Practical Machine Learning in R

Kyriakos Chatzidimitriou, Themistoklis Diamantopoulos, Michail

Papamichail and Andreas Symeonidis
This book is for sale at http://leanpub.com/practical-machine-learning-r

This version was published on 2018-06-17

This is a Leanpub book. Leanpub empowers authors and publishers with the Lean Publishing
process. Lean Publishing is the act of publishing an in-progress ebook using lightweight tools
and many iterations to get reader feedback, pivot until you have the right book and build
traction once you do.

© 2013 - 2018 Kyriakos Chatzidimitriou, Themistoklis Diamantopoulos, Michail Papamichail

and Andreas Symeonidis
Contents

Part I - Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Chapter 1 - Introduction to R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1 The basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Data frames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5 R Scripts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.6 Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.7 for loops . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.8 Making decisions (if & else) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.9 Datasets and statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.10 Factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.11 Challenge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

Chapter 2 - Introduction to Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . 27

2.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.2 Main categories and tasks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.3 Survival guide to machine learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

Part II - Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Chapter 3 - Classification with Decision trees . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.2 Splitting Criteria and Decision Tree Construction . . . . . . . . . . . . . . . . . . . . 32
3.3 Application with Pruning and Evaluation Metrics . . . . . . . . . . . . . . . . . . . . 36
3.4 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

Chapter 4 - Classification with Naive Bayes . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.2 Naive Bayes Model Construction and Classification . . . . . . . . . . . . . . . . . . . 39
4.3 Naive Bayes Application with Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . 42

Chapter 5 - Classification with k-Nearest Neighbors . . . . . . . . . . . . . . . . . . . . . 45

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
5.2 k-Nearest Neighbors Model Construction and Classification . . . . . . . . . . . . . . 45
5.3 k-Nearest Neighbors Real World Example . . . . . . . . . . . . . . . . . . . . . . . . . 47

Chapter 6 - Classification with Support Vector Machines . . . . . . . . . . . . . . . . . . 57

CONTENTS

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
6.2 Support Vector Machines Model Construction and Classification . . . . . . . . . . . 57
6.3 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

Part III - Data Processing . . . . . . . . . . . . . . . . . . . . . . . . . . 64

Chapter 7 - Feature Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
7.2 Filter Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
7.2 Wrapper Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

Chapter 8 - Dimensionality Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

8.1 Principal Components Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

Part III - Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

Chapter 9 - Centroid-based clustering and Evaluation . . . . . . . . . . . . . . . . . . . . 83
9.1 - k-Means in R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
9.2 - k-Medoids in R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
9.3 - Clustering Evaluation in R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
9.4 - k-Means clustering overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
9.5 - k-Means clustering and evaluation in real life Application . . . . . . . . . . . . . . 94
9.6 - k-Medoids Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99

Chapter 10 - Connectivity-based clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

10.1 - Hierarchical clustering in R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
10.2 - Hierarchical Clustering Application and Evaluation . . . . . . . . . . . . . . . . . 108

Chapter 11 - Density-based clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115

11.1 - DBSCAN in R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
11.2 - DBSCAN model construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
11.3 - DBSCAN calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
11.4 - DBSCAN clustering with R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
11.5 - Density-based Clustering Application . . . . . . . . . . . . . . . . . . . . . . . . . . 119

Chapter 12 - Distribution-based clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124

12.1 - Theoretical background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
12.2 - Modeling Gaussian Mixture Models using EM . . . . . . . . . . . . . . . . . . . . . 125
12.3 - GMMs Clustering Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
12.4 - GMMs Application with Information Criteria . . . . . . . . . . . . . . . . . . . . . 136

Part V - Extended Topics . . . . . . . . . . . . . . . . . . . . . . . . . . . 140

Chapter 13 - Association Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
Part I - Introduction
Chapter 1 - Introduction to R
1.1 The basics
R is an open source programming language and a free environment, mainly used for statistical
computing and graphics. You can find information about R in the official website1 . By searching
with the keyword R with other topic-specific words in sites like Google, one can find additional
information from sites, blog posts, tutorials, documents etc.
Even through R comes with its own environment: command line and graphical interfaces, one
can use the popular RStudio2 , which offers additional graphical functionalities.
When in the R environment (the R prompt is >) one can exit by calling the quit() function or
q() for short. When asked if you want to save the workspace, if you reply with a y for yes, all the
variables that you have during the current R session will be saved into a file named .Rdata, in
the current working directory. If you later start R in the same directory, the variables and their
names will be automatically loaded.
To check which is your current working directory, your can enter:

getwd()

To set the working directory one can use the setwd function. For example setting the working
directory to the Desktop directory in a Linux machine:

setwd("~/Desktop")

What you type at the R prompt is an expression, which R attempts to evaluate and type the result.
For example getwd() is an expression that is evaluated by calling the function getwd() with no
arguments. The same for 42

42

## [1] 42

Console Output
Whatever starts with ## in the book signifies what the reader should see in the console
output.

and the same for

1 https://www.r-project.org/
2 https://www.rstudio.com/
Chapter 1 - Introduction to R 3

(100 * 2 - 12 ^ 2) / 7 * 5 + 2

## [1] 42

There are also predefined constants like pi or e

sin(pi/2)

## [1] 1

To find out the documentation of a specific function you can enter ?sum or help(sum). To
search for functions, there is the help.search("sin") function to help you with that. For certain
functions one can see examples of use by typing the expression example(plot). Comments start
with #, while to assign values to variables you can use <- or =. For example:

a <- 42
b <- (42 + a) / 2
print(a)

## [1] 42

print(b)

## [1] 42

With ls() one can check all the variables existing in the current R session.

ls()

## [1] "a" "accuracy" "b"

## [4] "centers.matrix" "clustering" "col"
## [7] "control" "ctree" "ctrl"
## [10] "d"

ls() output
The output of the ls method will differ from the above, based on what are the current
variables in your session.

To delete all the variables in the current session you can use the call:
Chapter 1 - Introduction to R 4

rm(list=ls())

1.2 Vectors
A vector is a collection of similar types of elements. For example integers. Some things you can
do with vectors in R are shown below.

1. Create the vector a = (10, 5, 3, 100, -2, 5, -50):

a <- c(10, 5, 3, 100, -2, 5, -50)

a

## [1] 10 5 3 100 -2 5 -50

1. Select the elements of the vector a with indices 1, 3, 4, and 5:

a[c(1,3:4)]

## [1] 10 3 100

The above expression uses the c() function for combining values and the : operator that
generates sequences from:to with step 1. Another easy way of specifying sequences is to use
the seq function.

c(1, 2, 7, 10)

## [1] 1 2 7 10

1:10

## [1] 1 2 3 4 5 6 7 8 9 10

seq(1, 6, by=1)

## [1] 1 2 3 4 5 6

seq(1,6, by=2)
Chapter 1 - Introduction to R 5

## [1] 1 3 5

seq(1,by=2, length=6)

## [1] 1 3 5 7 9 11

Type ?seq to get to know the function.

1. To check the type of a variable there is the class function:

class(a)

## [1] "numeric"

1. To check which a elements have a value greater than 5:

a > 5

## [1] TRUE FALSE FALSE TRUE FALSE FALSE FALSE

1. To return the indices of vector a, for which the values are TRUE:

which(a>5)

## [1] 1 4

1. To get the positive elements of a:

b <- a > 0
positives <- a[b]
positives

## [1] 10 5 3 100 5

# or more succintly
positives <- a[a>0]
positives
Chapter 1 - Introduction to R 6

## [1] 10 5 3 100 5

1. To check the length of a vector:

length(a)

## [1] 7

1. One can also bind vectors by column (cbind()) or by row (rbind())

c <- 1:7
rbind(a,c)

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

## a 10 5 3 100 -2 5 -50
## c 1 2 3 4 5 6 7

cbind(a,c)

## a c
## [1,] 10 1
## [2,] 5 2
## [3,] 3 3
## [4,] 100 4
## [5,] -2 5
## [6,] 5 6
## [7,] -50 7

1.3 Matrices
To create matrics use the matrix() function:

matrix(10,3, 2)

## [,1] [,2]
## [1,] 10 10
## [2,] 10 10
## [3,] 10 10

# or
matrix(c(1,2,3,4,5,6), 3, 2)
Chapter 1 - Introduction to R 7

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

# or
matrix(c(1,2,3), 3, 2)

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

But let’s examine how are we calling the matrix function:

args(matrix)

## function (data = NA, nrow = 1, ncol = 1, byrow = FALSE, dimnames = NULL)

## NULL

So the first argument are the data, then with nrow or ncol arguments we can declare the number
of rows and columns and with the argument byrow we declare that we want to fill in the matrix
column-by-column if byrow=FALSE and row-by-row if byrow=TRUE. In the above calls we didn’t
use the byrow argument because the function matrix has a default value byrow=FALSE as we can
also check from the documentation, ?matrix.

m = matrix(1:9, byrow = TRUE, nrow=3)

m

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

## [1,] 1 2 3
## [2,] 4 5 6
## [3,] 7 8 9

Here we have filled in a matrix with values 1 to 9, by row, with the number of rows equal to
3. This gives us a square 3x3 matrix. R is pretty smart in knowing that the number of columns
should be 3 as well!
We can also call cbind and rbind and other functions like rowSums, colSums, mean and t for
transpose:

m2 <- rbind(m, m)
m2
Chapter 1 - Introduction to R 8

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

## [1,] 1 2 3
## [2,] 4 5 6
## [3,] 7 8 9
## [4,] 1 2 3
## [5,] 4 5 6
## [6,] 7 8 9

rowSums(m2)

## [1] 6 15 24 6 15 24

colSums(m2)

## [1] 24 30 36

mean(m2)

## [1] 5

For element wise multiplication on can use the * operator while for matrix multiplication you
can use the %*% operator.

am <- matrix(10:18, byrow = TRUE, nrow = 3)

am

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

## [1,] 10 11 12
## [2,] 13 14 15
## [3,] 16 17 18

bm <- matrix(c(3,6,7,10,8,1,2,3,2), byrow = TRUE, nrow = 3)

bm

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

## [1,] 3 6 7
## [2,] 10 8 1
## [3,] 2 3 2

am * bm
Chapter 1 - Introduction to R 9

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

## [1,] 30 66 84
## [2,] 130 112 15
## [3,] 32 51 36

am %*% bm

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

## [1,] 164 184 105
## [2,] 209 235 135
## [3,] 254 286 165

t(am)

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

## [1,] 10 13 16
## [2,] 11 14 17
## [3,] 12 15 18

1.4 Data frames

Unlike matrices, data frames can store values of different types in their columns. They are used
extensively in R for data analysis. As rows usually we have the observations (or samples) and
as columns we have the characteristics (or attributes or features). When we read from a file, the
result is read as a data frame. Download the zip file r-novice-inflammation.zip3 and unzip it in
the Desktop. Examine the file inflammation-01.csv with a text editor to see what we are going
to be loading. The read the file:

data <- read.csv(file = "data/inflammation-01.csv", header = FALSE)

Notice the use of the path including data/ since we previously set the working directory as the
Desktop with getwd().

dir()

3 http://swcarpentry.github.io/r-novice-inflammation/files/r-novice-inflammation-data.zip
Chapter 1 - Introduction to R 10

## [1] "_site.yml" "ar.html"

## [3] "ar.pdf" "ar.Rmd"
## [5] "breast-cancer-wisconsin.data" "cfs.html"
## [7] "cfs.Rmd" "classification.html"
## [9] "classification.Rmd" "clustering.html"
## [11] "clustering.Rmd" "CNAME"
## [13] "data" "data-eng.html"
## [15] "data-eng.Rmd" "dbscan_files"
## [17] "dbscan.html" "dbscan.pdf"
## [19] "dbscan.Rmd" "docs"
## [21] "dt_files" "dt.html"
## [23] "dt.pdf" "dt.Rmd"
## [25] "finder.html" "finder.Rmd"
## [27] "footer.html" "footer.tex"
## [29] "head.tex" "hierarchical_files"
## [31] "hierarchical.html" "hierarchical.pdf"
## [33] "hierarchical.Rmd" "index.html"
## [35] "index.Rmd" "intro.Rmd"
## [37] "kmeans.pdf" "kmeans.Rmd"
## [39] "knn.Rmd" "ml-tutorials.Rproj"
## [41] "nb.pdf" "nb.Rmd"
## [43] "nn.Rmd" "other.Rmd"
## [45] "outlier.pdf" "outlier.Rmd"
## [47] "pca.Rmd" "r.Rmd"
## [49] "README.md" "render.sh"
## [51] "retail.data.gz" "rfe.Rmd"
## [53] "site_libs" "style.css"
## [55] "uc.Rmd" "wrapper.Rmd"

The dir function return the files and directories of the file system. The argument header=FALSE
lets the read.csv function know that there is no header row to give the columns names.
With head(data) I can check if the data are loaded correctly. It return the first few rows:

head(data)

## V1 V2 V3 V4 V5 V6 V7 V8 V9 V10 V11 V12 V13 V14 V15 V16 V17 V18 V19 V20
## 1 0 0 1 3 1 2 4 7 8 3 3 3 10 5 7 4 7 7 12 18
## 2 0 1 2 1 2 1 3 2 2 6 10 11 5 9 4 4 7 16 8 6
## 3 0 1 1 3 3 2 6 2 5 9 5 7 4 5 4 15 5 11 9 10
## 4 0 0 2 0 4 2 2 1 6 7 10 7 9 13 8 8 15 10 10 7
## 5 0 1 1 3 3 1 3 5 2 4 4 7 6 5 3 10 8 10 6 17
## 6 0 0 1 2 2 4 2 1 6 4 7 6 6 9 9 15 4 16 18 12
## V21 V22 V23 V24 V25 V26 V27 V28 V29 V30 V31 V32 V33 V34 V35 V36 V37 V38
## 1 6 13 11 11 7 7 4 6 8 8 4 4 5 7 3 4 2 3
## 2 18 4 12 5 12 7 11 5 11 3 3 5 4 4 5 5 1 1
## 3 19 14 12 17 7 12 11 7 4 2 10 5 4 2 2 3 2 2
## 4 17 4 4 7 6 15 6 4 9 11 3 5 6 3 3 4 2 3
## 5 9 14 9 7 13 9 12 6 7 7 9 6 3 2 2 4 2 0
## 6 12 5 18 9 5 3 10 3 12 7 8 4 7 3 5 4 4 3
## V39 V40
## 1 0 0
## 2 0 1
## 3 1 1
Chapter 1 - Introduction to R 11

## 4 2 1
## 5 1 1
## 6 2 1

Other functions I can use are:

# type of the variable

class(data)

## [1] "data.frame"

# dimensions
dim(data)

## [1] 60 40

which tells me that I have a data frame with 60 observations (instances) and 40 variables.

# structure
str(data)

## 'data.frame': 60 obs. of 40 variables:

## $ V1 : int 0 0 0 0 0 0 0 0 0 0 ...
## $ V2 : int 0 1 1 0 1 0 0 0 0 1 ...
## $ V3 : int 1 2 1 2 1 1 2 1 0 1 ...
## $ V4 : int 3 1 3 0 3 2 2 2 3 2 ...
## $ V5 : int 1 2 3 4 3 2 4 3 1 1 ...
## $ V6 : int 2 1 2 2 1 4 2 1 5 3 ...
## $ V7 : int 4 3 6 2 3 2 2 2 6 5 ...
## $ V8 : int 7 2 2 1 5 1 5 3 5 3 ...
## $ V9 : int 8 2 5 6 2 6 5 5 5 5 ...
## $ V10: int 3 6 9 7 4 4 8 3 8 8 ...
## $ V11: int 3 10 5 10 4 7 6 7 2 6 ...
## $ V12: int 3 11 7 7 7 6 5 8 4 8 ...
## $ V13: int 10 5 4 9 6 6 11 8 11 12 ...
## $ V14: int 5 9 5 13 5 9 9 5 12 5 ...
## $ V15: int 7 4 4 8 3 9 4 10 10 13 ...
## $ V16: int 4 4 15 8 10 15 13 9 11 6 ...
## $ V17: int 7 7 5 15 8 4 5 15 9 13 ...
## $ V18: int 7 16 11 10 10 16 12 11 10 8 ...
## $ V19: int 12 8 9 10 6 18 10 18 17 16 ...
## $ V20: int 18 6 10 7 17 12 6 19 11 8 ...
## $ V21: int 6 18 19 17 9 12 9 20 6 18 ...
## $ V22: int 13 4 14 4 14 5 17 8 16 15 ...
## $ V23: int 11 12 12 4 9 18 15 5 12 16 ...
## $ V24: int 11 5 17 7 7 9 8 13 6 14 ...
## $ V25: int 7 12 7 6 13 5 9 15 8 12 ...
## $ V26: int 7 7 12 15 9 3 3 10 14 7 ...
## $ V27: int 4 11 11 6 12 10 13 6 6 3 ...
## $ V28: int 6 5 7 4 6 3 7 10 13 8 ...
Chapter 1 - Introduction to R 12

## $ V29: int 8 11 4 9 7 12 8 6 10 9 ...

## $ V30: int 8 3 2 11 7 7 2 7 11 11 ...
## $ V31: int 4 3 10 3 9 8 8 4 4 2 ...
## $ V32: int 4 5 5 5 6 4 8 9 6 5 ...
## $ V33: int 5 4 4 6 3 7 4 3 4 4 ...
## $ V34: int 7 4 2 3 2 3 2 5 7 5 ...
## $ V35: int 3 5 2 3 2 5 3 2 6 1 ...
## $ V36: int 4 5 3 4 4 4 5 5 3 4 ...
## $ V37: int 2 1 2 2 2 4 4 3 2 1 ...
## $ V38: int 3 1 2 3 0 3 1 2 1 2 ...
## $ V39: int 0 0 1 2 1 2 1 2 0 0 ...
## $ V40: int 0 1 1 1 1 1 1 1 0 0 ...

which returns the class and dimensions of the variable data, along with a list of the variables
with their type and their first values.

# statistical summarization of the data frame

summary(data)

## V1 V2 V3 V4 V5
## Min. :0 Min. :0.00 Min. :0.000 Min. :0.00 Min. :1.000
## 1st Qu.:0 1st Qu.:0.00 1st Qu.:1.000 1st Qu.:1.00 1st Qu.:1.000
## Median :0 Median :0.00 Median :1.000 Median :2.00 Median :2.000
## Mean :0 Mean :0.45 Mean :1.117 Mean :1.75 Mean :2.433
## 3rd Qu.:0 3rd Qu.:1.00 3rd Qu.:2.000 3rd Qu.:3.00 3rd Qu.:3.000
## Max. :0 Max. :1.00 Max. :2.000 Max. :3.00 Max. :4.000
## V6 V7 V8 V9
## Min. :1.00 Min. :1.0 Min. :1.000 Min. :2.000
## 1st Qu.:2.00 1st Qu.:2.0 1st Qu.:2.000 1st Qu.:4.000
## Median :3.00 Median :4.0 Median :4.000 Median :5.000
## Mean :3.15 Mean :3.8 Mean :3.883 Mean :5.233
## 3rd Qu.:4.00 3rd Qu.:5.0 3rd Qu.:5.250 3rd Qu.:7.000
## Max. :5.00 Max. :6.0 Max. :7.000 Max. :8.000
## V10 V11 V12 V13
## Min. :2.000 Min. : 2.00 Min. : 2.00 Min. : 3.00
## 1st Qu.:3.750 1st Qu.: 4.00 1st Qu.: 3.75 1st Qu.: 5.00
## Median :6.000 Median : 6.00 Median : 5.50 Median : 9.50
## Mean :5.517 Mean : 5.95 Mean : 5.90 Mean : 8.35
## 3rd Qu.:7.000 3rd Qu.: 9.00 3rd Qu.: 8.00 3rd Qu.:11.00
## Max. :9.000 Max. :10.00 Max. :11.00 Max. :12.00
## V14 V15 V16 V17
## Min. : 3.000 Min. : 3.000 Min. : 3.0 Min. : 4.000
## 1st Qu.: 5.000 1st Qu.: 5.000 1st Qu.: 6.0 1st Qu.: 6.750
## Median : 8.000 Median : 8.000 Median :10.0 Median : 8.500
## Mean : 7.733 Mean : 8.367 Mean : 9.5 Mean : 9.583
## 3rd Qu.:10.000 3rd Qu.:12.000 3rd Qu.:13.0 3rd Qu.:13.000
## Max. :13.000 Max. :14.000 Max. :15.0 Max. :16.000
## V18 V19 V20 V21
## Min. : 5.00 Min. : 5.00 Min. : 5.00 Min. : 5.00
## 1st Qu.: 7.75 1st Qu.: 8.00 1st Qu.: 8.75 1st Qu.: 9.00
## Median :11.00 Median :11.50 Median :13.00 Median :14.00
## Mean :10.63 Mean :11.57 Mean :12.35 Mean :13.25
## 3rd Qu.:13.00 3rd Qu.:15.00 3rd Qu.:16.00 3rd Qu.:16.25
## Max. :17.00 Max. :18.00 Max. :19.00 Max. :20.00
Chapter 1 - Introduction to R 13

## V22 V23 V24 V25

## Min. : 4.00 Min. : 4.00 Min. : 4.00 Min. : 4.0
## 1st Qu.: 8.00 1st Qu.: 7.75 1st Qu.: 6.75 1st Qu.: 7.0
## Median :13.00 Median :11.00 Median :10.00 Median :10.5
## Mean :11.97 Mean :11.03 Mean :10.17 Mean :10.0
## 3rd Qu.:15.25 3rd Qu.:15.00 3rd Qu.:13.25 3rd Qu.:13.0
## Max. :19.00 Max. :18.00 Max. :17.00 Max. :16.0
## V26 V27 V28 V29
## Min. : 3.000 Min. : 3.00 Min. : 3.00 Min. : 3.000
## 1st Qu.: 5.750 1st Qu.: 7.00 1st Qu.: 5.00 1st Qu.: 5.000
## Median : 9.000 Median :10.00 Median : 7.00 Median : 7.000
## Mean : 8.667 Mean : 9.15 Mean : 7.25 Mean : 7.333
## 3rd Qu.:12.000 3rd Qu.:12.00 3rd Qu.: 9.00 3rd Qu.:10.000
## Max. :15.000 Max. :14.00 Max. :13.00 Max. :12.000
## V30 V31 V32 V33
## Min. : 2.000 Min. : 2.000 Min. :2.00 Min. :2.000
## 1st Qu.: 3.000 1st Qu.: 4.000 1st Qu.:4.00 1st Qu.:4.000
## Median : 7.000 Median : 6.000 Median :6.00 Median :5.000
## Mean : 6.583 Mean : 6.067 Mean :5.95 Mean :5.117
## 3rd Qu.:10.000 3rd Qu.: 8.000 3rd Qu.:8.00 3rd Qu.:6.000
## Max. :11.000 Max. :10.000 Max. :9.00 Max. :8.000
## V34 V35 V36 V37 V38
## Min. :1.0 Min. :1.0 Min. :1.000 Min. :1.000 Min. :0.0
## 1st Qu.:2.0 1st Qu.:2.0 1st Qu.:2.000 1st Qu.:2.000 1st Qu.:0.0
## Median :4.0 Median :3.0 Median :4.000 Median :2.000 Median :1.0
## Mean :3.6 Mean :3.3 Mean :3.567 Mean :2.483 Mean :1.5
## 3rd Qu.:5.0 3rd Qu.:5.0 3rd Qu.:5.000 3rd Qu.:4.000 3rd Qu.:3.0
## Max. :7.0 Max. :6.0 Max. :5.000 Max. :4.000 Max. :3.0
## V39 V40
## Min. :0.000 Min. :0.0000
## 1st Qu.:0.000 1st Qu.:0.0000
## Median :1.000 Median :1.0000
## Mean :1.133 Mean :0.5667
## 3rd Qu.:2.000 3rd Qu.:1.0000
## Max. :2.000 Max. :1.0000

There are also different ways I can select slices of data from the data frame:

# first value in data

data[1, 1]

## [1] 0

# middle value in data

data[30, 20]

## [1] 16
Chapter 1 - Introduction to R 14

# first four rows and first ten columns

data[1:4, 1:10]

## V1 V2 V3 V4 V5 V6 V7 V8 V9 V10
## 1 0 0 1 3 1 2 4 7 8 3
## 2 0 1 2 1 2 1 3 2 2 6
## 3 0 1 1 3 3 2 6 2 5 9
## 4 0 0 2 0 4 2 2 1 6 7

# doesn't have to start from the beginning

data[5:10, 1:10]

## V1 V2 V3 V4 V5 V6 V7 V8 V9 V10
## 5 0 1 1 3 3 1 3 5 2 4
## 6 0 0 1 2 2 4 2 1 6 4
## 7 0 0 2 2 4 2 2 5 5 8
## 8 0 0 1 2 3 1 2 3 5 3
## 9 0 0 0 3 1 5 6 5 5 8
## 10 0 1 1 2 1 3 5 3 5 8

# specific rows and columns

data[c(3, 8, 37, 56), c(10, 14, 29)]

## V10 V14 V29

## 3 9 5 4
## 8 3 5 6
## 37 6 9 10
## 56 7 11 9

# All columns from row 5

data[5, ]

## V1 V2 V3 V4 V5 V6 V7 V8 V9 V10 V11 V12 V13 V14 V15 V16 V17 V18 V19 V20
## 5 0 1 1 3 3 1 3 5 2 4 4 7 6 5 3 10 8 10 6 17
## V21 V22 V23 V24 V25 V26 V27 V28 V29 V30 V31 V32 V33 V34 V35 V36 V37 V38
## 5 9 14 9 7 13 9 12 6 7 7 9 6 3 2 2 4 2 0
## V39 V40
## 5 1 1

# All rows from column 16

data[, 16]
Chapter 1 - Introduction to R 15

## [1] 4 4 15 8 10 15 13 9 11 6 3 8 12 3 5 10 11 4 11 13 15 5 14
## [24] 13 4 9 13 6 7 6 14 3 15 4 15 11 7 10 15 6 5 6 15 11 15 6
## [47] 11 15 14 4 10 15 11 6 13 8 4 13 12 9

A subtle point is that the last selection returned a vector instead of a data frame. This is because
we selected only a single column. If you don’t want this behavior do:

# All columns from row 5

d1 <- data[5, ]
class(d1)

## [1] "data.frame"

# All rows from column 16

d2 <- data[, 16]
class(d2)

## [1] "integer"

d3 <- data[, 16, drop=FALSE]

class(d3)

## [1] "data.frame"

Other functions you can call are min, max, mean, sd and median to get statistical values of interest:

# first row, all of the columns

patient_1 <- data[1, ]

# max inflammation for patient 1

max(patient_1)

## [1] 18

# max inflammation for patient 2

max(data[2, ])

## [1] 18

# minimum inflammation on day 7

min(data[, 7])
Chapter 1 - Introduction to R 16

## [1] 1

# mean inflammation on day 7

mean(data[, 7])

## [1] 3.8

# median inflammation on day 7

median(data[, 7])

## [1] 4

# standard deviation of inflammation on day 7

sd(data[, 7])

## [1] 1.725187

To do more complex calculations like the maximum inflammation for all patients, or
the average for each day? we need to apply the function max or mean per row or column
respectivelly. Luckily there is the function apply that applies a function for each one of the
“margins”, 1 for rows and 2 for columns:

args(apply) # args return NULL because it prints the information, but every funct\
ion must return something!

## function (X, MARGIN, FUN, ...)

## NULL

max_patient_inflammation <- apply(data, 1, max)

max_patient_inflammation

## [1] 18 18 19 17 17 18 17 20 17 18 18 18 17 16 17 18 19 19 17 19 19 16 17
## [24] 15 17 17 18 17 20 17 16 19 15 15 19 17 16 17 19 16 18 19 16 19 18 16
## [47] 19 15 16 18 14 20 17 15 17 16 17 19 18 18

avg_day_inflammation <- apply(data, 2, mean)

avg_day_inflammation
Chapter 1 - Introduction to R 17

## V1 V2 V3 V4 V5 V6
## 0.0000000 0.4500000 1.1166667 1.7500000 2.4333333 3.1500000
## V7 V8 V9 V10 V11 V12
## 3.8000000 3.8833333 5.2333333 5.5166667 5.9500000 5.9000000
## V13 V14 V15 V16 V17 V18
## 8.3500000 7.7333333 8.3666667 9.5000000 9.5833333 10.6333333
## V19 V20 V21 V22 V23 V24
## 11.5666667 12.3500000 13.2500000 11.9666667 11.0333333 10.1666667
## V25 V26 V27 V28 V29 V30
## 10.0000000 8.6666667 9.1500000 7.2500000 7.3333333 6.5833333
## V31 V32 V33 V34 V35 V36
## 6.0666667 5.9500000 5.1166667 3.6000000 3.3000000 3.5666667
## V37 V38 V39 V40
## 2.4833333 1.5000000 1.1333333 0.5666667

Now let’s do some plotting

plot(avg_day_inflammation)

max_day_inflammation <- apply(data, 2, max)

plot(max_day_inflammation)
Chapter 1 - Introduction to R 18

min_day_inflammation <- apply(data, 2, min)

plot(min_day_inflammation)

plot is a function with many arguments so you will probably need to study a lot of examples to
do what you want (change an axis, name an axis, change the plot points and/or lines, add title,
add grids, add legend, color the graph, add arrows and text etc.)
Chapter 1 - Introduction to R 19

1.5 R Scripts
So far we have been typing directly into the R command line. What we could also do is save a
sequence of commands in an R source file to run it at will. The way to do this is to have such a
file with an .R extension and use the function source to run it.
If the source file contains an analysis from the beginning to the end it is a good practice to always
clear your session of variables using rm(list=ls()). On the other hand if it is used as a library,
for example to load some functions you have created, then you probably should not do it. You
can also include other source files inside your current source file using the function (you guessed
it): source
Because it is a good practice to have coding guidelines/conventions for standardizing the way
you write your script to make it more readable, Google has some: Google’s R Sytel Guide4

1.6 Functions
Let’s learn how to create functions by creating a fuction fahr_to_kelvin that converts temper-
atures from Fahrenheit to Kelvin:

fahr_to_kelvin <- function(temp) {

kelvin <- ((temp - 32) * (5 / 9)) + 273.15
return(kelvin)
}

To run a function:

# freezing point of water

fahr_to_kelvin(32)

## [1] 273.15

# boiling point of water

fahr_to_kelvin(212)

## [1] 373.15

Let’s also create a function that converts Kelvin to Celcius:

4 https://google.github.io/styleguide/Rguide.xml
Chapter 1 - Introduction to R 20

kelvin_to_celsius <- function(temp) {

celsius <- temp - 273.15
return(celsius)
}

#absolute zero in Celsius

kelvin_to_celsius(0)

## [1] -273.15

We can also use functions inside functions

fahr_to_celsius <- function(temp) {

temp_k <- fahr_to_kelvin(temp)
result <- kelvin_to_celsius(temp_k)
return(result)
}

# freezing point of water in Celsius

fahr_to_celsius(32.0)

## [1] 0

or we can obtain this result by function chaining:

# freezing point of water in Celsius

kelvin_to_celsius(fahr_to_kelvin(32.0))

## [1] 0

1.7 for loops

Like all the programming language R also has for loops to do recurring tasks. In general the
syntax is:

for (variable in collection) {

do things with variable
}

Let’s do an example:
Chapter 1 - Introduction to R 21

best_practice <- c("Let", "the", "computer", "do", "the", "work")

print_words <- function(sentence) {

for (word in sentence) {
print(word)
}
}

print_words(best_practice)

## [1] "Let"
## [1] "the"
## [1] "computer"
## [1] "do"
## [1] "the"
## [1] "work"

or another example:

len <- 0
vowels <- c("a", "e", "i", "o", "u")
for (v in vowels) {
len <- len + 1
}
# Number of vowels
len

## [1] 5

1.8 Making decisions (if & else)

To make decisions in your R scripts, R provides you with the standard if-else conditional
statements

num <- 37
if (num > 100) {
print("greater")
} else {
print("not greater")
}

## [1] "not greater"

Chapter 1 - Introduction to R 22

# or let's create a function that uses conditionals

sign <- function(num) {
if (num > 0) {
return(1)
} else if (num == 0) {
return(0)
} else {
return(-1)
}
}

sign(-3)

## [1] -1

You can make decisions using the logical operators

• equal (==)
• greater than or equal to (>=),
• less than or equal to (<=),
• and not equal to (!=).

We can also combine tests. An ampersand, &, symbolizes the logical and. A vertical bar, |,
symbolizes the logical or.

1.9 Datasets and statistics

By running the function data() we can see some datasets that are currently included in the R
installation. To check for example the iris dataset, we can for example use the function:

head(iris)

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

## 1 5.1 3.5 1.4 0.2 setosa
## 2 4.9 3.0 1.4 0.2 setosa
## 3 4.7 3.2 1.3 0.2 setosa
## 4 4.6 3.1 1.5 0.2 setosa
## 5 5.0 3.6 1.4 0.2 setosa
## 6 5.4 3.9 1.7 0.4 setosa

To see the structure of it:

str(iris)
Chapter 1 - Introduction to R 23

## 'data.frame': 150 obs. of 5 variables:

## $ Sepal.Length: num 5.1 4.9 4.7 4.6 5 5.4 4.6 5 4.4 4.9 ...
## $ Sepal.Width : num 3.5 3 3.2 3.1 3.6 3.9 3.4 3.4 2.9 3.1 ...
## $ Petal.Length: num 1.4 1.4 1.3 1.5 1.4 1.7 1.4 1.5 1.4 1.5 ...
## $ Petal.Width : num 0.2 0.2 0.2 0.2 0.2 0.4 0.3 0.2 0.2 0.1 ...
## $ Species : Factor w/ 3 levels "setosa","versicolor",..: 1 1 1 1 1 1 1 1 \
1 1 ...

For a statistical summary

summary(iris)

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

## Min. :4.300 Min. :2.000 Min. :1.000 Min. :0.100
## 1st Qu.:5.100 1st Qu.:2.800 1st Qu.:1.600 1st Qu.:0.300
## Median :5.800 Median :3.000 Median :4.350 Median :1.300
## Mean :5.843 Mean :3.057 Mean :3.758 Mean :1.199
## 3rd Qu.:6.400 3rd Qu.:3.300 3rd Qu.:5.100 3rd Qu.:1.800
## Max. :7.900 Max. :4.400 Max. :6.900 Max. :2.500
## Species
## setosa :50
## versicolor:50
## virginica :50
##
##
##

For the Species column, one can observe that the summary function did not do the standard
statistical calculation like it did with the other variables. From the str function we can see that the
Species column is a special type of value called factor, which is what R uses to declare categorial
or ordinal values. More on that in the next section.
We can also get the full attribute of Sepal.Length by name through the use of the $ operator:

iris$Sepal.Length

## [1] 5.1 4.9 4.7 4.6 5.0 5.4 4.6 5.0 4.4 4.9 5.4 4.8 4.8 4.3 5.8 5.7 5.4
## [18] 5.1 5.7 5.1 5.4 5.1 4.6 5.1 4.8 5.0 5.0 5.2 5.2 4.7 4.8 5.4 5.2 5.5
## [35] 4.9 5.0 5.5 4.9 4.4 5.1 5.0 4.5 4.4 5.0 5.1 4.8 5.1 4.6 5.3 5.0 7.0
## [52] 6.4 6.9 5.5 6.5 5.7 6.3 4.9 6.6 5.2 5.0 5.9 6.0 6.1 5.6 6.7 5.6 5.8
## [69] 6.2 5.6 5.9 6.1 6.3 6.1 6.4 6.6 6.8 6.7 6.0 5.7 5.5 5.5 5.8 6.0 5.4
## [86] 6.0 6.7 6.3 5.6 5.5 5.5 6.1 5.8 5.0 5.6 5.7 5.7 6.2 5.1 5.7 6.3 5.8
## [103] 7.1 6.3 6.5 7.6 4.9 7.3 6.7 7.2 6.5 6.4 6.8 5.7 5.8 6.4 6.5 7.7 7.7
## [120] 6.0 6.9 5.6 7.7 6.3 6.7 7.2 6.2 6.1 6.4 7.2 7.4 7.9 6.4 6.3 6.1 7.7
## [137] 6.3 6.4 6.0 6.9 6.7 6.9 5.8 6.8 6.7 6.7 6.3 6.5 6.2 5.9

and run the standard statistics:

mean(iris$Sepal.Length)
Chapter 1 - Introduction to R 24

## [1] 5.843333

median(iris$Sepal.Length)

## [1] 5.8

min(iris$Sepal.Length)

## [1] 4.3

max(iris$Sepal.Length)

## [1] 7.9

sd(iris$Sepal.Length)

## [1] 0.8280661

var(iris$Sepal.Length)

## [1] 0.6856935

range(iris$Sepal.Length)

## [1] 4.3 7.9

or other functions like:

sort(iris$Sepal.Length)
Chapter 1 - Introduction to R 25

## [1] 4.3 4.4 4.4 4.4 4.5 4.6 4.6 4.6 4.6 4.7 4.7 4.8 4.8 4.8 4.8 4.8 4.9
## [18] 4.9 4.9 4.9 4.9 4.9 5.0 5.0 5.0 5.0 5.0 5.0 5.0 5.0 5.0 5.0 5.1 5.1
## [35] 5.1 5.1 5.1 5.1 5.1 5.1 5.1 5.2 5.2 5.2 5.2 5.3 5.4 5.4 5.4 5.4 5.4
## [52] 5.4 5.5 5.5 5.5 5.5 5.5 5.5 5.5 5.6 5.6 5.6 5.6 5.6 5.6 5.7 5.7 5.7
## [69] 5.7 5.7 5.7 5.7 5.7 5.8 5.8 5.8 5.8 5.8 5.8 5.8 5.9 5.9 5.9 6.0 6.0
## [86] 6.0 6.0 6.0 6.0 6.1 6.1 6.1 6.1 6.1 6.1 6.2 6.2 6.2 6.2 6.3 6.3 6.3
## [103] 6.3 6.3 6.3 6.3 6.3 6.3 6.4 6.4 6.4 6.4 6.4 6.4 6.4 6.5 6.5 6.5 6.5
## [120] 6.5 6.6 6.6 6.7 6.7 6.7 6.7 6.7 6.7 6.7 6.7 6.8 6.8 6.8 6.9 6.9 6.9
## [137] 6.9 7.0 7.1 7.2 7.2 7.2 7.3 7.4 7.6 7.7 7.7 7.7 7.7 7.9

length(iris$Sepal.Length)

## [1] 150

1.10 Factors
The factor() command is used to create and modify factors in R:

sex <- factor(c("male", "female", "female", "male"))

R will assign 1 to the level "female" and 2 to the level "male" (because f comes before m, even
though the first element in this vector is "male"). You can check this by using the function
levels(), and check the number of levels using nlevels():

levels(sex)

## [1] "female" "male"

nlevels(sex)

## [1] 2

Sometimes, the order of the factors does not matter, other times you might want to specify the
order because it is meaningful (e.g., “low”, “medium”, “high”) or it is required by particular type
of analysis. Additionally, specifying the order of the levels allows us to compare levels:

food <- factor(c("low", "high", "medium", "high", "low", "medium", "high"))

levels(food)

## [1] "high" "low" "medium"

Chapter 1 - Introduction to R 26

food <- factor(food, levels = c("low", "medium", "high"))

levels(food)

## [1] "low" "medium" "high"

min(food) ## doesn't work

## Error in Summary.factor(structure(c(1L, 3L, 2L, 3L, 1L, 2L, 3L), .Label = c("l\
ow", : 'min' not meaningful for factors

food <- factor(food, levels = c("low", "medium", "high"), ordered=TRUE)

levels(food)

## [1] "low" "medium" "high"

min(food) ## works!

## [1] low
## Levels: low < medium < high

Acknowledgments
Parts of this chapter were taken and adapted from Software Carpentry Lessons5 lessons
for R. You can check them out to gain a more in depth knowledge of R.

1.11 Challenge

Air Passengers dataset

Use the dataset AirPassengers that comes with R and refers to number of passengers
traveled every month from 1949 to 1960 in thousands. Because the dataset is of type
Time-Series or ts, you can make it a data.frame through the following commands:

dn = list(paste(“Y”, as.character(1949:1960), sep = “”), month.abb) airmat = ma-

trix(AirPassengers, 12, byrow = TRUE, dimnames = dn) air = as.data.frame(t(airmat))
Then try to answer the next questions/problems:

• Use the help functionality to try and learn about the functions used above.
• How many passengers traveled in average for the year 1951?
• Which is the maximum number of passengers for the months January and
February?
• Calculate the summation per year and assign the result to a vector.
• Plot the vector nicely (names in axes, point and lines for the graph, title the graph,
add grid lines)
• Repeat the last two bullets for every month for all the years.

Tip: to transform a row of the data frame to a vector you can use unlist (e.g.
unlist(air["Jan",])
5 https://software-carpentry.org/lessons/
Chapter 2 - Introduction to Machine
Learning
2.1 Definition
Machine Learning (ML) is a subset of Artificial Intelligence (AI) in the field of computer science
that often uses statistical techniques to give computers the ability to “learn” (i.e., progressively
improve performance on a specific task) with data, without being explicitly programmed.
Machine Learning is often closelly related, if not used as an alternate term, to fields like
Data Mining (the process of discovering patterns in large data sets involving methods at
the intersection of machine learning, statistics, and database systems), Pattern Recognition,
Statistical Inference or Statistical Learning. All these areas often employ the same methods
and perhaps the name changes based on the practitioner’s expertise or the application domain.

2.2 Main categories and tasks

The main ML tasks are typically classified into two broad categories, depending on whether there
is “feedback” or a “teacher” available to the learning system or not.

• Supervised Learning: The system is presented with example inputs and their desired outputs
provided by the “teacher” and the goal of the machine learning algorithm is to create a
mapping from the inputs to the outputs. The mapping can be thought of as a function that
if it is given as an input one of the training samples it should output the desired value.
• Unsupervised Learning: In the unsupervised learning case, the machine learning algorithm
is not given any examples of desired output, and is left on its own to find structure in its
input.

The main machine learning tasks and the ones we will examine in the present textbook, are
separated based on what the system tries to accomplish in the end:

• Classification: inputs are divided into two or more classes, and the learner must produce
a model that assigns unseen inputs to one or more (multi-label classification) of these
classes. This is typically tackled in a supervised manner. Spam filtering is an example of
classification, where the inputs are email (or other) messages and the classes are “spam”
and “not spam”.
• Regression: also a supervised problem, the outputs are continuous rather than discrete.
• Clustering: a set of inputs is to be divided into groups. Unlike in classification, the groups
are not known beforehand, making this typically an unsupervised task.
• Dimensionality Reduction: simplifies inputs by mapping them into a lower-dimensional
space. Topic modeling is a related problem, where a program is given a list of human
language documents and is tasked with finding out which documents cover similar topics.
Chapter 2 - Introduction to Machine Learning 28

• Association Rules learning (or dependency modelling): Searches for relationships between
inputs. For example, a supermarket might gather data on customer purchasing habits. Using
association rule learning, the supermarket can determine which products are frequently
bought together and use this information for marketing purposes. This is sometimes referred
to as market basket analysis.

2.3 Survival guide to machine learning

So you are given a problem. With what methodology are you going to approach it?
First of all you have to understand the problem you are trying to solve with the data you have
(or you are given) and ML algorithms. Is the problem defined formally? Unformally? Both? Are
there any assumptions that need to be made? How one can go and solve it without ML? What is
the motivation behind solving the problem?
The next step is to prepare the data to be used by your favorite ML algorithms since usually data
do not come with the format you need or have other problems. One should consider: 1) what data
are available, 2) if there are any missing values, 3) if any data need to be removed (for example
the student ID, which does not have any predictive value, if we want to predict the average GPA
of the student by the end of the semester), 4) to be organized, cleaned and sampled and 5) to be
transformed either to other values due to domain knowledge (for example image pixel values to
average image intensity or symmetry).
After the data is ready and prepared it is a good practice to select your validation strategy (split,
k-fold cross validation, etc.) and test 10 or more basic ML algorithms. This is useful for getting a
feeling on which algorithm can produce good results for the specific dataset.
From here on one can do algorithm tuning through hyper-parameter optimization methods,
perform more advanced feature selection methods, use ensembles and more…sky is the limit.
Finally, it is good that all the above steps to be logged into an online or offline notepad in order
to be disseminated to the stakeholders and be available for future use: what is the problem, what
is the solution you derived, what are the results, are there any limitations to your approach and
the final conclusions.
So to summarize the main steps are:

1. Problem definition
2. Data preparation
3. Algorithm application
4. Algorithm optimization
5. Result presentation

If someone wants more formally defined processes for working on data problems there are
CRISP-DM and OSEMN. The steps for CRISP-DM6 (check out the diagram below7 ) are:
6 https://en.wikipedia.org/wiki/Cross-industry_standard_process_for_data_mining
7 By Kenneth Jensen - Own work based on: ftp://public.dhe.ibm.com/software/analytics/spss/documentation/modeler/18.0/en/ModelerCRISPDM.pdf
(Figure 1), CC BY-SA 3.0, https://commons.wikimedia.org/w/index.php?curid=24930610
Chapter 2 - Introduction to Machine Learning 29

• Business understanding: understanding the project objectives and requirements from a

business perspective, and then converting this knowledge into a data mining problem
definition, and a preliminary plan designed to achieve the objectives.
• Data understanding: From an initial data collection and proceeds with activities in order
to:
– get familiar with the data,
– identify data quality problems,
– discover first insights into the data, or
– detect interesting subsets to form hypotheses for hidden information.
• Data preparation: This phase deals with the construction of the final dataset from the initial
raw data.
• Modeling: Application of various modeling techniques and parameter calibration to optimal
values.
• Evaluation: Thorough evaluation of the model before final deployment.
• Deployment: The deployment can have many forms, from a generated report to a repeatable
data mining process. Of ocurse we should mention that the creation of the model is
generally not the end of a project.

CRISP-DM diagram

The steps for OSEMN8 are:

• Obtain
• Scrub
• Explore
• Models
• Interpret

8 http://www.dataists.com/2010/09/a-taxonomy-of-data-science/
Part II - Classification
Chapter 3 - Classification with
Decision trees
3.1 Introduction

3.1.1 Introduction to Decision Trees

A Decision Tree is a Machine Learning algorithm that is used mainly for classification problems.
It can be applied to categorical and continuous data. The main concept of the algorithm
is simple; the data are split consecutively according to certain splitting criteria. Decision
trees are computationally efficient when training and quite fast for classifying new instances.
Furthermore, they are comprehensional when they are not too large, therefore they are frequently
used for data exploration. A common pitfall is that they are sensitive to overfitting, which can
be confronted by pruning or by adding further criteria when building the tree.

3.1.2 Decision Trees in R

We may use the rpart library to build decision trees in R:

library(rpart)

Furthermore, we shall use the library rpart.plot to plot the produced trees:

library(rpart.plot)

Building a new tree requires executing the rpart command:

model <- rpart(Target ~ ., method = "class", data = ..., minsplit = ..., minbucke\
t = ..., cp = ...)

We can plot the tree using plot(model) and text(model, use.n = TRUE), or, alternatively, using
the following command:

rpart.plot(model, extra = 104, nn = TRUE)

In order to check the parameters of rpart, we can run ?rpart. Additionally, if we execute
?rpart.control, we can also check the main parameters:

• minsplit: the minimum number of instances in a node so that it is split

• minbucket: the minimum allowed number of instances in each leaf of the tree
• maxdepth: the maximum depth of the tree
• cp: parameter that controls the complexity for a split and is set intuitively (the larger its
value, the more probable to apply pruning to the tree)
Chapter 3 - Classification with Decision trees 32

3.2 Splitting Criteria and Decision Tree Construction

To compute the values of different splitting criteria, we will use as an example the training data
of the following table, which refer to a problem of binary classification.

Outlook Temperature Humidity Play

Sunny Hot High No
Sunny Hot Low No
Rainy Hot Low Yes
Rainy Cool High Yes
Rainy Cool Low Yes
Rainy Hot Low No
Rainy Cool Low Yes
Sunny Hot High No
Sunny Cool Low Yes
Rainy Hot Low Yes
Sunny Cool Low Yes
Rainy Hot High Yes
Rainy Cool Low Yes
Sunny Cool High No

We are going to apply a decision tree algorithm on the above dataset. We will answer to the
following questions:
a) If we split using the Gini index, on which feature (out of Outlook, Temperature, and Humidity)
should we perform the first split?
b) If we split using the Information gain, on which feature (out of Outlook, Temperature, and
Humidity) should we perform the first split?
c) Build the full decision tree using the Gini index. Also, plot the tree.

3.2.1 Data and Library Imports

Initially, we read the data and import the required librares:

weather = read.csv("weather.txt")
library(rpart)
library(rpart.plot)

3.2.2 Splitting Criteria

To see the split for the variable Outlook, we execute the following command:

model <- rpart(Play ~ Outlook, method = "class", data = weather, minsplit = 1)

If we also run this command for Temperature and Humidity, and then plot all three with
rpart.plot(model, extra = 104, nn = TRUE)), then the following visualizations are provided:
Chapter 3 - Classification with Decision trees 33

Intuitively, we can already understand which of the three splits will be selected, right?

3.2.2.1 Gini Index

We compute the Gini index for Outlook using the following formulas:

GIN I(Sunny) =
1 − F req(P lay = N o|Outlook = Sunny)2 − F req(P lay = Y es|Outlook = Sunny)2

GIN I(Rainy) =
1 − F req(P lay = N o|Outlook = Rainy)2 − F req(P lay = Y es|Outlook = Rainy)2

GIN IOutlook =
F req(Outlook = Sunny) · GIN I(Sunny) + F req(Outlook = Rainy) · GIN I(Rainy)

The calculations provide:

GIN I(Sunny) = 1 − (4/6)2 − (2/6)2 = 0.444

GIN I(Rainy) = 1 − (1/8)2 − (7/8)2 = 0.219

GIN IOutlook = (6/14) · 0.444 + (8/14) · 0.219 = 0.315

Similarly, the GINI for Temperature and Humidity is GIN IT emp. = 0.367 and GIN IT emp. = 0.394,
respectively. Hence, answering to question (a), the optimal first split using the Gini index is on
Outlook.
We can also make these computations using R. For Outlook, we build the following frequency
arrays:
Chapter 3 - Classification with Decision trees 34

absfreq = table(weather[, c(1, 4)])

freq = prop.table(absfreq, 1)
freqSum = rowSums(prop.table(absfreq))

We calculate the Gini index for Sunny and Rainy:

GINI_Sunny = 1 - freq["Sunny", "No"]^2 - freq["Sunny", "Yes"]^2

GINI_Rainy = 1 - freq["Rainy", "No"]^2 - freq["Rainy", "Yes"]^2

The total Gini for Outlook is computed by:

GINI_Outlook = freqSum["Sunny"] * GINI_Sunny + freqSum["Rainy"] * GINI_Rainy

3.2.2.2 Information Gain

We compute the Information Gain for Outlook using the following formulas:

Entropy(All) = −F req(N o) · log(F req(N o)) − F req(Y es) · log(F req(Y es))

Entropy(Sunny) =
− F req(N o|Sunny) · log(F req(N o|Sunny)) − F req(Y es|Sunny) · log(F req(Y es|Sunny))

Entropy(Rainy) =
− F req(N o|Rainy) · log(F req(N o|Rainy)) − F req(Y es|Rainy) · log(F req(Y es|Rainy))

GAINOutlook =
Entropy(All) − F req(Sunny) · Entropy(Sunny) − F req(Rainy) · Entropy(Rainy)

The calculations provide:

Entropy(All) = −(5/14) · log(5/14) − (9/14) · log(9/14) = 0.652

Entropy(Sunny) = −(4/6) · log(4/6) − (2/6) · log(2/6) = 0.637

Entropy(Rainy) = −(1/8) · log(1/8) − (7/8) · log(7/8) = 0.377

GAINOutlook = 0.652 − (6/14) · 0.637 + (8/14) · 0.377 = 0.164

Similarly, the Information Gain for Temperature and Humidity is GAINT emperature = 0.105 and
GAINT emperature = 0.071, respectively. Hence, answering to question (a), the optimal first split
using the Information Gain is on Outlook.
Chapter 3 - Classification with Decision trees 35

We can also make these computations using R. Initially, we compute the total entropy of the
dataset:

freq = prop.table(table(weather[, c(4)]))

Entropy_All = - freq["No"] * log(freq["No"]) - freq["Yes"] * log(freq["Yes"])

Then, for Outlook, we build the following frequency arrays:

absfreq = table(weather[, c(1, 4)])

freq = prop.table(absfreq, 1)
freqSum = rowSums(prop.table(absfreq))

We calculate the entropy for Sunny and Rainy:

Entropy_Sunny =
- freq["Sunny", "No"] * log(freq["Sunny", "No"])
- freq["Sunny", "Yes"] * log(freq["Sunny", "Yes"])
Entropy_Rainy =
- freq["Rainy", "No"] * log(freq["Rainy", "No"])
- freq["Rainy", "Yes"] * log(freq["Rainy", "Yes"])

The total information gain for Outlook is computed by:

GAIN_Outlook =
Entropy_All
- freqSum["Sunny"] * Entropy_Sunny
- freqSum["Rainy"] * Entropy_Rainy

3.2.3 Building the Decision Tree

For question (c), we build a decision tree using rpart:

model <- rpart(

Play ~ Outlook + Temperature + Humidity,
method = "class",
data = weather,
minsplit = 1,
minbucket = 1,
cp = -1)

Furthermore, we can plot the tree with the following command:

rpart.plot(model, extra = 104, nn = TRUE)

The tree is shown in the following figure:

Chapter 3 - Classification with Decision trees 36

3.3 Application with Pruning and Evaluation Metrics

A decision tree classifier can also be applied on datasets with continuous predictor variables, such
as the iris dataset. From the iris dataset, we shall only keep the first 2 columns of the dataset and
drop the last 50 instances, so that we now have a binary classification problem with two predictor
features and one target class. We can perform these transformations in R as follows:
iris2 = iris[, c(1, 2, 5)]
iris2$Species[c(101:150)] = iris2$Species[c(21:70)]
iris2$Species = factor(iris2$Species)

After that, we split the dataset into training and testing data:
trainingdata = iris2[c(1:40, 51:90, 101:140),]
testdata = iris2[c(41:50, 91:100, 141:150),]

We will answer to the following questions:

a) Build a decision tree using the training data (the minsplit parameter of rpart should be set to
20, which is the default).
b) Apply the model on testdata and compute the precision, the recall, and the f-measure for the
two classes.
c) Build two decision trees, one for minsplit equal to 10 and one for minsplit equal to 30, apply
them on testdata and compute the precision, the recall, and the f-measure for the two classes.
d) Compare the three models given their f-measure for class versicolor.

3.3.1 Decision Tree Training and Testing

We train the decision tree and plot it using the following commands (questions (a), (c)):
Chapter 3 - Classification with Decision trees 37

model <- rpart(

Species ~ .,
method = "class",
data = trainingdata,
minsplit = 20)
rpart.plot(model, extra = 104, nn = TRUE)

The . in the formula stands for all the remaining variables in the data frame trainingdata. We
can also execute our model on the test set using the commands:

xtest = testdata[,1:2]
ytest = testdata[,3]
pred = predict(model, xtest, type="class")

3.3.2 Evaluation Metrics

We can produce the confusion matrix and compute useful metrics with the following commands
(questions (b), (c)):

library(MLmetrics)
cm = ConfusionMatrix(pred, ytest)
accuracy = Accuracy(pred, ytest)
precision = Precision(ytest, pred, 'versicolor')
recall = Recall(ytest, pred, 'versicolor')
f1 = F1_Score(ytest, pred, 'versicolor')
data.frame(precision, recall, f1)

(alternatively we can compute TP, FP, TN, and FN and find precision as TP/(TP + FP) and recall
as TP/(TP + FN))
Finally, the F-measure for the 3 models is shown in the following array (question (d)):

Decision Tree F-Measure

minsplit = 10 0.833
minsplit = 20 0.947
minsplit = 30 0.857

3.4 Exercise
You are given the training data of the following array for a binary classification problem.
Chapter 3 - Classification with Decision trees 38

CustomerID Sex CarType Budget Insurance

1 M Family Low No
2 M Sport Medium No
3 M Sport Medium No
4 M Sport High No
5 M Sport VeryHigh No
6 M Sport VeryHigh No
7 F Sport Low No
8 F Sport Low No
9 F Sport Medium No
10 F Sedan High No
11 M Family High Yes
12 M Family VeryHigh Yes
13 M Family Medium Yes
14 M Sedan VeryHigh Yes
15 F Sedan Low Yes
16 F Sedan Low Yes
17 F Sedan Medium Yes
18 F Sedan Medium Yes
19 F Sedan Medium Yes
20 F Sedan High Yes

Calculate the Gini index for:

a) all the training instances
b) the feature CustomerID
c) the feature Sex
d) the feature CarType using multiway split
e) the feature Budget using multiway split
Which of these should be used for the first split according to the Gini index? Hint-question:
Explain why CustomerID should not be used as a feature for splitting even though it has the
lowest Gini index.
Chapter 4 - Classification with Naive
Bayes
4.1 Introduction

4.1.1 Introduction to Naive Bayes

Naive Bayes is a machine learning algorithm that is based on the Bayes theorem as well as the
assumption of independence between the features of the dataset. Given a dataset with a set of fea-
tures x1 , x2 , . . . , xn and a class feature c, the Bayes theorem defines the probability that an instance
{x1 , x2 , . . . , xn } belongs to c as the combination of the probabilities P (x1 |c), P (x2 |c), . . . , P (xn |c):

P (x1 |c) · P (x2 |c) · · · · · P (xn |c) · P (c)

P (c|x) =
P (x1 ) · P (x2 ) · · · · · P (xn )

where we notice that the probability of the set of features is given by their product, as their
values are independent from one another.

4.1.2 Naive Bayes in R

We may use the e1071 library to build the probabilistic Naive Bayes model in R:

library(e1071)

Building a model requires executing the naiveBayes command:

model <- naiveBayes(Target ~ ., data = ..., laplace = ...)

where parameter laplace is used to select laplace smoothing.

Given a trained model, we can predict the class for a new instance using the command:

predict(model, trvalue)

while if we further add the parameter type = "raw", we are also provided with the posterior
probabilities.

4.2 Naive Bayes Model Construction and

Classification
We will use the data of the following table for a problem of binary classification:
Chapter 4 - Classification with Naive Bayes 40

Weather Day HighTraffic

Hot Vacation No
Cold Work Yes
Normal Work No
Cold Weekend Yes
Normal Weekend Yes
Cold Work No
Hot Work No
Hot Vacation Yes

We will answer to the following questions:

a) Using the Naive Bayes Classifier, in which class would you classify a new instance with values
(Weather, Day) = (Hot, Vacation)?
b) Using the Naive Bayes Classifier, in which class would you classify a new instance with values
(Weather, Day) = (Hot, Weekend)?
c) Repeat question (b) using Laplace smoothing.
d) Build the full model in R and repeat question (a).
e) Build the full model in R using laplace smoothing and repeat question (b).

4.2.1 Data and Library Imports

Initially, we read the data and import the required librares:

traffic = read.csv("traffic.txt")
library(e1071)

4.2.2 Probability Calculation

For question (a), we compute the probability using the following formulas:

P (Hot|Y es) · P (V acation|Y es) · P (Y es) 1/4 · 1/4 · 1/2

P (Y es|Hot, V acation) = = = 1/3
P (Hot) · P (V acation) 3/8 · 2/8

P (Hot|N o) · P (V acation|N o) · P (N o) 2/4 · 1/4 · 1/2

P (N o|Hot, V acation) = = = 2/3
P (Hot) · P (V acation) 3/8 · 2/8

Since P (N o|Hot, V acation) > P (Y es|Hot, V acation), the algorithm classifies the instance to No.
Similarly, for question (b):

P (Hot|Y es) · P (W eekend|Y es) · P (Y es) 1/4 · 2/4 · 1/2

P (Y es|Hot, W eekend) = = = 2/3
P (Hot) · P (W eekend) 3/8 · 2/8

P (Hot|N o) · P (W eekend|N o) · P (N o) 2/4 · 0/4 · 1/2

P (N o|Hot, W eekend) = = =0
P (Hot) · P (W eekend) 3/8 · 2/8

Since P (N o|Hot, W eekend) < P (Y es|Hot, W eekend), the algorithm classifies the instance to Yes.
Chapter 4 - Classification with Naive Bayes 41

Concerning question (c), we will re-compute the probabilities, however adding 1 as the Lapla-
cian parameter. So, given the new probabilities are P (Hot|Y es) = (1 + 1)/(4 + 3) = 2/7,
P (W eekend|Y es) = (2 + 1)/(4 + 3) = 3/7, P (Hot|N o) = (2 + 1)/(4 + 3) = 3/7, and P (W eekend|N o) =
(0 + 1)/(4 + 3) = 1/7, we compute:

P (Hot|Y es) · P (W eekend|Y es) · P (Y es) 2/7 · 3/7 · 1/2

P (Y es|Hot, W eekend) = = = 32/49
P (Hot) · P (W eekend) 3/8 · 2/8

P (Hot|N o) · P (W eekend|N o) · P (N o) 3/7 · 1/7 · 1/2

P (N o|Hot, W eekend) = = = 16/49
P (Hot) · P (W eekend) 3/8 · 2/8

Since P (N o|Hot, W eekend) < P (Y es|Hot, W eekend), the algorithm classifies the instance to Yes.

4.2.3 Building the Naive Bayes Model

4.2.3.1 Building Simple Model

For question (d), we build a naive bayes model:

model <- naiveBayes(HighTraffic ~ ., data = traffic)

Furthermore, we can print the model:

print(model)

The output is shown below:

Naive Bayes Classifier for Discrete Predictors

Call: naiveBayes.default(x = X, y = Y, laplace = laplace)

A-priori probabilities:
Y
No Yes
0.5 0.5

Conditional probabilities:
Weather
Y Cold Hot Normal
No 0.25 0.50 0.25
Yes 0.50 0.25 0.25

Day
Y Vacation Weekend Work
No 0.25 0.00 0.75
Yes 0.25 0.50 0.25

We may now re-compute question (a):

Chapter 4 - Classification with Naive Bayes 42

trvalue <- data.frame(Weather = factor("Hot", levels(traffic$Weather)), Day = fa\

ctor("Vacation", levels(traffic$Day)))
predict(model, trvalue)

We can also get the relevant probabilities:

predict(model, trvalue, type = "raw")

4.2.3.2 Building Model with Laplace Smoothing

For question (c), we build a naive bayes model with laplace smoothing:
model <- naiveBayes(HighTraffic ~ ., data = traffic, laplace = 1)

We may now re-compute question (a):

trvalue <- data.frame(Weather = factor("Hot", levels(traffic$Weather)), Day = fa\
ctor("Weekend", levels(traffic$Day)))
predict(model, trvalue)

We can also get the relevant probabilities:

predict(model, trvalue, type = "raw")

4.3 Naive Bayes Application with Evaluation

We will apply the Naive Bayes classifier on the dataset HouseVotes84. Initially, we import the
dataset, ignoring any instances with missing values:
data(HouseVotes84, package = "mlbench")
votes = na.omit(HouseVotes84)

Furthermore, we import the libraries required to execute and evaluate the model:
library(e1071)
library(MLmetrics)
library(ROCR)

After that, we split the dataset into training and testing data:
trainingdata = votes[1:180,]
testingdata = votes[181:232,]

We will answer to the following questions:

a) Build a Naive Bayes model using the training data and apply it on the test data.
b) Compute the precision, the recall and the F-measure for the test data for the class democrat.
c) Build the ROC curve for the model.

4.3.1 Building and Applying the Naive Bayes Model

For question (a), we train a the naive bayes model:
Chapter 4 - Classification with Naive Bayes 43

model <- naiveBayes(Class ~ ., data = trainingdata)

After that, we can apply the model on the test set:

xtest = testingdata[,-1]
ytest = testingdata[,1]
pred = predict(model, xtest)
predprob = predict(model, xtest, type = "raw")

4.3.2 Metrics Computation and ROC Curve Plotting

We can see the confusion matrix and calculate useful metrics with the following commands
(question (b)):

ConfusionMatrix(ytest, pred)
Precision(ytest, pred, "democrat")
Recall(ytest, pred)

Plotting the ROC curve (question (c)) initially requires computing TPR and FRP with the
following commands:

pred_obj = prediction(predprob[,1], ytest, label.ordering = c("republican", "demo\

crat"))
ROCcurve <- performance(pred_obj, "tpr", "fpr")

If we plot the ROCcurve object, we shall see TPR and FRP and the corresponding thresholds.
Finally, we can plot the curve as follows:

plot(ROCcurve, col = "blue")

abline(0,1, col = "grey")

And find the area under the curve using the command:

performance(pred_obj, "auc")
Chapter 4 - Classification with Naive Bayes 44

ROC curves
In statistics, a receiver operating characteristic (ROC), or ROC curve, is a graphical
plot that illustrates the performance of a binary classifier system as its discrimination
threshold is varied. The curve is created by plotting the true positive rate (TPR) against
the false positive rate (FPR) at various threshold settings
An ROC curve demonstrates several things:

• It shows the tradeoff between sensitivity and specificity (any increase in sensitivity
will be accompanied by a decrease in specificity).
• The closer the curve follows the left-hand border and then the top border of the
ROC space, the more accurate the classifier.
• The closer the curve comes to the 45-degree diagonal of the ROC space, the less
accurate the test.
• The slope of the tangent line at a cutpoint gives the likelihood ratio (LR) for that
value of the test.
• The area under the curve is a measure of text accuracy. This is discussed further
in the next section.

Accuracy is measured by the area under the ROC curve. An area of 1 represents a perfect
test; an area of .5 represents a worthless test. A rough guide for classifying the accuracy
of a diagnostic test is the traditional academic point system:

• .90-1 = excellent (A)

• .80-.90 = good (B)
• .70-.80 = fair (C)
• .60-.70 = poor (D)
• .50-.60 = fail (F)
Chapter 5 - Classification with
k-Nearest Neighbors
5.1 Introduction

5.1.1 Introduction to k-Nearest Neighbors

k-Nearest Neighbors (kNN) is a machine learning algorithm, where the prediction for a new
instance depends on its k nearest instances. As a result, the class for a new instance is determined
by majority vote given the class of its k nearest instances.

5.1.2 k-Nearest Neighbors in R

We may use the class library to apply the kNN model in R:

library(class)

A new instance may be classified using the command knn:

knn(X_train, X_test, Y_train, k = 1, prob = TRUE)

where we have to provide the training data (X_train, Y_train), the new instances (X_test), and
the value of k. If we also add the parameter prob = TRUE, then the probabilities for the class are
also given.

5.2 k-Nearest Neighbors Model Construction and

Classification
We will use the data of the following table for a problem of binary classification:

X1 X2 Y
0.7 0.7 A
0.7 0.8 A
0.6 0.6 A
0.5 0.5 A
0.5 0.6 A
0.5 0.7 A
0.5 0.8 A
0.7 0.5 B
0.8 0.7 B
0.8 0.5 B
Chapter 5 - Classification with k-Nearest Neighbors 46

X1 X2 Y
0.8 0.6 B
1.0 0.3 B
1.0 0.5 B
1.0 0.6 B

We will answer to the following questions:

a) Plot the data with different colors/symbols for each class.
b) Using kNN with k = 1, in which class would you classify a new instance with values (X1,
X2) = (0.7, 0.4)?

c) Repeat question (b) using k = 5.

d) Using kNN with k = 1, in which class would you classify a new instance with values (X1,
X2) = (0.7, 0.6)?

5.2.1 Data and Library Imports

Initially, we read the data and import the required librares:

knndata = read.csv("knndata.txt")
library(class)

5.2.2 Algorithm Application

We initially split the dataset as follows:

X_train = knndata[,c("X1","X2")]
Y_train = knndata$Y

For question (a) we execute the command:

plot(X_train, col = Y_train, pch = c("o","+")[Y_train])

Chapter 5 - Classification with k-Nearest Neighbors 47

When k = 1 (question (b)), we have to find the nearest instance (using euclidean distance) to
(0.7, 0.4), which is (0.7, 0.5) that belongs to class B. So, we execute the command:

knn(X_train, c(0.7, 0.4), Y_train, k = 1, prob = TRUE)

and the new instance is classified to class B with probability 1.

When k = 5 (question (c)), we have to find the 5 nearest instances to (0.7, 0.4), which are (0.7,
0.5), (0.8, 0.5), (0.6, 0.6), (0.8, 0.6), (0.7, 0.7) and belong to classes B, B, A, B, A respectively. We
note that 3 of them (out of 5) belong to class B. We may execute the command:

knn(X_train, c(0.7, 0.4), Y_train, k = 5, prob = TRUE)

and the new instance is classified to class B with probability 0.5.

For question (d), we would like to find the nearest instance to (0.7, 0.6). We observe that instances
(0.7, 0.5), (0.7, 0.5) (0.6, 0.6), (0.8, 0.6), (0.7, 0.7), (0.7, 0.7) (that belong to classes B, B, A, B, A, B
respectively) all have the same distance from (0.7, 0.6). So (and since the parameter use.all of
kNN is TRUE), the algorithm will use all of these instances. We note that 4 of them (out of 6)
belong to class B. We may execute the command:

knn(X_train, c(0.7, 0.6), Y_train, k = 1, prob = TRUE)

and the new instance is classified to class B with probability 0.66.

5.3 k-Nearest Neighbors Real World Example

In this sub-section we will apply k-NN in the breast cancer Wisconsin dataset.
Chapter 5 - Classification with k-Nearest Neighbors 48

Breast Cancer Wisconsin (Original) Data Set

For this example we are going to use the Breast Cancer Wisconsin (Original) Data Set9
and in particular the breast-cancer-wisconsin.data file10 from the UCI Machine Learning
Repository11 .

5.3.1 Data Handling

First we are going to load the dataset as a dataframe. We are assuming that the current working
directory is in the same directory where the dataset is stored. We add the sep option because the
default separator is the empty string. In addition, as one can observe from the dataset instructions,
the missing values are denoted with ?. To check the documentation of the read.table function
use the command ?read.table. Next we will split the dataset into training and validation
datasets. The steps for loading and splitting the dataset to training and validation are the
following:

rm(list=ls())
# download the dataset
fileURL <- "http://archive.ics.uci.edu/ml/machine-learning-databases/breast-cance\
r-wisconsin/breast-cancer-wisconsin.data"
download.file(fileURL, destfile="breast-cancer-wisconsin.data", method="curl")
# read the data
data <- read.table("breast-cancer-wisconsin.data", na.strings = "?", sep=",")
# remove the id column
data <- data[,-1]
# put names in the columns (attributes)
names(data) <- c("ClumpThickness",
"UniformityCellSize",
"UniformityCellShape",
"MarginalAdhesion",
"SingleEpithelialCellSize",
"BareNuclei",
"BlandChromatin",
"NormalNucleoli",
"Mitoses",
"Class")
# make the class a factor
data$Class <- factor(data$Class, levels=c(2,4), labels=c("benign", "malignant"))
# set the seed
set.seed(1234)
# split the dataset
ind <- sample(2, nrow(iris), replace=TRUE, prob=c(0.7, 0.3))
trainData <- data[ind==1,]
validationData <- data[ind==2,]

5.3.2 Training & Prediction

We load the k-NN algorithm as before.
9 http://archive.ics.uci.edu/ml/datasets/Breast+Cancer+Wisconsin+%28Original%29
10 http://archive.ics.uci.edu/ml/machine-learning-databases/breast-cancer-wisconsin/breast-cancer-wisconsin.data
11 http://archive.ics.uci.edu/ml/index.html
Chapter 5 - Classification with k-Nearest Neighbors 49

library(class)

Because all the attributes are between 1 and 10 there is no need to do normalization between 0
and 1, since no attribute will dominate the others in the distance calculation of kNN. Because
kNN accepts the training and testing datasets without the target column, which puts in a 3rd
argument, we are going to do some data manipulation to have the data the way the knn function
likes them (look the manual with ?knn). Also, because no missing values are allowed in kNN,
let’s remove those too.

trainData <- trainData[complete.cases(trainData),]

validationData <- validationData[complete.cases(validationData),]
trainDataX <- trainData[,-ncol(trainData)]
trainDataY <- trainData$Class
validationDataX <- validationData[,-ncol(trainData)]
validationDataY <- validationData$Class

Lets predict, since there is no need to training when using kNN. The training instances are the
model.

prediction = knn(trainDataX, validationDataX, trainDataY, k = 1)

You can play with the values of k to look for a better model.

5.3.3 Evaluation
Make the predictions for the validation dataset and print the confusion matrix:

cat("Confusion matrix:\n")
xtab = table(prediction, validationData$Class)
print(xtab)
cat("\nEvaluation:\n\n")
accuracy = sum(prediction == validationData$Class)/length(validationData$Class)
precision = xtab[1,1]/sum(xtab[,1])
recall = xtab[1,1]/sum(xtab[1,])
f = 2 * (precision * recall) / (precision + recall)
cat(paste("Accuracy:\t", format(accuracy, digits=2), "\n",sep=" "))
cat(paste("Precision:\t", format(precision, digits=2), "\n",sep=" "))
cat(paste("Recall:\t\t", format(recall, digits=2), "\n",sep=" "))
cat(paste("F-measure:\t", format(f, digits=2), "\n",sep=" "))

and the output:

Chapter 5 - Classification with k-Nearest Neighbors 50

## Confusion matrix:
##
## prediction benign malignant
## benign 99 4
## malignant 6 60
##
## Evaluation:
## Accuracy: 0.94
## Precision: 0.94
## Precision: 0.94
## F-measure: 0.95

5.3.4 Cross Validation

For performing cross-validation we will use the caret package. Here12 you can find a quick guide
to caret.

5.3.4.1 Split data in two groups

The function createDataParitiion does a stratified random split of the data. Similar to what
we did above ourselves (not stratified though). Then we will use the train function to build the
kNN model.

library(caret)
library(mlbench)
data(Sonar)
set.seed(107)
inTrain <- createDataPartition(y = Sonar$Class, p = .75, list = FALSE)
training <- Sonar[ inTrain,]
testing <- Sonar[-inTrain,]
kNNFit <- train(Class ~ .,
data = training,
method = "knn",
preProc = c("center", "scale"))
print(kNNFit)

## k-Nearest Neighbors
##
## 157 samples
## 60 predictor
## 2 classes: 'M', 'R'
##
## Pre-processing: centered (60), scaled (60)
## Resampling: Bootstrapped (25 reps)
## Summary of sample sizes: 157, 157, 157, 157, 157, 157, ...
## Resampling results across tuning parameters:
##
## k Accuracy Kappa
## 5 0.7683412 0.5331399
## 7 0.7559100 0.5053696
## 9 0.7378175 0.4715006
12 https://cran.r-project.org/web/packages/caret/vignettes/caret.pdf
Chapter 5 - Classification with k-Nearest Neighbors 51

##
## Accuracy was used to select the optimal model using the largest value.
## The final value used for the model was k = 5.

We can also search for the best k value given the training dataset.

kNNFit1 <- train(Class ~ .,

data = training,
method = "knn",
tuneLength = 15,
preProc = c("center", "scale"))
print(kNNFit1)

## k-Nearest Neighbors
##
## 157 samples
## 60 predictor
## 2 classes: 'M', 'R'
##
## Pre-processing: centered (60), scaled (60)
## Resampling: Bootstrapped (25 reps)
## Summary of sample sizes: 157, 157, 157, 157, 157, 157, ...
## Resampling results across tuning parameters:
##
## k Accuracy Kappa
## 5 0.7477453 0.4840235
## 7 0.7189901 0.4225515
## 9 0.7211797 0.4275156
## 11 0.7140987 0.4150135
## 13 0.7031182 0.3932055
## 15 0.7034819 0.3945755
## 17 0.6914316 0.3698916
## 19 0.6855830 0.3588189
## 21 0.6847459 0.3619330
## 23 0.6821917 0.3571894
## 25 0.6626137 0.3186673
## 27 0.6551801 0.3042504
## 29 0.6660760 0.3291024
## 31 0.6643681 0.3273283
## 33 0.6697744 0.3389183
##
## Accuracy was used to select the optimal model using the largest value.
## The final value used for the model was k = 5.

To create a 10-fold cross-validation based search of k, repeated 3 times we have to use the function
trainControl:
Chapter 5 - Classification with k-Nearest Neighbors 52

ctrl <- trainControl(method = "repeatedcv", repeats = 3)

kNNFit2 <- train(Class ~ .,
data = training,
method = "knn",
tuneLength = 15,
trControl = ctrl,
preProc = c("center", "scale"))
print(kNNFit2)
plot(kNNFit2)

the output:

## k-Nearest Neighbors
##
## 157 samples
## 60 predictor
## 2 classes: 'M', 'R'
##
## Pre-processing: centered (60), scaled (60)
## Resampling: Cross-Validated (10 fold, repeated 3 times)
## Summary of sample sizes: 142, 142, 142, 142, 141, 142, ...
## Resampling results across tuning parameters:
##
## k Accuracy Kappa
## 5 0.7820261 0.5548350
## 7 0.7709314 0.5321546
## 9 0.7605147 0.5111479
## 11 0.7504657 0.4924263
## 13 0.7396324 0.4705332
## 15 0.7247386 0.4401370
## 17 0.7058170 0.4006357
## 19 0.7179330 0.4246010
## 21 0.7055229 0.4003491
## 23 0.6930065 0.3779050
## 25 0.6933742 0.3791319
## 27 0.7052451 0.4028364
## 29 0.7163562 0.4277789
## 31 0.7332108 0.4629731
## 33 0.7176389 0.4335254
##
## Accuracy was used to select the optimal model using the largest value.
## The final value used for the model was k = 5.

and the plot:

Chapter 5 - Classification with k-Nearest Neighbors 53

For predictions:

knnPredict <- predict(kNNFit2, newdata = testing )

#Get the confusion matrix to see accuracy value and other parameter values
confusionMatrix(knnPredict, testing$Class )

## Confusion Matrix and Statistics

##
## Reference
## Prediction M R
## M 25 9
## R 2 15
##
## Accuracy : 0.7843
## 95% CI : (0.6468, 0.8871)
## No Information Rate : 0.5294
## P-Value [Acc > NIR] : 0.0001502
##
## Kappa : 0.56
## Mcnemar's Test P-Value : 0.0704404
##
## Sensitivity : 0.9259
## Specificity : 0.6250
## Pos Pred Value : 0.7353
## Neg Pred Value : 0.8824
## Prevalence : 0.5294
## Detection Rate : 0.4902
Chapter 5 - Classification with k-Nearest Neighbors 54

## Detection Prevalence : 0.6667

## Balanced Accuracy : 0.7755
##
## 'Positive' Class : M
##

Adding more information in the output:

ctrl <- trainControl(method = "repeatedcv", repeats = 3, classProbs=TRUE, summary\

Function = twoClassSummary)
kNNFit4 <- train(Class ~ .,
data = training,
method = "knn",
tuneLength = 15,
trControl = ctrl,
preProc = c("center", "scale"))
kNNFit4
plot(kNNFit4)

## k-Nearest Neighbors
##
## 157 samples
## 60 predictor
## 2 classes: 'M', 'R'
##
## Pre-processing: centered (60), scaled (60)
## Resampling: Cross-Validated (10 fold, repeated 3 times)
## Summary of sample sizes: 142, 142, 142, 141, 141, 141, ...
## Resampling results across tuning parameters:
##
## k ROC Sens Spec
## 5 0.8846768 0.8773148 0.6988095
## 7 0.8613467 0.8726852 0.6678571
## 9 0.8642361 0.8601852 0.6500000
## 11 0.8592634 0.8518519 0.6494048
## 13 0.8527364 0.8282407 0.6250000
## 15 0.8384011 0.8009259 0.5994048
## 17 0.8389550 0.8004630 0.5904762
## 19 0.8246280 0.8087963 0.5988095
## 21 0.8219783 0.8125000 0.6125000
## 23 0.8125951 0.8055556 0.6125000
## 25 0.8139716 0.7898148 0.6351190
## 27 0.8118676 0.7893519 0.6535714
## 29 0.8090112 0.7842593 0.6577381
## 31 0.8090939 0.7958333 0.6625000
## 33 0.7983300 0.7606481 0.6815476
##
## ROC was used to select the optimal model using the largest value.
## The final value used for the model was k = 5.
Chapter 5 - Classification with k-Nearest Neighbors 55

And finally how to use bootstrap with the caret package:

knnFit5 <- train(Class ~ .,

data = training,
method = "knn",
preProcess = c("center", "scale"),
tuneLength = 10,
trControl = trainControl(method = "boot"))
knnPredict2 <- predict(knnFit5, newdata = testing)
#Get the confusion matrix to see accuracy value and other parameter values
confusionMatrix(knnPredict2, testing$Class)

## Confusion Matrix and Statistics

##
## Reference
## Prediction M R
## M 25 9
## R 2 15
##
## Accuracy : 0.7843
## 95% CI : (0.6468, 0.8871)
## No Information Rate : 0.5294
## P-Value [Acc > NIR] : 0.0001502
##
## Kappa : 0.56
## Mcnemar's Test P-Value : 0.0704404
##
Chapter 5 - Classification with k-Nearest Neighbors 56

## Sensitivity : 0.9259
## Specificity : 0.6250
## Pos Pred Value : 0.7353
## Neg Pred Value : 0.8824
## Prevalence : 0.5294
## Detection Rate : 0.4902
## Detection Prevalence : 0.6667
## Balanced Accuracy : 0.7755
##
## 'Positive' Class : M
##

Reference
http://rstudio-pubs-static.s3.amazonaws.com/16444_caf85a306d564eb490eebdbaf0072df2.html13

13 http://rstudio-pubs-static.s3.amazonaws.com/16444_caf85a306d564eb490eebdbaf0072df2.html
Chapter 6 - Classification with
Support Vector Machines
6.1 Introduction

6.1.1 Introduction to Support Vector Machines

Support Vector Machines (SVM) constitute an algorithm that finds a linear hyperplane that
separates the data. The problem can be defined as maximizing the distance between the
hyperplane and the data (margin):
Maximize M argin = 2/||w2 || given that f (x) = {1if w · x + b ≥ 1, −1if w · x + b ≤ 1}
If the data are not linearly separable, then they can be mapped to a space of different (usually
larger) dimensionality, in order to make them linearly separable.

6.1.2 Support Vector Machines in R

We may use the e1071 library to build an SVM model in R:

library(e1071)

We can execute the command svm to build a model:

model = svm(y ~ ., data, kernel = "radial", type = "C-classification")

where parameter kernel controls which kernel is going to be used (one of linear, polynomial,
radial), while we may also define more parameters for each kernel, such as degree, gamma, coef0.
By executing ?svm, we can see all the parameters.
Having a model, a new instance may be classified using the command:

predict(model, test)

If we also want to view the probabilities for each class, then we have to add the parameter
probability = TRUE. Note that the parameter probability = TRUE must be added both while
training (command svm) and while running (command predict).

6.2 Support Vector Machines Model Construction

and Classification
We will use the data of the file alldata.txt for a problem of binary classification:
Chapter 6 - Classification with Support Vector Machines 58

X1 X2 y
Min. :-3.25322007 Min. :-4.664161 Min. :1.0
1st Qu.:-0.70900886 1st Qu.: 0.625361 1st Qu.:1.0
Median :-0.01162501 Median : 1.641370 Median :1.5
Mean : 0.03493570 Mean : 1.939284 Mean :1.5
3rd Qu.: 0.73337179 3rd Qu.: 3.115350 3rd Qu.:2.0
Max. : 3.63957363 Max. : 9.784076 Max. :2.0

Initially, we import the dataset, and split it into training and testing data::

alldata = read.csv("alldata.txt")
trainingdata = alldata[1:600, ]
testdata = alldata[601:800, ]

We will answer to the following questions:

a) Plot the training data with different colors/symbols for each class.
b) Apply SVM with RBF kernel and gamma equal to 1, and plot the hyperplane on the figure of
question (a).
c) Repeat question (b) for gamma value equal to 0.01 and equal to 100. What are your
observations?
Using the gamma values gammavalues = c(0.01, 0.1, 1, 10, 100, 1000, 10000, 100000,
1000000), we will also answer to the following questions:

d) Compute the training error and the testing error.

e) Plot the two errors on the same diagram. What are your observations?
f) Apply 10-fold cross validation to find the best value for gamma.

6.2.1 Library Imports

Initially, we import the required librares:

library(MLmetrics)
library(e1071)

6.2.2 Algorithm Application and Hyperplane Diagram

For question (a), we execute

plot(trainingdata[, c(1:2)], col = trainingdata$y, pch = c("o","+")[trainingdata$\

y])

For questions (b), (c), we initially build a grid, on top of which we are going to plot the
hyperplanes:
Chapter 6 - Classification with Support Vector Machines 59

X1 = seq(min(trainingdata[, 1]), max(trainingdata[, 1]), by = 0.1)

X2 = seq(min(trainingdata[, 2]), max(trainingdata[, 2]), by = 0.1)
mygrid = expand.grid(X1, X2)
colnames(mygrid) = colnames(trainingdata)[1:2]

For question (b), we train the SVM with gamma = 1 and apply it on the grid:

svm_model = svm(y ~ ., kernel="radial", type="C-classification", data = trainingd\

ata, gamma = 1)
pred = predict(svm_model, mygrid)
Y = matrix(pred, length(X1), length(X2))
contour(X1, X2, Y, add = TRUE, levels = 1.5, labels = "gamma = 1", col = "blue")

Similarly, for question (c), when gamma = 0.01:

svm_model = svm(y ~ ., kernel="radial", type="C-classification", data = trainingd\

ata, gamma = 0.01)
pred = predict(svm_model, mygrid)
Y = matrix(pred, length(X1), length(X2))
contour(X1, X2, Y, add = TRUE, levels = 1.5, labels = "gamma = 0.01", col = "red")

and when gamma = 100:

svm_model = svm(y ~ ., kernel="radial", type="C-classification", data = trainingd\

ata, gamma = 100)
pred = predict(svm_model, mygrid)
Y = matrix(pred, length(X1), length(X2))
contour(X1, X2, Y, add = TRUE, levels = 1.5, labels = "gamma = 100", col = "green\
")

Finally, the diagram is shown below:

Chapter 6 - Classification with Support Vector Machines 60

6.2.3 Training and Testing Error Curves

For question (d), we iterate over all gamma values and calculcate the training error:

training_error = c()
for (gamma in gammavalues) {
svm_model = svm(y ~ ., kernel="radial", type="C-classification", data = trainin\
gdata, gamma = gamma)
pred = predict(svm_model, trainingdata[, c(1:2)])
training_error = c(training_error, 1 - Accuracy(trainingdata$y, pred))
}

Similarly, for the testing error:

testing_error = c()
for (gamma in gammavalues) {
svm_model = svm(y ~ ., kernel="radial", type="C-classification", data = trainin\
gdata, gamma = gamma)
pred = predict(svm_model, testdata[, c(1:2)])
testing_error = c(testing_error, 1 - Accuracy(testdata$y, pred))
}

We may plot the two errors on the same diagram using the commands:
Chapter 6 - Classification with Support Vector Machines 61

plot(training_error, type = "l", col="blue", ylim = c(0, 0.5), xlab = "Gamma", yl\
ab = "Error", xaxt = "n")
axis(1, at = 1:length(gammavalues), labels = gammavalues)
lines(testing_error, col="red")
legend("right", c("Training Error", "Testing Error"), pch = c("-","-"), col = c(\
"blue", "red"))

Finally, we are provided with the following diagram:

6.2.4 Cross-validation
We will apply k-fold cross validation to compute the best value for gamma (question (f)). Initially,
we construct k folds:

k = 10
dsize = nrow(trainingdata)
folds = split(sample(1:dsize), ceiling(seq(dsize) * k / dsize))

After that, we make a loop over gamma, and inside the loop we create a new (nested) loop over
the folds:
Chapter 6 - Classification with Support Vector Machines 62

accuracies <- c()

for (gamma in gammavalues) {
predictions <- data.frame()
testsets <- data.frame()
for(i in 1:k){
# Select 9 out of 10 folds for training and 1 for validation
trainingset <- trainingdata[unlist(folds[-i]),]
validationset <- trainingdata[unlist(folds[i]),]
# Train and apply the model
svm_model = svm(y ~ ., kernel="radial", type="C-classification", data = train\
ingset, gamma = gamma)
pred = predict(svm_model, validationset[, c(1:2)])
# Save predictions and testsets
predictions <- rbind(predictions, as.data.frame(pred))
testsets <- rbind(testsets, as.data.frame(validationset[,3]))
}
# Calculate the new accuracy and add it to the previous ones
accuracies = c(accuracies, Accuracy(predictions, testsets))
}

We calculate the accuracy for each fold, and, finally, select the gamma value with the maximum
accuracy:

print(accuracies)
bestgamma = gammavalues[which.max(accuracies)]

6.3 Exercise
You are given the following data where Y is the class vector of a binary classification problem.

X1 X2 Y
2 2 1
2 -2 1
-2 -2 1
-2 2 1
1 1 2
1 -1 2
-1 -1 2
-1 1 2

You can enter the data in R using the following commands:

X1 = c(2, 2, -2, -2, 1, 1, -1, -1)
X2 = c(2, -2, -2, 2, 1, -1, -1, 1)
Y = c(1, 1, 1, 1, 2, 2, 2, 2)
alldata = data.frame(X1, X2, Y)

Answer to the following questions:

a) Plot the training data with different colors/symbols for each class.
b) Apply SVM with RBF kernel, and plot the hyperplane on the figure of question (a).
Chapter 6 - Classification with Support Vector Machines 63

c) Using your jmodel, in which class would you classify a new instance with values (4, 5)?
Part III - Data Processing
Chapter 7 - Feature Selection
7.1 Introduction
What is feature selection? What are filter methods? What are wrapper methods?

7.2 Filter Methods

In this section we will use the Correlation Feature Selection (CFS) method to select attributes
to be used when training and testing the machine learning models. We will use the Iris dataset
for this purpose.

data = iris
# split into training and validation datasets
set.seed(1234)
ind <- sample(2, nrow(data), replace=TRUE, prob=c(0.7, 0.3))
trainData <- data[ind==1,]
validationData <- data[ind==2,]
# keep only instances that do not have missing values.
trainData <- trainData[complete.cases(trainData),]
validationData <- validationData[complete.cases(validationData),]

7.2.1 Using the CFS method

The CFS method is found in the FSelector package and to use it we call the cfs method.

library(FSelector)
subset <- cfs(Species ~ ., trainData)
f <- as.simple.formula(subset, "Species")
print(f)

## Species ~ Petal.Length + Petal.Width

## <environment: 0x114e18a70>

The output is a formula that we can use in various classification algorithms and says that
according to the CFS algorithm and the training dataset, the best features in order to predict
the Species of the Iris flowers are the Petal Length and the Petal Width.
For example in the Naive Bayes algorithm, we will use both the formula that includes all the
attributes for predicting the Species and the formula derived from CFS.
Chapter 7 - Feature Selection 66

library(e1071)
model <- naiveBayes(Species ~ ., data=trainData, laplace = 1)
simpler_model <- naiveBayes(f, data=trainData, laplace = 1)

pred <- predict(model, validationData)

simpler_pred <- predict(simpler_model, validationData)

library(MLmetrics)
train_pred <- predict(model, trainData)
train_simpler_pred <- predict(simpler_model, trainData)
paste("Accuracy in training all attributes",
Accuracy(train_pred, trainData$Species), sep=" - ")

## [1] "Accuracy in training all attributes - 0.964285714285714"

paste("Accuracy in training CFS attributes",

Accuracy(train_simpler_pred, trainData$Species), sep=" - ")

## [1] "Accuracy in training CFS attributes - 0.955357142857143"

paste("Accuracy in validation all attributes",

Accuracy(pred, validationData$Species), sep=" - ")

## [1] "Accuracy in validation all attributes - 0.947368421052632"

paste("Accuracy in validation CFS attributes",

Accuracy(simpler_pred, validationData$Species), sep=" - ")

## [1] "Accuracy in validation CFS attributes - 0.973684210526316"

The accuracy in the training set is increased when using all the attrbitues, but decreased in the
validation set in relation to the simpler model that only used 2 attributes.

7.2 Wrapper Methods

7.2.1 Recurrent Feature Elimination

As a demonstration of a wrapper method, we will use the Recurrent Feature Elimination (RFE)
method to select attributes in the Pima Indians Diabetes dataset. For this purpose the library
caret will be used.

First, as always, we will split the dataset in train and validation datasets.
Chapter 7 - Feature Selection 67

# ensure the results are repeatable

set.seed(1234)
# load the required libraries
library(mlbench)
library(caret)
# load the data
data(PimaIndiansDiabetes)
data = PimaIndiansDiabetes
# split into training and validation datasets
set.seed(1234)
ind <- sample(2, nrow(data), replace=TRUE, prob=c(0.7, 0.3))
trainData <- data[ind==1,]
validationData <- data[ind==2,]
trainData <- trainData[complete.cases(trainData),]
validationData <- validationData[complete.cases(validationData),]

Next we will use the rfe method of the caret package, setting it up using the rfeControl method.

# define the control using a random forest selection function

control <- rfeControl(functions=nbFuncs, method="cv", number=10)
# run the RFE algorithm
results <- rfe(trainData[,1:8], trainData[,9], sizes=c(1:8), rfeControl=control)

Let’s see the results.

# summarize the results

print(results)

##
## Recursive feature selection
##
## Outer resampling method: Cross-Validated (10 fold)
##
## Resampling performance over subset size:
##
## Variables Accuracy Kappa AccuracySD KappaSD Selected
## 1 0.7453 0.3681 0.04390 0.12449
## 2 0.7546 0.4182 0.04598 0.12317
## 3 0.7715 0.4707 0.05759 0.14388 *
## 4 0.7584 0.4479 0.04145 0.10478
## 5 0.7546 0.4457 0.04429 0.10095
## 6 0.7490 0.4344 0.04393 0.09394
## 7 0.7509 0.4290 0.04686 0.11122
## 8 0.7491 0.4201 0.03507 0.07767
##
## The top 3 variables (out of 3):
## glucose, age, mass

The RFE out of the 8 variables selected three (glucose, age, mass) that provided the best accuracy
under 10-fold cross-validation.
Chapter 7 - Feature Selection 68

# list the chosen features

predictors(results)

## [1] "glucose" "age" "mass"

# plot the results

plot(results, type=c("g", "o"))

We will use the 3 top variables that come out of RFE in the Naive Bayes algorithm:

library(e1071)
(f <- as.formula(paste("diabetes", paste(results$optVariables, collapse=" + "), s\
ep=" ~ ")))

## diabetes ~ glucose + age + mass

Outer parentheses
Note the usage of the outer parenthesis. We both set the variable f equal to the Τ>
formula and print it to the console.
Chapter 7 - Feature Selection 69

model <- naiveBayes(diabetes ~ ., data=trainData, laplace = 1)

simpler_model <- naiveBayes(f, data=trainData, laplace = 1)

pred <- predict(model, validationData)

simpler_pred <- predict(simpler_model, validationData)

library(MLmetrics)
train_pred <- predict(model, trainData)
train_simpler_pred <- predict(simpler_model, trainData)
paste("Accuracy in training all attributes",
Accuracy(train_pred, trainData$diabetes), sep=" - ")

## [1] "Accuracy in training all attributes - 0.760299625468165"

paste("Accuracy in training RFE attributes",

Accuracy(train_simpler_pred, trainData$diabetes), sep=" - ")

## [1] "Accuracy in training RFE attributes - 0.764044943820225"

paste("Accuracy in validation all attributes",

Accuracy(pred, validationData$diabetes), sep=" - ")

## [1] "Accuracy in validation all attributes - 0.764957264957265"

paste("Accuracy in validation RFE attributes",

Accuracy(simpler_pred, validationData$diabetes), sep=" - ")

## [1] "Accuracy in validation RFE attributes - 0.799145299145299"

Using a simpler formula with 3 out of 8 attibutes, we were able to get better generalization results
in the validation dataset.

7.2.2 Forward Search

The standard wrapper methods are forward and backward search. In this sub-section we will
apply forward search and decision trees as a model in the breast cancer Wisconsin dataset (see
Chapter 5).
Chapter 7 - Feature Selection 70

# Downloading the file

fileURL <- "http://archive.ics.uci.edu/ml/machine-learning-databases/breast-cance\
r-wisconsin/breast-cancer-wisconsin.data"
download.file(fileURL, destfile="breast-cancer-wisconsin.data", method="curl")
data <- read.table('breast-cancer-wisconsin.data', na.strings = "?", sep=",")
data <- data[,-1]
names(data) <- c("ClumpThickness",
"UniformityCellSize",
"UniformityCellShape",
"MarginalAdhesion",
"SingleEpithelialCellSize",
"BareNuclei",
"BlandChromatin",
"NormalNucleoli",
"Mitoses",
"Class")
data$Class <- factor(data$Class, levels=c(2,4), labels=c("benign", "malignant"))
set.seed(1234)
ind <- sample(2, nrow(data), replace=TRUE, prob=c(0.7, 0.3))
trainData <- data[ind==1,]
validationData <- data[ind==2,]
# remove cases with missing data
trainData <- trainData[complete.cases(trainData),]
validationData <- validationData[complete.cases(validationData),]

We will use the forward.search method of the FSelector package and the rpart decision trees
training method of the homonymous package. The forward.search method needs an evaluation
function that evaluates a subset of attributes. In our case the evaluator function performs 5-fold
cross-validation on the training dataset.

library(FSelector)
library(rpart)

evaluator <- function(subset) {

#k-fold cross validation
k <- 5
splits <- runif(nrow(trainData))
results = sapply(1:k, function(i) {
test.idx <- (splits >= (i - 1) / k) & (splits < i / k)
train.idx <- !test.idx
test <- trainData[test.idx, , drop=FALSE]
train <- trainData[train.idx, , drop=FALSE]
tree <- rpart(as.simple.formula(subset, "Class"), train)
error.rate = sum(test$Class != predict(tree, test, type="c")) / nrow(test)
return(1 - error.rate)
})
print(subset)
print(mean(results))
return(mean(results))
}

subset <- forward.search(names(trainData)[-10], evaluator)

Chapter 7 - Feature Selection 71

## [1] "ClumpThickness"
## [1] 0.8535402
## [1] "UniformityCellSize"
## [1] 0.8982188
## [1] "UniformityCellShape"
## [1] 0.9132406
## [1] "MarginalAdhesion"
## [1] 0.8564007
## [1] "SingleEpithelialCellSize"
## [1] 0.9055005
## [1] "BareNuclei"
## [1] 0.9079083
## [1] "BlandChromatin"
## [1] 0.9027608
## [1] "NormalNucleoli"
## [1] 0.895089
## [1] "Mitoses"
## [1] 0.7942549
## [1] "ClumpThickness" "UniformityCellShape"
## [1] 0.9297849
## [1] "UniformityCellSize" "UniformityCellShape"
## [1] 0.9335359
## [1] "UniformityCellShape" "MarginalAdhesion"
## [1] 0.9167509
## [1] "UniformityCellShape" "SingleEpithelialCellSize"
## [1] 0.9404055
## [1] "UniformityCellShape" "BareNuclei"
## [1] 0.9384218
## [1] "UniformityCellShape" "BlandChromatin"
## [1] 0.9283328
## [1] "UniformityCellShape" "NormalNucleoli"
## [1] 0.944158
## [1] "UniformityCellShape" "Mitoses"
## [1] 0.9162663
## [1] "ClumpThickness" "UniformityCellShape" "NormalNucleoli"
## [1] 0.9405466
## [1] "UniformityCellSize" "UniformityCellShape" "NormalNucleoli"
## [1] 0.9322648
## [1] "UniformityCellShape" "MarginalAdhesion" "NormalNucleoli"
## [1] 0.9387149
## [1] "UniformityCellShape" "SingleEpithelialCellSize"
## [3] "NormalNucleoli"
## [1] 0.9314805
## [1] "UniformityCellShape" "BareNuclei" "NormalNucleoli"
## [1] 0.9470806
## [1] "UniformityCellShape" "BlandChromatin" "NormalNucleoli"
## [1] 0.9267064
## [1] "UniformityCellShape" "NormalNucleoli" "Mitoses"
## [1] 0.9401233
## [1] "ClumpThickness" "UniformityCellShape" "BareNuclei"
## [4] "NormalNucleoli"
## [1] 0.9335856
## [1] "UniformityCellSize" "UniformityCellShape" "BareNuclei"
## [4] "NormalNucleoli"
## [1] 0.9498017
## [1] "UniformityCellShape" "MarginalAdhesion" "BareNuclei"
Chapter 7 - Feature Selection 72

## [4] "NormalNucleoli"
## [1] 0.9418352
## [1] "UniformityCellShape" "SingleEpithelialCellSize"
## [3] "BareNuclei" "NormalNucleoli"
## [1] 0.9415623
## [1] "UniformityCellShape" "BareNuclei" "BlandChromatin"
## [4] "NormalNucleoli"
## [1] 0.9506024
## [1] "UniformityCellShape" "BareNuclei" "NormalNucleoli"
## [4] "Mitoses"
## [1] 0.9382617
## [1] "ClumpThickness" "UniformityCellShape" "BareNuclei"
## [4] "BlandChromatin" "NormalNucleoli"
## [1] 0.9515459
## [1] "UniformityCellSize" "UniformityCellShape" "BareNuclei"
## [4] "BlandChromatin" "NormalNucleoli"
## [1] 0.930654
## [1] "UniformityCellShape" "MarginalAdhesion" "BareNuclei"
## [4] "BlandChromatin" "NormalNucleoli"
## [1] 0.9454745
## [1] "UniformityCellShape" "SingleEpithelialCellSize"
## [3] "BareNuclei" "BlandChromatin"
## [5] "NormalNucleoli"
## [1] 0.9385037
## [1] "UniformityCellShape" "BareNuclei" "BlandChromatin"
## [4] "NormalNucleoli" "Mitoses"
## [1] 0.935793
## [1] "ClumpThickness" "UniformityCellSize" "UniformityCellShape"
## [4] "BareNuclei" "BlandChromatin" "NormalNucleoli"
## [1] 0.9460072
## [1] "ClumpThickness" "UniformityCellShape" "MarginalAdhesion"
## [4] "BareNuclei" "BlandChromatin" "NormalNucleoli"
## [1] 0.9494458
## [1] "ClumpThickness" "UniformityCellShape"
## [3] "SingleEpithelialCellSize" "BareNuclei"
## [5] "BlandChromatin" "NormalNucleoli"
## [1] 0.9506376
## [1] "ClumpThickness" "UniformityCellShape" "BareNuclei"
## [4] "BlandChromatin" "NormalNucleoli" "Mitoses"
## [1] 0.939423

After the search we get the following formula, where 5 out of the 9 variables were kept.
f <- as.simple.formula(subset, "Class")
print(f)

## Class ~ ClumpThickness + UniformityCellShape + BareNuclei + BlandChromatin +

## NormalNucleoli
## <environment: 0x132db8c38>

The fact that we performed forward search using decision trees in order to get a formula with
a subset of attributes, doesn’t stop us from using another classification model for training and
prediction. For example as in the previous examples in tehi chapter, we can use the Naive Bayes
algorithm to evaluate the forward selection algorithm both in the training and the validation
datasets under the accuracy metric.
Chapter 7 - Feature Selection 73

library(e1071)
model <- naiveBayes(Class ~ ., data=trainData, laplace = 1)
simpler_model <- naiveBayes(f, data=trainData, laplace = 1)

pred <- predict(model, validationData)

simpler_pred <- predict(simpler_model, validationData)

library(MLmetrics)
train_pred <- predict(model, trainData)
train_simpler_pred <- predict(simpler_model, trainData)
paste("Accuracy in training all attributes",
Accuracy(train_pred, trainData$Class), sep=" - ")

## [1] "Accuracy in training all attributes - 0.957805907172996"

paste("Accuracy in training forward search attributes",

Accuracy(train_simpler_pred, trainData$Class), sep=" - ")

## [1] "Accuracy in training forward search attributes - 0.953586497890295"

paste("Accuracy in validation all attributes",

Accuracy(pred, validationData$Class), sep=" - ")

## [1] "Accuracy in validation all attributes - 0.976076555023923"

paste("Accuracy in validation forward search attributes",

Accuracy(simpler_pred, validationData$Class), sep=" - ")

## [1] "Accuracy in validation forward search attributes - 0.971291866028708"

In the breast cancer Wisconsin dataset, the feature selection algorithm did not outperform the use
of all attributes. The obvious cause for this is that in the dataset we have 9 attributes, handpicked
by domain experts and have indeed a combined predictive power that cannot outperform any
subset of them. So removing features in this case some does not produce better results.

Backward Search
One can also alter the forward.search method with backward.search to perform the
backward search wrapper selection method. How many attributes are kept in the case
of Backward Search?
Chapter 8 - Dimensionality
Reduction
8.1 Principal Components Analysis

8.1.1 “Raw” Principal Components Analysis**

First we will try to perform Principal Components analysis (PCA) without using a ready-made
R function from a library.
Step 1: Data
First we are going to make a dataset and plot it.

d <- c(2.5, 2.4, 0.5, 0.7, 2.2, 2.9, 1.9, 2.2, 3.1, 3.0, 2.3,
2.7, 2, 1.6, 1, 1.1, 1.5, 1.6, 1.1, 0.9)
data <- matrix(d, ncol=2, byrow = T)
plot(data, xlab="x1", ylab="x2")

Step 2: Subtract the mean

Chapter 8 - Dimensionality Reduction 75

We subtract the mean of each attribute (x1, x2) from the respective values (centering the data)
using the function scale.

data_norm <- scale(data, scale=F)

plot(data_norm, xlab="x1", ylab="x2")

Step 3: Calculate the covariance matrix

We calculate the covariance matrix using the function cov.

S <- cov(data_norm)
print(S)

## [,1] [,2]
## [1,] 0.6165556 0.6154444
## [2,] 0.6154444 0.7165556

Step 4: Computer the eigenvectors of the covariance matrix

Using the svd method (Singular Value Decomposition) we derive the eigenvectors of the
covariance matrix. First we print the decomposition:
Chapter 8 - Dimensionality Reduction 76

udv <- svd(S)

print(udv)

## $d
## [1] 1.2840277 0.0490834
##
## $u
## [,1] [,2]
## [1,] -0.6778734 -0.7351787
## [2,] -0.7351787 0.6778734
##
## $v
## [,1] [,2]
## [1,] -0.6778734 -0.7351787
## [2,] -0.7351787 0.6778734

and then we plot the eigenvectors:

# asp=1 keep the aspect ratio to 1 so that the eigenvectors are

# displayed appropriatelly, perpendicular
plot(data_norm, asp=1, xlab="x1", ylab="x2")
arrows(0,0,udv$u[1,1],udv$u[2,1], lwd=1)
arrows(0,0,udv$u[1,2],udv$u[2,2], lwd=0.5)

Step 5: Choosing components

We can use a barplot to display the variance accounted for each component.
Chapter 8 - Dimensionality Reduction 77

barplot(udv$d)
print(cumsum(udv$d)/sum(udv$d))

We can see that the 1st component accounts for more than 95% of the variance.
Step 6: Picking the 1st component
We transform the 2D dataset into a 1D dataset using just the 1st Principal Component (PC).

data_new <- t(udv$u[,1,drop=FALSE]) %*% t(data_norm)

data_new

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

## [1,] -0.8279702 1.77758 -0.9921975 -0.2742104 -1.675801 -0.9129491
## [,7] [,8] [,9] [,10]
## [1,] 0.09910944 1.144572 0.4380461 1.223821

plot(data_new,data_new,asp=1,xlab="x1", ylab="x2")
Chapter 8 - Dimensionality Reduction 78

8.1.2 PCA using an R function

In this subsection, we will use the prcomp function that performs PCA in one step. We will apply
it on the Iris dataset and test the gains or losses in terms of accuracy using the kNN algorithm.

# download the file

data = iris
set.seed(1234)
ind <- sample(2, nrow(data), replace=TRUE, prob=c(0.7, 0.3))
trainData <- data[ind==1,]
validationData <- data[ind==2,]

library(class)
library(stats)

trainData <- trainData[complete.cases(trainData),]

validationData <- validationData[complete.cases(validationData),]

trainDataX <- trainData[,-ncol(trainData)]

logTrainDataX <- log(trainDataX)
train.pca <- prcomp(logTrainDataX, center = TRUE, scale. = TRUE)
summary(train.pca)
Chapter 8 - Dimensionality Reduction 79

## Importance of components:
## PC1 PC2 PC3 PC4
## Standard deviation 1.7226 0.9313 0.36484 0.17916
## Proportion of Variance 0.7419 0.2168 0.03328 0.00802
## Cumulative Proportion 0.7419 0.9587 0.99198 1.00000

From the summary, we observe that with 2 PCs, we can explain more than 95% of the variance.

trainDataY <- trainData$Species

validationDataX <- validationData[,-ncol(trainData)]
# Let's also transform the validation data
logValidationDataX <- log(validationDataX)
validation.pca <- predict(train.pca, newdata=logValidationDataX)
validationDataY <- validationData$Species

# no pca prediction
prediction = knn(trainDataX, validationDataX, trainDataY, k = 3)
# So let's predict using only the 7 principal components
prediction_pca = knn(train.pca$x[,1:2], validation.pca[,1:2], trainDataY, k = 3)

cat("Confusion matrix:\n")

## Confusion matrix:

xtab = table(prediction, validationData$Species)

print(xtab)

##
## prediction setosa versicolor virginica
## setosa 10 0 0
## versicolor 0 12 1
## virginica 0 0 15

cat("\nEvaluation:\n\n")

##
## Evaluation:

accuracy = sum(prediction == validationData$Species)/length(validationData$Specie\

s)
cat(paste("Accuracy:\t", format(accuracy, digits=2), "\n",sep=" "))

## Accuracy: 0.97
Chapter 8 - Dimensionality Reduction 80

cat("Confusion matrix PCA:\n")

## Confusion matrix PCA:

xtab = table(prediction_pca, validationData$Species)

print(xtab)

##
## prediction_pca setosa versicolor virginica
## setosa 10 0 0
## versicolor 0 10 3
## virginica 0 2 13

cat("\nEvaluation PCA:\n\n")

##
## Evaluation PCA:

accuracy = sum(prediction_pca == validationData$Species)/length(validationData$Sp\

ecies)
cat(paste("Accuracy PCA:\t", format(accuracy, digits=2), "\n",sep=" "))

## Accuracy PCA: 0.87

Even though the validation dataset is small, one can observe that using 50% less predictor
attributes, the algorithm made only 4 extra mistakes. Of course in real world scenarios, cross-
validation should be used with such small datasets.

7.2.3 PCA for Plotting

Using the PCA we can also plot in 2 dimensions, high-dimensional data, by using just the first
two components.

plot(train.pca$x[trainData$Species == 'setosa',1:2], col="blue", ylim = c(-3, 3),\

xlim=c(-3,3), asp=1)
points(train.pca$x[trainData$Species == 'versicolor',1:2], pch = 3, col="red")
points(train.pca$x[trainData$Species == 'virginica',1:2], pch = 4, col="green")
Chapter 8 - Dimensionality Reduction 81

In this case the PCA is not the best method for dimensionality reduction to improve the accuracy.
From the plot one cas observe that the versicolor and virginica species are intermixed in 2D
and thus the k-NN algorithm will have trouble classifying between the two.
Part III - Clustering
Chapter 9 - Centroid-based
clustering and Evaluation
Centroid-based clustering is a subset of clustering techniques where clusters are represented
by a central vector, which may not necessarily be a member of the data set. When the number
of clusters is fixed to a given number k, k-Means clustering gives a formal definition of the
following optimization problem:
Given a dataset containing N samples:

X = {x1 , x2 , x3 , . . . , xN }

and K as the number of selected clusters:

Clusters : {C1 , C2 , . . . , CK } with Centers : {m1 , m2 , . . . , mK }

k-Means Algorithm Purpose: Find the distribution of all samples among clusters so as to
minimize the sum of squares of the distances between each sample and the center of the cluster
it belongs. This metric corresponds to the Within cluster Sum of Squares – WSS or SSE and
is given by the following formula:

∑
K ∑
2
SSE = ‖x − mi ‖
i−1 x∈Ci

In order to solve the aforementioned optimization problem, k-Mean algorithm involves two
distinc steps:

1. Given certain centers, formulate clusters by assigning each sample to the closest cluster
2. Calculate the new centers based on the formulate clusters

In the initialization step, the algorithm randomly selects K samples as the initial centers and
then using these centers assigned each sample in the closest cluster. Once every sample has been
assigned to a certain cluster, the new center of each cluster is calculated as the mean of every
sample that belongs to the cluster. This iterative process is continuoued until there is no change
in the clusters in to consecutive iterations.
The most famous and used variation of the k-Means algorithm is k-Medoids, which is mainly
used for categorical data and allows the use of various different distance functions. It is worth
noticing that unlike k-Means, k-Medoids always uses a certain sample as the medoid of each
cluster.
Chapter 9 - Centroid-based clustering and Evaluation 84

9.1 - k-Means in R
For performing k-Means clustering in R, we can use the following command:

# Perform k-Means to form 3 clusters

model = kmeans(data, centers = 3)

The centers parameter can either take the number of the initial centers (and thus the number of
clusters) or the centers themselves as argument.
We can use the following commands in order to show the final centroids and the distribution of
the samples into the formulated clusters:

# Get centroids
model$centers

# Get distribution of samples into clusters

model$cluster

9.2 - k-Medoids in R
In order to perform k-Medoids clustering in R, we need to import the library cluster:

library(cluster)

and then we can use the following command:

# Perform k-Medoids to form 3 clusters

model = pam(conferences, 3)

We can use the following commands in order to show the final medoids and the distribution of
the samples into the formulated clusters:

# Get centroids
model$medoids

# Get distribution of samples into clusters

model$clustering

If the dataset contains categorical variables, then we can show the medoids using the respective
variable with the following command:

data[model$id.med,]
Chapter 9 - Centroid-based clustering and Evaluation 85

9.3 - Clustering Evaluation in R

In order to evaluate our clustering in R, we can use the library cluster which enables us to
easily calculate several evaluation metrics. Once we have perform the clutering and get the final
distribution of the samples into the clusters, then we can compute a list of metrics that enable us
evaluate the result. Some of those metrics are the following:

• Cohesion (or within-cluster sum of squares)

∑
K ∑
2
W SS = ‖x − mi ‖
i−1 x∈Ci

• Separation (or between-cluster sum of squares)

∑
K
2
BSS = |Ci |‖m − mi ‖
i−1

• Silhouette coefficient

b(i) − a(i)
Silhouette(i) =
max{a(i), b(i)}

where a(i) is the average distance between the i-th sample and all other samples within the
same cluster and b(i) is the lowest average distance of the i-th sample to all samples in any
other clusters, of which i is not a member.
We can compute the aforementioned metrics using the following R commands:

# Compute WSS
cohesion = model$tot.withinss

# Compute BSS
separation = model$betweenss

# Compute Silhouette
model_silhouette = silhouette(model$cluster, dist(data))

In the computation of silhouette coefficient, data stands for the original data. In order to plot the
silhouette coeffiecient and the mean silhouette we can use the following commands:
Chapter 9 - Centroid-based clustering and Evaluation 86

# Plot Silhouette
plot(model_silhouette)

# Compute mean silhouette

mean(model_silhouette[, 3])

Another way that can provide us with an overview of the clustering and enables us to evaluate
its performance is the construction of heatmaps. At first, we can order out data samples using
the following command

data_ord = data[order(model$cluster),]

and then constuct our heatmap

heatmap(as.matrix(dist(data_ord)), Rowv = NA, Colv = NA,

col = heat.colors(256), revC = TRUE)

The following figure illustrates a sample heatmap constructed by using data samples that are
distributed among 3 clusters. The color of the heatmap refers to the distance of the ordered data
samples. The red color denotes small distance while yellow high distance.
Chapter 9 - Centroid-based clustering and Evaluation 87

9.4 - k-Means clustering overview

The following table contains the dataset that we are going to use in order to solve a clustering
problem using k-Means algorithm.

X Y
x1 7 1
x2 3 4
x3 1 5
x4 5 8
x5 1 3
x6 7 8
x7 8 2
x8 5 9
Chapter 9 - Centroid-based clustering and Evaluation 88

Using the above dataset, we will answer the following questions:

(a) Plot the data in a two dimentions.
(b) Apply k-Means algorithm in order to distribute the data points into 3 clusters. Initialize the
algorithm by using the data points x1, x2 and x3 as the initial centers.
(c) Calculate the cohesion and the separation of the clustering performed in question (b)
(d) Plot the clustering results using a different color for the data points included in each cluster
and also mark the centroids

9.4.1 - Data Construction

First of all, we need to properly construct our data frame containing the above dataset. In order
to do so, we can use the following commands:

# Construct vector containing X-axis data

X = c(7, 3, 1, 5, 1, 7, 8, 5)

# Construct vector containing Y-axis data

Y = c(1, 4, 5, 8, 3, 8, 2, 9)

# Construct vector containing the data labels which correspond to the row names
rnames = c("x1", "x2", "x3", "x4", "x5", "x6", "x7", "x8")

# Create the data frame

kdata = data.frame(X, Y, row.names = rnames)

Once having created our data frame, we can plot our data using the following commands:

# Plot the data

plot(kdata, pch = 15)

# Add the data labels

text(hdata, labels = row.names(hdata),
pos = c(4, 2, 4, 4, 4, 4, 4, 4))

Our data is illutrated in the following figure:

Chapter 9 - Centroid-based clustering and Evaluation 89

9.4.2 - Clustering using k-Means algorithm

In order to answer question (b), we apply k-Means algorithm setting the data points x1, x2 and
x3 as the initial centers:
[ ] [ ] [ ]
7 3 1
C10 = C20 = C30 =
1 4 5

1st iteration:
At first, we will compute the distances between all data points contained in the dataset and the
initial centers. These distances are given in the following distance matrix:

C10 C20 C30

x1 0.00 5.00 7.21

x2 5.00 0.00 2.24
x3 7.21 2.24 0.00
x4 7.28 4.47 5.00
x5 6.32 2.24 2.00
x6 7.00 5.66 6.71
x7 1.41 5.39 7.62
x8 8.25 5.39 5.66

After having computed the distances, we find the minimum distance for each data point in order
to assign it into the respective cluster. Once all data point are assigned to one cluster, we compute
the new centers.
Chapter 9 - Centroid-based clustering and Evaluation 90

[ ]
x1 + x7 7.5
C11 = =
2 1.5

[ ] [ ]
x2 + x4 + x6 + x8 7 x3 + x5 1
C21 = = C31 = =
4 7.25 2 4

2nd iteration:
The following table contains the distances of the data points from the updated centers.

C11 C21 C31

x1 0.71 6.56 6.71

x2 5.15 3.82 2.00
x3 7.38 4.59 1.00
x4 6.96 0.75 5.66
x5 6.67 5.84 1.00
x6 6.52 2.14 7.21
x7 0.71 6.05 7.28
x8 7.91 1.75 6.40

The new centers are:

[ ]
x 1 + x 7 7.5
C12 = =
2 1.5

[ ] [ ]
x4 + x6 + x8 5.67 x2 + x3 + x5 1.67
C22 = = C32 = =
3 8.33 3 4

Given that the centers have changed, we the execution of the algorithm continuoues in the next
iteration
3rd iteration:
Once again, we compute the updated distances.

C12 C22 C33

x1 0.71 7.45 6.12

x2 5.15 5.09 1.33
x3 7.38 5.73 1.20
x4 6.96 0.75 5.21
x5 6.67 7.09 1.20
x6 6.52 1.37 6.67
x7 0.71 6.75 6.64
x8 7.91 0.94 6.01

The new centers are:

Chapter 9 - Centroid-based clustering and Evaluation 91

[ ]
x1 + x7 7.5
C13 = =
2 1.5

[ ] [ ]
x4 + x6 + x8 5.67 x2 + x3 + x5 1.67
C23 = = C33 = =
3 8.33 3 4

As we can see, the centers have not changes and thus the algorithm converges after the 3rd
iteration.
We can simply perform the aforementioned manual procedure with R using the following
commands:

# Perform k-Means clustering

model = kmeans(kdata, centers = kdata[1:3,])

# Get final centers

model$centers

# Get the distribution of data points among clusters

model$cluster

9.4.3 - Compute Cohesion and Separation

In order to answer the question (c), we have to compute the cohesion and the sepeartion of the
clustering performed in the previous subsection.
At first, we will compute these metrics manually and then we will see how they can be computed
using R. In the computation, we will use the above clusters.
[ ] [ ] [ ]
7.5 5.67 1.67
C1 : m1 = C2 : m2 = C3 : m3 =
1.5 8.33 4

• Cohesion computation

The Cohesion metric is given by the following equation:

∑
K ∑
2
W SS = ‖x − mi ‖
i−1 x∈Ci

At this point, we will compute the cohesion for every cluster:

1st Cluster:

∑ 2 2 2
W SS(C1 ) = ‖x − m1 ‖ = ‖x1 − m1 ‖ + ‖x7 − m1 ‖
x∈C1

√ 2 √ 2
= (7 − 7.5)2 + (1 − 1.5)2 + (8 − 7.5)2 + (2 − 1.5)2
Chapter 9 - Centroid-based clustering and Evaluation 92

= 0.25 + 0.25 + 0.25 + 0.25 = 1

2nd Cluster:

∑ 2 2 2 2
W SS(C2 ) = ‖x − m2 ‖ = ‖x4 − m2 ‖ + ‖x6 − m2 ‖ + ‖x8 − m2 ‖
x∈C2

√ 2 √ 2
= (5 − 5.67)2 + (8 − 8.33)2 + (7 − 5.67)2 + (8 − 8.33)2

√ 2
+ (5 − 5.67)2 + (9 − 8.33)2

= 0.4489 + 0.1089 + 1.7689 + 0.1089 + 0.4489 + 0.4489 = 3.3334

3rd Cluster:

∑ 2 2 2 2
W SS(C3 ) = ‖x − m3 ‖ = ‖x2 − m2 ‖ + ‖x3 − m2 ‖ + ‖x5 − m2 ‖
x∈C3

√ 2 √ 2
= (3 − 1.67)2 + (4 − 4)2 + (1 − 1.67)2 + (5 − 4)2

√ 2
+ (1 − 1.67)2 + (3 − 4)2

= 1.7689 + 0 + 0.4489 + 1 + 0.4489 + 1 = 4.6667

The total cohesion is given by the following equation:

∑
W SST otal = W SS(Ci ) = W SS(C1 ) + W SS(C2 ) + W SS(C3 )
i

= 1 + 3.3334 + 4.6667 = 9

• Separation computation

The Separation (Between cluster Sum of Squares) metric is given by the following equation:

∑
K
2
BSST otal = |Ci | · ‖x − mi ‖
i−1

So, our first setp is to compute the center of our dataset, which is given by the mean of all data
points.
Chapter 9 - Centroid-based clustering and Evaluation 93

[ ]
x1 + x2 + x3 + x4 + x5 + x6 + x7 + x8 4.625
m= =
8 5

√ 2
BSST otal = 2 · (4.625 − 7.5)2 + (5 − 1.5)2

√ 2
+ 3· (4.625 − 5.67)2 + (5 − 8.33)2

√ 2
+ 3· (4.625 − 1.67)2 + (5 − 4)2

= 41.031 + 36.589 + 29.255 = 106.875

We can easily compute the above metrics using R with the following commands:

# Perform k-Means clustering

model = kmeans(kdata, centers = kdata[1:3,])

# Compute cohesion
cohesion = model$tot.withinss

# Compute separation
separation = model$betweenss

9.4.4 - Visualize k-Means Clustering

Finally, in order to answer question (d), we can use the following commands:

# Plot data using a different color for each cluster

plot(kdata, col = model$cluster, pch = 15)

# Add data labels

text(kdata, labels = row.names(kdata),
pos = c(4, 4, 4, 4, 4, 4, 2, 4))

# Display centers
points(model$centers, col = 1:length(model$centers),
pch = "+", cex = 2)

The above command will result in the following figure:

Chapter 9 - Centroid-based clustering and Evaluation 94

9.5 - k-Means clustering and evaluation in real life

Application
In this section, we are going to demonstrate the application of k-Means algorithm in a real
life application example. In this context, we have a dataset that contains 300 different 2D data
points that originate from three different distributions. A summary of the dataset is given in the
following table:

X1 X2
Min. -4.720 -3.974
1st Qu. -1.237 -1.059
Median 1.757 2.096
Mean 1.322 1.532
3rd Qu. 3.592 3.675
Max. 6.077 6.689

We will use the aforementioned data in order to answer the following questions:
(a) Plot the data using a different color for each cluster, apply the k-Means algorithm and identify
the optimal number of clusters based on the SSE metric.
(b) Apply k-Means in order to formulate 3 clusters (plot the outcome using a different color for
each cluster along with the centroids) and compute the cohesion and the separation.
(c) Calculate silhouette coeffiecient and create the silhouette plot.
Chapter 9 - Centroid-based clustering and Evaluation 95

(d) Plot the heatmap for the clustered data.

9.5.1 - Identify optimal number of clusters

At first, we will read the given dataset using the following commands:

# Import cluster library

library(cluster)

# Read data
cdata = read.csv("cdata.txt")

# The third column contains the cluster for each data point
target = cdata[, 3]

# The first two columns contain the data coordinates

cdata = cdata[, 1:2]

After that, we can plot the dataset using the following command:

# Plot data
plot(cdata, col = target)

The output is the following figure, where tha data point for each cluster are illustrated with
different color (red, black, green):

In order to identify the optimal number of clusters for the given dataset, we will be based on
the values of SSE metric. To that end, we will perform k-Means clustering setting the number of
Chapter 9 - Centroid-based clustering and Evaluation 96

clusters from 1 to 10 and calculate SSE for each different clustering approach. This process can
be done using the following commands:

# Initialize a vector to hold the SSE values

SSE <- (nrow(cdata) - 1) * sum(apply(cdata, 2, var))

# Calculate SSE values for the different clusterings

for (i in 2:10)
SSE[i] <- kmeans(cdata, centers = i)$tot.withinss

Once finished, we can plot the SSE values using the command:

plot(1:10, SSE, type="b", xlab="Number of Clusters",

ylab="SSE")

As we can see from the SSE values, following the elbow-method, the optimal number of clusters
for the given dataset is 2 or 3.

9.5.2 - k-Means modeling

In order to answer question (b), we can use the following commands:
Chapter 9 - Centroid-based clustering and Evaluation 97

# Create k-Means model

model = kmeans(cdata, centers = 3)

# Get clusters' centroids

model$centers

# Get distribution of data point into clusters

model$cluster

# Calculate cohesion
cohesion = model$tot.withinss

# Calculate separation
separation = model$betweenss

# Plot clustering result

plot(cdata, col = model$cluster)
points(model$centers, col = 4, pch = "+", cex = 2)

The following figure depicts the clustering outcome:

9.5.3 - Silhouette computation

For computing and plotting the silhouette coefficient (question (c)), we can use the following
commands:
Chapter 9 - Centroid-based clustering and Evaluation 98

# Compute silhouette
model_silhouette = silhouette(model$cluster, dist(cdata))

# Plot silhouette
plot(model_silhouette)

In addition, we can compute the mean silhouette using the following command:

mean_silhouette = mean(model_silhouette[, 3])

9.5.3 - Heatmap construction

In order to construct the heatmap, we first sort the data based on the cluster they belong using
the following command:
Chapter 9 - Centroid-based clustering and Evaluation 99

cdata_ord = cdata[order(model$cluster),]

Then, we plot the heatmap using the command:

heatmap(as.matrix(dist(cdata_ord)), Rowv = NA, Colv = NA,

col = heat.colors(256), revC = TRUE)

9.6 - k-Medoids Application

For applying the k-Medoids algorithm, we will use the data contained in the following table:

Rank Topic
Conference_1 High SE
Conference_2 Low SE
Conference_3 High ML
Conference_4 Low DM
Conference_5 Low ML
Conference_6 High SE

The above dataset presents the rank and the topic of 6 different conferences. SE satnds for
Software Engineering, ML for Machine Learning and DM for Data Mining.
At first, we will construct our data frame containing the above dataset with the following
commands:
Chapter 9 - Centroid-based clustering and Evaluation 100

# Vector that holds Rank data

Rank = c("High", "Low", "High", "Low", "Low", "High")

# Vector that holds Topic data

Topic = c("SE", "SE", "ML", "DM", "ML", "SE")

# Data frame
conferences = data.frame(Rank, Topic)

We can apply k-Medoids using the following commands:

# Import library
library(cluster)

# Create k-Medoids model

model = pam(conferences, 3)

We can display the medoids and the distribution of the data instances into clusters using the
following commands:

# Get medoids
model$medoids

# Get distribution of samples into clusters

model$clustering

We can also display the medoids by their names using the following command:

conferences[model$id.med,]

Finally, we can plot the data using the following commands:

# Get the different ranks

L1 = levels(conferences$Rank)

# Get the different topics

L2 = levels(conferences$Topic)

# Plot data
plot(model$data, xaxt = "n", yaxt = "n", pch = 15, col = model$cluster)

# Insert axes labels

axis(1, at = 1:length(L1), labels = L1)
axis(2, at = 1:length(L2), labels = L2)

# Display medoids
points(conferences[model$id.med,], col = 1:3, pch = "o", cex = 2)
Chapter 9 - Centroid-based clustering and Evaluation 101
Chapter 10 - Connectivity-based
clustering
Connectivity-based clustering, also known as hierarchical clustering, is based on the core
idea that every instance of a given dataset is related with every other instance. The main idea
behind the aforementioned assumption is that the degree of this relation is stronger for the
“nearby” objects than to objects farther away. The term nearby implies that these algorithms
connect data instances to form clusters based on their distance. A cluster can be described largely
by the maximum distance needed to connect parts of the cluster. At different distances, different
clusters will form, which can be represented using a dendrogram, which explains where the
common name “hierarchical clustering” comes from: these algorithms do not provide a single
partitioning of the data set, but instead provide an extensive hierarchy of clusters that merge
with each other at certain distances.

10.1 - Hierarchical clustering in R

10.1.1 - Data Construction

In order to provide a detailed overview of how hierarchical clustering can be performed using
R packages, let’s consider the following simple dataset containing five instances of the 2-
dimentional space.

X Y
x1 2 0
x2 8 4
x3 0 6
x4 7 2
x5 6 1

First of all, we need to properly construct our data frame containing the above dataset. In order
to do so, we can use the following commands:

# Construct vector containing X-axis data

X = c(2, 8, 0, 7, 6)

# Construct vector containing Y-axis data

Y = c(0, 4, 6, 2, 1)

# Construct vector containing the data labels which correspond to the row names
rnames = c("x1", "x2", "x3", "x4", "x5")

# Create the data frame

hdata = data.frame(X, Y, row.names = rnames)
Chapter 10 - Connectivity-based clustering 103

Once having created our data frame, we can plot our data using the following commands:

# Plot the data

plot(hdata, pch = 15)

# Add the data labels

text(hdata, labels = row.names(hdata), pos = c(4, 2, 4, 4, 4))

Our data is illutrated in the following figure:

10.1.2 - Hierarchical Clustering using Single and Complete

Linkage
In this subsection, we will perform hierarchical clustering using single and complete linkage
using the aforementioned dataset.
As already noted, hierarchical clustering is an iterative process that starts with the individual
elements of a given dataset in order to progressively form clusters. These clusters are formed by
merging elements that appear to have a strong connection. This connection is expressed using
a distance metric. A very common distance metric is the Euclidean distance, which is given by
the following equation:
√∑
‖x1 − x2 ‖2 = (x1 − x2 )2
i

At first, we will go through the single linkage algorithm:

Chapter 10 - Connectivity-based clustering 104

1st iteration:
The following table presents the distnce matrix for the aforementioned dataset.

x1 x2 x3 x4 x5
x1 0 7.21 6.32 5.39 4.12
x2 7.21 0 8.25 2.24 3.61
x3 6.32 8.25 0 8.06 7.81
x4 5.39 2.24 8.06 0 1.41
x5 4.12 3.61 7.81 1.41 0

As shown in the table, the points x4 and x5 are the ones having the smallest distance. As a result,
we link them together and thus we formulate our first cluster.
2nd iteration:
Since we are using single linkage, the distance between the cluster containing the points x4 and
x5 and every other point is the minimum of the distance between this point and the points x4
and x5.
The following table presents the updated distance matrix.

x1 x2 x3 x4 - x5
x1 0 7.21 6.32 4.12
x2 7.21 0 8.25 2.24
x3 6.32 8.25 0 7.81
x4 - x5 4.12 2.24 7.81 0

Again, using the same strategy, we link the point x2 with the cluster (x4 - x5).
3rd iteration:
In the next iteration, the distance matrix is updated again and the point x1 is linked with the
cluster (x2 - x4 - x5).
x1 x2 - x4 - x5 x3
x1 0 4.12 6.32
x2 - x4 - x5 4.12 0 7.81
x3 6.32 7.81 0

4th iteration:
Finally, the clustering is completed in the 4th iteration where the final point x3 is linked with the
cluster (x1 - x2 - x4 - x5).
x1 - x2 - x4 - x5 x3
x1 - x2 - x4 - x5 0 6.32
x3 6.32 0

At this point we will go through the same iterative process again, but for the case of complete
linkage.
Chapter 10 - Connectivity-based clustering 105

1st iteration:
The first step is the same as in the case of single linkage, where we compute the distance matrix
of our original dataset and link the points x4 and x5 which are the ones having the smallest
distance.
x1 x2 x3 x4 x5
x1 0 7.21 6.32 5.39 4.12
x2 7.21 0 8.25 2.24 3.61
x3 6.32 8.25 0 8.06 7.81
x4 5.39 2.24 8.06 0 1.41
x5 4.12 3.61 7.81 1.41 0

2nd iteration:
As opposed to the case of single linkage, in complete linkage the distance between the cluster
containing the points x4 and x5 and every other point is the maximum of the distance between
this point and the points x4 and x5.
The following table presents the updated distance matrix.

x1 x2 x3 x4 - x5
x1 0 7.21 6.32 5.39
x2 7.21 0 8.25 3.61
x3 6.32 8.25 0 8.06
x4 - x5 5.39 3.61 8.06 0

Based on the updated distances, we link the point x2 with the cluster (x4 - x5). Up to this point,
the results are the same with the case of single linkage.
3rd iteration:
In the next iteration, the distance matrix is updated again and this time the closest point are
found to be x1 and x3. As a result, they are linked together and formulate the cluster (x1 - x3)

x1 x2 - x4 - x5 x3
x1 0 7.21 6.32
x2 - x4 - x5 7.21 0 8.25
x3 6.32 8.25 0

4th iteration:
Finally, the clustering is completed in the 4th iteration where the clusters (x1 - x3) and (x2 - x4 -
x5) are linked together.

x1 - x3 x2 - x4 - x5
x1 - x2 - x4 - x5 0 8.25
x3 8.25 0
Chapter 10 - Connectivity-based clustering 106

10.1.3 - Modeling with R

We can use the following R command in order to compute the distance matrix.

d = dist(hdata)

Have in mind that hdata is the already constructed data frame that contains our dataset (see
Data Construction section).
We can use R commands in order to perform hierarchical clustering and plot the respective
dendrogram that illustrates the formulated clusters.

# Perform clustering using single linkage

hc_single = hclust(d, method = "single")

# Plot the respective dendrogram

plot(hc_single)

The following figure depicts the produced dengrogram.

We can use the same commands for the case of complete linkage.
Chapter 10 - Connectivity-based clustering 107

# Perform clustering using complete linkage

hc_single = hclust(d, method = "complete")

# Plot the respective dendrogram

plot(hc_single)

The following figure depicts the produced dengrogram for the case of complete linkage.

It is worth noticing that the Y-axis of the dendrograms refers to the distance metric.
Once having performed the hierarchical clustering, We can split the original data into a number
of clusters based on our preference using the following command:

# Split data into clusters. k refers to the number of clusters

clusters = cutree(hc_single, k = 2)
# -------- OR --------
clusters = cutree(hc_complete, k = 2)

In order to demonstrate the formulated clusters from the two aforementioned approaches (single
and complete linkage), we can plot using the following commands:
Chapter 10 - Connectivity-based clustering 108

# Single linkage result

plot(hdata, col = clusters, pch = 15, main = "Single Linkage")
text(hdata, labels = row.names(hdata), pos = c(4, 2, 4, 4, 4))

# Complete linkage result

plot(hdata, col = clusters, pch = 15, main = "Complete Linkage")
text(hdata, labels = row.names(hdata), pos = c(4, 2, 4, 4, 4))

The following figure depicts the two formulated clusters for both the cases of single and complete
linkage.

10.2 - Hierarchical Clustering Application and

Evaluation
In order to apply hierarchical clustering into a real world example, we will use the dataset
europe.txt. This dataset contains statistics regarding the Gross Domestic Product (GDP), the
Inflation and the Unemployment in 28 EU countries. The first 6 entries are presented in the
table below:
GDP Inflation Unemployment
Austria 41600 3.5 4.2
Belgium 37800 3.5 7.2
Bulgaria 13800 4.2 9.6
Croatia 18000 2.3 17.7
Czech Republic 27100 1.9 8.5
Denmark 37000 2.8 6.1
… … … …

We will this dataset in order to answer the following questions:

(a) Apply hierarchical clustering using complete linkage and plot the respective dendrogram
(b) Select the appropriate number of clusters based on silhouette coefficient
Chapter 10 - Connectivity-based clustering 109

(c) Use the given data in order to form 7 clusters and produce the respective dendrogram that
illustrates the results
(d) Create a 3D plot of the data using a different color for each cluster
(e) Calculate and plot the silhouette coefficient

10.2.1 - Create Hierarchical Clustering Model

At first, in order to answer (a), we will perform hierarchical clustering using the given data.
Our initial step involves importing the libraries necessary for answering the above questions
along with importing the given data. This is done using the following commands:

# Import libraries
library(cluster)
library(scatterplot3d)

# Read data from file

europe = read.csv("europe.txt")

After reading the given dataset, we scale the data and we compute the distance matrix using
Eucledean distance as our distance metric.

# Get distance matrix

d = dist(scale(europe))

After the computation of the distance matrix, we are ready to perform hierarchical clustering
using complete linkage.

# Perform hierarchical clustering using complete linkage

hc <- hclust(d, method = 'complete')

We can plot the respective dendrogram (illustrated in the following Figure) using the following
command:

# Plot dendrogram
plot(hc)
Chapter 10 - Connectivity-based clustering 110

10.2.2 - Selecting Optimal Number of Clusters

Our next goal is to determine the optimal number of clusters. Our selection will be based on the
values of silhouette coefficient which is used as a measure of how similar an object is to its own
cluster (cohesion) compared to other clusters (separation).
In order to determine the optimal number of clusters, we will use the constructed (from question
(a)) hierarchical clustering model in order to form 2, … , 20 clusters and then select the case that
maximizes silhouette. The following R code can be used to create the different splits and store
the silhouette coeffiecient values:

# Initialize a vector that will hold the silhouette values

slc = c()

# Iterate over the number of clusters

for (i in 2:20){
# Create clusters
clusters = cutree(hc, k = i)

# Calculate and store silhouette values

slc [i-1] = mean(silhouette(clusters, d)[, 3])
}

After the computation of the silhouette values for the different clusterings, we can plot them
using the following command:

# Initialize a vector that will hold the silhouette values

plot(2:20, slc, type="b", xlab="Number of Clusters",
ylab="Silhouette")

The following figure depicts the mean silhouette values for the different number of clusters. As
shown form the graph, the optimal number of clusters that maximizes silhouette coefficient is 7.
Chapter 10 - Connectivity-based clustering 111

10.2.3 - Visualizing Clusters

In order to answer question (c), we will first use the constructed hierarchical clustering model in
order to form 7 clusters and then use these clusters to visualize the data of the original dataset
into a 3D plot.

# Form the 7 clusters

clusters = cutree(hc, k = 7)

Uusing the aforementioned command, we have split our data into 7 clusters. The variable
clusters holds the information of how the clusters are distributed among the data instances.

In order to display the constructed clusters on the produced dendrogram, we can use the
following commands:

# Plot dendrogram
plot(hc)

# Display clusters
rect.hclust(hc, k = 7)

The following figure is the output of the aforementioned commands, where each red rectangle
refers to each cluster.
Chapter 10 - Connectivity-based clustering 112

We can also create a 3D plot of the clusters using the following commands:

# Plot dendrogram
s3d = scatterplot3d(europe, angle = 125, scale.y = 1.5,
color = clusters)

coords <- s3d$xyz.convert(europe)

text(coords$x, coords$y, labels=row.names(europe),

pos=sample(1:4), col = clusters)
Chapter 10 - Connectivity-based clustering 113

10.2.4 - Silhouette Plot

Finally, in order to calculate and plot the silhouette coefficient (question (d)), we can use the
following commands:

# Calculate silhouette
model_silhouette = silhouette(clusters, d)

# Plot silhouette
plot(model_silhouette)
Chapter 10 - Connectivity-based clustering 114
Chapter 11 - Density-based
clustering
Density-based clustering is based on the ground truth that given a cetrain dataset, clusters
can be defined as areas of higher density than the remainder of the dataset. Data instances
within the sparse areas, which are required to separate clusters, are usually considered to be
noise. The most famous and widely used density-based algorithm is DBSCAN (Density-based
Spatial Clustering of Applications with Noise)which was introduced by Martin Ester, Hans-Peter
Kriegel, Jorg Sander and Xiaowei Xu in 1996. The main idea behind the DBSCAN algorithm is
that given a set of points in some n-dimantional space, points that are closely packed together
(points with many nearby neighbors and thus of higher density) are grouped into clusters, while
points that lie alone in low-density regions are considered as outliers.

11.1 - DBSCAN in R
In order to use DBSCAN in R, we need to use the cluster library which can be imported with the
following command:

library(dbscan)

In order to perform clustering using DBSCAN and get the distribution of the input dataset into
clusters, we can use the following commands:

# Perform clustering using DBSCAN

model = dbscan(ddata, eps = 1.0, minPts = 4)

# Get distribution of the input dataset into clusters

clusters = model$cluster

DBSCAN algorithm requires setting 2 main parameters:

• epsilon (eps), which specifies the maximum distance that two data points can have in
order to be considered as part of the same cluster.

and

• minPoints (minPts), which specifies the minimum number of data points required to form
a cluster.

11.2 - DBSCAN model construction

In an effort to construct a density-based clustering model using DBSCAN algorithm, we will use
the data of the following table:
Chapter 11 - Density-based clustering 116

X Y
x1 2 10
x2 2 5
x3 8 4
x4 5 8
x5 7 5
x6 6 4
x7 1 2
x8 4 9

In order to visualize the data using R, we will first need to create the resperctive data frame usinf
the following commands:

# Construct vector containing X-axis data

X = c(2, 2, 8, 5, 7, 6, 1, 4)

# Construct vector containing Y-axis data

Y = c(10, 5, 4, 8, 5, 4, 2, 9)

# Construct vector containing the data labels which correspond to the row names
rnames = c("x1", "x2", "x3", "x4", "x5", "x6", "x7", "x8")

# Create the data frame

ddata = data.frame(X, Y, row.names = rnames)

Once having created our data frame, we can plot our data using the following commands:

# Plot the data

plot(ddata, pch = 15)

# Add the data labels

text(ddata, labels = row.names(hdata),
pos = c(4, 2, 4, 4, 4, 4, 4, 4))

Our data is illutrated in the following figure:

Chapter 11 - Density-based clustering 117

11.3 - DBSCAN calculation

In this subsection, we will apply DBSCAN algorithm using minPoints = 2 and epsilon = {2, 3.5}
in order to distribute the given data into clusters.

• minPoints = 2 and epsilon = 2

We first need to compute the distance matrix for the given dataset which is given in the following
table:
x1 x2 x3 x4 x5 x6 x7 x8
x1 0.00
x2 5.00 0.00
x3 8.49 6.08 0.00
x4 3.61 4.24 5.00 0.00
x5 7.07 5.00 1.41 3.61 0.00
x6 7.21 4.12 2.00 4.12 1.41 0.00
x7 8.06 3.16 7.28 7.21 6.71 5.39 0.00
x8 2.24 4.47 6.40 1.41 5.00 5.39 7.62 0.00

Given the fact that minPoints value is 2, each cluster should contain at least two points. In
addition, based on the epsilon values, which is 2, DBSCAN forms the following clusters:

C1 : {x3 , x5 , x6 } and C2 : {x4 , x8 }

Chapter 11 - Density-based clustering 118

The remaining points are considered as noise.

• minPoints = 2 and epsilon = 3.5

x1 x2 x3 x4 x5 x6 x7 x8
x1 0.00
x2 5.00 0.00
x3 8.49 6.08 0.00
x4 3.61 4.24 5.00 0.00
x5 7.07 5.00 1.41 3.61 0.00
x6 7.21 4.12 2.00 4.12 1.41 0.00
x7 8.06 3.16 7.28 7.21 6.71 5.39 0.00
x8 2.24 4.47 6.40 1.41 5.00 5.39 7.62 0.00

In the case of epsilon = 3.5, DBSCAN forms the following clusters:

C1 : {x1 , x4 , x8 }, C2 : {x2 , x7 } and C3 : {x3 , x5 , x6 }

As we can see no point is considered as noise.

11.4 - DBSCAN clustering with R

In order to perform clustering using DBSCAN with R, we can use the following command:

model = dbscan(ddata, eps = 2, minPts = 2)

In order to get the distribution of the data point into clusters, we can use the following command:

clusters = model$cluster

Finally, we can plot the clustering results using the following commands:

# Plot clusters
plot(ddata, col=clusters+1, pch=15, main="DBSCAN(eps = 2, minPts = 2)")

# Add data labels

text(ddata, labels = row.names(ddata), pos = 4)

In the following figure, we can see the formulated clusters from the previous subsection.
Chapter 11 - Density-based clustering 119

11.5 - Density-based Clustering Application

In this section, we are going to demonstrate the application of DBSCAN algorithm in a real
life application example. In this context, we have a dataset that contains 2,000 different 2D data
points. A summary of the dataset is given in the following table:

X1 X2
Min. 5.09 7.54
1st Qu. 10.45 9.78
Median 13.30 12.01
Mean 13.28 12.03
3rd Qu. 16.13 14.27
Max. 21.43 16.48

We will use the aforementioned data in order to answer the following questions:
(a) Plot the given data points
(b) Apply k-Means algorithm in order to create two clusters and plot the result using different
color for each cluster. Is the result satisfying? Do the created clusters correspond to the actual
ones?
(c) Optimize esp parameter for DBSCAN using kNN distance (use k = 10).
(d) Apply DBSCAN using eps = 0.4 and minPts = 10 into the given dataset. Plot the result using
different color for each cluster and black for the data points considered as noise.

11.5.1 - Modeling using k-Means

Our first step is to read and plot the given dataset using the following commands:
Chapter 11 - Density-based clustering 120

# Plot library
library(dbscan)

# Read data
mdata = read.csv("mdata.txt")

# Plot data
plot(mdata)

The following figure illustrates the given dataset. As we can see from the plot, we can identify
two distinguished clusters.

Then, we apply k-Means and plot the clustering results using the following commands:

# Construct k-Means model

model = kmeans(mdata, 2)

# Plot data
plot(mdata, col = model$cluster + 1)
Chapter 11 - Density-based clustering 121

As we can see, k-Means is not able to suffieciently split the data into clusters. This result
originates from the fact that k-Means relies on the Euclidean distance and thus cannot distingush
clusters organized in complex structures.

11.5.2 - Optimizing parameter eps

In order to optimize the parameter eps of the DBSCAN algorithm (quaestion (c)), we will use
kNN distance. We can use the following R command in order to compute the destance of every
data point with its 10 nearest neighbours:

knndist = kNNdist(mdata, k = 10)

Then, we can plot the respective graph after sorting the data into ascending order:

# Get the distances of the 10 nearest neighbours for each point

kdist = knndist[, 10]

# Plot distances
plot(sort(kdist), type = 'l', xlab = "Points sorted by distance",
ylab = "10-NN distance")
Chapter 11 - Density-based clustering 122

In order to optimize eps, we follow the elbow method. As we can see from the graph, the
optimized eps valus are in the interval [0.35, 0.45].

11.5.3 - Modeling using DBSCAN

Lastly, we can answer question (d) by applying DBSCAN on the given dataset using eps = 0.4
and minPoints = 10.

# Construct DBSCAN model

model = dbscan(mdata, eps = 0.4, minPts = 10)

# Plot distances
plot(mdata, col = model$cluster + 1, pch = ifelse(model$cluster, 1, 4))

The following figure depicts the clustering result. As we can see, DBSCAN is able to efficiently
split the given dataset into clusters.
Chapter 11 - Density-based clustering 123
Chapter 12 - Distribution-based
clustering
12.1 - Theoretical background
Distribution-based clustering is closely related to statistics and is based on the fact that
clusters can easily be defined as objects including data instances belonging most likely to the
same distribution. While the theoretical foundation of the methods that apply to this clustering
category is excellent, they suffer from one key problem known as overfitting, unless cetrain
constraints are set into the complexity of the model. A more complex model will usually be able
to explain the data better, which makes choosing the appropriate level of complexity inherently
difficult. One prominent and widely used method is known as Gaussian Mixture Models or
GMMS (using the Expectation-Maximization algorithm or EM).

12.1.1 - The Expectation-Maximization algorithm

In statistics, an Expectation-Maximization (EM) algorithm is an iterative method to find
maximum likelihood or maximum a posteriori (MAP) estimates of parameters in statistical
models, where the model depends on unobserved latent variables. In practice EM algorithm
maximizes the following expression:

X : dataset, Z : latent variables, θ : set of parameters

∑
ln P (X/θ) = ln P (X, Z/θ)
Z

The EM iteration alternates between performing an expectation step (E-step), which creates
a function for the expectation of the log-likelihood evaluated using the current estimate for
the parameters, and a maximization step (M-step), which computes parameters maximizing
the expected log-likelihood found on the E step. These parameter estimates are then used to
determine the distribution of the latent variables in the next E-step.
E-step:

Given θold : f ind P (Z/X, θold )

M-step:

∑
θnew = argmaxθ ( P (Z/X, θold ) ln P (X, Z/θ))
Z
Chapter 12 - Distribution-based clustering 125

The EM algorithm reaches convergence when the log-likelihood value is no longer updated or
the updat is smaller than a given threshold.

12.1.2 - Gaussian Mixture Models - GMMs

A Gaussian Mixture Model is a probabilistic model that assumes all the data points are generated
from a mixture of a finite number of Gaussian distributions with unknown parameters. One can
think of mixture models as generalizing k-means clustering to incorporate information about the
covariance structure of the data as well as the centers of the latent Gaussians.
Generally, a mixture of k normal distributions is given by the following equation:

∑
K
P (x) = πk N (x/µk , Σk )
k=1

where

∑
K
πk : M ixing Coef f icients ∈ [0, 1] and πk = 1
k=1

µk , Σk : M ean and Covariance of distribution k

1 Σ −1 (x−µ)
e− 2 (x−µ)
1 T
N : N (x/µ, Σ) = 1/2
2π n/2 |Σ |

GMMs use EM as the optimization algorithm that enables calculating the latent variables of the
distributions that describe the goven data. As a result, the EM algorithm in the case of GMMs
maximizes the following quantity:

∑
N ∑
K
ln P (X/π, µ, Σ) = ln( πk N (xn /µk , Σk ))
n=1 k=1

where X = {x1 , x2 , . . . , xn } −→ Given dataset

12.2 - Modeling Gaussian Mixture Models using EM

In order to perform clustering using GMMs we will use a dataset that contains 1,000 1D data
instances that originate from two different distributions. Regarding the distributions, we know
that their standard deviation is 1. A summary of the data is given in the following table:
Chapter 12 - Distribution-based clustering 126

D1_X D2_X
Min. -1.905 3.992
1st Qu. 0.223 6.322
Median 0.907 6.944
Mean 1.000 6.985
3rd Qu. 1.664 7.681
Max. 3.659 10.810

Using the above data, we will answer the following questions:

(a) Plot the data (1 one dimention) along with the probability density function using a different
color for each distribution.
(b) Calculate the means (µ1 , µ2 ) and the latent variables λ1 , λ2 for each one of the normal
distributions.
(d) Plot the estimated and the real probability density function of the given data in the same
figure.

12.2.1 - Data Plot

In order to answer the first question, we will first read the data using the following commands:

# Read data
gdata = read.csv("gdata.txt")

# Save 1D data in the vector x

x = gdata[, 1]

# Save cluster index in the vector y

y = gdata[, 2]

In order to plot the data along with the probability density function, we can use the following
commands:

# Plot data
plot(data.frame(x, 0), ylim = c(-0.01, 0.25), col = y,
xlab = "Data",
ylab = "Density")

# Plot probability density function

lines(density(x), col = y)

The following figure illustrates the given dataset along with the probability density function.
The data from each cluster are illustrated with a different color (black and red)
Chapter 12 - Distribution-based clustering 127

12.2.2 - Manual calcultion of the parameters of the normal

distibutions
In this subsection, we will use the given dataset in order to calculate the parameters of the
normal distributions using EM algorithm. At first, we will choose random initial values for the
parameters µ and λ:

Initialization : µ1 = 0, µ2 = 1, λ1 = 0.5, λ2 = 0.5

Θ = {θ1 , θ2 } = {(0, 1, 0.5), (1, 1, 0.5)}

E-step:
In this step, we want to calculate the probability for a given data point to originate from a certain
normal distibution. This probability is given by the following equation:

λj · P (xi /θj )
P (Dj /xi , θj ) =
λ1 · P (xi /θ1 ) + λ2 · P (xi /θ2 )

where : j = 1, 2 λinit = {λ1 , λ2 } = {0.5, 0.5}

At this point, we will compute the above probability for the given point xi = 2:

(x −µ )2
1 − i2σ j
P (xi /θj ) = √ e j
2πσj
Chapter 12 - Distribution-based clustering 128

1 (2−0)2 1
P (2/θ1 ) = √ e− 2·1 = √ e−2 = 0.054
2π · 1 2π

1 (2−1)2 1
P (2/θ2 ) = √ e− 2·1 = √ e−1/2 = 0.242
2π · 1 2π

The probabilities of the given point to originate from distributions 1 and 2 are given by the
following two equations:

0.5 · 0.054
P (D1 /2, θ1 ) = = 0.182
0.5 · 0.054 + 0.5 · 0.242

0.5 · 0.242
P (D2 /2, θ2 ) = = 0.818
0.5 · 0.054 + 0.5 · 0.242

Given the aforementioned results, the point xi = 2 originates from distribution 1 with probability
0.182 and from distribution 2 with probability 0.818. Once having computed the aforementioned
probabilities for all 1,000 given data points, we will proceed in the maximization step of the EM
algorithm.
M-step:
In this step, we will calculate the parameters µ1 , µ2 that maximize the probabilities P (D1 /Xi , θ1 ), P (D2 /Xi , θ2 ).
In order to compute the parameters µ1 , µ2 , we will use the following equations:

∑
1000
xi · P (D1 /xi , θ1 )
i=1
µ1 =
∑
1000
P (D1 /xi , θ1 )
i=1

∑
1000
xi · P (D2 /xi , θ2 )
i=1
µ2 =
∑
1000
P (D2 /xi , θ2 )
i=1

The above equations calculate the weighted average of all data points, where the weights are
actually the probabilities for each data point to originate from one of the two distributions.
The mixing coefficients are given by the following equations:

∑
1000
P (D1 /xi , θ1 )
i=1
λ1 =
1000

∑
1000
P (D2 /xi , θ2 )
i=1
λ2 =
1000
Chapter 12 - Distribution-based clustering 129

After computing the values of the parameters µ1 , µ2 , λ1 , λ2 , the log-likelihood of the given dataset
is given by the following equation:

∑
1000 ∑
2
ln(P (X/θ)) = ln( λj · P (xi /θj ))
i=1 j=1

∑
1000
= ln(λ1 · P (xi /θ1 ) + λ2 · P (xi /θ2 ))
i=1

Once having computed the log-likelihood of the dataset, we compare it with the value of the
previous iteration. If the value has not changed more than a given threshold (espilon value),
then the algorithm converges. Otherwise, we return back in the E-step to re-calculate the new
parameters’ values and continue the execution.

12.2.3 - Calculting the parameters of the normal distibutions

using R
The R algorithm that enables us to compute the parameters of the normal distibutions is the
following and involves two main steps, the initialization and the loop until convergence:

• Step 1: Initialize the means and the latent variables

# Initialize means
mu = c(0, 1)

# Initialize lambdas
lambda = c(0.5, 0.5)

# Set stopping criterion

epsilon = 1e-08

# Compute initial log-likelihood

log_likelihood = sum(log(lambda[1] * dnorm(x, mean = mu[1], sd = 1) +
lambda[2] * dnorm(x, mean = mu[2], sd = 1)))

• Step 2: Loop until convergence

repeat {
# ------ Expectation step ------
# Find distributions given mu, lambda (and sigma)
T1 <- dnorm(x, mean = mu[1], sd = 1)
T2 <- dnorm(x, mean = mu[2], sd = 1)
P1 <- lambda[1] * T1 / (lambda[1] * T1 + lambda[2] * T2)
P2 <- lambda[2] * T2 / (lambda[1] * T1 + lambda[2] * T2)

# ------ Maximization step ------

# Find mu, lambda (and sigma) given the distributions
mu[1] <- sum(P1 * x) / sum(P1)
Chapter 12 - Distribution-based clustering 130

mu[2] <- sum(P2 * x) / sum(P2)

lambda[1] <- mean(P1)
lambda[2] <- mean(P2)

# Calculate the new log likelihood (to be maximized)

new_log_likelihood = sum(log(lambda[1] * dnorm(x, mean = mu[1], sd = 1) +
lambda[2] * dnorm(x, mean = mu[2], sd = 1)))

# Print the current parameters and the log likelihood

cat("mu =", mu, " lambda =", lambda, " log_likelihood =", new_log_likelihood,\
"\n")

# Break if the algorithm converges

if (new_log_likelihood - log_likelihood <= epsilon) break
log_likelihood = new_log_likelihood
}

12.2.4 - GMMs model construction using mixtools

In order to answer question (c), we will construct a GMMs clustering model using R. The library
we are going to use is mixtools:

library(mixtools)

We can use the expectation maximization (EM) algorithm and overview the final calculated
parameters using the following commands:

# Apply EM using the aforementioned initialization

model <- normalmixEM(x, mu = c(0,1), sd.constr = c(1,1))

# Get means of the distrubitions

model$mu

# Get lambda values

model$lambda

# Get final log-likelihood

model$loglik

We can plot both the original and the estimated probability density functions (into the same plot)
using the following commands:

# Plot estimated probability density function

plot(model, which = 2)

# Plot original probability density function

lines(density(x), lty = 2, lwd = 2)

The following figure illustrates the probability density functions. The dashed line refers to the
actual density, while the straight line to the estimated. As we can see, the estimation lies very
close to the actual density.
Chapter 12 - Distribution-based clustering 131

12.3 - GMMs Clustering Application

In this section, we are going to demonstrate the application of GMMs in a real life application
example. In this context, we have a dataset that contains 300 different 2D data points that
originate from three different distributions. A summary of the two distibutions is given in the
following table:

D1_X1 D1_X2 D1_X1 D1_X2

Min. 3.33 5.57 17.64 5.47
1st Qu. 8.29 9.01 22.43 8.56
Median 10.07 10.22 24.91 10.02
Mean 1.32 1.53 -4.72 -3.97
3rd Qu. 11.88 11.38 27.44 11.48
Max. 17.32 14.80 31.27 23.47

Then, we will use the given dataset in order to answer the following questions:
(a) Plot the given dataset using a different color for each cluster.
(b) Constuct a GMMs model inorder to cluster the provided dataset into 3 clusters. Set epsilon
value to 0.1.
(c) Show how the EM algorithm converges and plot the data using the estimated probability
density function
Chapter 12 - Distribution-based clustering 132

(d) Assign each data point to the cluster with the highest probability and plot the formulated
clusters along with their centroids.
(e) Calculate and plot silhouette coefficient
(f) Show the heatmap of the data after sorting them based on the above clustering results

12.3.1 - Dataset construction

At first, we will read the given dataset using the following commands:
# Read data
gsdata = read.csv("gsdata.txt")

# The third column contains the cluster for each data point
target = gsdata[, 3]

# The first two columns contain the data coordinates

gsdata = gsdata[, 1:2]

We can plot tha data using the following command:

plot(gsdata, col = target)

The output is the following figure, where tha data points for each cluster are illustrated with
different colors (red, black, green):

12.3.2 - GMMs model construction

In order to construct the GMMs model, we will use the following commands:
Chapter 12 - Distribution-based clustering 133

# Import library mixtools

library(mixtools)

# Construct model
model = mvnormalmixEM(gsdata, k = 3 , epsilon = 0.1)

Once we have constructed the model, we can use the following command to demonstrate how
the algorithm converges:

plot(model, which = 1)

The convergence can be shown in the following figure:

As shown in the graph, it is worth noticing that the EM algorith converges very fast. Since the
2nd iteration, the log-likelihood is almost the same.
Finally, we can plot the given dataset along with the probability density function for each cluster
using the following command:

plot(model, which = 2)

The output is given in the following figure:

Chapter 12 - Distribution-based clustering 134

Using the model$posterior command, we can see the probability for each data point to originate
from each one of the three clusters (soft-assigments). However, in order to answer this question,
we have made hard-assigments using the following commands:

# Assign each point to the cluster with the highest probability

clusters = max.col(model$posterior)

# Calculate centers
centers = matrix(unlist(model$mu), byrow = TRUE, ncol = 2)

Using the above hard-assigmentsl, we can compute the silhouette coeffiecient and construct the
heatmap.

12.3.3 - Silhouette computation

In order to calculate and plot silhouette, we can use the following commands:

# Calculate silhouette
model_silhouette = silhouette(clusters, dist(gsdata))

# Plot silhouette
plot(model_silhouette)
Chapter 12 - Distribution-based clustering 135

12.3.4 Heatmap constuction

In order to construct the heatmap, we can use the following R commands:

# Order data based on the cluster value

gsdata_ord = gsdata[order(clusters),]

# Construct heatmap
heatmap(as.matrix(dist(gsdata _ord)), Rowv = NA, Colv = NA,
col = heat.colors(256), revC = TRUE)
Chapter 12 - Distribution-based clustering 136

12.4 - GMMs Application with Information Criteria

In this section, we will perfrom clustering using GMMs based on information criteria in order
to select the optimal number of clusters. Towards this direction, we will use a dataset containing
300 1D data points. A summary of the given dataset is given in the following table:

X
Min. -2.2239
1st Qu. 0.6396
Median 5.1139
Mean 5.0336
3rd Qu. 9.1177
Max. 12.0908

We will use the given dataset in order to answer the following questions:
(a) Plot the data along with the probability density function
(b) Perform clustering using the EM algorithm setting the number of clusters to 2, 3, 4 and 5. For
each case, plot the probability density function and calculate the information criteria AIC and
BIC.
(c) Plot the AIC and BIC values for the different number of clusters and select the optimal number
of clusters
Chapter 12 - Distribution-based clustering 137

12.4.1 - Data Construction

At first, we will read the given dataset using the following commands:

# Import mixtools library

library(mixtools)

# Read data
icdata = read.csv("icdata.txt")

# The first column contains the value of each data point

x = icdata[, 1]

# The second column contains the cluster for each data point
y = icdata[, 2]

After that, we can plot the dataset along with the probability density function using the following
command:

# Plot data
plot(data.frame(x, 0), ylim = c(-0.01, 0.1))

# Plot probability density function

lines(density(x))

12.4.2 - Usage of Information Criteria for selecting the optimal

number of clusters
In order to answer question (b), we need to compute the Akaike Information Criterion (AIC) and
the Bayesian Information Criterion (BIC). For a given model that has t parameters, is applied in
Chapter 12 - Distribution-based clustering 138

N data samples and has log-likelihood L, the aforementioned criteria are given by the following
equations:

AIC = 2 · t − 2 · ln(L)

BIC = t · ln(N ) − 2 · ln(L)

We can cluster the data using 2, 3, 4 and 5 distributions using the following code:

# Vector for holding AIC values

AIC = c()

# Vector for holding BIC values

BIC = c()

# Set a 2x2 plot grid

par(mfrow = c(2, 2))

# Iterate over the number of clusters

for (k in 2:5) {
# Construct model
model <- normalmixEM(x, k = k, epsilon = 0.0001)

# Plot clustering results

plot(model, which = 2 , main2 = paste("Density (k = ", k, ")", sep = ""))

# Calculate number of parameters

numparams = length(model$mu) + length(model$stdev) + length(model$lambda)

# Store AIC and BIC values

AIC[k] = 2 * numparams - 2 * model$loglik
BIC[k] = log(length(x)) * numparams - 2 * model$loglik
}

The following figure depicts the clustering results:

Chapter 12 - Distribution-based clustering 139

Finally, we can plot AIC and BIC values using the following commands:

# Plot AIC values

plot(AIC[2:5], type = 'l', xaxt = "n", main = "AIC", ylab = "AIC")
axis(1, at = 1:4, labels = 2:5)

# Plot BIC values

plot(BIC[2:5], type = 'l', xaxt = "n", main = "BIC", ylab = "BIC")
axis(1, at = 1:4, labels = 2:5)

The AIC and the BIC values in relation to the number of clusters are shown in the following
figure:
Part V - Extended Topics
Chapter 13 - Association Rules
The association rules extraction algorithm is included in the arules library. The supermarkt
transaction data to be used for executing market basket analysis can be found in the Grocery
Shopping datasets page of ACM RecSys. More specifically we will use the Belgium retail market
dataset.
First let’s load the data, read it with the read.transactions function and inspect the first 10 of
them:

library(arules)
fileURL <- "http://fimi.ua.ac.be/data/retail.dat.gz"
download.file(fileURL, destfile="retail.data.gz", method="curl")
# Read the data in basket format
trans = read.transactions("retail.data.gz", format = "basket", sep=" ");
inspect(trans[1:10])

## items
## [1] {0,
## 1,
## 10,
## 11,
## 12,
## 13,
## 14,
## 15,
## 16,
## 17,
## 18,
## 19,
## 2,
## 20,
## 21,
## 22,
## 23,
## 24,
## 25,
## 26,
## 27,
## 28,
## 29,
## 3,
## 4,
## 5,
## 6,
## 7,
## 8,
## 9}
## [2] {30,
Chapter 13 - Association Rules 142

## 31,
## 32}
## [3] {33,
## 34,
## 35}
## [4] {36,
## 37,
## 38,
## 39,
## 40,
## 41,
## 42,
## 43,
## 44,
## 45,
## 46}
## [5] {38,
## 39,
## 47,
## 48}
## [6] {38,
## 39,
## 48,
## 49,
## 50,
## 51,
## 52,
## 53,
## 54,
## 55,
## 56,
## 57,
## 58}
## [7] {32,
## 41,
## 59,
## 60,
## 61,
## 62}
## [8] {3,
## 39,
## 48}
## [9] {63,
## 64,
## 65,
## 66,
## 67,
## 68}
## [10] {32,
## 69}

One can see that each transaction containsa list of item IDs. The function summary given a dataset
read by the read.tansactions function will provide a specialized summary for transactional
data:
Chapter 13 - Association Rules 143

summary(trans)

## transactions as itemMatrix in sparse format with

## 88162 rows (elements/itemsets/transactions) and
## 16470 columns (items) and a density of 0.0006257289
##
## most frequent items:
## 39 48 38 32 41 (Other)
## 50675 42135 15596 15167 14945 770058
##
## element (itemset/transaction) length distribution:
## sizes
## 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
## 3016 5516 6919 7210 6814 6163 5746 5143 4660 4086 3751 3285 2866 2620 2310
## 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
## 2115 1874 1645 1469 1290 1205 981 887 819 684 586 582 472 480 355
## 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45
## 310 303 272 234 194 136 153 123 115 112 76 66 71 60 50
## 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
## 44 37 37 33 22 24 21 21 10 11 10 9 11 4 9
## 61 62 63 64 65 66 67 68 71 73 74 76
## 7 4 5 2 2 5 3 3 1 1 1 1
##
## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 1.00 4.00 8.00 10.31 14.00 76.00
##
## includes extended item information - examples:
## labels
## 1 0
## 2 1
## 3 10

After reading succesfully the transactions we go ahead with our analysis:

rules <- apriori(trans, parameter = list(support = 0.01, confidence = 0.6))

## Apriori
##
## Parameter specification:
## confidence minval smax arem aval originalSupport maxtime support minlen
## 0.6 0.1 1 none FALSE TRUE 5 0.01 1
## maxlen target ext
## 10 rules FALSE
##
## Algorithmic control:
## filter tree heap memopt load sort verbose
## 0.1 TRUE TRUE FALSE TRUE 2 TRUE
##
## Absolute minimum support count: 881
##
## set item appearances ...[0 item(s)] done [0.00s].
## set transactions ...[16470 item(s), 88162 transaction(s)] done [0.22s].
## sorting and recoding items ... [70 item(s)] done [0.01s].
Chapter 13 - Association Rules 144

## creating transaction tree ... done [0.06s].

## checking subsets of size 1 2 3 4 done [0.01s].
## writing ... [84 rule(s)] done [0.00s].
## creating S4 object ... done [0.03s].

quality(rules) <- round(quality(rules), digits=3)

rules

## set of 84 rules

84 rules were generated. To see them use inspect.

inspect(rules)

## lhs rhs support confidence lift count

## [1] {37} => {38} 0.012 0.974 5.505 1046
## [2] {286} => {38} 0.013 0.943 5.333 1116
## [3] {12925} => {39} 0.011 0.639 1.112 938
## [4] {1146} => {39} 0.011 0.689 1.199 983
## [5] {79} => {39} 0.013 0.694 1.208 1111
## [6] {1327} => {39} 0.013 0.647 1.126 1156
## [7] {438} => {39} 0.014 0.676 1.177 1260
## [8] {60} => {39} 0.011 0.660 1.149 983
## [9] {255} => {48} 0.012 0.717 1.500 1057
## [10] {255} => {39} 0.012 0.717 1.248 1057
## [11] {533} => {39} 0.010 0.620 1.079 922
## [12] {270} => {39} 0.014 0.689 1.198 1194
## [13] {2238} => {39} 0.015 0.750 1.306 1287
## [14] {110} => {38} 0.031 0.975 5.513 2725
## [15] {110} => {39} 0.020 0.630 1.095 1759
## [16] {147} => {39} 0.013 0.639 1.112 1137
## [17] {271} => {39} 0.016 0.685 1.191 1434
## [18] {413} => {48} 0.013 0.604 1.263 1135
## [19] {413} => {39} 0.013 0.601 1.046 1130
## [20] {36} => {38} 0.032 0.950 5.372 2790
## [21] {36} => {39} 0.023 0.694 1.207 2037
## [22] {475} => {48} 0.016 0.659 1.379 1428
## [23] {475} => {39} 0.017 0.692 1.204 1500
## [24] {170} => {38} 0.034 0.978 5.529 3031
## [25] {170} => {39} 0.023 0.664 1.156 2059
## [26] {101} => {39} 0.016 0.626 1.089 1400
## [27] {310} => {48} 0.019 0.652 1.365 1692
## [28] {310} => {39} 0.021 0.714 1.242 1852
## [29] {237} => {39} 0.022 0.636 1.107 1929
## [30] {225} => {39} 0.027 0.722 1.256 2351
## [31] {89} => {48} 0.032 0.729 1.526 2798
## [32] {89} => {39} 0.031 0.716 1.246 2749
## [33] {65} => {39} 0.032 0.623 1.084 2787
## [34] {38} => {39} 0.117 0.663 1.154 10345
## [35] {41} => {48} 0.102 0.603 1.263 9018
## [36] {41} => {39} 0.129 0.764 1.329 11414
## [37] {48} => {39} 0.331 0.692 1.203 29142
Chapter 13 - Association Rules 145

## [38] {110,48} => {38} 0.015 0.986 5.575 1361

## [39] {110,38} => {39} 0.020 0.639 1.111 1740
## [40] {110,39} => {38} 0.020 0.989 5.592 1740
## [41] {110,48} => {39} 0.012 0.751 1.307 1037
## [42] {36,48} => {38} 0.015 0.960 5.429 1360
## [43] {36,38} => {39} 0.022 0.697 1.213 1945
## [44] {36,39} => {38} 0.022 0.955 5.398 1945
## [45] {36,48} => {39} 0.013 0.788 1.371 1116
## [46] {475,48} => {39} 0.012 0.765 1.330 1092
## [47] {39,475} => {48} 0.012 0.728 1.523 1092
## [48] {170,48} => {38} 0.017 0.988 5.584 1538
## [49] {170,38} => {39} 0.023 0.666 1.159 2019
## [50] {170,39} => {38} 0.023 0.981 5.543 2019
## [51] {170,48} => {39} 0.014 0.775 1.348 1206
## [52] {101,48} => {39} 0.011 0.722 1.255 946
## [53] {101,39} => {48} 0.011 0.676 1.414 946
## [54] {310,48} => {39} 0.015 0.796 1.385 1347
## [55] {310,39} => {48} 0.015 0.727 1.522 1347
## [56] {237,48} => {39} 0.014 0.740 1.287 1244
## [57] {237,39} => {48} 0.014 0.645 1.349 1244
## [58] {225,48} => {39} 0.016 0.806 1.403 1400
## [59] {48,89} => {39} 0.024 0.759 1.321 2125
## [60] {39,89} => {48} 0.024 0.773 1.617 2125
## [61] {48,65} => {39} 0.020 0.711 1.236 1797
## [62] {39,65} => {48} 0.020 0.645 1.349 1797
## [63] {32,38} => {39} 0.021 0.649 1.130 1840
## [64] {38,41} => {48} 0.027 0.609 1.275 2374
## [65] {38,41} => {39} 0.035 0.783 1.362 3051
## [66] {38,48} => {39} 0.069 0.768 1.336 6102
## [67] {32,41} => {48} 0.023 0.645 1.351 2063
## [68] {32,41} => {39} 0.027 0.738 1.284 2359
## [69] {32,48} => {39} 0.061 0.672 1.170 5402
## [70] {32,39} => {48} 0.061 0.639 1.337 5402
## [71] {41,48} => {39} 0.084 0.817 1.421 7366
## [72] {39,41} => {48} 0.084 0.645 1.350 7366
## [73] {110,38,48} => {39} 0.012 0.758 1.318 1031
## [74] {110,39,48} => {38} 0.012 0.994 5.620 1031
## [75] {36,38,48} => {39} 0.012 0.794 1.382 1080
## [76] {36,39,48} => {38} 0.012 0.968 5.471 1080
## [77] {170,38,48} => {39} 0.014 0.776 1.349 1193
## [78] {170,39,48} => {38} 0.014 0.989 5.592 1193
## [79] {32,38,48} => {39} 0.014 0.751 1.306 1236
## [80] {32,38,39} => {48} 0.014 0.672 1.406 1236
## [81] {38,41,48} => {39} 0.023 0.839 1.459 1991
## [82] {38,39,41} => {48} 0.023 0.653 1.365 1991
## [83] {32,41,48} => {39} 0.019 0.798 1.388 1646
## [84] {32,39,41} => {48} 0.019 0.698 1.460 1646

For example the first rule {37} => {38} says that if someone buys item with code 37, he or she
will probably buy item with code 38 with support 1.2% and confidence 97.4% and a lift of 5.5.
For a more detailed analysis of the arules library check the Introduction to arules14 vignette.

14 https://mran.revolutionanalytics.com/web/packages/arules/vignettes/arules.pdf