averages at the beginning
are more biased. The bias correction can help to achieve a better
estimate:

𝑉𝑡
𝑉𝑡𝑐𝑜𝑟𝑟𝑒𝑐𝑡𝑒𝑑 =
1 − 𝛽𝑡

𝑉1
𝑉1𝑐𝑜𝑟𝑟𝑒𝑐𝑡𝑒𝑑 = = 𝜃1
1 − 0.95

𝑉2
𝑉2𝑐𝑜𝑟𝑟𝑒𝑐𝑡𝑒𝑑 = = 0.4872𝜃1 + 0.5128𝜃2
1 − 0.952
284 12 Deep Learning

For 𝛽 = 0.95, the origional 𝑉2 = 0.0475𝜃1 + 0.05𝜃2 which is a

small fraction of both 𝑡ℎ𝑒𝑡𝑎1 and 𝑡ℎ𝑒𝑡𝑎2 . That is why it starts so
much lower with big bias. After correction, 𝑉2𝑐𝑜𝑟𝑟𝑒𝑐𝑡𝑒𝑑 = 0.4872𝜃1 +
0.5128𝜃2 is a weighted average with two weights added up to 1
which revmoves the bias. Notice that as 𝑡 increases, 𝛽 𝑡 converges
to 0 and 𝑉 𝑐𝑜𝑟𝑟𝑒𝑐𝑡𝑒𝑑 converges to 𝑉 𝑡 .
The following code simulates a set of temperature data and applies
exponentially weighted average with and without correction using
various 𝛽 values (0.5, 0.8, 0.95).
Without Correction With Correction
100

100
80

80
Tempreture

60

60
40

40
20

20

beta1=0.95 beta1=0.95
beta2=0.8 beta2=0.8
beta3=0.5 beta3=0.5
0

0 20 40 60 80 100 0 20 40 60 80 100

Days Days

FIGURE 12.8: Exponentially weighted averages with and with-

out corrrection

How do we apply the exponentially weighted average to optimiza-

tion? Instead of using the gradients (𝑑𝑤 and 𝑑𝑏) to update the
parameter, we use the gradients’ exponentially weighted average.
There are various optimizers built on top of this idea. We will look
at three optimizers, Momentum, Root Mean Square Propagation
(RMSprop), and Adaptive Moment Estimation (Adam).
Momentum
12.2 Feedforward Neural Network 285

The momentum algorithm uses the exponentially weighted average

of gradients to update the parameters. On iteration t, compute dw,
db using samples in one mini-batch and update the parameters as
follows:

𝑉𝑑𝑤 = 𝛽𝑉𝑑𝑤 + (1 − 𝛽)𝑑𝑤

𝑉𝑑𝑏 = 𝛽𝑉𝑑𝑏 + (1 − 𝛽)𝑑𝑏

𝑤 = 𝑤 − 𝛼𝑉𝑑𝑤 ; 𝑏 = 𝑏 − 𝛼𝑉𝑑𝑏

The learning rate 𝛼 and weighted average parameter 𝛽 are hy-

perparameters. The most common and robust choice is 𝛽 = 0.9.
This algorithm does not use the bias correction because the aver-
age will warm up after a dozen iterations and no longer be biased.
The momentum algorithm, in general, works better than the orig-
inal gradient descent without any average.
Root Mean Square Propagation (RMSprop)
The Root Mean Square Propagation (RMSprop) is another al-
gorithm that applies the idea of exponentially weighted average.
On iteration t, compute dw and db using the current mini-batch.
Instead of V, it calculates the weighted average of the squared
gradient descents. When dw and db are vectors, the squaring is an
element-wise operation.

𝑆𝑑𝑤 = 𝛽𝑆𝑑𝑤 + (1 − 𝛽)𝑑𝑤2

𝑆𝑑𝑏 = 𝛽𝑆𝑑𝑏 + (1 − 𝛽)𝑑𝑏2

Then, update the parameters as follows:

𝑑𝑤 𝑑𝑏
𝑤=𝑤−𝛼 ; 𝑏 =𝑏−𝛼
√𝑆𝑑𝑤 √𝑆𝑑𝑏

The RMSprop algorithm divides the learning rate for a parameter

286 12 Deep Learning

by a weighted average of recent gradients’ magnitudes for that

parameter. The goal is still to adjust the learning speed. Recall the
example that illustrates the intuition behind it. When parameter
b is close to its target value, we want to decrease the oscillations
along the vertical direction.
Adaptive Moment Estimation (Adam)
The Adaptive Moment Estimation (Adam) algorithm is, in some
way, a combination of momentum and RMSprop. On iteration t,
compute dw, db using the current mini-batch. Then calculate both
V and S using the gradient descents.

𝑉𝑑𝑤 = 𝛽1 𝑉𝑑𝑤 + (1 − 𝛽1 )𝑑𝑤

{ 𝑚𝑜𝑚𝑎𝑛𝑡𝑢𝑚 𝑢𝑝𝑑𝑎𝑡𝑒 𝛽1
𝑉𝑑𝑏 = 𝛽1 𝑉𝑑𝑏 + (1 − 𝛽1 )𝑑𝑏

𝑆𝑑𝑤 = 𝛽2 𝑆𝑑𝑤 + (1 − 𝛽2 )𝑑𝑤2

{ 𝑅𝑀 𝑆𝑝𝑟𝑜𝑝 𝑢𝑝𝑑𝑎𝑡𝑒 𝛽2
𝑆𝑑𝑏 = 𝛽2 𝑆𝑑𝑏 + (1 − 𝛽2 )𝑑𝑏2

The Adam algorithm implements bias correction.

𝑐𝑜𝑟𝑟𝑒𝑐𝑡𝑒𝑑 𝑉𝑑𝑤 𝑐𝑜𝑟𝑟𝑒𝑐𝑡𝑒𝑑 𝑆𝑑𝑤

𝑉𝑑𝑤 = 1−𝛽1𝑡
𝑆𝑑𝑤 = 1−𝛽2𝑡
{ 𝑐𝑜𝑟𝑟𝑒𝑐𝑡𝑒𝑑 𝑉𝑑𝑏 ; { 𝑐𝑜𝑟𝑟𝑒𝑐𝑡𝑒𝑑 𝑆𝑑𝑏
𝑉𝑑𝑏 = 1−𝛽1𝑡
𝑆𝑑𝑏 = 1−𝛽2𝑡

And it updates the parameter using both corrected V and S. 𝜖

here is a tiny positive number to make sure the denominator is
larger than zero. The choice of 𝜖 doesn’t matter much, and the
inventors of the Adam algorithm recommended 𝜖 = 10−8 . For hy-
perparameter 𝛽1 and 𝛽2 , the common settings are 𝛽1 = 0.9 and
𝛽2 = 0.999.
𝑐𝑜𝑟𝑟𝑒𝑐𝑡𝑒𝑑 𝑐𝑜𝑟𝑟𝑒𝑐𝑡𝑒𝑑
𝑉𝑑𝑤 𝑉𝑑𝑏
𝑤=𝑤−𝛼 ; 𝑏 =𝑏−𝛼
𝑐𝑜𝑟𝑟𝑒𝑐𝑡𝑒𝑑
√𝑆𝑑𝑤 𝑐𝑜𝑟𝑟𝑒𝑐𝑡𝑒𝑑
√𝑆𝑑𝑏
+𝜖 +𝜖
12.2 Feedforward Neural Network 287

12.2.6 Deal with Overfitting

The biggest problem for deep learning is overfitting. It happens
when the model learns too much from the data. We discussed
this in more detail in section 7.1. A common way to diagnose the
problem is to use cross-validation (section 7.2). You can recognize
the problem when the model fits great on the training data but
gives poor predictions on the testing data. One way to prevent the
model from over learning the data is to limit model complexity.
There are several approaches to that.
12.2.6.1 Regularization

For logistic regression,

1 𝑚
𝑚𝑖𝑛𝐽 (𝑤, 𝑏) = Σ 𝐿(𝑦(𝑖)
̂ , 𝑦(𝑖) ) + 𝑝𝑒𝑛𝑎𝑙𝑡𝑦
𝑤,𝑏 𝑚 𝑖=1

Common penalties are L1 or L2 as follows:

𝜆 𝜆 𝑛𝑥 2
𝐿2 𝑝𝑒𝑛𝑎𝑙𝑡𝑦 = ∥ 𝑤 ∥22 = Σ 𝑤
2𝑚 2𝑚 𝑖=1 𝑖

𝜆 𝑛𝑥
𝐿1 𝑝𝑒𝑛𝑎𝑙𝑡𝑦 = Σ |𝑤|
𝑚 𝑖=1
For neural network,

1 𝑚 𝜆
𝐽 (𝑤[1] , 𝑏[1] , … , 𝑤[𝐿] , 𝑏[𝐿] ) = Σ𝑖=1 𝐿(𝑦(𝑖)
̂ , 𝑦(𝑖) ) + Σ𝐿 ∥ 𝑤[𝑙] ∥2𝐹
𝑚 2 𝑙=1

where
[𝑙] [𝑙−1] [𝑙]
∥ 𝑤[𝑙] ∥2𝐹 = Σ𝑛𝑖=1 Σ𝑛𝑗=1 (𝑤𝑖𝑗 )2

Many people call it “Frobenius Norm” instead of L2-norm.

12.2.6.2 Dropout

Another powerful regularization technique is “dropout.” In chapter

@ref(tree model), we mentioned that the random forest model de-
288 12 Deep Learning

correlates the trees by randomly choosing a subset of features.

Dropout uses a similar idea in the parameter estimation process.
It temporally freezes a randomly selected subset of nodes at a spe-
cific layer in the neural network during the optimization process
to reduce overfitting. When applying dropout to a particular layer
and mini-batch, we pre-set a percentage, for example, 30%, and
randomly remove 30% of the layer’s nodes. The output from the
30% removed nodes will be zero. During backpropagation, we up-
date the remaining parameters for a much-diminished network. We
need to do the random drop out again for the next mini-batch.
To normalize the output with all nodes in this layer, we need to
scale up the output accordingly to the same percentage to make
sure the dropped-out nodes do not impact the overall signal. Please
note, the dropout process will randomly turn-off different nodes for
each mini-batch. Dropout is more efficient in reducing overfitting
in deep learning than L1 or L2 regularizations.

12.2.7 Image Recognition Using FFNN

In this section, we will walk through a toy example of image clas-
sification problem using keras package. We use R in the section to
illustrate the process and also provide the python notebook on the
book website. Please check the keras R package website1 for the
most recent development.
What is an image as data? You can consider a digital image as
a set of points on 2-d or 3-d space. Each each point has a value
between 0 to 255 which is considered as a pixel. Figure 12.9 shows
an example of grayscale image. It is a set of pixels on 2-d space
and each pixel has a value between 0 to 255. You can process
the image as a 2-d array input if you use a Convolutional Neural
Network(CNN). Or, you can vectorize the array as the input for
FFNN as shown in the figure.
A color image is a set of pixels on 3-d space and each pixel has a
value between 0 to 255. There are three 2-d panels which repre-
1
https://keras.rstudio.com/
12.2 Feedforward Neural Network 289

FIGURE 12.9: Grayscale image is a set of pixels on 2-d space.

Each pixel has a value range from 0 to 255.

sent the color red, blue and green accordingly. Similarly, You can
process the image as a 3-d array. Or you can vectorize the array
as shown in figure 12.10.

FIGURE 12.10: Color image is a set of pixels on 3-d space. Each

pixel has a value range from 0 to 255.

Let’s look at how to use the keras R package for a toy exam-
 290 12 Deep Learning

ple in deep learning with the handwritten digits image dataset

(i.e. MNIST). keras has many dependent packages, so it takes a
few minutes to install. Be patient! In a production cloud environ-
ment such as the paid version of Databricks, you can save what
you have and resume from where you left.

install.packages("keras")

As keras is just an interface to popular deep learning frameworks,

we have to install the deep learning backend. The default and
recommended backend is TensorFlow. By calling install_keras(),
it will install all the needed dependencies for TensorFlow.

library(keras)
install_keras()

You can run the code in this section in the Databrick community
edition with R as the interface. Refer to section 4.3 for how to
set up an account, create a notebook (R or Python) and start
a cluster. For an audience with a statistical background, using a
well-managed cloud environment has the following benefit:
• Minimum language barrier in coding for most statisticians
• Zero setups to save time using the cloud environment
• Get familiar with the current trend of cloud computing in the
industrial context
You can also run the code on your local machine with R and the re-
quired Python packages (keras uses the Python TensorFlow back-
end engine). Different versions of Python may cause some errors
when running install_keras(). Here are the things you could do
when you encounter the Python backend issue in your local ma-
chine:
• Run reticulate::py_config() to check the current Python config-
uration to see if anything needs to be changed.
• By default, install_keras() uses virtual environment
~/.virtualenvs/r-reticulate. If you don’t know how to set
 12.2 Feedforward Neural Network 291

the right environment, try to set the installation method as

conda (install_keras(method = "conda"))
• Refer to this document for more details on how to install keras
and the TensorFlow backend2 .
Now we are all set to explore deep learning! As simple as three
lines of R code, but there are quite a lot going on behind the
scene. If you are using a cloud environment, you do not need to
worry about these behind scene setup and maintenance.
We will use the widely used MNIST handwritten digit image
dataset. More information about the dataset and benchmark
results from various machine learning methods can be found
at http://yann.lecun.com/exdb/mnist/ and https://en.wikipedia.org/
wiki/MNIST_database.

This dataset is already included in the keras/TensorFlow instal-

lation and we can simply load the dataset as described in the
following cell. It takes less than a minute to load the dataset.

mnist <- dataset_mnist()

The data structure of the MNIST dataset is straight forward and

well prepared for R, which has two pieces:

(1) training set: x (i.e. features): 60000x28x28 tensor which

corresponds to 60000 28x28 pixel greyscale images (i.e. all
the values are integers between 0 and 255 in each 28x28
matrix), and y (i.e. responses): a length 60000 vector
which contains the corresponding digits with integer val-
ues between 0 and 9.
(2) testing set: same as the training set, but with only 10000
images and responses. The detailed structure for the
dataset can be seen with str(mnist) below.

2
https://tensorflow.rstudio.com/reference/keras/install_keras/
292 12 Deep Learning

str(mnist)

List of 2
$ train:List of 2
..$ x: int [1:60000, 1:28, 1:28] 0 0 0 0 0 0 0 0 0 0 ...
..$ y: int [1:60000(1d)] 5 0 4 1 9 2 1 3 1 4 ...
$ test :List of 2
..$ x: int [1:10000, 1:28, 1:28] 0 0 0 0 0 0 0 0 0 0 ...
..$ y: int [1:10000(1d)] 7 2 1 0 4 1 4 9 5 9 ...

Now we prepare the features (x) and the response variable (y)
for both the training and testing dataset, and we can check the
structure of the x_train and y_train using str() function.

x_train <- mnist$train$x

y_train <- mnist$train$y
x_test <- mnist$test$x
y_test <- mnist$test$y

str(x_train)
str(y_train)

int [1:60000, 1:28, 1:28] 0 0 0 0 0 0 0 0 0 0 ...

int [1:60000(1d)] 5 0 4 1 9 2 1 3 1 4 ...

Now let’s plot a chosen 28x28 matrix as an image using R’s image()
function. In R’s image() function, the way of showing an image
is rotated 90 degree from the matrix representation. So there is
additional steps to rearrange the matrix such that we can use
image() function to show it in the actual orientation.

index_image = 28 ## change this index to see different image.

input_matrix <- x_train[index_image, 1:28, 1:28]
output_matrix <- apply(input_matrix, 2, rev)
12.2 Feedforward Neural Network 293

output_matrix <- t(output_matrix)

image(1:28, 1:28, output_matrix, col = gray.colors(256),
xlab = paste("Image for digit of: ", y_train[index_image]),
ylab = "")

Here is the original 28x28 matrix for the above image:

input_matrix

[,1] [,2] [,3] [,4] [,5] [,6] [,7] [,8] [,9] [,10] [,11] [,12] [,13]
[1,] 0 0 0 0 0 0 0 0 0 0 0 0 0
[2,] 0 0 0 0 0 0 0 0 0 0 0 0 0
[3,] 0 0 0 0 0 0 0 0 0 0 0 0 0
[4,] 0 0 0 0 0 0 0 0 0 0 0 0 0
[5,] 0 0 0 0 0 0 0 0 0 0 0 0 0
[6,] 0 0 0 0 0 0 0 0 0 0 9 80 207
[7,] 0 0 0 0 0 0 39 158 158 158 168 253 253
[8,] 0 0 0 0 0 0 226 253 253 253 253 253 253
[9,] 0 0 0 0 0 0 139 253 253 253 238 113 215
294 12 Deep Learning

[10,] 0 0 0 0 0 0 39 34 34 34 30 0 31
[11,] 0 0 0 0 0 0 91 0 0 0 0 0 0
[12,] 0 0 0 0 0 0 0 0 0 0 0 11 33
[13,] 0 0 0 0 0 0 0 0 0 0 11 167 253
[14,] 0 0 0 0 0 0 0 0 0 0 27 253 253
[15,] 0 0 0 0 0 0 0 0 0 0 18 201 253
[16,] 0 0 0 0 0 0 0 0 0 0 0 36 87
...

There are multiple deep learning methods to solve the handwritten

digits problem and we will start from the simple and generic one,
feedforward neural network (FFNN). FFNN contains a few fully
connected layers and information is flowing from a front layer to
a back layer without any feedback loop from the back layer to the
front layer. It is the most common deep learning models to start
with.

12.2.7.1 Data preprocessing

In this section, we will walk through the needed steps of data

preprocessing. For the MNIST dataset that we just loaded, some
preprocessing is already done. So we have a relatively “clean” data,
but before we feed the data into FFNN, we still need to do some
additional preparations.
First, for each digits, we have a scalar response and a 28x28 integer
matrix with value between 0 and 255. To use the out of box DNN
functions, for each response, all the features are one row of all
features. For an image in MNIST dataet, the input for one response
y is a 28x28 matrix, not a single row of many columns and we need
to convet the 28x28 matrix into a single row by appending every
row of the matrix to the first row using reshape() function.
In addition, we also need to scale all features to be between (0, 1)
or (-1, 1) or close to (-1, 1) range. Scale or normalize every feature
will improve numerical stability in the optimization procedure as
there are a lot of parameters to be optimized.
We first reshape the 28x28 image for each digit (i.e each row)
12.2 Feedforward Neural Network 295

into 784 columns (i.e. features), and then rescale the value to be
between 0 and 1 by dividing the original pixel value by 255, as
described in the cell below.

# step 1: reshape
x_train <- array_reshape(x_train,
c(nrow(x_train), 784))
x_test <- array_reshape(x_test,
c(nrow(x_test), 784))

# step 2: rescale
x_train <- x_train / 255
x_test <- x_test / 255

And here is the structure of the reshaped and rescaled features

for training and testing dataset. Now for each digit, there are 784
columns of features.

str(x_train)
str(x_test)

num [1:60000, 1:784] 0 0 0 0 0 0 0 0 0 0 ...

num [1:10000, 1:784] 0 0 0 0 0 0 0 0 0 0 ...

In this example, though the response variable is an integer (i.e. the

corresponding digits for an image), there is no order or rank for
these integers and they are just an indication of one of the 10 cate-
gories. So we also convert the response variable y to be categorical.

y_train <- to_categorical(y_train, 10)

y_test <- to_categorical(y_test, 10)
str(y_train)

num [1:60000, 1:10] 0 1 0 0 0 0 0 0 0 0 ...

296 12 Deep Learning

12.2.7.2 Fit model

Now we are ready to fit the model. It is straight forward to build

a deep neural network using keras. For this example, the number
of input features is 784 (i.e. scaled value of each pixel in the 28x28
image) and the number of class for the output is 10 (i.e. one of the
ten categories). So the input size for the first layer is 784 and the
output size for the last layer is 10. And we can add any number of
compatible layers in between.
In keras, it is easy to define a DNN model: (1) use
keras_model_sequential() to initiate a model placeholder and all
model structures are attached to this model object, (2) layers are
added in sequence by calling the layer_dense() function, (3) add
arbitrary layers to the model based on the sequence of calling
layer_dense(). For a dense layer, all the nodes from the previous
layer are connected with each and every node to the next layer. In
layer_dense() function, we can define how many nodes in that layer
through the units parameter. The activation function can be de-
fined through the activation parameter. For the first layer, we also
need to define the input features’ dimension through input_shape
parameter. For our preprocessed MNIST dataset, there are 784
columns in the input data. A common way to reduce overfitting is
to use the dropout method, which randomly drops a proportion of
the nodes in a layer. We can define the dropout proportion through
layer_dropout() function immediately after the layer_dense() func-
tion.

dnn_model <- keras_model_sequential()

dnn_model %>%
layer_dense(units = 256, activation = 'relu', input_shape = c(784)) %>%
layer_dropout(rate = 0.4) %>%
layer_dense(units = 128, activation = 'relu') %>%
layer_dropout(rate = 0.3) %>%
layer_dense(units = 64, activation = 'relu') %>%
layer_dropout(rate = 0.3) %>%
layer_dense(units = 10, activation = 'softmax')
12.2 Feedforward Neural Network 297

The above dnn_model has 4 layers with first layer 256 nodes, 2nd
layer 128 nodes, 3rd layer 64 nodes, and last layer 10 nodes. The
activation function for the first 3 layers is relu and the activation
function for the last layer is softmax which is typical for classifi-
cation problems. The model detail can be obtained through sum-
mary() function. The number of parameter of each layer can be cal-
culated as: (number of input features +1) times (numbe of nodes
in the layer). For example, the first layer has (784+1)x256=200960
parameters; the 2nd layer has (256+1)x128=32896 parameters.
Please note, dropout only randomly drop certain proportion of
parameters for each batch, it will not reduce the number of pa-
rameters in the model. The total number of parameters for the
dnn_model we just defined has 242762 parameters to be estimated.

summary(dnn_model)

________________________________________________________________________________
Layer (type) Output Shape Param #
================================================================================
dense_1 (Dense) (None, 256) 200960
________________________________________________________________________________
dropout_1 (Dropout) (None, 256) 0
________________________________________________________________________________
dense_2 (Dense) (None, 128) 32896
________________________________________________________________________________
dropout_2 (Dropout) (None, 128) 0
________________________________________________________________________________
dense_3 (Dense) (None, 64) 8256
________________________________________________________________________________
dropout_3 (Dropout) (None, 64) 0
________________________________________________________________________________
dense_4 (Dense) (None, 10) 650
================================================================================
Total params: 242,762
Trainable params: 242,762
298 12 Deep Learning

Non-trainable params: 0
________________________________________________________________________________

Once a model is defined, we need to compile the model with a few

other hyper-parameters including (1) loss function, (2) optimizer,
and (3) performance metrics. For multi-class classification prob-
lems, people usually use the categorical_crossentropy loss function
and optimizer_rmsprop() as the optimizer which performs batch gra-
dient descent.

dnn_model %>% compile(

loss = 'categorical_crossentropy',
optimizer = optimizer_rmsprop(),
metrics = c('accuracy')
)

Now we can feed data (x and y) into the neural network that
we just built to estimate all the parameters in the model. Here
we define three hyperparameters: epochs, batch_size, and valida-
tion_split, for this model. It just takes a couple of minutes to
finish.

dnn_history <- dnn_model %>% fit(

x_train, y_train,
epochs = 15, batch_size = 128,
validation_split = 0.2
)

There is some useful information stored in the output object

dnn_history and the details can be shown by using str(). We can
plot the training and validation accuracy and loss as function of
epoch by simply calling plot(dnn_history).

str(dnn_history)
12.2 Feedforward Neural Network 299

List of 2
$ params :List of 8
..$ metrics : chr [1:4] "loss" "acc" "val_loss" "val_acc"
..$ epochs : int 15
..$ steps : NULL
..$ do_validation : logi TRUE
..$ samples : int 48000
..$ batch_size : int 128
..$ verbose : int 1
..$ validation_samples: int 12000
$ metrics:List of 4
..$ acc : num [1:15] 0.83 0.929 0.945 0.954 0.959 ...
..$ loss : num [1:15] 0.559 0.254 0.195 0.165 0.148 ...
..$ val_acc : num [1:15] 0.946 0.961 0.967 0.969 0.973 ...
..$ val_loss: num [1:15] 0.182 0.137 0.122 0.113 0.104 ...
- attr(*, "class")= chr "keras_training_history"

plot(dnn_history)
300 12 Deep Learning

12.2.7.3 Prediction

dnn_model %>% evaluate(x_test, y_test)

32/10000 [..............................] - ETA: 3s

416/10000 [>.............................] - ETA: 1s
800/10000 [=>............................] - ETA: 1s
1184/10000 [==>...........................] - ETA: 1s
1472/10000 [===>..........................] - ETA: 1s
1728/10000 [====>.........................] - ETA: 1s
2016/10000 [=====>........................] - ETA: 1s
2464/10000 [======>.......................] - ETA: 1s
2912/10000 [=======>......................] - ETA: 1s
3360/10000 [=========>....................] - ETA: 0s
3744/10000 [==========>...................] - ETA: 0s
4192/10000 [===========>..................] - ETA: 0s
4704/10000 [=============>................] - ETA: 0s
5088/10000 [==============>...............] - ETA: 0s
5568/10000 [===============>..............] - ETA: 0s
6080/10000 [=================>............] - ETA: 0s
6592/10000 [==================>...........] - ETA: 0s
7008/10000 [====================>.........] - ETA: 0s
7488/10000 [=====================>........] - ETA: 0s
7968/10000 [======================>.......] - ETA: 0s
8288/10000 [=======================>......] - ETA: 0s
8704/10000 [=========================>....] - ETA: 0s
9120/10000 [==========================>...] - ETA: 0s
9600/10000 [===========================>..] - ETA: 0s
10000/10000 [==============================] - 1s 128us/step
$loss
[1] 0.093325

$acc
[1] 0.981
12.2 Feedforward Neural Network 301

dnn_pred <- dnn_model %>%

predict_classes(x_test)
head(dnn_pred)

[1] 7 2 1 0 4 1

Let’s check a few misclassified images. A number of misclassified

images can be found using the following code. And we can plot
these misclassified images to see whether a human can correctly
read it out.

## total number of mis-classcified images

sum(dnn_pred != mnist$test$y)

[1] 190

missed_image = mnist$test$x[dnn_pred != mnist$test$y,,]

missed_digit = mnist$test$y[dnn_pred != mnist$test$y]
missed_pred = dnn_pred[dnn_pred != mnist$test$y]

index_image = 34

## change this index to see different image.

input_matrix <- missed_image[index_image,1:28,1:28]
output_matrix <- apply(input_matrix, 2, rev)
output_matrix <- t(output_matrix)

image(1:28, 1:28, output_matrix, col = gray.colors(256),

xlab = paste("Image for digit ", missed_digit[index_image],
", wrongly predicted as ", missed_pred[index_image]),
ylab = "")
302 12 Deep Learning

Now we finish this simple tutorial of using deep neural networks

for handwritten digit recognition using the MNIST dataset. We
illustrate how to reshape the original data into the right format
and scaling; how to define a deep neural network with arbitrary
number of layers; how to choose activation function, optimizer,
and loss function; how to use dropout to limit overfitting; how
to setup hyperparameters; and how to fit the model and using
a fitted model to predict. Finally, we illustrate how to plot the
accuracy/loss as functions of the epoch. It shows the end-to-end
cycle of how to fit a deep neural network model.
On the other hand, the image can be better dealt with Convolu-
tional Neural Network (CNN) and we are going to walk through
the exact same problem using CNN in the next section.

12.3 Convolutional Neural Network

There are some challenges using a feedforward neural network to
solve computer vision problems. One of the challenges is that the
inputs can get really big after you vectorize the image array. A
 12.3 Convolutional Neural Network 303

64 x 64 x 3 image gives you an input vector of 12288! And that

is a very small image. So you can expect the number of param-
eters grows fast and it is difficult to get enough data to fit the
model. Also as the input image size grows, the computational re-
quirements to train a feedforward neural network will soon become
infeasible. Also, after vectorization, you lose most of the spacial in-
formation of the image. To overcome these, people instead use the
convolutional neural network for computer vision problems.
This section introduces the Convolutional Neural Network (CNN),
the deep learning model that is almost universally used in com-
puter vision applications. Computer vision has been advancing
rapidly which enables many new applications such as self-driving
cars and unlocking a phone using face. The application is not lim-
ited to the tech industry but some traditional industries such as
agriculture and healthcare. Precision agriculture utilizes advanced
hardware and computer vision algorithms to increase efficiency and
reduce costs for farmers. For example, analyze images from cam-
eras in the greenhouse to track plant growth state. Use sensory
data from drones, satellites, and tractors to track soil conditions,
detect herbs and pests, automate irrigation, etc. In health care,
computer vision helps clinicians to diagnose disease, identify can-
cer sites with high accuracy (Kwak and Hui, 2019). Even if you
don’t work on computer vision, you may find some of the ideas
inspiring and borrow them into your area.
Some popular computer vision problems are:
• Image classification (or image recognition): Recognize the object
in the image. Is there a cat in the image? Who is this person?
• Object detection: Detect the position and boarder of a specific
object. For example, if you are building a self-driving car, you
need to know the positions of other cars around you.
• Neural Style Transfer (NST): Given a “content” image and a
“style” image, generate an image that merges the two.
304 12 Deep Learning

12.3.1 Convolution Layer

A fundamental building block of the convolution neural network is,
as the name indicates, the convolution operation. In this chapter,
we illustrate the fundamentals of CNN using the example of image
classification.
How do you do convolution? For example, you have a 5 x 5 2-d
image (figure 12.11. You apply a 3 x 3 filter and convolute over
the image. The output of this convolution operator will be a 3 x 3
matrix, which you can consider as a 3 x 3 image and visualize it
(top right of figure 12.11).

FIGURE 12.11: There are an input image (left), a filter (middel),

and an output image (right).

You start from the top left corner of the image and put the fil-
ter on the top left 3 x3 sub-matrix of the input image and take
the element-wise product. Then you add up the 9 numbers. Move
forward one step each time until it gets to the bottom right. The
detailed process is shown in figure 12.12.
Let’s use edge detection as an example to see how convolution
operation works. Given a picture as the left of figure 12.13, you
want to detect the vertical lines. For example, there are vertical
lines along with the hair and the edges of the bookcase. How do
you do that? There are standard filters for operations like blurring,
12.3 Convolutional Neural Network 305

FIGURE 12.12: Convolution step by step

sharpening, and edge detection. To get the edge, you can use the
following 3 x 3 filter to convolute over the image.

kernel_vertical = matrix(c(1, 1, 1, 0, 0, 0, -1, -1, -1),

nrow = 3, ncol = 3)

kernel_vertical

## [,1] [,2] [,3]

## [1,] 1 0 -1
## [2,] 1 0 -1
## [3,] 1 0 -1

The following code implements the convolution process. The result

is shown as the middle of figure 12.13.

image = magick::image_read("http://bit.ly/2Nh5ANX")
kernel_vertical = matrix(c(1, 1, 1, 0, 0, 0, -1, -1, -1),
nrow = 3, ncol = 3)

kernel_horizontal = matrix(c(1, 1, 1, 0, 0, 0, -1, -1, -1),

nrow = 3, ncol = 3, byrow = T)

image_edge_vertical = magick::image_convolve(image, kernel_vertical)

image_edge_horizontal = magick::image_convolve(image, kernel_horizontal)
306 12 Deep Learning

par(mfrow = c(1, 3))

plot(image)
plot(image_edge_vertical)
plot(image_edge_horizontal)

FIGURE 12.13: Edge Detection Example

Why can kernel_vertical detect vertical edge? Let’s look at a sim-

pler example. The following 8 x 8 matrix where half of the matrix
is 10 and the other half is 0. The corresponding image is shown as
the left of the figure 12.14.

input_image = matrix(rep(c(200, 200, 200, 200, 0, 0, 0, 0), 8),

nrow = 8, byrow = T)
input_image

## [,1] [,2] [,3] [,4] [,5] [,6] [,7] [,8]

## [1,] 200 200 200 200 0 0 0 0
## [2,] 200 200 200 200 0 0 0 0
## [3,] 200 200 200 200 0 0 0 0
## [4,] 200 200 200 200 0 0 0 0
## [5,] 200 200 200 200 0 0 0 0
## [6,] 200 200 200 200 0 0 0 0
## [7,] 200 200 200 200 0 0 0 0
## [8,] 200 200 200 200 0 0 0 0

If we use the above filter kernel_vertical , the output matrix is

12.3 Convolutional Neural Network 307

shown below and the corresponding image is shown as the right of

the figure 12.14.

output_image = matrix(rep(c(0, 0, 200, 200, 0, 0), 6),

nrow = 6, byrow = T)
output_image

## [,1] [,2] [,3] [,4] [,5] [,6]

## [1,] 0 0 200 200 0 0
## [2,] 0 0 200 200 0 0
## [3,] 0 0 200 200 0 0
## [4,] 0 0 200 200 0 0
## [5,] 0 0 200 200 0 0
## [6,] 0 0 200 200 0 0
1.0

1.0
0.8

0.8
0.6

0.6
0.4
0.4

0.2
0.2

0.0
0.0

0.0 0.4 0.8 0.0 0.4 0.8

FIGURE 12.14: Simple Edge Detection Example

So the output image has a lighter region in the middle that cor-
responds to the vertical edge of the input image. When the input
image is large, such as the image in figure 12.13 is 1020 x 711,
the edge will not seem as thick as it is in this small example. To
detect the horizontal edge, you only need to rotate the filter by 90
degrees. The right image in figure 12.13 shows the horizontal edge
detection result. You can see how convolution operator detects a
specific feature from the image.
308 12 Deep Learning

The parameters for the convolution operation are the elements in

the filter. For a 3x3 filter shown below, the parameters to estimate
are 𝑤1 to 𝑤9 . So far, we move the filter one step each time when
we convolve. You can do more than 1 step as well. For example,
you can hop 2 steps each time after the sum of the element-wise
product. It is called strided-convolution. Use stride 𝑠 means the
output is downsized by a factor of 𝑠. It is rarely used in practice
but it is good to be familiar with the concept.

FIGURE 12.15: Convolution Layer Parameters

12.3.2 Padding Layer

Assume the stride is 𝑠 and we have a 𝑛×𝑛 input image to convolve
with a 𝑓 × 𝑓 filter, the output image is ( 𝑛−𝑓 𝑛−𝑓
𝑠 + 1) × ( 𝑠 + 1). After
each convolution, the dimension of the output shrinks. Depending
on the size of the input image, the output size may get too small
after a few rounds. Also, the pixel at the corner is used less than
the pixel in the middle. So it overlooks some information in the
image. To overcome these problems, you can add a padding layer
during the process. To keep the output the same size as the input
image in figure 12.15, you can pad two pixels on each side of the
image with 0 (figure 12.16 ).
If the stride is 𝑠 and we have a 𝑛 × 𝑛 input image to convolve with
a 𝑓 × 𝑓 filter. This time we pad 𝑝 pixels in each side, then the
output size becomes ( 𝑛+2𝑝−𝑓
𝑠 + 1) × ( 𝑛+2𝑝−𝑓
𝑠 + 1). You can specify
12.3 Convolutional Neural Network 309

FIGURE 12.16: Padding Layer

the value for p and also the pixel value used. Or you can just use
0 to pad and make the output the same size with input.

12.3.3 Pooling Layer

People sometimes use the pooling layer to reduce the size of the
representation and make some of the feature detection more ro-
bust. If you have a 4×4 input, the max and mean pooling operation
are shown in the figure 12.17. The process is quite simple. In the
example, the filter is 2 × 2 and stride is 2, so break the input into
four 2 × 2 regions (shown in the figure with different shaded col-
ors). For max pooling, each of the outputs is the maximum from
the corresponding shaded sub-region. Mean pooling layer works in
the same way except for getting the mean instead of maximum of
the sub-region. The pooling layer has hyperparameters (𝑓 and 𝑠)
but it has no parameter for gradient descent to learn.
Let’s go through an example of pooling a 2D grayscale image. Hope
it gives you some intuition behind what it does. Read the image
and convert the original color image (a 3D array) to grayscale (a
2D matrix).
310 12 Deep Learning

FIGURE 12.17: Pooling Layer

library(EBImage)
library(dplyr)

eggshell <- readImage("https://scientistcafe.com/images/eggshell.jpeg") %>%

# make it smaller
resize(560, 420) %>%
# rotate image
rotate(90)

# convert to 2D grayscale
gray_eggshell = apply(eggshell, c(1,2), mean)

The following function takes an image matrix or array, and apply

pooling operation.

pooling <- function(type = "max", image, filter, stride) {

f <- filter
s <- stride
12.3 Convolutional Neural Network 311

if (length(dim(image)) == 3) {
# get image dimensions
col <- dim(image[, , 1])[2]
row <- dim(image[, , 1])[1]
# calculate new dimension size
c <- (col - f)/s + 1
r <- (row - f)/s + 1
# create new image object
newImage <- array(0, c(c, r, 3))
# loops in RGB layers
for (rgb in 1:3) {
m <- image[, , rgb]
m3 <- matrix(0, ncol = c, nrow = r)
i <- 1
if (type == "mean")
for (ii in 1:r) {
j <- 1
for (jj in 1:c) {
m3[ii, jj] <- mean(as.numeric(m[i:(i +
(f - 1)), j:(j + (f - 1))]))
j <- j + s
}
i <- i + s
} else for (ii in 1:r) {
j = 1
for (jj in 1:c) {
m3[ii, jj] <- max(as.numeric(m[i:(i +
(f - 1)), j:(j + (f - 1))]))
j <- j + s
}
i <- i + s
}
newImage[, , rgb] <- m3
}
} else if (length(dim(image)) == 2) {
312 12 Deep Learning

# get image dimensions

col <- dim(image)[2]
row <- dim(image)[1]
# calculate new dimension size
c <- (col - f)/s + 1
r <- (row - f)/s + 1
m3 <- matrix(0, ncol = c, nrow = r)
i <- 1
if (type == "mean")
for (ii in 1:r) {
j <- 1
for (jj in 1:c) {
m3[ii, jj] <- mean(as.numeric(image[i:(i +
(f - 1)), j:(j + (f - 1))]))
j <- j + s
}
i <- i + s
} else for (ii in 1:r) {
j = 1
for (jj in 1:c) {
m3[ii, jj] <- max(as.numeric(image[i:(i +
(f - 1)), j:(j + (f - 1))]))
j <- j + s
}
i <- i + s
}
newImage <- m3
}
return(newImage)
}

Let’s apply both max and mean pooling with filter size 10 (𝑓 = 10)
and stride 10 (𝑠 = 10).
12.3 Convolutional Neural Network 313

gray_eggshell_max = pooling(type = "max",

image = gray_eggshell,
filter = 10, stride = 10)

gray_eggshell_mean = pooling(type = "mean",

image = gray_eggshell,
filter = 10, stride = 10)

You can see the result by plotting the output image (figure 12.18).
The top left is the original color picture. The top right is the 2D
grayscale picture. The bottom left is the result of max pooling.
The bottom right is the result of mean pooling. The max-pooling
gives you the value of the largest pixel and the mean-pooling gives
the average of the patch. You can consider it as a representation
of features, looking at the maximal or average presence of differ-
ent features. In general, max-pooling works better. You can gain
some intuition from the example (figure 12.18). The max-pooling
“picks” more distinct features and average-pooling blurs out fea-
tures evenly.

par(mfrow = c(2,2), oma = c(1, 1, 1, 1))

plot(eggshell)
plot(as.Image(gray_eggshell))
plot(as.Image(gray_eggshell_max))
plot(as.Image(gray_eggshell_mean))

12.3.4 Convolution Over Volume

So far, we have shown different types of layers on 2D inputs. If you
have a 3D input (such as a color image), then the filters will have
3 channels too. For example, if you have a 6 × 6 color image, the
input dimension is 6×6×3. We call them the height, width, and the
number of channels. The filter itself has 3 channels corresponding
to the red, green, and blue channels of the input. You can consider
it as a 3D cube with 27 parameters. Apply each channel of the filter
to the corresponding channel of the input. Multiply each of the 27
314 12 Deep Learning

FIGURE 12.18: Example of max and mean pooling

numbers with the corresponding numbers from the top left region
of the color input image and add them up. Add a bias parameter
and apply an activation function which gives you the first number
of the output image. Then slide it over to calculate the next one.
The final output is 2D 4 × 4. If you want to detect features in the
red channel only, you can use a filter with the second and third
channels to be all 0s. With different choices of the parameters,
you can get different feature detectors. You can use more than one
filter and each filter has multiple channels. For example, you can
use one 3 × 3 × 3 filter to detect the horizontal edge and another
to detect the vertical edge. Figure 12.18 shows an example of one
layer with two filters. Each filter has a dimension of 3 × 3 × 3. The
output dimension is 4×4×2. The output has two channels because
we have two filters on the layer. The total number of parameters
is 58 (each filter has one bias parameter 𝑏).
We use the following notations for layer 𝑙:
 12.3 Convolutional Neural Network 315

FIGURE 12.19: Example of convolution over volume

[𝑙] [𝑙] [𝑙]

• Use 𝑛𝑊 , 𝑛𝐻 and 𝑛𝐶 to denote the width, height and number of
channels of the input
• 𝑓 [𝑙] is the filter size
• 𝑝[𝑙] is the padding size
• 𝑠[𝑙] is the stride
The number of filters for layer 𝑙 is the number of channels of the
output of layer 𝑙. Since the output of layer 𝑙 is the input of layer
[𝑙+1]
𝑙 + 1, the number of filters for layer 𝑙 is 𝑛𝐶 . The number of
channels of the filter of layer 𝑙 should be equal to the number of
channels of the input of layer 𝑙 which is the output of layer 𝑙 − 1
[𝑙−1]
(i.e. 𝑛𝐶 ). So the dimensions of some key elements of layer 𝑙 are:
[𝑙−1]
• Each filter: 𝑓 [𝑙] × 𝑓 [𝑙] × 𝑛𝐶
[𝑙] [𝑙] [𝑙]
• Activations: 𝑎[𝑙] → 𝑛𝐻 × 𝑛𝑊 × 𝑛𝐶
[𝑙−1] [𝑙]
• Weights: 𝑓 [𝑙] × 𝑓 [𝑙] × 𝑛𝐶 × 𝑛𝐶
[𝑙]
• bias: 𝑏𝐶
[𝑙−1] [𝑙−1] [𝑙−1]
• Input: 𝑛𝐻 × 𝑛𝑊 × 𝑛𝐶
[𝑙] [𝑙] [𝑙]
• Output: 𝑛𝐻 × 𝑛𝑊 × 𝑛𝐶
After a series of 3D convolutional layers, we need to ‘flatten’ the
3D tensor to a 1D tensor, and add one or several dense layers to
connect the output to the response variable.
Now you know the basic building blocks of CNN. Let’s look at how
to use the keras R package to solve the same handwritten digits
316 12 Deep Learning

image recognition problem as in section 12.2.7. You will see the

CNN is better at handling image recognition problem.

12.3.5 Image Recognition Using CNN

CNN leverages the relationship among neighbor pixels in the 2D
image for better performance. It also avoids generating thousands
or millions of features for high resolution images with full color.
Now let’s import the MNIST dataset again as we have done some
preprocessing specifically for FFNN before. CNN requires different
preprocessing steps. Let’s start with a few parameters to be used
later.

# Load the mnist data's training and testing dataset

mnist <- dataset_mnist()
x_train <- mnist$train$x
y_train <- mnist$train$y
x_test <- mnist$test$x
y_test <- mnist$test$y

# Define a few parameters to be used in the CNN model

batch_size <- 128
num_classes <- 10
epochs <- 10

# Input image dimensions

img_rows <- 28
img_cols <- 28

12.3.5.1 Data preprocessing

For CNN, the input is a 𝑛𝐻 ×𝑛𝑊 ×𝑛𝐶 3D array with 𝑛𝐶 channels.

For example, a greyscale 𝑛𝐻 × 𝑛𝑊 image has only one channel,
and the input is 𝑛𝐻 × 𝑛𝑊 × 1 tensor. A 𝑛𝐻 × 𝑛𝑊 8-bit per channel
RGB image has three channels with 3 𝑛𝐻 × 𝑛𝑊 array with values
between 0 and 255, so the input is 𝑛𝐻 × 𝑛𝑊 × 3 tensor. For the
problem that we have now, the image is greyscaled, but we need to
specifically define there is one channel by reshaping the 2D array
12.3 Convolutional Neural Network 317

into 3D tensor using array_reshape(). The input_shape variable will

be used in the CNN model later. For an RGB color image, the
number of channels is 3 and we need to replace “1” with “3” for
the code cell below if the input image is RGB format.

x_train <- array_reshape(x_train,

c(nrow(x_train), img_rows, img_cols, 1))
x_test <- array_reshape(x_test,
c(nrow(x_test), img_rows, img_cols, 1))
input_shape <- c(img_rows, img_cols, 1)

Here is the structure of the reshaped image, the first dimension

is the image index, the 2-4 dimension is a 3D tensor even though
there is only one channel.

str(x_train)

int [1:60000, 1:28, 1:28, 1] 0 0 0 0 0 0 0 0 0 0 ...

Same as the FFNN model, we scale the input values to be be-

tween 0 and 1 for the same numerical stability consideration in
the optimization process.

x_train <- x_train / 255

x_test <- x_test / 255

Encode the response variable to binary vectors.

# Convert class vectors to binary class matrices

y_train <- to_categorical(y_train, num_classes)
y_test <- to_categorical(y_test, num_classes)

12.3.5.2 Fit model

CNN model contains a series of 3D convolutional layers which

contains a few parameters:
318 12 Deep Learning

(1) the kernal_size which is typically 3x3 or 5x5;

(2) the number of filters, which is equal to the number of
channels of the output;
(3) activation function.

For the first layer, there is also an input_shape parameter which is

the input image size and channel. To prevent overfitting and speed
up computation, a pooling layer is usually applied after one or a
few convolutional layers. A maximum pooling layer with pool_size
= 2x2 reduces the size to half. Dropout can be used as well in ad-
dition to pooling neighbor values. After a few 3D convolutional
layers, we also need to ‘flatten’ the 3D tensor output into 1D ten-
sor, and then add one or a couple of dense layers to connect the
output to the target response classes.
Let’s define the CNN model structure. Now we define a CNN model
with two convolutional layers, two max-pooling layers, and two
dropout layers to mediate overfitting. There are three dense layers
after flattening the 3D tensor. The last layer is a dense layer that
connects to the response.

# define model structure

cnn_model <- keras_model_sequential() %>%
layer_conv_2d(filters = 32,
kernel_size = c(5,5),
activation = 'relu',
input_shape = input_shape) %>%
layer_max_pooling_2d(pool_size = c(2, 2)) %>%
layer_conv_2d(filters = 64,
kernel_size = c(5,5),
activation = 'relu') %>%
layer_max_pooling_2d(pool_size = c(2, 2)) %>%
layer_dropout(rate = 0.2) %>%
layer_flatten() %>%
layer_dense(units = 120, activation = 'relu') %>%
layer_dropout(rate = 0.5) %>%
12.3 Convolutional Neural Network 319

layer_dense(units = 84, activation = 'relu') %>%

layer_dense(units = num_classes, activation = 'softmax')

summary(cnn_model)

Similar to before, we need to compile the defined CNN model.

cnn_model %>% compile(

loss = loss_categorical_crossentropy,
optimizer = optimizer_adadelta(),
metrics = c('accuracy')
)

Train the model and save each epochs’s history. Please note, as
we are not using GPU, it takes a few minutes to finish. Please
be patient while waiting for the results. The training time can be
significantly reduced if running on GPU.
320 12 Deep Learning

cnn_history <- cnn_model %>% fit(

x_train, y_train,
batch_size = batch_size,
epochs = epochs,
validation_split = 0.2
)

The trained model accuracy can be evaluated on the testing

dataset which is pretty good.

cnn_model %>% evaluate(x_test, y_test)

## loss accuracy
## 0.02153786 0.99290001

12.3.5.3 Prediction

We can apply the trained model to predict new image.

12.3 Convolutional Neural Network 321

# model prediction
cnn_pred <- cnn_model %>%
predict_classes(x_test)
head(cnn_pred)

[1] 7 2 1 0 4 1

Now let’s check a few misclassified images to see whether a human

can do a better job than this simple CNN model.

## number of mis-classcified images

sum(cnn_pred != mnist$test$y)

[1] 72

missed_image = mnist$test$x[cnn_pred != mnist$test$y,,]

missed_digit = mnist$test$y[cnn_pred != mnist$test$y]
missed_pred = cnn_pred[cnn_pred != mnist$test$y]

index_image = 10 ## change this index to see different image.

input_matrix <- missed_image[index_image,1:28,1:28]
output_matrix <- apply(input_matrix, 2, rev)
output_matrix <- t(output_matrix)
image(1:28, 1:28, output_matrix, col=gray.colors(256),
xlab=paste('Image for digit ', missed_digit[index_image], ',
wrongly predicted as ', missed_pred[index_image]), ylab="")
 322 12 Deep Learning

12.4 Recurrent Neural Network

Traditional neural networks don’t have a framework that can han-
dle sequential events where the later events are based on the previ-
ous ones. For example, map an input audio clip to a text transcript
where the input is voice over time, and the output is the corre-
sponding sequence of words over time. Recurrent Neural Network
is a deep-learning model that can process this type of sequential
data.
The recurrent neural network allows information to flow from one
step to the next with a repetitive structure. Figure 12.20 shows the
basic chunk of an RNN network. You combine the activated neuro
from the previous step with the current input 𝑥<𝑡> to produce an
output 𝑦<𝑡>
̂ and an updated activated neuro to support the next
input at 𝑡 + 1.
So the whole process repeats a similar pattern. If we unroll the
loop:
The chain-like recurrent nature makes it the natural architecture
for sequential data. There is incredible success applying RNNs to
this type of problems:
• Machine translation
• Voice recognition
 12.4 Recurrent Neural Network 323

FIGURE 12.20: Recurrent Neural Network Unit

FIGURE 12.21: An Unrolled Recurrent Neural Network

• Music generation
• Sentiment analysis
A trained CNN accepts a fixed-sized vector as input (such as
28 × 28 image) and produces a fixed-sized vector as output (such
as the probabilities of being one the ten digits). RNN has a much
more flexible structure. It can operate over sequences of vectors
and produces sequences of outputs and they can vary in size. To
understand what it means, let’s look at some RNN structure ex-
amples.
 324 12 Deep Learning

The rectangle represents a vector and the arrow represents matrix

multiplications. The input vector is in green and the output vector
is in blue. The red rectangle holds the intermediate state. From
left to right:
• one-to-one: model takes a fixed size input and produces a fixed
size output, such as CNN. it is not sequential.
• one-to-many: model takes one input and generate a sequence of
output, such as the music generation.
• many-to-one: model takes a sequence of input and produces a
single output, such as sentiment analysis.
• many-to-many: model takes a sequence of input and produces a
sequence of output. The input size can be the same with the out-
put size (such as name entity recognition) or it can be different
(such as machine translation).

12.4.1 RNN Model

To further understand the RNN model, let’s look at an entity
recognition example. Assume you want to build a sequence model
to recognize the company or computer language names in a sen-
tence like this: “Use Netlify and Hugo”. It is a name recognition
problem which is used by the research company to index different
company names in the articles. For tasks like this, we need a model
that can learn and share the learning across different texts. The
position of the word has important information about the word.
For example, the word before “to” is more likely to be a verb than
a noun. It is also used in material science to tag chemicals men-
tioned in the most recent journals to find any indication of the
next research topic.
12.4 Recurrent Neural Network 325

Given input sentence x, you want a model to produce one output

for each word in x that tells you if that word is the name for
something. So in this example, the input is a sequence of 5 words
including the period in the end. The output is a sequence of 0/1
with the same length that indicates whether the input word is
a name (1) or not (0). We use superscript < 𝑡 > to denote the
element position of input and output; use superscript (𝑖) to denote
the 𝑖𝑡ℎ sample (you will have different sentences in the training
(𝑖) (𝑖)
data); Use 𝑇𝑥 to denote the length of 𝑖𝑡ℎ input, 𝑇𝑦 for output.
(𝑖) (𝑖)
In this case, 𝑇𝑥 is equal to 𝑇𝑦 .

Before we build a neural network, we need to decide a way to

represent individual words in numbers. What should 𝑥<1> be? In
practice, people use word embedding which we will discuss in the
later section. Here, for illustration, we use the one-hot encoding
word representation. Assume we have a dictionary of 10,000 unique
words. You can build the dictionary by finding the top 10,000
occurring words in your training set. Each word in your training
set will have a position in the dictionary sequence. For example,
“use” is the 8320th element of the dictionary sequence. So 𝑥<1>
is a vector with all zeros except for a one on position 8320. Using
this one-hot representation, each input 𝑥<𝑡> is a vector with all
zeros except for one element.
326 12 Deep Learning

Given this representation of input words, the goal is to learn a

sequence model that maps the input words to output y, indicating
if the word is an entity (1) or not (0). Let us build a one-layer
recurrent neural network. The model starts from the first word
“use” (𝑥<1> ) and build a neural network to predict the output.
To start the process, we also need to initialize the activation at
time 0, 𝑎0 . The most common choice is to use a vector of zeros.
The common activation function for the intermediate layer is the
Hyperbolic Tangent Function (tanh). RNN uses other methods to
prevent the vanishing gradient problem discussed in section 12.4.3.
Similar to FFNN, The output layer activation function depends on
the output type. The current example is a binary classification, so
we use the sigmoid function (𝜎).

𝑎<0> = 0; 𝑎<1> = 𝑡𝑎𝑛ℎ(𝑊𝑎𝑎 𝑎<0> + 𝑊𝑎𝑥 𝑥<1> + 𝑏𝑎 )

𝑦<1>
̂ = 𝜎(𝑊𝑦𝑎 𝑎<1> + 𝑏𝑦 )
And when it takes the second word 𝑥<2> , it also gets information
from the previous step using the non-activated neurons.

𝑎<2> = 𝑡𝑎𝑛ℎ(𝑊𝑎𝑎 𝑎<1> + 𝑊𝑎𝑥 𝑥<2> + 𝑏𝑎 )

𝑦<2>
̂ = 𝜎(𝑊𝑦𝑎 𝑎<2> + 𝑏𝑦 )

For the 𝑡𝑡ℎ word:

𝑎<𝑡> = 𝑡𝑎𝑛ℎ(𝑊𝑎𝑎 𝑎<𝑡−1> + 𝑊𝑎𝑥 𝑥<𝑡> + 𝑏𝑎 )

𝑦<𝑡>
̂ = 𝜎(𝑊𝑦𝑎 𝑎<𝑡> + 𝑏𝑦 )
12.4 Recurrent Neural Network 327

The information flows from one step to the next with a repetitive
structure until the last time step input 𝑥<𝑇𝑥 > and then it out-
puts 𝑦<𝑇
̂ 𝑦 > . In this example, 𝑇𝑥 = 𝑇𝑦 . The architecture changes
when 𝑇𝑥 and 𝑇𝑦 are not the same. The model shares parameters,
𝑊𝑦𝑎 , 𝑊𝑎𝑎 , 𝑊𝑎𝑥 , 𝑏𝑎 , 𝑏𝑦 , for all time steps of the input.

Calculate the loss function:

𝐿<𝑡> (𝑦<𝑡>
̂ ) = −𝑦<𝑡> 𝑙𝑜𝑔(𝑦<𝑡>
̂ ) − (1 − 𝑦<𝑡> )𝑙𝑜𝑔(1 − 𝑦<𝑡>
̂ )
𝑇
𝑦
𝐿(𝑦,̂ 𝑦) = Σ𝑡=1 𝐿<𝑡> (𝑦,̂ 𝑦)
The above defines the forward process. Same as before, the back-
ward propagation computes the gradient descent for the parame-
ters by the chain rule for differentiation.
In this RNN structure, the information only flows from the left to
the right. So at any position, it only uses data from earlier in the
sequence to make a prediction. It does not work when predicting
the current word needs information from later words. For example,
consider the following two sentences:

1. Do you like April Kepner in Grey’s Anatomy?

2. Do you like April in Los Angeles? It is not too hot.

Given just the first three words is not enough to know if the word
“April” is part of a person’s name. It is a person’s name in 1 but
not 2. The two sentences have the same first three words. In this
case, we need a model that allows the information to flow in both
directions. A bidirectional RNN takes data from both earlier and
later in the sequence. The disadvantage is that it needs the entire
328 12 Deep Learning

word sequence to predict at any position. For a speech recognition

application that requires capturing the speech in real-time, we need
a more complex method called the attention model. We will not
get into those models here. Deep Learning with R (Chollet and
Allaire, 2018) provides a high-level introduction of bidirectional
RNN with applicable codes. It teaches both intuition and practical,
computational usage of deep learning models. For python users,
refer to Deep Learning with Python (Chollet, 2017). A standard
text with heavy mathematics is Deep Learning (Goodfellow et al.,
2016).

12.4.2 Word Embedding

So far, we have been using one-hot encoding to represent the words.
This representation is sparse and doesn’t capture the relationship
between the words. For example, if your model learns from the
training data that the word after “pumpkin” in the first sentence
is “pie,” can it fill the second sentence’s blank with “pie”?

1. [training data] My favorite Christmas dessert is pumpkin

pie.
2. [testing data] My favorite Christmas dessert is apple
____.

The algorithm can not learn the relationship between “pumpkin”

and “apple” by the distance between the one-hot encoding for the
two words. If we can find a way to create features to represent the
words, we may teach the model to learn that pumpkin and apple
is food. And the distance between the feature representations of
these two words is closer than, for example, “apple” and “nurse.”
So when the model sees “apple” in a new sentence and needs to
predict the word after, it is more likely to choose “pie” since it sees
“pumpkin pie” in the training data. The idea of word embedding is
to learn a set of features to represent words. For example, you can
score each word in the dictionary according to a group of features
like this:
12.4 Recurrent Neural Network 329

The word “male” has a score of -1 for the “gender” feature, “fe-
male” has a score of 1. Both “Apple” and “pumpkin” have a high
score for the “food” feature and much lower scores for the rest. You
can set the number of features to learn, usually more than what
we list in the above figure. If you use 200 features to represent the
words, then the learned embedding for each word is a vector with
a length of 200.
For language-related applications, text embedding is the most crit-
ical step. It converts raw text into a meaningful vector representa-
tion. Once we have a vector representation, it is easy to calculate
typical numerical metrics such as cosine similarity. There are many
pre-trained text embeddings available for us to use. We will briefly
introduce some of these popular embeddings.
The first widely used embedding is word2vec. It was first intro-
duced in 2013 and was trained by a large collection of text in
an unsupervised way. Training the word2vec embedding vector
uses bag-of-words or skip-gram. In the bag-of-words architecture,
the model predicts the current word based on a window of sur-
rounding context words. In skip-gram architecture, the model uses
the current word to predict the surrounding window of context
words. There are pre-trained word2vec embeddings based on a
large amount of text (such as wiki pages, news reports, etc.) for
general applications.
GloVe (Global Vectors) embedding is an extension of word2vec
and performs better. It uses a unique version of the square loss
function. However, words are composite of meaningful components
330 12 Deep Learning

such as radicals.
“eat” and “eaten” are different forms of the same word. Both
word2vec and GloVe use word-level information, and they treat
each word uniquely based on its context.
The fastText embedding is introduced to use the word’s internal
structure to make the process more efficient. It uses morphologi-
cal information to extend the skip-gram model. New words that
are not in the training data can be repressed well. It also sup-
ports 150+ different languages. The above-mentioned embeddings
(word2vec, GloVe, and fastText) do not consider the words’ con-
text (i.e., the same word has the same embedding vector). How-
ever, the same word may have different meanings in a different con-
text. BERT (Bidirectional Encoder Representations from Trans-
formers) is introduced to add context-level information in text-
related applications. As of early 2021, BERT is generally consid-
ered the best language model for common application tasks.

12.4.3 Long Short Term Memory

The sequence in RNN can be very long, which leads to the van-
ishing gradient problem even when the RNN network is not deep.
Think about the following examples:

1. The girl walked away, sat down in the shade of a tree,

and began to read a new book which she bought the day
before.
2. The boy walked away, sat down in the shade of a tree,
and began to read a new book which he bought the day
before.

For sentence 1, you need to use “she” in the adjective clause af-
ter “which” because it is a girl. For sentence 2, you need to use
“he” because it is a boy. This is a long-term dependency example
where the information at the beginning can affect what needs to
come much later in the sentence. RNN needs to forward propa-
gate information from left to right and then backpropagate from
right to left. It can be difficult for the error associated with the
12.4 Recurrent Neural Network 331

later sequence to affect the optimization earlier. So in practice, it

means the model might fail to do the task mentioned above. Peo-
ple came up with different methods to mitigate this issue, such
as the Greater Recurrent Units (GRU) (Chung et al., 2014) and
Long Short Term Memory Units (LSTM) (Hochreiter and Schmid-
huber, 1997). The goal is to help the model memorize information
in the earlier sequence. We are going to walk through LSTM step
by step.
The first step of LSTM is to decide what information to forget.
This decision is made by “forget gate”, a sigmoid function (Γ𝑓 ). It
looks at 𝑎<𝑡−1> and 𝑥𝑡 and outputs a number between 0 and 1 for
each number in the cell state 𝑐𝑡−1 . A value 1 means “completely
remember the state”, while 0 means “completely forget the state”.

The next step is to decide what new information we’re going to

add to the cell state. This step includes two parts:

1. input gate (Γ𝑢 ): a sigmoid function that decides how

much we want to update
2. a vector of new candidate value (𝑐<𝑡>
̃ )
332 12 Deep Learning

The multiplication of the above two parts Γ𝑢 ∗ 𝑐<𝑡>

̃ is the new
candidate scaled by the input gate. We then combine the results
we get so far to get new cell state 𝑐<𝑡> .

Finally, we need to decide what we are going to output. The output

is a filtered version of the new cell state 𝑐<𝑡> .
12.4 Recurrent Neural Network 333

12.4.4 Sentiment Analysis Using RNN

In this section, we will walk through an example of text sentiment
analysis using RNN. Refer to section 4.3 to set up an account,
create a notebook (R or Python) and start a cluster. Refer to
section 12.2.7 for package installation.
We will use the IMDB movie review data. It is one of the most used
datasets for text-related machine learning methods. The datasets’
inputs are movie reviews published at IMDB in its raw text format,
and the output is a binary sentiment indicator( “1” for positive and
“0” for negative) created through human evaluation. The training
and testing data have 25,000 records each. Each review varies in
length.

12.4.4.1 Data preprocessing

Machine learning algorithms can not deal with raw text, and we
have to convert text into numbers before feeding it into an algo-
rithm. Tokenization is one way to convert text data into a nu-
merical representation. For example, suppose we have 500 unique
words for all reviews in the training dataset. We can label each
word by the rank (i.e., from 1 to 500) of their frequency in the
334 12 Deep Learning

training data. Then each word is replaced by an integer between

1 to 500. This way, we can map each movie review from its raw
text format to a sequence of integers.
As reviews can have different lengths, sequences of integers will
have different sizes too. So another important step is to make sure
each input has the same length by padding or truncating. For
example, we can set a length of 50 words, and for any reviews less
than 50 words, we can pad 0 to make it 50 in length; and for reviews
with more than 50 words, we can truncate the sequence to 50 by
keeping only the first 50 words. After padding and truncating, we
have a typical data frame, each row is an observation, and each
column is a feature. The number of features is the number of words
designed for each review (i.e., 50 in this example).
After tokenization, the numerical input is just a naive mapping
to the original words, and the integers do not have their usual
numerical meanings. We need to use embedding to convert these
categorical integers to more meaningful representations. The word
embedding captures the inherited relationship of words and dra-
matically reduces the input dimension (see section 12.4.2). The
dimension is a vector space representing the entire vocabulary. It
can be 128 or 256, and the vector space dimension is the same
when the vocabulary changes. It has a lower dimension, and each
vector is filled with real numbers. The embedding vectors can be
learned from the training data, or we can use pre-trained embed-
ding models. There are many pre-trained embeddings for us to use,
such as Word2Vec, BIRD.
12.4 Recurrent Neural Network 335

12.4.4.2 R code for IMDB dataset

The IMDB dataset is preloaded for keras and we can call

dataset_imdb() to load a partially pre-processed dataset into a data
frame. We can define a few parameters in that function. num_words
is the number of words in each review to be used. All the unique
words are ranked by their frequency counts in the training dataset.
The dataset_imdb() function keeps the top num_words words and re-
places other words with a default value of 2, and using integers to
represent text (i.e., top frequency word will be replaced by 3 and
0, 1, 2 are reserved for “padding,” “start of the sequence,” and
“unknown.” ).

# Load `keras` package

library(keras)

# consider only the top 10,000 words in the dataset

max_unique_word <- 2500
# cut off reviews after 100 words
max_review_len <- 100

Now we load the IMDB dataset, and we can check the structure
of the loaded object by using str() command.
336 12 Deep Learning

my_imdb <- dataset_imdb(num_words = max_unique_word)

str(my_imdb)

Downloading data from

https://storage.googleapis.com/tensorflow/tf-keras-datasets/imdb.npz

8192/17464789 [..............................] - ETA: 0s

811008/17464789 [>.............................] - ETA: 1s
4202496/17464789 [======>.......................] - ETA: 0s
11476992/17464789 [==================>...........] - ETA: 0s
17465344/17464789 [==============================] - 0s 0us/step
List of 2
$ train:List of 2
..$ x:List of 25000
.. ..$ : int [1:218] 1 14 22 16 43 530 973 1622 1385 65 ...
.. ..$ : int [1:189] 1 194 1153 194 2 78 228 5 6 1463 ...

*** skipped some output ***

x_train <- my_imdb$train$x

y_train <- my_imdb$train$y
x_test <- my_imdb$test$x
y_test <- my_imdb$test$y

Next, we do the padding and truncating process.

x_train <- pad_sequences(x_train, maxlen = max_review_len)

x_test <- pad_sequences(x_test, maxlen = max_review_len)

The x_train and x_test are numerical data frames ready to be used
for recurrent neural network models.
Simple Recurrent Neurel Network
Like DNN and CNN models we trained in the past, RNN mod-
12.4 Recurrent Neural Network 337

els are relatively easy to train using keras after the pre-processing
stage. In the following example, we use layer_embedding() to fit
an embedding layer based on the training dataset, which has
two parameters: input_dim (the number of unique words) and out-
put_dim (the length of dense vectors). Then, we add a simple RNN
layer by calling layer_simple_rnn() and followed by a dense layer
layer_dense() to connect to the response binary variable.

rnn_model <- keras_model_sequential()

rnn_model %>%
layer_embedding(input_dim = max_unique_word, output_dim = 128) %>%
layer_simple_rnn(units = 64, dropout = 0.2,
recurrent_dropout = 0.2) %>%
layer_dense(units = 1, activation = 'sigmoid')

We compile the RNN model by defining the loss function, optimizer

to use, and metrics to track the same way as DNN and CNN
models.

rnn_model %>% compile(

loss = 'binary_crossentropy',
optimizer = 'adam',
metrics = c('accuracy')
)

Let us define a few more variables before fitting the model:

batch_size, epochs, and validation_split. These variables have the
same meaning as DNN and CNN models we see in the past.

batch_size = 128
epochs = 5
validation_split = 0.2

rnn_history <- rnn_model %>% fit(

x_train, y_train,
batch_size = batch_size,
338 12 Deep Learning

epochs = epochs,
validation_split = validation_split
)

Epoch 1/5

1/157 [...............] - ETA: 0s - loss: 0.7348 - accuracy: 0.4766

2/157 [...............] - ETA: 7s - loss: 0.7279 - accuracy: 0.4961
3/157 [...............] - ETA: 10s - loss: 0.7290 - accuracy: 0.4896

*** skipped some output ***

154/157 [=============>.] - ETA: 0s - loss: 0.4456 - accuracy: 0.7991

155/157 [=============>.] - ETA: 0s - loss: 0.4460 - accuracy: 0.7991
156/157 [=============>.] - ETA: 0s - loss: 0.4457 - accuracy: 0.7995
157/157 [===============] - 17s 109ms/step
loss: 0.4457 - accuracy: 0.7995 - val_loss: 0.4444 - val_accuracy: 0.7908

plot(rnn_history)
12.4 Recurrent Neural Network 339

rnn_model %>%
evaluate(x_test, y_test)

1/782 [...............] - ETA: 0s - loss: 0.3019 - accuracy: 0.8438

8/782 [...............] - ETA: 5s - loss: 0.4328 - accuracy: 0.8008
15/782 [...............] - ETA: 5s - loss: 0.4415 - accuracy: 0.7937
23/782 [...............] - ETA: 5s - loss: 0.4247 - accuracy: 0.8043

*** skipped some output ***

775/782 [=============>.] - ETA: 0s - loss: 0.4371 - accuracy: 0.8007

782/782 [===============] - ETA: 0s - loss: 0.4365 - accuracy: 0.8010
782/782 [===============] - 6s 8ms/step - loss: 0.4365 - accuracy: 0.8010
340 12 Deep Learning

loss accuracy
0.4365373 0.8010000

LSTM RNN Model

A simple RNN layer is a good starting point for learning RNN, but
the performance is usually not that good because these long-term
dependencies are impossible to learn due to vanishing gradient.
Long Short Term Memory RNN model (LSTM) can carry useful
information from the earlier words to later words. In keras, it is
easy to replace a simple RNN layer with an LSTM layer by using
layer_lstm().

lstm_model <- keras_model_sequential()

lstm_model %>%
layer_embedding(input_dim = max_unique_word, output_dim = 128) %>%
layer_lstm(units = 64, dropout = 0.2, recurrent_dropout = 0.2) %>%
layer_dense(units = 1, activation = 'sigmoid')

lstm_model %>% compile(

loss = 'binary_crossentropy',
optimizer = 'adam',
metrics = c('accuracy')
)

batch_size = 128
epochs = 5
validation_split = 0.2

lstm_history <- lstm_model %>% fit(

x_train, y_train,
batch_size = batch_size,
epochs = epochs,
validation_split = validation_split
)
12.4 Recurrent Neural Network 341

Epoch 1/5

1/157 [.........] - ETA: 0s - loss: 0.6939 - accuracy: 0.4766

2/157 [.........] - ETA: 31s - loss: 0.6940 - accuracy: 0.4766
3/157 [.........] - ETA: 43s - loss: 0.6937 - accuracy: 0.4896

*** skipped some output ***

155/157 [=======>.] - ETA: 0s - loss: 0.2610 - accuracy: 0.8918

156/157 [=======>.] - ETA: 0s - loss: 0.2607 - accuracy: 0.8920
157/157 [=========] - ETA: 0s - loss: 0.2609 - accuracy: 0.8917
157/157 [=========] - 67s 424ms/step
loss: 0.2609 accuracy: 0.8917 val_loss: 0.3754 val_accuracy: 0.8328

plot(lstm_history)
342 12 Deep Learning

lstm_model %>%
evaluate(x_test, y_test)

1/782 [.........] - ETA: 0s - loss: 0.2332 - accuracy: 0.9062

4/782 [.........] - ETA: 12s - loss: 0.3536 - accuracy: 0.8750
7/782 [.........] - ETA: 14s - loss: 0.3409 - accuracy: 0.8705
10/782 [.........] - ETA: 14s - loss: 0.3508 - accuracy: 0.8625

*** skipped some output ***

775/782 [=======>.] - ETA: 0s - loss: 0.3640 - accuracy: 0.8415

778/782 [=======>.] - ETA: 0s - loss: 0.3637 - accuracy: 0.8417
780/782 [=======>.] - ETA: 0s - loss: 0.3632 - accuracy: 0.8419
782/782 [=========] - ETA: 0s - loss: 0.3631 - accuracy: 0.8420
.0 Recurrent Neural Network 343

782/782 [=========] - 18s 22ms/step - loss: 0.3631 - accuracy: 0.8420

loss accuracy
0.3631141 0.8420000

This simple example shows that LSTM’s performance has im-

proved dramatically from the simple RNN model. The compu-
tation time for LSTM is roughly doubled when compared with the
simple RNN model for this small dataset.
A
Handling Large Local Data

When the data is too large to fit in a computer’s memory, we

can use some big data analytics engine like Spark on a cloud plat-
form (see Chapter 4). However, even when the data can fit in the
memory, there may be a situation where it is slow to read and ma-
nipulate due to a relatively large size. Some R packages can make
the process faster with the cost of familiarity, especially for data
wrangling. But it avoids the hurdle of setting up Spark cluster and
working in an unfamiliar environment. This section presents some
of the alternative R packages to read, write and wrangle a data
set that is relatively large but not too big to fit in the memory.
Load the R packages first:

# install packages from CRAN if you haven't

library(readr)
library(data.table)

A.1 readr

You must be familiar with read.csv(), read.table() and write.csv()

in base R. Here we will introduce a more efficient package for read-
ing and writing data: readr package. The corresponding functions
are read_csv(), read_table() and write_csv(). The commands look
quite similar, but readr is different in the following respects:

1. It is 10x faster. The trick is that readr uses C++ to process

the data quickly.

345
 346 A Handling Large Local Data

2. It doesn’t change the column names. The names can start

with a number and “.” will not be substituted to “_”. For
example:

read_csv("2015,2016,2017
1,2,3
4,5,6")

## # A tibble: 2 x 3
## `2015` `2016` `2017`
## <dbl> <dbl> <dbl>
## 1 1 2 3
## 2 4 5 6

1. readr functions do not convert strings to factors by de-

fault, are able to parse dates and times and can automat-
ically determine the data types in each column.
2. The killing character, in my opinion, is that readr provides
progress bar. What makes you feel worse than waiting
is not knowing how long you have to wait.

The major functions of readr is to turn flat files into data frames:
• read_csv():
reads comma delimited files
• read_csv2(): reads semicolon separated files (common in countries
where , is used as the decimal place)
• read_tsv(): reads tab delimited files
• read_delim(): reads in files with any delimiter
• read_fwf(): reads fixed width files. You can specify fields ei-
ther by their widths with fwf_widths() or their position with
fwf_positions()

• read_table():
reads a common variation of fixed width files where
columns are separated by white space
 A.1 readr 347

• read_log(): reads Apache style log files

The good thing is that those functions have similar syntax. Once
you learn one, the others become easy. Here we will focus on
read_csv().

The most important information for read_csv() is the path to your

data:

sim.dat <- read_csv("http://bit.ly/2P5gTw4")

head(sim.dat)

# A tibble: 6 x 19
age gender income house store_exp online_exp store_trans online_trans Q1
<int> <chr> <dbl> <chr> <dbl> <dbl> <int> <int> <int>
1 57 Female 1.21e5 Yes 529. 304. 2 2 4
2 63 Female 1.22e5 Yes 478. 110. 4 2 4
3 59 Male 1.14e5 Yes 491. 279. 7 2 5
4 60 Male 1.14e5 Yes 348. 142. 10 2 5
5 51 Male 1.24e5 Yes 380. 112. 4 4 4
6 59 Male 1.08e5 Yes 338. 196. 4 5 4
# ... with 10 more variables: Q2 <int>, Q3 <int>, Q4 <int>, Q5 <int>, Q6 <int>,
# Q7 <int>, Q8 <int>, Q9 <int>, Q10 <int>, segment <chr>

The function reads the file to R as a tibble. You can consider tibble
as next iteration of the data frame. They are different with data
frame for the following aspects:
• It never changes an input’s type (i.e., no more stringsAsFactors
= FALSE!)
• It never adjusts the names of variables
• It has a refined print method that shows only the first 10 rows
and all the columns that fit on the screen. You can also control
the default print behavior by setting options.
Refer to http://r4ds.had.co.nz/tibbles.html for more information
about ‘tibble’.
When you run read_csv() it prints out a column specification that
gives the name and type of each column. To better understanding
348 A Handling Large Local Data

how readr works, it is helpful to type in some baby data set and
check the results:

dat <- read_csv("2015,2016,2017

100,200,300
canola,soybean,corn")
print(dat)

## # A tibble: 2 x 3
## `2015` `2016` `2017`
## <chr> <chr> <chr>
## 1 100 200 300
## 2 canola soybean corn

You can also add comments on the top and tell R to skip those
lines:

dat <- read_csv("# I will never let you know that

# my favorite food is carrot
Date,Food,Mood
Monday,carrot,happy
Tuesday,carrot,happy
Wednesday,carrot,happy
Thursday,carrot,happy
Friday,carrot,happy
Saturday,carrot,extremely happy
Sunday,carrot,extremely happy",
skip = 2)
print(dat)

## # A tibble: 7 x 3
## Date Food Mood
## <chr> <chr> <chr>
## 1 Monday carrot happy
## 2 Tuesday carrot happy
## 3 Wednesday carrot happy
## 4 Thursday carrot happy
A.1 readr 349

## 5 Friday carrot happy

## 6 Saturday carrot extremely happy
## 7 Sunday carrot extremely happy

If you don’t have column names, set col_names = FALSE then R will
assign names “X1”,“X2”… to the columns:

dat <- read_csv("Saturday,carrot,extremely happy

Sunday,carrot,extremely happy", col_names = FALSE)
print(dat)

## # A tibble: 2 x 3
## X1 X2 X3
## <chr> <chr> <chr>
## 1 Saturday carrot extremely happy
## 2 Sunday carrot extremely happy

You can also pass col_names a character vector which will be

used as the column names. Try to replace col_names=FALSE with
col_names=c("Date","Food","Mood") and see what happen.

As mentioned before, you can use read_csv2() to read semicolon

separated files:

dat <- read_csv2("Saturday; carrot; extremely happy \n

Sunday; carrot; extremely happy", col_names = FALSE)
print(dat)

## # A tibble: 2 x 3
## X1 X2 X3
## <chr> <chr> <chr>
## 1 Saturday carrot extremely happy
## 2 Sunday carrot extremely happy

Here “\n” is a convenient shortcut for adding a new line.

You can use read_tsv() to read tab delimited files:
350 A Handling Large Local Data

dat <- read_tsv("every\tman\tis\ta\tpoet\twhen\the\tis\tin\tlove\n",

col_names = FALSE)
print(dat)

## # A tibble: 1 x 10
## X1 X2 X3 X4 X5 X6 X7 X8 X9
## <chr> <chr> <chr> <chr> <chr> <chr> <chr> <chr> <chr>
## 1 every man is a poet when he is in
## # ... with 1 more variable: X10 <chr>

Or more generally, you can use read_delim() and assign separating

character:

dat <- read_delim("THE|UNBEARABLE|RANDOMNESS|OF|LIFE\n",

delim = "|", col_names = FALSE)
print(dat)

## # A tibble: 1 x 5
## X1 X2 X3 X4 X5
## <chr> <chr> <chr> <chr> <chr>
## 1 THE UNBEARABLE RANDOMNESS OF LIFE

Another situation you will often run into is the missing value. In
marketing survey, people like to use “99” to represent missing. You
can tell R to set all observation with value “99” as missing when
you read the data:

dat <- read_csv("Q1,Q2,Q3

5, 4,99",
na = "99")
print(dat)

## # A tibble: 1 x 3
## Q1 Q2 Q3
## <dbl> <dbl> <lgl>
## 1 5 4 NA
 A.2 data.table— enhanced data.frame 351

For writing data back to disk, you can use write_csv() and
write_tsv(). The following two characters of the two functions in-
crease the chances of the output file being read back in correctly:
• Encode strings in UTF-8
• Save dates and date-times in ISO8601 format so they are easily
parsed elsewhere
For example:

write_csv(sim.dat, "sim_dat.csv")

For other data types, you can use the following packages:
• Haven: SPSS, Stata and SAS data
• Readxl and xlsx: excel data(.xls and .xlsx)
• DBI: given data base, such as RMySQL, RSQLite and RPost-
greSQL, read data directly from the database using SQL
Some other useful materials:
• For getting data from the internet, you can refer to the book
“XML and Web Technologies for Data Sciences with R”.

• R data import/export manual1

• rio package:https://github.com/leeper/rio

A.2 data.table— enhanced data.frame

What is data.table? It is an R package that provides an enhanced

version of data.frame. The most used object in R is data frame.
Before we move on, let’s briefly review some basic characters and
manipulations of data.frame:
• It is a set of rows and columns.
1
https://cran.r-project.org/doc/manuals/r-release/R-data.html#
Acknowledgements
 352 A Handling Large Local Data

• Each row is of the same length and data type

• Every column is of the same length but can be of differing data
types
• It has characteristics of both a matrix and a list
• It uses [] to subset data
We will use the clothes customer data to illustrate. There are two
dimensions in []. The first one indicates the row and second one
indicates column. It uses a comma to separate them.

# read data
sim.dat <- readr::read_csv("http://bit.ly/2P5gTw4")

## Parsed with column specification:

## cols(
## age = col_double(),
## gender = col_character(),
## income = col_double(),
## house = col_character(),
## store_exp = col_double(),
## online_exp = col_double(),
## store_trans = col_double(),
## online_trans = col_double(),
## Q1 = col_double(),
## Q2 = col_double(),
## Q3 = col_double(),
## Q4 = col_double(),
## Q5 = col_double(),
## Q6 = col_double(),
## Q7 = col_double(),
## Q8 = col_double(),
## Q9 = col_double(),
## Q10 = col_double(),
## segment = col_character()
## )
 A.2 data.table— enhanced data.frame 353

# subset the first two rows

sim.dat[1:2, ]
# subset the first two rows and column 3 and 5
sim.dat[1:2, c(3, 5)]
# get all rows with age>70
sim.dat[sim.dat$age > 70, ]
# get rows with age> 60 and gender is Male select column 3 and 4
sim.dat[sim.dat$age > 68 & sim.dat$gender == "Male", 3:4]

Remember that there are usually different ways to conduct the

same manipulation. For example, the following code presents three
ways to calculate an average number of online transactions for male
and female:

tapply(sim.dat$online_trans, sim.dat$gender, mean)

aggregate(online_trans ~ gender, data = sim.dat, mean)

sim.dat %>%
group_by(gender) %>%
summarise(Avg_online_trans = mean(online_trans))

There is no gold standard to choose a specific function to manipu-

late data. The goal is to solve the real problem, not the tool itself.
So just use whatever tool that is convenient for you.
The way to use [] is straightforward. But the manipulations are
limited. If you need more complicated data reshaping or aggrega-
tion, there are other packages to use such as dplyr, reshape2, tidyr
etc. But the usage of those packages are not as straightforward as
[]. You often need to change functions. Keeping related operations
together, such as subset, group, update, join etc, will allow for:
• concise, consistent and readable syntax irrespective of the set of
operations you would like to perform to achieve your end goal
• performing data manipulation fluidly without the cognitive bur-
den of having to change among different functions
 354 A Handling Large Local Data

• by knowing precisely the data required for each operation, you

can automatically optimize operations effectively
data.tableis the package for that. If you are not familiar with
other data manipulating packages and are interested in reducing
programming time tremendously, then this package is for you.
Other than extending the function of [], data.table has the follow-
ing advantages:
• Offers fast import, subset, grouping, update, and joins for large
data files
• It is easy to turn data frame to data table
• Can behave just like a data frame
You need to install and load the package:
Use data.table() to convert the existing data frame sim.dat to data
table:

dt <- data.table(sim.dat)
class(dt)

## [1] "data.table" "data.frame"

Calculate mean for counts of online transactions:

dt[, mean(online_trans)]

## [1] 13.55

You can’t do the same thing using data frame:

sim.dat[,mean(online_trans)]

Error in mean(online_trans) : object 'online_trans' not found

If you want to calculate mean by group as before, set “by =” argu-

ment:
 A.2 data.table— enhanced data.frame 355

dt[ , mean(online_trans), by = gender]

## gender V1
## 1: Female 15.38
## 2: Male 11.26

You can group by more than one variables. For example, group by
“gender” and “house”:

dt[ , mean(online_trans), by = .(gender, house)]

## gender house V1
## 1: Female Yes 11.312
## 2: Male Yes 8.772
## 3: Female No 19.146
## 4: Male No 16.486

Assign column names for aggregated variables:

dt[ , .(avg = mean(online_trans)), by = .(gender, house)]

## gender house avg

## 1: Female Yes 11.312
## 2: Male Yes 8.772
## 3: Female No 19.146
## 4: Male No 16.486

data.tablecan accomplish all operations that aggregate() and tap-

ply()can do for data frame.

• General setting of data.table

Different from data frame, there are three arguments for data ta-
ble:
356 A Handling Large Local Data

It is analogous to SQL. You don’t have to know SQL to learn data

table. But experience with SQL will help you understand data
table. In SQL, you select column j (use command SELECT) for row
i (using command WHERE). GROUP BY in SQL will assign the variable
to group the observations.

Let’s review our previous code:

dt[ , mean(online_trans), by = gender]

The code above is equal to the following SQL:

SELECT
gender,
avg(online_trans)
FROM
sim.dat
GROUP BY
gender

R code:
 A.2 data.table— enhanced data.frame 357

dt[ , .(avg = mean(online_trans)), by = .(gender, house)]

is equal to SQL:

SELECT
gender,
house,
avg(online_trans) AS avg
FROM
sim.dat
GROUP BY
gender,
house

R code:

dt[ age < 40, .(avg = mean(online_trans)), by = .(gender, house)]

is equal to SQL:

SELECT
gender,
house,
avg(online_trans) AS avg
FROM
sim.dat
WHERE
age < 40
GROUP BY
gender,
house

You can see the analogy between data.table and SQL. Now let’s
focus on operations in data table.
• select row
 358 A Handling Large Local Data

# select rows with age<20 and income > 80000

dt[age < 20 & income > 80000]

## age gender income house store_exp online_exp

## 1: 19 Female 83535 No 227.7 1491
## 2: 18 Female 89416 Yes 209.5 1926
## 3: 19 Female 92813 No 186.7 1042
## store_trans online_trans Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8 Q9
## 1: 1 22 2 1 1 2 4 1 4 2 4
## 2: 3 28 2 1 1 1 4 1 4 2 4
## 3: 2 18 3 1 1 2 4 1 4 3 4
## Q10 segment
## 1: 1 Style
## 2: 1 Style
## 3: 1 Style

# select the first two rows

dt[1:2]

## age gender income house store_exp online_exp

## 1: 57 Female 120963 Yes 529.1 303.5
## 2: 63 Female 122008 Yes 478.0 109.5
## store_trans online_trans Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8 Q9
## 1: 2 2 4 2 1 2 1 4 1 4 2
## 2: 4 2 4 1 1 2 1 4 1 4 1
## Q10 segment
## 1: 4 Price
## 2: 4 Price

• select column
Selecting columns in data.table don’t need $:

# select column “age” but return it as a vector

# the argument for row is empty so the result
# will return all observations
 A.2 data.table— enhanced data.frame 359

ans <- dt[, age]

head(ans)

## [1] 57 63 59 60 51 59

To return data.table object, put column names in list():

# Select age and online_exp columns

# and return as a data.table instead
ans <- dt[, list(age, online_exp)]
head(ans)

Or you can also put column names in .():

ans <- dt[, .(age, online_exp)]

To select all columns from “age” to “income”:

ans <- dt[, age:income, with = FALSE]

head(ans,2)

## age gender income

## 1: 57 Female 120963
## 2: 63 Female 122008

Delete columns using - or !:

# delete columns from age to online_exp

ans <- dt[, -(age:online_exp), with = FALSE]
ans <- dt[, !(age:online_exp), with = FALSE]

• tabulation
In data table. .N means to count�
360 A Handling Large Local Data

# row count
dt[, .N]

## [1] 1000

If you assign the group variable, then it will count by groups:

# counts by gender
dt[, .N, by= gender]

## gender N
## 1: Female 554
## 2: Male 446

# for those younger than 30, count by gender

dt[age < 30, .(count=.N), by= gender]

## gender count
## 1: Female 292
## 2: Male 86

Order table:

# get records with the highest 5 online expense:

head(dt[order(-online_exp)],5)

age gender income house store_exp online_exp store_trans ...

1: 40 Female 217599.7 No 7023.684 9479.442 10
2: 41 Female NA Yes 3786.740 8638.239 14
3: 36 Male 228550.1 Yes 3279.621 8220.555 8
4: 31 Female 159508.1 Yes 5177.081 8005.932 11
5: 43 Female 190407.4 Yes 4694.922 7875.562 6
...

Since data table keep some characters of data frame, they share
some operations:
A.2 data.table— enhanced data.frame 361

dt[order(-online_exp)][1:5]

You can also order the table by more than one variable. The fol-
lowing code will order the table by gender, then order within gender
by online_exp:

dt[order(gender, -online_exp)][1:5]

• Use fread() to import dat

Other than read.csv in base R, we have introduced ‘read_csv’ in
‘readr’. read_csv is much faster and will provide progress bar which
makes user feel much better (at least make me feel better). fread()
in data.table further increase the efficiency of reading data. The
following are three examples of reading the same data file topic.csv.
The file includes text data scraped from an agriculture forum with
209670 rows and 6 columns:

system.time(topic <- read.csv("http://bit.ly/2zam5ny"))

user system elapsed

3.561 0.051 4.888

system.time(topic <- readr::read_csv("http://bit.ly/2zam5ny"))

user system elapsed

0.409 0.032 2.233

system.time(topic <- data.table::fread("http://bit.ly/2zam5ny"))

user system elapsed

0.276 0.096 1.117
362 A Handling Large Local Data

It is clear that read_csv() is much faster than read.csv(). fread()

is a little faster than read_csv(). As the size increasing, the differ-
ence will become for significant. Note that fread() will read file as
data.table by default.
B
R code for data simulation

B.1 Customer Data for Clothing Company

The simulation is not very straightforward and we will break it
into three parts:

1. Define data structure: variable names, variable distribu-

tion, customer segment names, segment size
2. Variable distribution parameters: mean and variance
3. Iterate across segments and variables. Simulate data ac-
cording to specific parameters assigned

By organizing code this way, it makes easy for us to change specific

parts of the simulation. For example, if we want to change the
distribution of one variable, we can just change the corresponding
part of the code.
Here is code to define data structure:

# set a random number seed to

# make the process repeatable
set.seed(12345)
# define the number of observations
ncust <- 1000
# create a data frmae for simulated data
seg_dat <- data.frame(id = as.factor(c(1:ncust)))
# assign the variable names
vars <- c("age", "gender", "income", "house", "store_exp",
"online_exp", "store_trans", "online_trans")

363
364 B R code for data simulation

# assign distribution for each variable

vartype <- c("norm", "binom", "norm", "binom", "norm", "norm",
"pois", "pois")
# names of 4 segments
group_name <- c("Price", "Conspicuous", "Quality", "Style")
# size of each segments
group_size <- c(250, 200, 200, 350)

The next step is to define variable distribution parameters. There

are 4 segments of customers and 8 parameters. Different segments
correspond to different parameters. Let’s store the parameters in
a 4×8 matrix:

# matrix for mean

mus <- matrix( c(
# Price
60, 0.5, 120000,0.9, 500,200,5,2,
# Conspicuous
40, 0.7, 200000,0.9, 5000,5000,10,10,
# Quality
36, 0.5, 70000, 0.4, 300, 2000,2,15,
# Style
25, 0.2, 90000, 0.2, 200, 2000,2,20),
ncol=length(vars), byrow=TRUE)

# matrix for variance

sds<- matrix( c(
# Price
3,NA,8000,NA,100,50,NA,NA,
# Conspicuous
5,NA,50000,NA,1000,1500,NA,NA,
# Quality
7,NA,10000,NA,50,200,NA,NA,
# Style
B.1 Customer Data for Clothing Company 365

2,NA,5000,NA,10,500,NA,NA),
ncol=length(vars), byrow=TRUE)

Now we are ready to simulate data using the parameters defined

above:

# simulate non-survey data

sim.dat <- NULL
set.seed(2016)
# loop on customer segment (i)
for (i in seq_along(group_name)) {

# add this line in order to moniter the process

cat(i, group_name[i], "\n")

# create an empty matrix to store relevent data

seg <- data.frame(matrix(NA, nrow = group_size[i],
ncol = length(vars)))

# Simulate data within segment i

for (j in seq_along(vars)) {

# loop on every variable (j)

if (vartype[j] == "norm") {
# simulate normal distribution
seg[, j] <- rnorm(group_size[i], mean = mus[i,
j], sd = sds[i, j])
} else if (vartype[j] == "pois") {
# simulate poisson distribution
seg[, j] <- rpois(group_size[i], lambda = mus[i,
j])
} else if (vartype[j] == "binom") {
# simulate binomial distribution
seg[, j] <- rbinom(group_size[i], size = 1,
prob = mus[i, j])
} else {
366 B R code for data simulation

# if the distribution name is not one of the above, stop

# and return a message
stop("Don't have type:", vartype[j])
}
}
sim.dat <- rbind(sim.dat, seg)
}

Now let’s edit the data we just simulated a little by adding tags
to 0/1 binomial variables:

# assign variable names

names(sim.dat) <- vars
# assign factor levels to segment variable
sim.dat$segment <- factor(rep(group_name, times = group_size))
# recode gender and house variable
sim.dat$gender <- factor(sim.dat$gender, labels = c("Female",
"Male"))
sim.dat$house <- factor(sim.dat$house, labels = c("No",
"Yes"))
# store_trans and online_trans are at least 1
sim.dat$store_trans <- sim.dat$store_trans + 1
sim.dat$online_trans <- sim.dat$online_trans + 1
# age is integer
sim.dat$age <- floor(sim.dat$age)

In the real world, the data always includes some noise such as
missing, wrong imputation. So we will add some noise to the data:

# add missing values

idxm <- as.logical(rbinom(ncust, size = 1, prob = sim.dat$age/200))
sim.dat$income[idxm] <- NA
# add wrong imputations and outliers
set.seed(123)
idx <- sample(1:ncust, 5)
B.1 Customer Data for Clothing Company 367

sim.dat$age[idx[1]] <- 300

sim.dat$store_exp[idx[2]] <- -500
sim.dat$store_exp[idx[3:5]] <- c(50000, 30000, 30000)

So far we have created part of the data. You can check it using
summary(sim.dat). Next, we will move on to simulate survey data.

# number of survey questions

nq <- 10

# mean matrix for different segments

mus2 <- matrix( c( 5,2,1,3,1,4,1,4,2,4, # Price
1,4,5,4,4,4,4,1,4,2, # Conspicuous
5,2,3,4,3,2,4,2,3,3, # Quality
3,1,1,2,4,1,5,3,4,2), # Style
ncol=nq, byrow=TRUE)

# assume the variance is 0.2 for all

sd2 <- 0.2
sim.dat2 <- NULL
set.seed(1000)
# loop for customer segment (i)
for (i in seq_along(group_name)) {
# the following line is used for checking the
# progress cat (i, group_name[i],'\n') create an
# empty data frame to store data
seg <- data.frame(matrix(NA, nrow = group_size[i],
ncol = nq))
# simulate data within segment
for (j in 1:nq) {
# simulate normal distribution
res <- rnorm(group_size[i], mean = mus2[i,
j], sd = sd2)
# set upper and lower limit
res[res > 5] <- 5
res[res < 1] <- 1
368 B R code for data simulation

# convert continuous values to discrete integers

seg[, j] <- floor(res)
}
sim.dat2 <- rbind(sim.dat2, seg)
}

names(sim.dat2) <- paste("Q", 1:10, sep = "")

sim.dat <- cbind(sim.dat, sim.dat2)
sim.dat$segment <- factor(rep(group_name, times = group_size))

B.2 Swine Disease Breakout Data

# sim1_da1.csv the 1st simulated data similar

# sim1_da2 and sim1_da3 sim1.csv simulated data,
# the first simulation dummy.sim1.csv dummy
# variables for the first simulated data with all
# the baseline code for simulation

nf <- 800
for (j in 1:20) {
set.seed(19870 + j)
x <- c("A", "B", "C")
sim.da1 <- NULL
for (i in 1:nf) {
# sample(x, 120, replace=TRUE)->sam
sim.da1 <- rbind(sim.da1, sample(x, 120, replace = TRUE))
}

sim.da1 <- data.frame(sim.da1)

col <- paste("Q", 1:120, sep = "")
row <- paste("Farm", 1:nf, sep = "")
colnames(sim.da1) <- col
B.2 Swine Disease Breakout Data 369

rownames(sim.da1) <- row

# use class.ind() function in nnet package to encode

# dummy variables
library(nnet)
dummy.sim1 <- NULL
for (k in 1:ncol(sim.da1)) {
tmp = class.ind(sim.da1[, k])
colnames(tmp) = paste(col[k], colnames(tmp))
dummy.sim1 = cbind(dummy.sim1, tmp)
}
dummy.sim1 <- data.frame(dummy.sim1)

# set 'C' as the baseline delete baseline dummy variable

base.idx <- 3 * c(1:120)

dummy1 <- dummy.sim1[, -base.idx]

# simulate independent variable for different values of

# r simulate based on one value of r each time r=0.1,
# get the link function

s1 <- c(rep(c(1/10, 0, -1/10), 40),

rep(c(1/10, 0, 0), 40),
rep(c(0, 0, 0), 40))
link1 <- as.matrix(dummy.sim1) %*% s1 - 40/3/10

# Other settings ---------------------------

# r = 0.25
# s1 <- c(rep(c(1/4, 0, -1/4), 40),
# rep(c(1/4, 0, 0), 40),
# rep(c(0, 0, 0), 40))
# link1 <- as.matrix(dummy.sim1) %*% s1 - 40/3/4

# r = 0.5
# s1 <- c(rep(c(1/2, 0, -1/2), 40),
370 B R code for data simulation

# rep(c(1/2, 0, 0), 40),

# rep(c(0, 0, 0), 40))
# link1 <- as.matrix(dummy.sim1) %*% s1 - 40/3/2

# r = 1
# s1 <- c(rep(c(1, 0, -1), 40),
# rep(c(1, 0, 0), 40),
# rep(c(0, 0, 0), 40))
# link1 <- as.matrix(dummy.sim1) %*% s1 - 40/3

# r = 2
# s1 <- c(rep(c(2, 0, -2), 40),
# rep(c(2, 0, 0), 40),
# rep(c(0, 0, 0), 40))
#
# link1 <- as.matrix(dummy.sim1) %*% s1 - 40/3/0.5

# calculate the outbreak probability

hp1 <- exp(link1)/(exp(link1) + 1)

# based on the probability hp1, simulate response

# variable: res
res <- rep(9, nf)
for (i in 1:nf) {
res[i] <- sample(c(1, 0), 1, prob = c(hp1[i], 1 -
hp1[i]))
}

# da1 with response variable, without group indicator

# da2 without response variable, with group indicator
# da3 without response variable, without group indicator

dummy1$y <- res

da1 <- dummy1
y <- da1$y
ind <- NULL
B.2 Swine Disease Breakout Data 371

for (i in 1:120) {
ind <- c(ind, rep(i, 2))
}

da2 <- rbind(da1[, 1:240], ind)

da3 <- da1[, 1:240]

# save simulated data

write.csv(da1, paste("sim", j, "_da", 1, ".csv", sep = ""),
row.names = F)
write.csv(da2, paste("sim", j, "_da", 2, ".csv", sep = ""),
row.names = F)
write.csv(da3, paste("sim", j, "_da", 3, ".csv", sep = ""),
row.names = F)
write.csv(sim.da1, paste("sim", j, ".csv", sep = ""),
row.names = F)
write.csv(dummy.sim1, paste("dummy.sim", j, ".csv",
sep = ""), row.names = F)
}
Bibliography

A, A. and J, A. A. (1984). On the existence of the maximum

likelihood estimates in logistic regression models. Biometrika,
71(1):1–10.
B, E. and R, T. (1986). Bootstrap methods for standard errors,
confidence intervals, and other measures of statistical accuracy.
Statistical Science, pages 54–75.
Bauer, E. and Kohavi, R. (1999). An empirical comparison of vot-
ing classification algorithms: Bagging, boosting, and variants.
Machine Learning, 36:105–142.
Box G, C. D. (1964). An analysis of transformations. Journal of
the Royal Statistical Society, pages 211–252.
Breiman, L. (1998). Arcing classifiers. The Annals of Statistics,
26:123–140.
Breiman, L. (2001a). Random forests. Machine Learning, 45:5–32.
Breiman, L. (2001b). Statistical modeling: The two cultures. Sta-
tistical Science, 16(3):199231.
Breiman, L., Friedman, J. H., Olshen, R. A., and Stone, C. J.
(1984). Classification and regression trees. Number ISBN 978-
0412048418. CRC.
Cestnik, B. and Bratko, I. (1991). Estimating probabilities in tree
pruning. EWSL, pages 138–150.
Chollet, F. (2017). Deep Learning with Python. Manning.
Chollet, F. and Allaire, J. J. (2018). Deep Learning with R. Man-
ning.

373
374 B Bibliography

Chun, H. and Keleş, S. (2010). Sparse partial least squares re-

gression for simultaneous dimension reduction and variable se-
lection. Journal of the Royal Statistical Society: Series B (Sta-
tistical Methodology), 72(1):3–25.
Chung, J., Gulcehre, C., Cho, K., and Bengio, Y. (2014). Empir-
ical evaluation of gated recurrent neural networks on sequence
modeling.
D, M. (1989). Analyzing a portion of the roc curve. Medical
Decision Making, 9:190–195.
de Waal, T., Pannekoek, J., and Scholtus, S. (2011). Handbook of
Statistical Data Editing and Imputation. John Wiley and Sons.
Dudoit S, F. J. and T, S. (2002). Comparison of discrimina-
tion meth- ods for the classification of tumors using gene ex-
pression data. Journal of the American Statistical Association,
97(457):77–87.
Efron, B. (1983). Estimating the error rate of a prediction rule:
Improvement on cross-validation. Journal of the American Sta-
tistical Association, pages 316–331.
Efron, B. and Tibshirani, R. (1986). Bootstrap methods for stan-
dard errors, confidence intervals, and other measures of statis-
tical accuracy. Statistical Science, pages 54–75.
Efron, B. and Tibshirani, R. (1997). Improvements on cross-
validation: The 632+ bootstrap method. Journal of the Ameri-
can Statistical Association, 92(438):548–560.
E.R. DeLong, D.M. DeLong, D. C.-P. (1988). Comparing the areas
under two or more correlated receiver operating characteristics
curves: a nonparametric approach. Biometrics, 44:837–845.
et. al, B. A. (2000). Tissue classification with gene expression
profiles. Journal of Computational Biology, 7(3):559–583.
et. al, B. J. (2006). Aggregate features and adaboost for music
classification. Machine Learning, 65:473–484.
F. Espoito, D. M. and Semeraro, G. (1997). A comparative analysis
B.2 Bibliography 375

of methods for pruning decision trees. IEEE Transactions on

Pattern Analysis and Machine Intelligence, 19(5):476–491.
Freund, Y. and Schapire, R. (1997). A decision-theoretic gen-
eralization of online learning and an application to boosting.
Journal of Computer and System Sciences, 55:119–139.
Friedman, J., Hastie, T., and Tibshirani, R. (2000). Additive logis-
tic regression: A statistical view of boosting. Annals of Statistics,
38:337–374.
Gareth James, Daniela Witten, T. H. and Tibshirani, R. (2015).
An Introduction to Statistical Learning. Springer, 6th edition.
Geladi P, K. B. (1986). Partial least squares regression: A tutorial.
Analytica Chimica Acta, (185):1–17.
Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep Learn-
ing. MIT Press.
Hall P, Hyndman R, F. Y. (2004). Nonparametric confidence in-
tervals for receiver operating characteristic curves. Biometrika,
91:743–750.
Hand D, T. R. (2001). A simple generalisation of the area under
the roc curve for multiple class classification problems. Machine
Learning, 45(2):171–186.
Hastie T, Tibshirani R, F. J. (2008). The Elements of Statistical
Learning: Data Mining, Inference and Prediction. Springer, 2nd
edition.
Hochreiter, S. and Schmidhuber, J. (1997). Long short-term mem-
ory. Neural Computation, 9(8):1735–1780.
Hoerl, A. and Kennard, R. (1970). Ridge regression: Biased estima-
tion for nonorthogonal problems. Technometrics, 12(1):55–67.
HSSINA, B., MERBOUHA, A., EZZIKOURI, H., and ERRITALI,
M. (2014). A comparative study of decision tree id3 and c4.5.
International Journal of Advanced Computer Science and Ap-
plications(IJACSA), Special Issue on Advances in Vehicular Ad
Hoc Networking and Applications 2014, 4(2).
376 B Bibliography

Hyndman, R. and Athanasopoulos, G. (2013). Forecasting: Prin-

ciples and Practice, volume Section 2/5. OTect: Melbourne,
Australia.
Iglewicz, B. and Hoaglin, D. (1993). How to detect and handle out-
liers. The ASQC Basic References in Quality Control: Statistical
Techniques, 16.
J, C. (1960). A coefficient of agreement for nominal data. Educa-
tional and Psychological Measurement, 20:37–46.
Jolliffe, I. (2002). Principla component analysis. Springer, 2nd
edition.
Kuhn, M. and Johnston, K. (2013). Applied Predictive Modeling.
Springer.
Kwak, G. H. J. and Hui, P. (2019). Deephealth: Deep learning
for health informatics reviews, challenges, and opportunities on
medical imaging, electronic health records, genomics, sensing,
and online communication health.
L, B. (1966a). Bagging predictors. Machine Learning, 24(2):123–
140.
L, V. (1984). A theory of the learnable. Communications of the
ACM, 27:1134–1142.
L Meier, S. v. d. G. and Buhlmann, P. (2008). The group lasso
for logistic regression. J. R. Stat. Soc. Ser. B Stat. Methodol,
70:53–71.
Lachiche N, F. P. (2003). Improving accuracy and cost of two–
class and multi–class probabilistic classifiers using roc curves.
In “Proceed- ings of the Twentieth International Conference on
Machine Learning, 20(416–424).
Landis JR, K. G. (1977). The measurement of observer agreement
for categorical data. Biometrics, 33:159–174.
Li J, F. J. (2008). Roc analysis with multiple classes and multiple
tests: Methodology and its application in microarray studies.
Biostatistics, 9(3):566–576.
B.2 Bibliography 377

Line Clemmensen, Trevor Hastie, D. W. and Ersbøll, B. (2011).

Sparse discriminant analysis. Technometrics, 53(4):406–413.
M, K. and L, V. (1989). Cryptographic limitations on learning
boolean formulae and finite automata. In In “Proceedings of the
Twenty-First Annual ACM Symposium on Theory of Comput-
ing.
M, S. T. and F, P. (2007b). Handling missing values when applying
classification models. Journal of Machine Learning Research,
8:1625–1657.
McElreath, R. (2020). Statistical Rethinking: A Bayesian Course
with Examples in R and STAN. Chapman and Hall/CRC.
Mulaik, S. (2009). Foundations of factor analysis. Chapman Hal-
l/CRC, 2nd edition.
Patel, N. and Upadhyay, S. (2012). Study of various decision tree
pruning methods with their empirical comparison in weka. In-
ternational Journal of Computer Applications, 60(12).
Provost F, Fawcett T, K. R. (1998). The case against accuracy
esti- mation for comparing induction algorithms. Proceedings
of the Fifteenth International Conference on Machine Learning,
pages 445–453.
Quinlan, J. (1999). Simplifying decision trees. International Jour-
nal of Human-Computer Studies, 61(2).
R, T. (1996). Regression shrinkage and selection via the lasso.
Journal of the Royal Statistical Society Series B (Methodologi-
cal), 58(1):267–288.
Ronald L. Wassersteina, N. A. L. (2016). Position on p-values:
context, process, and purpose.
Serneels S, Nolf ED, E. P. (2006). Spatial sign preprocessing:
A simple way to impart moderate robustness to multivariate
estimators. Journal of Chemical Information and Modeling,
46(3):1402–1409.
T, D. (2000). An experimental comparison of three methods for
378 B Bibliography

constructing ensembles of decision trees: Bagging, boosting, and

randomization. Machine Learning, 40:139–158.
T, F. (2006). An introduction to roc analysis. Pattern Recognition
Letters, 27(8):861–874.
T, H. (1998). The random subspace method for constructing de-
cision forests. IEEE Transactions on Pattern Analysis and Ma-
chine Intelligence, 13:340–354.
Varmuza K, H. P. and K, F. (2003). Boosting applied to classifica-
tion of mass spectral data. Journal of Data Science, 1(391–404).
W, M. (1965). Principal components regression in exploratory sta-
tistical research. Journal of the American Statistical Association,
60:234–246.
Wedderburn, R. W. M. (1976). On the existence and uniqueness of
the maximum likelihood estimates for certain generalized linear
models. Biometrika, 63:27–32.
Willett, P. (2004). Dissimilarity-based algorithms for selecting
structurally diverse sets of compounds. Journal of Computa-
tional Biology, 6(3-4)(doi:10.1089/106652799318382):447–457.
Wold, H. (1973). Nonlinear iterative partial least squares (nipals)
modelling: Some current developments. Academic Press, pages
383–407.
Wold, H. and Jöreskog, K. G. (1982). Systems Under Indirect
Observation: Causality, Structure, Prediction. North Holland,
Amsterdam.
Y, A. and D, G. (1997). Shape quantization and recognition with
randomized trees. Neural Computation, 9:1545–1588.
Y. Kim, J. K. and Kim, Y. (2006). Blockwise sparse regression.
Statist. Sin, 16:375–390.
Yeo, G. and Burge, C. (2004). Maximum entropy modeling of
short sequence motifs with applications to rna splicing signals.
Journal of Computational Biology, pages 475–494.
B.2 Bibliography 379

YFR, S. (1999). Adaptive game playing using multiplicative

weights. Games and Economic Behavior, 29:79–103.
Yuan, M. and Lin, Y. (2007). Model selection and estimation in
regression with grouped variables. J. R. Stat. Soc. Ser. B Stat.
Methodol, 68:49–67.
Zhang, C., Bengio, S., Hardt, M., Recht, B., and Vinyals, O.
(2017). Understanding deep learning requires rethinking gen-
eralization. arXiv :1611.03530.
Zou, H. and Hastie, T. (2005). Regularization and variable selec-
tion via the elastic net. Journal of the Royal Statistical Society,
Series B,, 67(2):301–320.
Index

Best Linear Unbiased Estimate Partial least square regression

(BLUE), 174 (PLS), 179
BLUE, 174 PCR, 179, 182
PLS, 179, 182
CART, 3 principal component analysis
classification and regression (PCA), 178
tree, 3 Principal component regression
Cook’s distance, 175 (PCR), 178
deep learning, 4 Q-Q Plot, 175
ETL, 6 random forest, 3, 4
extract, transform, and load, 7 residual plot, 174
generalized linear model, 3 RMSE, 182
GLM, 3 Spark, 7
GMB, 4 sparklyr, xi
Hadoop, 7 Standardized residuals plot,
175
Kafka, 7

linear regression, 2, 174

logistic regression, 3

Nonlinear Iterative Partial

Least Squares
(NIPALS), 179

OLS, 174
Ordinary Least Square (OLS),
174

P-value, 171

381