P. 1
Customer Segmentation and Profiling Thesis

Customer Segmentation and Profiling Thesis

|Views: 459|Likes:
Publicado porAnkesh Srivastava

More info:

Published by: Ankesh Srivastava on Jan 13, 2011
Copyright:Attribution Non-commercial

Availability:

Read on Scribd mobile: iPhone, iPad and Android.
download as PDF, TXT or read online from Scribd
See more
See less

05/29/2013

pdf

text

original

Sections

  • Introduction
  • 1.1 Customer segmentation and customer pro-
  • 1.1.1 Customer segmentation
  • 1.1.2 Customer profiling
  • 1.2 Data mining
  • 1.3 Structure of the report
  • 2.1 Data warehouse
  • 2.1.1 Selecting the customers
  • 2.1.2 Call detail data
  • 2.1.3 Customer data
  • 2.2 Data preparation
  • Clustering
  • 3.1 Cluster analysis
  • 3.1.1 The data
  • 3.1.2 The clusters
  • 3.1.3 Cluster partition
  • 3.2 Cluster algorithms
  • 3.2.1 K-means
  • 3.2.2 K-medoid
  • 3.2.3 Fuzzy C-means
  • 3.2.4 The Gustafson-Kessel algorithm
  • 3.2.5 The Gath Geva algorithm
  • 3.3 Validation
  • 3.4 Visualization
  • 3.4.1 Principal Component Analysis
  • 3.4.2 Sammon mapping
  • 3.4.3 Fuzzy Sammon mapping
  • 4.1 Determining the optimal number of clusters
  • 4.2 Comparing the clustering algorithms
  • 4.3 Designing the segments
  • Support Vector Machines
  • 5.1 The separating hyperplane
  • 5.2 The maximum-margin hyperplane
  • 5.3 The soft margin
  • 5.4 The kernel functions
  • 5.5 Multi class classification
  • 6.1 K-fold cross validation
  • 6.2 Parameter setting
  • 6.3 Feature Validation
  • Conclusions and discussion
  • 7.1 Conclusions
  • 7.2 Recommendations for future work
  • Model of data warehouse

Customer Segmentation and Customer Profiling

for a Mobile Telecommunications Company
Based on Usage Behavior
A Vodafone Case Study
S.M.H. Jansen
July 17, 2007
Acknowledgments
This Master thesis was written to complete the study Operations Research at
the University of Maastricht (UM). The research took place at the Department
of Mathematics of UM and at the Department of Information Management of
Vodafone Maastricht. During this research, I had the privilege to work together
with several people. I would like to express my gratitude to all those people for
giving me the support to complete this thesis. I want to thank the Department
of Information Management for giving me permission to commence this thesis
in the first instance, to do the necessary research work and to use departmental
data.
I am deeply indebted to my supervisor Dr. Ronald Westra, whose help, stimu-
lating suggestions and encouragement helped me in all the time of research for
and writing of this thesis. Furthermore, I would like to give my special thanks
to my second supervisor Dr. Ralf Peeters, whose patience and enthusiasm en-
abled me to complete this work. I have also to thank my thesis instructor, Drs.
Annette Schade, for her stimulating support and encouraging me to go ahead
with my thesis.
My former colleagues from the Department of Information Management sup-
ported me in my research work. I want to thank them for all their help, support,
interest and valuable hints. Especially I am obliged to Drs. Philippe Theunen
and Laurens Alberts, MSc.
Finally, I would like to thank the people, who looked closely at the final ver-
sion of the thesis for English style and grammar, correcting both and offering
suggestions for improvement.
1
Contents
1 Introduction 8
1.1 Customer segmentation and customer profiling . . . . . . . . . . 9
1.1.1 Customer segmentation . . . . . . . . . . . . . . . . . . . 9
1.1.2 Customer profiling . . . . . . . . . . . . . . . . . . . . . . 10
1.2 Data mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.3 Structure of the report . . . . . . . . . . . . . . . . . . . . . . . . 13
2 Data collection and preparation 14
2.1 Data warehouse . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.1.1 Selecting the customers . . . . . . . . . . . . . . . . . . . 14
2.1.2 Call detail data . . . . . . . . . . . . . . . . . . . . . . . . 15
2.1.3 Customer data . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2 Data preparation . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3 Clustering 22
3.1 Cluster analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.1.1 The data . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.1.2 The clusters . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.1.3 Cluster partition . . . . . . . . . . . . . . . . . . . . . . . 24
3.2 Cluster algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2.1 K-means . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2.2 K-medoid . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.2.3 Fuzzy C-means . . . . . . . . . . . . . . . . . . . . . . . . 28
3.2.4 The Gustafson-Kessel algorithm . . . . . . . . . . . . . . 29
3.2.5 The Gath Geva algorithm . . . . . . . . . . . . . . . . . . 30
3.3 Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.4 Visualization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.4.1 Principal Component Analysis . . . . . . . . . . . . . . . 33
3.4.2 Sammon mapping . . . . . . . . . . . . . . . . . . . . . . 34
3.4.3 Fuzzy Sammon mapping . . . . . . . . . . . . . . . . . . . 35
4 Experiments and results of customer segmentation 37
4.1 Determining the optimal number of clusters . . . . . . . . . . . . 37
4.2 Comparing the clustering algorithms . . . . . . . . . . . . . . . . 42
2
4.3 Designing the segments . . . . . . . . . . . . . . . . . . . . . . . 45
5 Support Vector Machines 53
5.1 The separating hyperplane . . . . . . . . . . . . . . . . . . . . . . 53
5.2 The maximum-margin hyperplane . . . . . . . . . . . . . . . . . 55
5.3 The soft margin . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5.4 The kernel functions . . . . . . . . . . . . . . . . . . . . . . . . . 56
5.5 Multi class classification . . . . . . . . . . . . . . . . . . . . . . . 59
6 Experiments and results of classifying the customer segments 60
6.1 K-fold cross validation . . . . . . . . . . . . . . . . . . . . . . . . 60
6.2 Parameter setting . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
6.3 Feature Validation . . . . . . . . . . . . . . . . . . . . . . . . . . 65
7 Conclusions and discussion 66
7.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
7.2 Recommendations for future work . . . . . . . . . . . . . . . . . 68
Bibliography 68
A Model of data warehouse 71
B Extra results for optimal number of clusters 73
3
List of Figures
1.1 A taxonomy of data mining tasks . . . . . . . . . . . . . . . . . . 12
2.1 Structure of customers by Vodafone . . . . . . . . . . . . . . . . 15
2.2 Visualization of phone calls per hour . . . . . . . . . . . . . . . . 17
2.3 Histograms of feature values . . . . . . . . . . . . . . . . . . . . . 18
2.4 Relation between originated and received calls . . . . . . . . . . . 18
2.5 Relation between daytime and weekday calls . . . . . . . . . . . 19
3.1 Example of clustering data . . . . . . . . . . . . . . . . . . . . . 22
3.2 Different cluster shapes in R
2
. . . . . . . . . . . . . . . . . . . . 24
3.3 Hard and fuzzy clustering . . . . . . . . . . . . . . . . . . . . . . 25
4.1 Values of Partition Index, Separation Index and the Xie Beni Index 38
4.2 Values of Dunn’s Index and the Alternative Dunn Index . . . . . 39
4.3 Values of Partition coefficient and Classification Entropy with
Gustafson-Kessel clustering . . . . . . . . . . . . . . . . . . . . . 40
4.4 Values of Partition Index, Separation Index and the Xie Beni
Index with Gustafson-Kessel clustering . . . . . . . . . . . . . . . 41
4.5 Values of Dunn’s Index and Alternative Dunn Index with Gustafson-
Kessel clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.6 Result of K-means algorithm . . . . . . . . . . . . . . . . . . . . 43
4.7 Result of K-medoid algorithm . . . . . . . . . . . . . . . . . . . . 44
4.8 Result of Fuzzy C-means algorithm . . . . . . . . . . . . . . . . . 44
4.9 Result of Gustafson-Kessel algorithm . . . . . . . . . . . . . . . . 44
4.10 Result of Gath-Geva algorithm . . . . . . . . . . . . . . . . . . . 45
4.11 Distribution of distances from cluster centers within clusters for
the Gath-Geva algorithm with c = 4 . . . . . . . . . . . . . . . . 46
4.12 Distribution of distances from cluster centers within clusters for
the Gustafson-Kessel algorithm with c = 6 . . . . . . . . . . . . . 46
4.13 Cluster profiles for c = 4 . . . . . . . . . . . . . . . . . . . . . . . 47
4.14 Cluster profiles for c = 6 . . . . . . . . . . . . . . . . . . . . . . . 48
4.15 Cluster profiles of centers for c = 4 . . . . . . . . . . . . . . . . . 49
4.16 Cluster profiles of centers for c = 6 . . . . . . . . . . . . . . . . . 50
5.1 Two-dimensional customer data of segment 1 and segment 2 . . . 54
4
5.2 Separating hyperplanes in different dimensions . . . . . . . . . . 54
5.3 Demonstration of the maximum-margin hyperplane . . . . . . . . 55
5.4 Demonstration of the soft margin . . . . . . . . . . . . . . . . . . 56
5.5 Demonstration of kernels . . . . . . . . . . . . . . . . . . . . . . . 57
5.6 Examples of separation with kernels . . . . . . . . . . . . . . . . 58
5.7 A separation of classes with complex boundaries . . . . . . . . . 59
6.1 Under fitting and over fitting . . . . . . . . . . . . . . . . . . . . 60
6.2 Determining the stopping point of training the SVM . . . . . . . 61
6.3 A K-fold partition of the dataset . . . . . . . . . . . . . . . . . . 61
6.4 Results while leaving out one of the features with 4 segments . . 65
6.5 Results while leaving out one of the features with 6 segments . . 65
A.1 Model of the Vodafone data warehouse . . . . . . . . . . . . . . . 72
B.1 Partition index and Separation index of K-medoid . . . . . . . . 73
B.2 Dunn’s index and Alternative Dunn’s index of K-medoid . . . . . 74
B.3 Partition coefficient and Classification Entropy of Fuzzy C-means 74
B.4 Partition index, Separation index and Xie Beni index of Fuzzy
C-means . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
B.5 Dunn’s index and Alternative Dunn’s index of Fuzzy C-means . . 75
5
List of Tables
2.1 Proportions within the different classification groups . . . . . . . 20
4.1 The values of all the validation measures with K-means clustering 39
4.2 The values of all the validation measures with Gustafson-Kessel
clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.3 The numerical values of validation measures for c = 4 . . . . . . 42
4.4 The numerical values of validation measures for c = 6 . . . . . . 43
4.5 Segmentation results . . . . . . . . . . . . . . . . . . . . . . . . . 51
6.1 Linear Kernel, 4 segments . . . . . . . . . . . . . . . . . . . . . . 61
6.2 Linear Kernel, 6 segments . . . . . . . . . . . . . . . . . . . . . . 62
6.3 Average C-value for polynomial kernel, 4 segments . . . . . . . . 62
6.4 Average C-value for polynomial kernel, 6 segments . . . . . . . . 62
6.5 Polynomial kernel, 4 segments . . . . . . . . . . . . . . . . . . . . 62
6.6 Polynomial kernel, 6 segments . . . . . . . . . . . . . . . . . . . . 62
6.7 Radial basis function, 4 segments . . . . . . . . . . . . . . . . . . 63
6.8 Radial basis function, 6 segments . . . . . . . . . . . . . . . . . . 63
6.9 Sigmoid function, 4 segments . . . . . . . . . . . . . . . . . . . . 63
6.10 Sigmoid function, 6 segments . . . . . . . . . . . . . . . . . . . . 64
6.11 Confusion matrix, 4 segments . . . . . . . . . . . . . . . . . . . . 64
6.12 Confusion matrix, 6 segments . . . . . . . . . . . . . . . . . . . . 64
6
Abstract
Vodafone, an International mobile telecommunications company, has accumu-
lated vast amounts of data on consumer mobile phone behavior in a data ware-
house. The magnitude of this data is so huge that manual analysis of data is
not feasible. However, this data holds valuable information that can be applied
for operational and strategical purposes. Therefore, in order to extract such in-
formation from this data, automatic analysis is essential, by means of advanced
data mining techniques. These data mining techniques search and analyze the
data in order to find implicit and useful information, without direct knowledge
of human experts. This research will address the question how to perform cus-
tomer segmentation and customer profiling with data mining techniques. In
our context, ’customer segmentation’ is a term used to describe the process of
dividing customers into homogeneous groups on the basis of shared or common
attributes (habits, tastes, etc). ’Customer profiling’ is describing customers
by their attributes, such as age, gender, income and lifestyles. Having these
two components, managers can decide which marketing actions to take for each
segment. In this research, the customer segmentation is based on usage call
behavior, i.e. the behavior of a customer measured in the amounts of incoming
or outgoing communication of whichever form. This thesis describes the process
of selecting and preparing the accurate data from the data warehouse, in order
to perform customer segmentation and to profile the customer. A number of
advanced and state-of-the-art clustering algorithms are modified and applied
for creating customer segments. An optimality criterion is constructed in order
to measure their performance. The best i.e. most optimal in the sense of the
optimality criterion clustering technique will be used to perform customer seg-
mentation. Each segment will be described and analyzed. Customer profiling
can be accomplished with information from the data warehouse, such as age,
gender and residential area information. Finally, with a recent data mining
technique, called Support Vector Machines, the segment of a customer will be
estimated based on the customers profile. Different kernel functions with dif-
ferent parameters will be examined and analyzed. The customer segmentation
will lead to two solutions. One solution with four segments and one solution
with six segments. With the Support Vector Machine approach it is possible in
80.3% of the cases to classify the segment of a customer based on its profile for
the situation with four segments. With six segments, a correct classification of
78.5% is obtained.
7
Chapter 1
Introduction
Vodafone is world’s leading mobile telecommunications company, with approx-
imately 4.1 million customers in The Netherlands. From all these customers a
tremendous amount of data is stored. These data include, among others, call de-
tail data, network data and customer data. Call detail data gives a description
of the calls that traverse the telecommunication networks, while the network
data gives a description of the state of the hardware and software components
in the network. The customer data contains information of the telecommunica-
tion customers. The amount of data is so great that manual analysis of data is
difficult, if not impossible [22]. The need to handle such large volumes of data
led to the development of knowledge-based expert systems [17, 22]. These auto-
mated systems perform important functions such as identifying network faults
and detecting fraudulent phone calls. A disadvantage of this approach is that
it is based on knowledge from human experts.
Obtaining knowledge from human experts is a time consuming process, and
in many cases, the experts do not have the requisite knowledge [2]. Solutions
to these problems were promised by data mining techniques. Data mining is
the process of searching and analyzing data in order to find implicit, but po-
tentially useful, information [12]. Within the telecommunication branch, many
data mining tasks can be distinguished. Examples of main problems for market-
ing and sales departments of telecommunication operators are churn prediction,
fraud detection, identifying trends in customer behavior and cross selling and
up-selling.
Vodafone is interested in a complete different issue, namely customer segmenta-
tion and customer profiling and the relation between them. Customer segmen-
tation is a term used to describe the process of dividing customers into homoge-
neous groups on the basis of shared or common attributes (habits, tastes, etc)
[10]. Customer profiling is describing customers by their attributes, such as age,
gender, income and lifestyles [1, 10]. Having these two components, marketers
can decide which marketing actions to take for each segment and then allocate
scarce resources to segments in order to meet specific business objectives.
A basic way to perform customer segmentation is to define segmentations in
8
advance with knowledge of an expert, and dividing the customers over these
segmentations by their best fits. This research will deal with the problem of
making customer segmentations without knowledge of an expert and without
defining the segmentations in advance. The segmentations will be determined
based on (call) usage behavior. To realize this, different data mining techniques,
called clustering techniques, will be developed, tested, validated and compared
to each other. In this report, the principals of the clustering techniques will be
described and the process of determining the best technique will be discussed.
Once the segmentations are obtained, for each customer a profile will be de-
termined with the customer data. To find a relation between the profile and
the segments, a data mining technique called Support Vector Machines (SVM)
will be used. A Support Vector machine is able to estimate the segment of a
customer by personal information, such as age, gender and lifestyle. Based on
the combination of the personal information (the customer profile), the segment
can be estimated and the usage behavior of the customer profile can be deter-
mined. In this research, different settings of the Support Vector Machines will
be examined and the best working estimation model will be used.
1.1 Customer segmentation and customer pro-
filing
To compete with other providers of mobile telecommunications it is important
to know enough about your customers and to know the wants and needs of your
customers [15]. To realize this, it is needed to divide customers in segments
and to profile the customers. Another key benefit of utilizing the customer
profile is making effective marketing strategies. Customer profiling is done by
building a customer’s behavior model and estimating its parameters. Customer
profiling is a way of applying external data to a population of possible customers.
Depending on data available, it can be used to prospect new customers or to
recognize existing bad customers. The goal is to predict behavior based on
the information we have on each customer [18]. Profiling is performed after
customer segmentation.
1.1.1 Customer segmentation
Segmentation is a way to have more targeted communication with the customers.
The process of segmentation describes the characteristics of the customer groups
(called segments or clusters) within the data. Segmenting means putting the
population in to segments according to their affinity or similar characteristics.
Customer segmentation is a preparation step for classifying each customer ac-
cording to the customer groups that have been defined.
Segmentation is essential to cope with today’s dynamically fragmenting con-
sumer marketplace. By using segmentation, marketers are more effective in
channeling resources and discovering opportunities. The construction of user
9
segmentations is not an easy task. Difficulties in making good segmentation are
[18]:
• Relevance and quality of data are essential to develop meaningful seg-
ments. If the company has insufficient customer data, the meaning of a
customer segmentation in unreliable and almost worthless. Alternatively,
too much data can lead to complex and time-consuming analysis. Poorly
organize data (different formats, different source systems) makes it also
difficult to extract interesting information. Furthermore, the resulting
segmentation can be too complicated for the organization to implement
effectively. In particular, the use of too many segmentation variables can
be confusing and result in segments which are unfit for management deci-
sion making. On the other hand, apparently effective variables may not be
identifiable. Many of these problems are due to an inadequate customer
database.
• Intuition: Although data can be highly informative, data analysts need
to be continuously developing segmentation hypotheses in order to identify
the ’right’ data for analysis.
• Continuous process: Segmentation demands continuous development
and updating as new customer data is acquired. In addition, effective seg-
mentation strategies will influence the behavior of the customers affected
by them; thereby necessitating revision and reclassification of customers.
Moreover, in an e-commerce environment where feedback is almost imme-
diate, segmentation would require almost a daily update.
• Over-segmentation: A segment can become too small and/or insuffi-
ciently distinct to justify treatment as separate segments.
One solution to construct segments can be provided by data mining methods
that belong to the category of clustering algorithms. In this report, several
clustering algorithms will be discussed and compared to each other.
1.1.2 Customer profiling
Customer profiling provides a basis for marketers to ’communicate’ with existing
customers in order to offer them better services and retaining them. This is done
by assembling collected information on the customer such as demographic and
personal data. Customer profiling is also used to prospect new customers using
external sources, such as demographic data purchased from various sources.
This data is used to find a relation with the customer segmentations that were
constructed before. This makes it possible to estimate for each profile (the
combination of demographic and personal information) the related segment and
visa versa. More directly, for each profile, an estimation of the usage behavior
can be obtained.
Depending on the goal, one has to select what is the profile that will be relevant
to the project. A simple customer profile is a file that contains at least age and
10
gender. If one needs profiles for specific products, the file would contain product
information and/or volume of money spent. Customer features one can use for
profiling, are described in [2, 10, 19]:
• Geographic. Are they grouped regionally, nationally or globally
• Cultural and ethnic. What languages do they speak? Does ethnicity affect
their tastes or buying behaviors?
• Economic conditions, income and/or purchasing power. What is the av-
erage household incom or power of the customers? Do they have any
payment difficulty? How much or how often does a customer spend on
each product?
• Age and gender. What is the predominant age group of your target buyers?
How many children and what age are in the family? Are more female or
males using a certain service or product?
• Values, attitudes and beliefs. What is the customers’ attitude toward your
kind of product or service?
• Life cycle. How long has the customer been regularly purchasing products?
• Knowledge and awareness. How much knowledge do customers have about
a product,service, or industry? How much education is needed? How much
brand building advertising is needed to make a pool of customers aware
of offer?
• Lifestyle. How many lifestyle characteristics about purchasers are useful?
• Recruitment method. How was the customer recruited?
The choice of the features depends also on the availability of the data. With
these features, an estimation model can be made. This can be realized by a
data mining method called Support Vector Machines (SVM). This report gives
an description of SVM’s and it will be researched under which circumstances
and parameters a SVM works best in this case.
1.2 Data mining
In section 1.1, the term data mining was used. Data mining is the process of
searching and analyzing data in order to find implicit, but potentially useful,
information [12]. It involves selecting, exploring and modeling large amounts of
data to uncover previously unknown patterns, and ultimately comprehensible
information, from large databases. Data mining uses a broad family of computa-
tional methods that include statistical analysis, decision trees, neural networks,
rule induction and refinement, and graphic visualization. Although, data min-
ing tools have been available for a long time, the advances in computer hardware
and software, particularly exploratory tools like data visualization and neural
11
networks, have made data mining more attractive and practical. The typical
data mining process consist of the following steps [4]:
• problem formulation
• data preparation
• model building
• interpretation and evaluation of the results
Pattern extraction is an important component of any data mining activity and
it deals with relationships between subsets of data. Formally, a pattern is de-
fined as [4]:
A statement S in L that describes relationships among a subsets of facts F
s
of a given set of facts F, with some certainty C, such that S is simpler than the
enumeration of all facts in F
s
.
Data mining tasks are used to extract patterns from large data sets. The vari-
ous data mining tasks can be broadly divided into six categories as summarized
in Figure 1.1. The taxonomy reflects the emerging role of data visualization as
Figure 1.1: A taxonomy of data mining tasks
a separate data mining task, even as it is used to support other data mining
tasks. Validation of the results is also a data mining task. By the fact that the
validation supports the other data mining tasks and is always necessary within
a research, this task was not mentioned as a separate one. Different data mining
tasks are grouped into categories depending on the type of knowledge extracted
by the tasks. The identification of patterns in a large data set is the first step to
gaining useful marketing insights and marking critical marketing decisions. The
data mining tasks generate an assortment of customer and market knowledge
which form the core of knowledge management process. The specific tasks to
be used in this research are Clustering (for the customer segmentation), Classi-
fication (for estimating the segment) and Data visualization.
Clustering algorithms produce classes that maximize similarity within clusters
but minimize similarity between classes. A drawback of this method is that the
number of clusters has to be given in advance. The advantage of clustering is
that expert knowledge is not required. For example, based on user behavior
data, clustering algorithms can classify the Vodafone customers into ”call only”
users, ”international callers”, ”SMS only” users etc.
Classification algorithms groups customers in predefined classes. For example,
12
Vodafone can classify its customers based on their age, gender and type of sub-
scription and then target its user behavior.
Data visualization allow data miners to view complex patterns in their cus-
tomer data as visual objects complete in three or two dimensions and colors.
In some cases it is needed to reduce high dimensional data into three or two
dimensions. To realize this, algorithms as Principal Component Analysis and
Sammon’s Mapping (discussed in Section 3.4) can be used. To provide varying
levels of details of observed patterns, data miners use applications that provide
advanced manipulation capabilities to slice, rotate or zoom the objects.
1.3 Structure of the report
The report comprises 6 chapters and several appendices. In addition to to this
introductory chapter, Chapter 2 describes the process of selecting the right data
from the data ware house. It provides information about the structure of the
data and the data ware house. Furthermore, it gives an overview of the data
that is used to perform customer segmentation and customer profiling. It ends
with an explanation of the preprocessing techniques that were used to prepare
the data for further usage.
In Chapter 3 the process of clustering is discussed. Clustering is a data mining
technique, that in this research is used to determine the customer segmenta-
tions. The chapter starts with explaining the general process of clustering.
Different cluster algorithms will be studied. It also focuses on validation meth-
ods, which can be used to determine the optimal number of clusters and to
measure the performance of the different cluster algorithms. The chapter ends
with a description of visualization methods. These methods are used to analyze
the results of the clustering.
Chapter 4 analyzes the different cluster algorithms of Chapter 3. This will be
tested with the prepared call detail data as described in Chapter 2 For each
algorithm, the optimal numbers of cluster will be determined. Then, the cluster
algorithms will be compared to each other and the best algorithm will be chosen
to determine the segments. Multiple plots and figures will show the working of
the different cluster methods and the meaning of each segment will be described.
Once the segments are determined, with the customer data of Chapter 2, a pro-
file can be made. Chapter 5 delves into a data mining technique called Support
Vector Machines. This technique will be used to classify the right segment for
each customer profile. Different parameter settings of the Support Vector Ma-
chines will be researched and examined in Chapter 6 to find the best working
model. Finally, in Chapter 7, the research will be discussed. Conclusions and
recommendations are given and future work is proposed.
13
Chapter 2
Data collection and
preparation
The first step (after the problem formulation) in the data mining process is
to understand the data. Without such an understanding, useful applications
cannot be developed. All data of Vodafone is stored in a data warehouse. In
this chapter, the process of collecting the right data from this data ware house,
will be described. Furthermore, the process of preparing the data for customer
segmentation and customer profiling will be explained.
2.1 Data warehouse
Vodafone has stored vast amounts of data in a Teradata data warehouse. This
data warehouse exists off more than 200 tables. A simplified model of the data
warehouse can be found in Appendix A.
2.1.1 Selecting the customers
Vodafone Maastricht is interested in customer segmentation and customer pro-
filing for (postpaid) business customers. In general, business customers can be
seen as employees of a business that have a subscription by Vodafone in re-
lation with that business. A more precisely view can be found in Figure 2.1.
It is clear to see, that prepaid users are always consumers. In the postpaid
group, there are captive and non captive users. A non-captive customer is using
the Vodafone network but has not a Vodafone subscription or prepaid (called
roaming). Vodafone has made an accomplishment with two other telecommuni-
cations companies, Debitel and InterCity Mobile Communications (ICMC), that
their customers can use the Vodafone network. Debitel customers are always
consumers and ICMC customers are always business customers. The ICMC cus-
tomers will also be involved in this research. A captive customer has a business
account if his telephone or subscription is bought in relation with the business
14
Figure 2.1: Structure of customers by Vodafone
he works. These customers are called business users. In some cases, customers
with a consumer account, can have a subscription that is under normal circum-
stances only available for business users. These customers also count as business
users. The total number of (postpaid) business users at Vodafone is more than
800,000. The next sections describe which data of these customers is needed for
customer segmentation and profiling.
2.1.2 Call detail data
Every time a call is placed on the telecommunications network of Vodafone,
descriptive information about the call is saved as a call detail record. The num-
ber of call detail records that are generated and stored is huge. For example,
Vodafone customers generate over 20 million call detail records per day. Given
that 12 months of call detail data is typically kept on line, this means that
hundreds of millions of call detail data will need to be stored at any time. Call
detail records include sufficient information to describe the important charac-
teristics of each call. At a minimum, each call detail record will include the
originating and terminating phone numbers, the date and the time of the call
and the duration of the call. Call detail records are generated in two or three
days after the day the calls were made, and will be available almost immediately
for data mining. This is in contrast with billing data, which is typically made
available only once per month. Call detail records can not be used directly
for data mining, since the goal of data applications is to extract knowledge at
the customer level, not at the level of individual phone calls [7, 8]. Thus, the
call detail records associated with a customer must be summarized into a single
record that describes the customer’s calling behavior. The choice of summary
variables (features) is critical in order to obtain a useful description of the cus-
tomer []. To define the features, one can think of the smallest set of variables
that describe the complete behavior of a customer. Keywords like what, when,
where, how often, who, etc. can help with this process:
15
• How?: How can a customer cause a call detail record? By making a voice
call, or sending an SMS (there are more possibilities, but their appearances
are so rare that they were not used during this research). The customer
can also receive an SMS or voice call.
• Who?: Who is the customer calling? Does he call to fixed lines? Does
he call to Vodafone mobiles?
• What?: What is the location of the customer and the recipient? They
can make international phone calls.
• When?: When does a customer call? A business customer can call during
office daytime, or in private time in the evening or at night and during
the weekend.
• Where?: Where is the customer calling? Is he calling abroad?
• How long?: How long is the customer calling?
• How often?: How often does a customer call or receive a call?
Based on these keywords and based on proposed features in the literature [1,
15, 19, 20] , a list of features that can be used as a summary description of a
customer based on the calls they originate and receive over some time period P
is obtained:
• 1. average call duration
• 2. average # calls received per day
• 3. average # calls originated per day
• 4. % daytime calls (9am - 6pm)
• 5. % of weekday calls (Monday - Friday)
• 6. % of calls to mobile phones
• 7. average # sms received per day
• 8. average # sms originated per day
• 9. % international calls
• 10. % of outgoing calls within the same operator
• 11. # unique area codes called during P
• 12. # different numbers called during P
16
These twelve features can be used to build customer segments. Such a segment
describes a certain behavior of group of customers. For examples, customers
who use their telephone only at their office could be in a different segment then
users that use their telephone also for private purposes. In that case, the seg-
mentation was based on the percentage weekday and daytime calls. Most of the
twelve features listed above can be generated in a straightforward manner from
the underlying data of the data ware house, but some features require a little
more creativity and operations on the data.
It may be clear that generating useful features, including summary features, is a
critical step within the data mining process. Should poor features be generated,
data mining will not be successful. Although the construction of these features
may be guided by common sense, it should include exploratory data analysis.
For example, the use of the time period 9am-6pm in the fourth feature is not
based on the commonsense knowledge that the typical workday on a office is
from 9am to 5pm. More detailed exploratory data analysis, shown in Figure
2.2 indicates that the period from 9am to 6pm is actually more appropriate for
this purpose. Furthermore, for each summary feature, there should be sufficient
Figure 2.2: Visualization of phone calls per hour
variance within the data, otherwise distinguish between customers is not possi-
ble and the feature is not useful. On the other hand, to much variance hampers
the process of segmentation. For some features values, the variance is visible in
the following histograms. Figure 2.3 shows that the average call duration, the
number of weekday and daytime calls and the originated calls have sufficient
variance. Note that the histograms resemble well known distributions. This
also indicates that the chosen features are suited for the customer segmenta-
tion. Interesting to see is the relation between the number of calls originated
and received. First of all, in general, customers originating more calls than
receiving. Figure 2.4 demonstrates this, values above the blue line represent
customers with more originating calls than receiving calls. In Figure 2.4 is also
visible that the customers that originated more calls, receive also more calls in
proportion. Another aspect that is simple to figure out is the fact that customer
17
(a) Call duration (b) Weekday calls
(c) Daytime calls (d) Originated calls
Figure 2.3: Histograms of feature values
Figure 2.4: Relation between originated and received calls
18
that make more weekday calls also call more at daytime (in proportion). This is
plotted in Figure 2.5. It is clear to see that the chosen features contain sufficient
variance and that certain relations and different customer behavior are already
visible. The chosen features appear to be well chosen and useful for customer
segmentation.
Figure 2.5: Relation between daytime and weekday calls
2.1.3 Customer data
To profile the customer, customer data is needed. The proposed data in Section
1.1.2 is not completely available. Information about lifestyles and income is
missing. However, with some creativity, some information can be subtracted
from the data ware house. The information that Vodafone stored in the data
ware house include name and address information and also include other in-
formation such as service plan, contract information and telephone equipment
information. With this information, the following variables can be used to define
a customers profile:
• Age group: <25, 25-40 40-55 >55
• Gender: male, female
• Type telephone: simple, basic, advanced
• Type subscription: basic, advance, expanded
• Company size: small, intermediate, big
• Living area: (big) city, small city /town
19
Because a relative small difference in age between customers should show close
relationships, the age of the customers has to be grouped. Otherwise, the result
of the classification algorithm is too specific to the trainings data [14]. In general,
the goal of grouping variables is to reduce the number of variables to a more
manageable size and to remove the correlations between each variable. The
composition of the groups should be chosen with care. It is of high importance
that the sizes of the groups are almost equal (if this is possible) [22]. If there is
one group with a sufficient higher amount of customers than other groups, this
feature will not increase the performance of the classification. This is caused
by the fact that from each segment a relative high number of customers is
represented in this group. Based on this feature, the segment of a customer can
not be determined. Table 2.1 shows the percentages of customers within the
chosen groups. It is clear to see that sizes of the groups were chosen with care
Age: <25 25-40 40-55 >55
21.2% 29.5% 27.9% 21.4%
Gender: Male Female
60.2% 39.8%
Telephone type: simple basic advanced
33.5% 38.7% 27.8%
Type of subscription: simple advanced expanded
34.9% 36.0% 29.1%
Company size: small intermediate big
31.5% 34.3% 34.2%
Living area: (big) city small city/town
42.0% 58.0%
Table 2.1: Proportions within the different classification groups
and the values can be used for defining the customers profile.With this profile, a
Support Vector Machine will be used to estimate the segment of the customer.
Chapter 5 and Chapter 6 contain information and results of this method.
2.2 Data preparation
Before the data can be used for the actual data mining process, it need to
cleaned and prepared in a required format. These tasks are [7]:
• Discovering and repairing inconsistent data formats and inconsistent data
encoding, spelling errors, abbreviations and punctuation.
• Deleting unwanted data fields. Data may contain many meaningless fields
from an analysis point of view, such as production keys and version num-
bers.
• Interpreting codes into text or replacing text into meaningful numbers.
20
Data may contain cryptic codes. These codes has to be augmented and
replaced by recognizable and equivalent text.
• Combining data, for instance the customer data, from multiple tables into
one common variable.
• Finding multiple used fields. A possible way to determine is to count or
list all the distinct variables of a field.
The following data preparations were needed during this research:
• Checking abnormal, out of bounds or ambiguous values. Some of these
outliers may be correct but this is highly unusual, thus almost impossible
to explain.
• Checking missing data fields or fields that have been replaced by a default
value.
• Adding computed fields as inputs or targets.
• Mapping continuous values into ranges, e.g. decision trees.
• Normalization of the variables. There are two types of normalization. The
first type is to normalize the values between [0,1]. The second type is to
normalize the variance to one.
• Converting nominal data (for example yes/no answers) to metric scales.
• Converting from textual to numeral or numeric data.
New fields can be generated through combinations of e.g. frequencies, aver-
ages and minimum/maximum values. The goal of this approach is to reduce
the number of variables to a more manageable size while also the correlations
between each variable will be removed. Techniques used for this purpose are of-
ten referred to as factor analysis, correspondence analysis and conjoint analysis
[14]. When there is a large amount of data, it is also useful to apply data reduc-
tion techniques (data cube aggregation, dimension and numerosity reduction,
discretization and concept hierarchy generation). Dimension reduction means
that one has to select relevant feature to a minimum set of attributes such that
the resulting probability distribution of data classes is a close as possible to the
original distribution given the values of all features. For this additional tools
may be needed, e.g. exhaustive, random or heuristic search, clustering, decision
trees or associations rules.
21
Chapter 3
Clustering
In this chapter, the used techniques for the cluster segmentation will be ex-
plained.
3.1 Cluster analysis
The objective of cluster analysis is the organization of objects into groups, ac-
cording to similarities among them [13]. Clustering can be considered the most
important unsupervised learning method. As every other unsupervised method,
it does not use prior class identifiers to detect the underlying structure in a
collection of data. A cluster can be defined as a collection of objects which are
”similar” between them and ”dissimilar” to the objects belonging to other clus-
ters. Figure 3.1 shows this with a simple graphical example. In this case the 3
Figure 3.1: Example of clustering data
clusters into which the data can be divided were easily identified. The similarity
criterion that was used in this case is distance: two or more objects belong to
the same cluster if they are ”close” according to a given distance (in this case
geometrical distance). This is called distance-based clustering. Another way of
clustering is conceptual clustering. Within this method, two or more objects
22
belong to the same cluster if this one defines a concept common to all that
objects. In other words, objects are grouped according to their fit to descriptive
concepts, not according to simple similarity measures. In this research, only
distance-based clustering algorithms were used.
3.1.1 The data
One can apply clustering techniques to quantitative (numerical) data, qualita-
tive (categoric) data, or a mixture of both. In this research, the clustering of
quantitative data is considered. The data, as described in Section 2.1.2, are
typically summarized observations of a physical process (call behavior of a cus-
tomer). Each observation of the customers calling behavior consists of n mea-
sured values, grouped into an n-dimensional row vector x
k
= [x
k1
, x
k2
, ..., x
kn
]
T
,
where x
k
∈ R
n
. A set of N observations is denoted by X = {x
k
|k = 1, 2, ..., N},
and is represented as an N x n matrix:
X =

¸
¸
¸
¸
x
11
x
12
· · · x
1n
x
21
x
22
· · · x
2n
.
.
.
.
.
.
.
.
.
.
.
.
x
N1
x
N2
· · · x
Nn
¸

. (3.1)
In pattern recognition terminology, the rows of X are called patterns or objects,
the columns are called the features or attributes, and X is called the pattern
matrix. In this research, X will be referred to the data matrix. The rows of
X represent the customers, and the columns are the feature variables of their
behavior as described in Section 2.1.2. As mentioned before, the purpose of
clustering is to find relationships between independent system variables, called
the regressors, and future values of dependent variables, called the regressands.
However, one should realize, that the relations revealed by clustering are not
more than associations among the data vectors. And therefore, they will not
automatically constitute a prediction model of the given system. To obtain such
a model, additional steps are needed.
3.1.2 The clusters
The definition of a cluster can be formulated in various ways, depending on
the objective of the clustering. In general, one can accept the definition that a
cluster is a group of objects that are more similar to another than to members
of other clusters. The term ”similarity” can be interpreted as mathematical
similarity, measured in some well-defined sense. In metric spaces, similarity is
often defined by means of a distance norm, or distance measure. Distance can
be measured in different ways. The first possibility is to measure among the
data vectors themselves. A second way is to measure the distance form the
data vector to some prototypical object of the cluster. The cluster centers are
usually (and also in this research) not known a priori, and will be calculated
by the clustering algorithms simultaneously with the partitioning of the data.
23
The cluster centers may be vectors of the same dimensions as the data objects,
but can also be defined as ”higher-level” geometrical objects, such as linear or
nonlinear subspaces or functions.
Data can reveal clusters of different geometrical shapes, sizes and densities as
demonstrated in Figure 3.2 Clusters can be spherical, elongated and also be
(a) Elongated (b) Spherical
(c) Hollow (d) Hollow
Figure 3.2: Different cluster shapes in R
2
hollow. Cluster can be found in any n-dimensional space. Clusters a,c and d
can be characterized as linear and non linear subspaces of the data space (R
2
in
this case). Clustering algorithms are able to detect subspaces of the data space,
and therefore reliable for identification. The performance of most clustering
algorithms is influenced not only by the geometrical shapes and densities of the
individual clusters, but also by the spatial relations and distances among the
clusters. Clusters can be well-separated, continuously connected to each other,
or overlapping each other.
3.1.3 Cluster partition
Clusters can formally be seen as subsets of the data set. One can distinguish
two possible outcomes of the classification of clustering methods. Subsets can
24
either be fuzzy or crisp (hard). Hard clustering methods are based on the clas-
sical set theory, which requires that an object either does or does not belong
to a cluster. Hard clustering in a data set X means partitioning the data into
a specified number of exclusive subsets of X. The number of subsets (clusters)
is denoted by c. Fuzzy clustering methods allow objects to belong to several
clusters simultaneously, with different degrees of membership. The data set X
is thus partitioned into c fuzzy subsets. In many real situations, fuzzy cluster-
ing is more natural than hard clustering, as objects on the boundaries between
several classes are not forced to fully belong to one of the classes, but rather
are assigned membership degrees between 0 and 1 indicating their partial mem-
berships (illustrated by Figure 3.3 The discrete nature of hard partitioning also
Figure 3.3: Hard and fuzzy clustering
causes analytical and algorithmic intractability of algorithms based on analytic
functionals, since these functionals are not differentiable. The structure of the
partition matrix U = [µ
ik
]:
U =

¸
¸
¸
¸
µ
1,1
µ
1,2
· · · µ
1,c
µ
2,1
µ
2,2
· · · µ
2,c
.
.
.
.
.
.
.
.
.
.
.
.
µ
N,1
µ
N,2
· · · µ
N,c
¸

. (3.2)
Hard partition
The objective of clustering is to partition the data set X into c clusters. Assume
that c is known, e.g. based on prior knowledge, or it is a trial value, of witch
partition results must be validated. Using classical sets, a hard partition can be
seen as a family of subsets {A
i
|1 ≤ i ≤ c ⊂ P(X)}, its properties can be defined
as follows:
c
¸
i=1
A
i
= X, (3.3)
A
i
∩ A
j
, 1 ≤ i = j ≤ c, (3.4)
Ø ⊂ A
i
⊂ X, 1 ≤ i ≤ c. (3.5)
25
These conditions imply that the subsets A
i
contain all the data in X, they must
be disjoint and none of them is empty nor contains all the data in X. Expressed
in the terms of membership functions:
c
¸
i=1
µ
Ai
= 1, (3.6)
µ
Ai
∨ µ
Ai
, 1 ≤ i = j ≤ c, (3.7)
0 ≤ µ
Ai
< 1, 1 ≤ i ≤ c. (3.8)
Where µ
Ai
represents the characteristic function of the subset A
i
which value
is zero or one. To simplify these notations, µ
i
will be used instead of µ
Ai
,
and denoting µ
i
(x
k
) by µ
ik
, partitions can be represented in a matrix notation.
U = [µ
ik
], a Nxc matrix, is a representation of the hard partition if and only if
its elements satisfy:
µ
ij
∈ {0, 1}, 1 ≤ i ≤ N, 1 ≤ k ≤ c, (3.9)
c
¸
k=1
µ
ik
= 1, 1 ≤ i ≤ N, (3.10)
0 <
N
¸
i=1
µ
ik
< N, 1 ≤ k ≤ c. (3.11)
A definition of a hard partitioning space can be defined as follows:
Let X be a finite data set and the number of clusters 2 ≤ c < N ∈ N. Then,
the hard partitioning space for X can be seen as the set:
M
hc
= {U ∈ R
Nxc

ik
∈ {0, 1}, ∀i, k;
c
¸
k=1
µ
ik
= 1, ∀i; 0 <
N
¸
i=1
µ
ik
< N, ∀k}.
(3.12)
Fuzzy partition
Fuzzy partition can be defined as a generalization of hard partitioning, in this
case µ
ik
is allowed to acquire all real values between zero and 1. Consider the
matrix U = [µ
ik
], containing the fuzzy partitions, its conditions are given by:
µ
ij
∈ [0, 1], 1 ≤ i ≤ N, 1 ≤ k ≤ c, (3.13)
c
¸
k=1
µ
ik
= 1, 1 ≤ i ≤ N, (3.14)
0 <
N
¸
i=1
µ
ik
< N, 1 ≤ k ≤ c. (3.15)
Note that there is only one difference with the conditions of the hard partition-
ing. Also the definition of the fuzzy partitioning space will not much differ with
26
the definition of the hard partitioning space. It can be defined as follows: Let
X be a finite data set and the number of clusters 2 ≤ c < N ∈ N. Then, the
fuzzy partitioning space for X can be seen as the set:
M
fc
= {U ∈ R
Nxc

ik
∈ [0, 1], ∀i, k;
c
¸
k=1
µ
ik
= 1, ∀i; 0 <
N
¸
i=1
µ
ik
< N, ∀k}.
(3.16)
The i-th column of U contains values of the membership functions of the i-th
fuzzy subset of X. Equation (1.14) implies that the sum of each column should
be 1, which means that the total membership of each x
k
in X equals one. There
are no constraints on the distribution of memberships among the fuzzy clusters.
This research will focus on hard partitioning. However, fuzzy cluster algorithms
will be applied as well. To deal with the problem of fuzzy memberships, the
cluster with the highest degree of membership will be chosen as the cluster were
the object belongs to. This method will result into hard partitioned clusters.
The possibilistic partition will not be used in this researched and will not be
discussed here.
3.2 Cluster algorithms
This section gives an overview of the clustering algorithms that were used during
the research.
3.2.1 K-means
K-means is one of the simplest unsupervised learning algorithms that solves
the clustering problem. However, the results of this hard partitioning method
are not always reliable and this algorithm has numerical problems as well. The
procedure follows an easy way to classify a given N x n data set through a certain
numbers of c clusters defined in advance. The K-means algorithm allocates each
data point to one of the c clusters to minimize the within sum of squares:
c
¸
i=1
sum
k∈Ai
||x
k
−v
i
||
2
. (3.17)
A
i
represents a set of data points in the i-th cluster and v
i
is the average of the
data points in cluster i. Note that ||x
k
−v
i
||
2
is actually a chosen distance norm.
Within the cluster algorithms, v
i
is the cluster center (also called prototype) of
cluster i:
v
i
=
¸
Ni
k=1
x
k
N
i
, x
k
∈ A
i
, (3.18)
where N
i
is the number of data points in A
i
.
27
3.2.2 K-medoid
K-medoid clustering, also a hard partitioning algorithm, uses the same equations
as the K-means algorithm. The only difference is that in K-medoid the cluster
centers are the nearest data points to the mean of the data in one cluster V =
{v
i
∈ X|1 ≤ i ≤ c}. This can be useful when, for example, there is no continuity
in the data space. This implies that a mean of the points in one cluster does
actually not exist.
3.2.3 Fuzzy C-means
The Fuzzy C-means algorithm (FCM) minimizes an objective function, called
C-means functional, to define the clusters. The C-means functional, invented
by Dunn, is defined as follows:
J(X; U, V ) =
c
¸
i=1
N
¸
k=1

ik
)
m
||x
k
−v
i
||
2
A
, (3.19)
with
V = [v
1
, v
2
, ..., v
c
], v
i
∈ R
n
. (3.20)
V denotes the vector with the cluster centers that has to be determined. The
distance norm ||x
k
−v
i
||
2
A
is called a squared inner-product distance norm and
is defined by:
D
2
ikA
= ||k
k
−v
i
||
2
A
= (x
k
−v
i
)
T
A(x
k
−v
i
). (3.21)
On a statistical point of view, equation 3.19 measures the total number of vari-
ance of x
k
from v
i
. The minimization of the C-means functional can be seen as a
non linear optimization problem, that can be solved by a variety of methods. Ex-
amples of methods that can solve non linear optimization problems are grouped
coordinate minimization and genetic algorithms. The simplest method to solve
this problem is utilizing the Picard iteration through the first-order conditions
for the stationary points of equation 3.19. This method is called the fuzzy c-
means algorithm. To find the stationary points of the c-means functional, one
can adjoint the constrained in 3.14 to J by means of Lagrange multipliers:
¯
J(X; U, V, λ) =
c
¸
i=1
N
¸
k=1

ik
)
m
D
2
ikA
+
N
¸
k=1
λ
k

c
¸
i=1
µ
ik
−1

, (3.22)
and by setting the gradients of (
ˆ
J), with respect to U, V and λ, to zero. When
D
2
ikA
> 0, ∀i, k and m > 1, then the C-means functional may only be minimized
by (U, V ) ∈ M
fc
xR
nxc
if
µ
ik
=
1
¸
c
j=1
(D
ikA
/D
jkA
)
2/(m−1)
, 1 ≤ i ≤ c, 1 ≤ k ≤ N, (3.23)
28
and
v
i
=
¸
N
k=1
µ
m
ik
x
k
¸
N
k=1
µ
m
i,k
, 1 ≤ i ≤ c. (3.24)
The solution of these equations are satisfying the constraints that were given in
equation (3.13) and (3.15). Remark that the v
i
of equation (3.24) is the weighted
average of the data points that belong to a cluster and the weights represents the
membership degrees. This explains why the name of the algorithm is c-means.
The Fuzzy C-means algorithm is actually an iteration between the equations
(3.23) and (3.24). The FCM algorithm uses the standard Euclidean distance for
its computations. Therefor, it is able to define hyper spherical clusters. Note
that it can only detect clusters with the same shape, caused by the common
choice of the norm inducing matrix A = I. The norm inducing matrix can also
be chosen as an nxn diagonal matrix of the form:
A
D
=

¸
¸
¸
¸
(1/σ
1
)
2
0 · · · 0
0 (1/σ
2
)
2
· · · 0
.
.
.
.
.
.
.
.
.
.
.
.
0 0 · · · (1/σ
n
)
2
¸

. (3.25)
This matrix accounts for different variances in the directions of the coordinate
axes of X. Another possibility is to choose A as the inverse of the nxn covariance
matrix A = F
−1
, where
F =
1
N
N
¸
k=1
(x
k
− ˆ x)(x
k
− ˆ x)
T
(3.26)
and ˆ x denotes the mean of the data. Hence that, in this case, matrix A is based
on the Mahalanobis distance norm.
3.2.4 The Gustafson-Kessel algorithm
The Gustafson and Kessel (GK) algorithm is a variation on the Fuzzy c-means
algorithm [11]. It employs a different and adaptive distance norm to recognize
geometrical shapes in the data. Each cluster will have its own norm-inducing
matrix A
i
, satisfying the following inner-product norm:
D
2
ikA
= (x
k
−v
i
)
T
· A
i
(x
k
−v
i
), where 1 ≤ i ≤ c and 1 ≤ k ≤ N. (3.27)
The matrices A
i
are used as optimization variables in the c-means functional.
This implies that each cluster is allowed to adapt the distance norm to the local
topological structure of the data. A c-tuple of the norm-inducing matrices is
defined by A, where A = (A
1
, A
2
, ..., A
c
). The objective functional of the GK
algorithm can be calculated by:
J(X; U, V, A) =
c
¸
i=1
N
¸
k=1
(u
ik
)
m
D
2
ikAi
. (3.28)
29
If A is fixed, the conditions under (3.13), (3.14) and (3.15) can be applied
without any problems. Unfortunately, equation (3.28) can not be minimized
in a straight forward manner, since it is linear in A
i
. This implies that J can
be made as small as desired by making A
i
less positive definite. To avoid this,
A
i
has to be constrained to obtain a feasible solution. A general way to this
is by constraining the determinant of the matrix. A varying A
i
with a fixed
determinant relates to the optimization of the cluster whit a fixed volume:
||A
i
|| = ρ
i
, ρ > 0. (3.29)
Here ρ is a remaining constant for each cluster. In combination with the La-
grange multiplier, A
i
can be expressed in the following way:
A
i
= [ρ
i
det(F
i
)]
1/n
F
−1
i
, (3.30)
with
F
i
=
¸
N
k=1

ik
)
m
(x
k
−v
i
)(x
k
−v
i
)
T
¸
N
k=1

ik
)
m
. (3.31)
F
i
is also called the fuzzy covariance matrix. Hence that this equation in com-
bination with equation (3.30) can be substituted into equation (3.27). The
outcome of the inner-product norm of (3.27) is a generalized squared Maha-
lanobis norm between the data points and the cluster center. The covariance is
weighted by the membership degrees of U.
3.2.5 The Gath Geva algorithm
Bezdek and Dunn [5] proposed a fuzzy maximum likelihood estimation (FMLE)
algorithm with a corresponding distance norm:
D
ik
(x
k
, v
i
) =

det(F
wi
)
α
i

1
2
(x
k
−v
(l)
i
)
T
F
−1
wi
(x
k
−v
(l)
i
)

, (3.32)
Comparing this with the Gustafson-Kessel algorithm, the distance norm includes
an exponentional term. This implies that this distance norm will decrease faster
than the inner-product norm. In this case, the fuzzy covariance matrix Fi is
defined by:
F
wi
=
¸
N
k=1

ik
)
w
(x
k
−v
i
)(x
k
−v
i
)
T
¸
N
k=1

ik
)
w
, 1 ≤ i ≤ c. (3.33)
The reason for using the w variable is to generalize this expression. In the origi-
nal FMLE algorithm, w = 1. In this research, w will be set to 2, to compensate
the exponential term and obtain clusters that are more fuzzy. Because of the
generalization, two weighted covariance matrices arise. The variable α
i
in equa-
tion (3.32) is the prior probability of selecting cluster i. α
i
can be defines as
follows:
α
i
=
1
N
N
¸
k=1
µ
ik
. (3.34)
30
Gath and Geva [9] discovered that the FMLE algorithm is able to detect clusters
of different shapes, sizes and densities and that the clusters are not constrained
in volume. The main drawback of this algorithm is the robustness, since the
exponential distance norm can converge to a local optimum. Furthermore, it is
not know how reliable the results of this algorithm are.
3.3 Validation
Cluster validation refers to the problem whether a found partition is correct and
how to measure the correctness of a partition. A clustering algorithm is designed
to parameterize clusters in a way that it gives the best fit. However, this does
not apply that the best fit is meaningful at all. The number of clusters might
not be correct or the cluster shapes do not correspond to the actual groups in
the data. In the worst case, the data can not be grouped in a meaningful way at
all. One can distinguish two main approaches to determine the correct number
of clusters in the data:
• Start with a sufficiently large number of clusters, and successively reducing
this number by combining clusters that have the same properties.
• Cluster the data for different values of c and validate the correctness of
the obtained clusters with validation measures.
To be able to perform the second approach, validation measures has to be de-
signed. Different validation methods have been proposed in the literature, how-
ever, none of them is perfect by oneself. Therefore, in this research are used
several indexes, which are described below:
• Partition Coefficient (PC): measures the amount of ”overlapping” be-
tween clusters. It is defined by Bezdek [5] as follows:
PC(c) =
1
N
c
¸
i=1
N
¸
j=1
(u
ij
)
2
(3.35)
where u
ij
is the membership of data point j in cluster i. The main draw-
back of this validity measure is the lack of direct connection to the data
itself. The optimal number of clusters can be found by the maximum
value.
• Classification Entropy (CE): measures only the fuzziness of the cluster,
which is a slightly variation on the Partition Coefficient.
CE(c) = −
1
N
c
¸
i=1
N
¸
j=1
u
ij
log(u
ij
) (3.36)
31
• Partition Index (PI): expresses the ratio of the sum of compactness and
separation of the clusters. Each individual cluster is measured with the
cluster validation method. This value is normalized by dividing it by the
fuzzy cardinality of the cluster. To receive the Partition index, the sum
of the value for each individual cluster is used.
PI(c) =
c
¸
i=1
¸
N
j=1
(u
ij
)
m
||x
j
−v
i
||
2
N
i
¸
c
k=1
||v
k
−v
i
||
2
(3.37)
PI is mainly used for the comparing of different partitions with the same
number of clusters. A minor value of a SC means a better partitioning.
• Separation Index (SI): in contrast with the partition index (PI), the
separation index uses a minimum-distance separation to validate the par-
titioning.
SI(c) =
¸
c
i=1
¸
N
j=1
(u
ij
)
2
||x
j
−v
i
||
2
N min
i,k
||v
k
−v
i
||
2
(3.38)
• Xie and Beni’s Index (XB): is a method to quantify the ratio of the
total variation within the clusters and the separations of the clusters [3].
XB(c) =
¸
c
i=1
¸
N
j=1
(u
ij
)
m
||x
j
−v
i
||
2
N min
i,j
||x
j
−v
i
||
2
(3.39)
The lowest value of the XB index should indicate the optimal number of
clusters.
• Dunn’s Index (DI): this index was originally designed for the identifica-
tion of hard partitioning clustering. Therefor, the result of the clustering
has to be recalculated.
DI(c) = min
i∈c
{ min
j∈c,i=j
{
min
x∈Ci,y∈Cj
d(x, y)
max
k∈c
{max
x,y∈C
d(x, y)}
}} (3.40)
The main disadvantage of the Dunn’s index is the very expansive compu-
tational complexity as c and N increase.
• Alternative Dunn Index (ADI):To simplify the calculation of the
Dunn index, the Alternative Dunn Index was designed. This will be the
case when the dissimilarity between two clusters, measured with min
x∈Ci,y∈Cj
d(x, y),
is rated in under bound by the triangle-inequality:
d(x, y) ≥ |d(y, v
j
) −d(x, v
j
)| (3.41)
were v
j
represents the cluster center of the j-th cluster.
ADI(c) = min
i∈c
{ min
j∈c,i=j
{
min
xi∈Ci,xj∈Cj
|d(y, v
j
) −d(x
i
, v
j
)|
max
k∈c
{max
x,y∈C
d(x, y)}
}} (3.42)
32
Note, that the Partition Coefficient and the Classification Entropy are only
useful for fuzzy partitioned clustering. In case of fuzzy clusters the values of the
Dunn’s Index and the Alternative Dunn Index are not reliable. This is caused
by the repartitioning of the results with the hard partition method.
3.4 Visualization
To understand the data and the results of the clustering methods, it is useful
to visualize the data and the results. However, the used data set is a high-
dimensional data set, which can not be plotted and visualized directly. This
section describes three methods that can map the data points into a lower
dimensional space.
In this research, the three mapping methods will be used for the visualization
of the clustering results. The first method is the Principal Component Analysis
(PCA), a standard and a most widely method to map high-dimensional data
into a lower dimensional space. Then, this report will focus on the Sammon
mapping method. The advantage of the Sammon mapping is the ability to
preserve inter pattern distances. This kind of mapping of distances is much
closer related to the proposition of clustering than saving the variances (which
will be done by PCA). However, the Sammon mapping application has two main
drawbacks:
• Sammon mapping is a projection method, which is based on the preser-
vation of the Euclidean inter point distance norm. This implies that the
Sammon mapping only can be applied on clustering algorithms that use
the Euclidean distance norm during the calculations of the clusters.
• The Sammon mapping method aims to find in a high n-dimensional space
N points in a lower q-dimensional subspace, such in a way the inter
point distances correspond to the distances measured in the n-dimensional
space. To achieve this, a computational expensive algorithm is needed, be-
cause in every iteration step a computation of N(N − 1)/2 distances is
required.
To avoid these problems of the Sammon mapping method, a modified algorithm,
called the Fuzzy Sammon mapping, is used during this research. A draw back
of this Fuzzy Sammon mapping is the loose of precision in distance, since only
the distance between the data points and the cluster centers considered to be
important.
The three visualisation methods will be explained in more detail in the following
subsections.
3.4.1 Principal Component Analysis
Principal component analysis (PCA) include a mathematical procedure that
maps a number of correlated variables into a smaller set of uncorrelated vari-
ables, called the principal components. The first principal component represents
33
as much of the variability in the data as possible. The succeeding components
describe the remaining variability. The main goals of the PCA method are:
• Identifying new meaningful underlying variables.
• Discovering and/or reducing the dimensionality of a data set.
In a mathematical way, the principal components will be achieved by analyzing
the eigenvectors and eigenvalues. The direction of the first principal component
is diverted from the eigenvector with the largest eigenvalue. The eigenvalue
associated with the second largest eigenvalue correspond to the second principal
component, etc. In this research, the second objective is used. In this case, the
covariance matrix of the data set can be described by:
F =
1
N
(x
k
−v)(x
k
−v)
T
, (3.43)
where v = ¯ x
k
. Principal Component Analysis is based on the projection of
correlated high-dimensional data onto a hyperplane [3]. This methods uses
only the first q nonzero eigenvalues and the corresponding eigenvectors of the
covariance matrix:
F
i
= U
i
Λ
i
U
T
i
. (3.44)
With Λ
i
as a matrix that contains the eigenvalues λ
i,j
of F
i
in its diagonal in
decreasing order and U
i
is a matrix containing the eigenvectors corresponding
to the eigenvalues in its columns. Furthermore, there is a q-dimensional reduced
vector that represents the vector x
k
of X, which can be defined as follows:
y
i,k
= W
−1
i
(x
k
) = W
T
i
(x
k
). (3.45)
The weight matrix W
i
contains the q principal orthonormal axes in its column:
W
i
= U
i,q
Λ
1
2
i,q
. (3.46)
3.4.2 Sammon mapping
As mentioned before,the Sammon mapping uses inter point distance measures to
find N points in a q-dimensional data space, which are representative for a higher
n-dimensional data set. The inter point distance measure of the n-dimensional
space, defined by d
ij
= d(x
i
, x
j
) correspond to the inter point distances in the
q-dimensional space, given by d

ij
= d

(y
i
, y
j
). This is achieved by Sammon’s
stress, a minimization criterion of the error:
E =
1
λ
N−1
¸
i=1
N
¸
j=i+1
(d
ij
−d

ij
)
2
d
ij
, (3.47)
where λ is a constant:
λ =
¸
i<j
d
ij
=
N−1
¸
i=1
N
¸
j=i+1
d
ij
. (3.48)
34
Note that there is no need to maintain λ, since a constant does not change
the result of the optimization process. The minimization of the error E is
an optimization problem in the Nxq variables y
il
, with i ∈ {1, 2, ..., N} and
l ∈ {1, 2, ..., q} which implies that y
i
= [y
i1
, ..., y
iq
]
T
. The rating of y
il
at the
t-th iteration can defined by:
y
il
(t + 1) = y
il
(t) −α

∂E(t)
∂y
il
(t)

2
E(t)
∂y
2
il
(t)
¸
¸
, (3.49)
where α is a nonnegative scalar constant, with a recommended value α 0.3 −
0.4. This scalar constant represents the step size for gradient search in the
direction of
∂E(t)
∂y
il
(t)
= −
2
λ
N
¸
k=1,k=i
¸
d
ki
−d

ki
d
ki
d

ki

(y
il
−y
kl
) (3.50)

2
E(t)
∂y
2
il
(t)
= −
2
λ
N
¸
k=1,k=i
1
d
ki
d

ki
¸
(d
ki
−d

ki
) −

(y
il
−y
kl
)
2
d

ki

1 +
d
ki
−d

ki
d
ki

(3.51)
With this gradient-descent method, it is not possible to reach a local minimum in
the error surface, while searching for the minimum of E. This is a disadvantage,
because multiple experiments with different random initializations are necessary
to find the minimum. However, it is possible to estimate the correct initialization
based on the information which is obtained from the data.
3.4.3 Fuzzy Sammon mapping
As mentioned in the introduction of this section, Sammon’s mapping has several
drawbacks. To avoid this drawbacks, a modified mapping method is designed
which takes into account the basic properties of fuzzy clustering algorithms
where only the distance between the data points and the clustering centers are
considered to be important [3]. The modified algorithm, called Fuzzy Sammon
mapping, uses only N∗c distances, weighted by the membership values similarly
to equation (3.19):
E
fuzz
=
c
¸
i=1
N
¸
k=1

ki
)
m
(d(x
k
, v
i
) −d

ki
)
2
, (3.52)
with d(x
k
, v
i
) representing the distance between data point x
k
and the cluster
center v
i
in the original n-dimensional space. The Euclidean distance between
the cluster center z
i
and the data y
k
of the projected q-dimensional space is
represented by d

(y
k
, z
i
). According to this information, in a projected two
dimensional space every cluster is represented by a single point, independently
to the shape of the original cluster. The Fuzzy Sammon mapping algorithm is
similar to the original Sammon mapping, but in this case the projected cluster
35
center will be recalculated in every iteration after the adaption of the projected
data points. The recalculation will be based on the weighted mean formula of
the fuzzy clustering algorithms, described in Section 3.2.3 (equation 3.19).
The membership values of the projected data can be plotted based on the stan-
dard equation for the calculation of the membership values:
µ

ki
=
1
¸
c
j=1

d

(x
k
,ηi)
d

(x
k
,vj)
2
m−1
, (3.53)
where U

= [µ

ki
] is the partition matrix with the recalculated memberships.
The plot will only give an approximation of the high dimensional clustering in
a two dimensional space. To measure the quality of this rating, an evaluation
function that determines the mean square error between the original and the
recalculated error can be defined as follows:
P = ||U −U

||. (3.54)
In the next chapter, the cluster algorithms will be tested and evaluated. The
PCA and the (Fuzzy) Sammon mapping methods will be used to visualize the
data and the clusters.
36
Chapter 4
Experiments and results of
customer segmentation
In this chapter, the cluster algorithms will be tested and their performance will
be measured with the proposed validation methods of the previous chapter. The
best working cluster method will be used to determine the segments. The chap-
ter ends with an evaluation of the segments.
4.1 Determining the optimal number of clusters
The disadvantage of the proposed cluster algorithms is the number of clusters
that has to be given in advance. In this research the number of clusters is not
known. Therefor, the optimal number of clusters has to be searched with the
given validation methods of Section 3.3. For each algorithm, calculations for
each cluster, c ∈ [215], were executed. To find the optimal number of clusters,
a process called Elbow Criterion is used. The elbow criterion is a common rule
of thumb to determine what number of clusters should be chosen. The elbow
criterion says that one should choose a number of clusters so that adding an-
other cluster does not add sufficient information. More precisely, by graphing
a validation measure explained by the clusters against the number of clusters,
the first clusters will add much information (explain a lot of variance), but at
some point the marginal gain will drop, giving an angle in the graph (the el-
bow). Unfortunately, this elbow can not always be unambiguously identified.
To demonstrate the working of the elbow criterion, the feature values that rep-
resent the call behavior of the customers, as described in Section 2.1.2, are used
as input for the cluster algorithms. From the 800,000 business customers of
Vodafone, 25,000 customers were randomly selected for the experiments. More
customers would lead to computational problems. First, the K-means algorithm
will be evaluated. The values of the validation methods depending on the num-
ber of clusters will be plotted. The value of the Partition Coefficient is for all
37
clusters 1, and the classification entropy is always ’NaN’. This is caused by the
fact that these 2 measures were designed for fuzzy partitioning methods, and
in this case the hard partitioning algorithm K-means is used. In Figure 4.1,
the values of the Partition Index, Separation Index and Xie and Beni’s Index
are shown. Mention again, that no validation index is reliable only by itself.
Figure 4.1: Values of Partition Index, Separation Index and the Xie Beni Index
Therefor, all the validation indexes are shown. The optimum could differ by
using different validation methods. This means that the optimum only could
be detected by the comparison of all the results. To find the optimal number of
cluster, partitions with less clusters are considered better, when the difference
between the values of the validation measure are small. Figure 4.1 shows that for
the PI and SI, the number of clusters easily could be rated to 4. For the Xie and
Beni index, this is much harder. The elbow can be found at c = 3, c = 6, c = 9
or c = 13, depending on the definition and parameters of an elbow. In Figure
4.2 there are more informative plots shown. The Dunn’s index and the Alterna-
tive Dunn’s index confirm that the optimal number of clusters for the K-means
algorithm should be chosen to 4. The values of all the validation measures for
the K-means algorithm, are embraced in table 4.1
38
Figure 4.2: Values of Dunn’s Index and the Alternative Dunn Index
c 2 3 4 5 6 7 8
PC 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000
CE NaN NaN NaN NaN NaN NaN NaN
PI 3.8318 1.9109 1.1571 1.0443 1.2907 0.9386 0.8828
SI 0.0005 0.0003 0.0002 0.0002 0.0002 0.0002 0.0002
XBI 5.4626 4.9519 5.0034 4.3353 3.9253 4.2214 3.9079
DI 0.0082 0.0041 0.0034 0.0065 0.0063 0.0072 0.0071
ADI 0.0018 0.0013 0.0002 0.0001 0.0001 0.0001 0.0000
c 9 10 11 12 13 14 15
PC 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000
CE NaN NaN NaN NaN NaN NaN NaN
PI 0.8362 0.8261 0.8384 0.7783 0.7696 0.7557 0.7489
SI 0.0002 0.0002 0.0002 0.0001 0.0001 0.0001 0.0001
XBI 3.7225 3.8620 3.8080 3.8758 3.4379 3.3998 3.5737
DI 0.0071 0.0052 0.0061 0.0070 0.0061 0.0061 0.0061
ADI 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
Table 4.1: The values of all the validation measures with K-means clustering
39
It is also possible to define the optimal numbers of clusters for fuzzy cluster-
ing algorithms with this method. To illustrate this, the results of the Gustafson-
Kessel algorithm will be shown. In Figure 4.3 the results of the Partition Index
and the Classification Entropy are plotted. Compared to the hard clustering
methods, the validation methods can be used now for the fuzzy clustering. How-
ever, the main drawback of PC is the monotonic decreasing with c, which makes
it hardly to detect the optimal number of cluster. The same problem holds for
CE: monotonic increasing, caused by the lack of direct connection to the data.
The optimal number of cluster can not be rated based on those two validation
methods. Figure 4.4 gives more information about the optimal number of clus-
Figure 4.3: Values of Partition coefficient and Classification Entropy with
Gustafson-Kessel clustering
ters. For the PI and the SI, the local minimum is reached at c = 6. Again,
for the XBI, it is difficult to find the optimal number of clusters. The points
at c = 3, c = 6 and c = 11, can be seen as an elbow. In Figure 4.5, the Dunn
index also indicates that the optimal number of clusters should be at c = 6. On
the other hand, the Alternative Dunn index, has an elbow at the point c = 3.
However, for the Alternative Dunn Index is not known how reliable its results
are, so the optimal number of clusters for the Gustafson-Kessel algorithm will be
six. The results of the validation measures for the Gustafson-Kessel algorithm
are written in table 4.2. This process can be repeated for all other cluster algo-
rithms. The results can be found in Appendix B. For the K-means, K-medoid
and the Gath-Geva,the optimal number of clusters is chosen at c = 4. For the
other algorithms, the optimal number of clusters is located at c = 6.
40
Figure 4.4: Values of Partition Index, Separation Index and the Xie Beni Index
with Gustafson-Kessel clustering
Figure 4.5: Values of Dunn’s Index and Alternative Dunn Index with Gustafson-
Kessel clustering
41
c 2 3 4 5 6 7 8
PC 0.6462 0.5085 0.3983 0.3209 0.3044 0.2741 0.2024
CE 0.5303 0.8218 1.0009 1.2489 1.4293 1.5512 1.7575
PI 0.9305 1.2057 1.5930 1.9205 0.8903 0.7797 0.8536
SI 0.0002 0.0003 0.0007 0.0004 0.0001 0.0001 0.0002
XBI 2.3550 1.6882 1.4183 1.1573 0.9203 0.9019 0.7233
DI 0.0092 0.0082 0.0083 0.0062 0.0029 0.0041 0.0046
ADI 0.0263 0.0063 0.0039 0.0018 0.0007 0.0001 0.0009
c 9 10 11 12 13 14 15
PC 0.2066 0.1611 0.1479 0.1702 0.1410 0.1149 0.1469
CE 1.8128 2.0012 2.0852 2.0853 2.2189 2.3500 2.3046
PI 0.9364 0.7293 0.7447 0.7813 0.7149 0.6620 0.7688
SI 0.0002 0.0001 0.0002 0.0002 0.0001 0.0001 0.0001
XBI 0.5978 0.5131 0.4684 0.5819 0.5603 0.5675 0.5547
DI 0.0039 0.0030 0.0028 0.0027 0.0017 0.0015 0.0006
ADI 0.0003 0.0002 0.0004 0.0002 0.0000 0.0001 0.0000
Table 4.2: The values of all the validation measures with Gustafson-Kessel
clustering
4.2 Comparing the clustering algorithms
The optimal number of cluster can be determined with the validation methods,
as mentioned in the previous section. The validation measures can also be
used to compare the different cluster methods. As examined in the previous
section, the optimal number of clusters was found at c = 4 or c = 6, depending
on the clustering algorithm. The validation measures for c = 4 and c = 6 of
all the clustering methods are collected in the tables 4.3 and 4.4. Table 4.3
PC CE PI SI XBI DI ADI
K-means 1 NaN 1.1571 0.0002 5.0034 0.0034 0.0002
K-medoid 1 NaN 0.2366 0.0001 Inf 0.0084 0.0002
FCM 0.2800 1.3863 0.0002 42.2737 1.0867 0.0102 0.0063
GK 0.3983 1.0009 1.5930 0.0007 1.4183 0.0083 0.0039
GG 0.4982 1.5034 0.0001 0.0001 1.0644 0.0029 00030
Table 4.3: The numerical values of validation measures for c = 4
and 4.4 show that the PC and CE are useless for the hard clustering methods
K-means and K-medoid. On the score of the values of the three most used
indexes, Separation index, Xie and Beni’s index and Dunn’s index, one can
conclude that for c = 4 the Gath-Geva algorithm has the best results and
for c = 6 the Gustafson-Kessel algorithm. To visualize the clustering results,
the validation methods that are described in Section 3.4 can be used. With
these visualization methods, the dataset can be reduced to a 2-dimensional
space. To avoid visibility problems (plotting too much values will cause one
42
PC CE PI SI XBI DI ADI
K-means 1 NaN 1.2907 0.0002 3.9253 0.0063 0.0001
K-medoid 1 NaN 0.1238 0.0001 Inf 0.0070 0.0008
FCM 0.1667 1.7918 0.0001 19.4613 0.9245 0.0102 0.0008
GK 0.3044 1.4293 0.8903 0.0001 0.9203 0.0029 0.0007
GG 0.3773 1.6490 0.1043 0.0008 1.0457 0.0099 0.0009
Table 4.4: The numerical values of validation measures for c = 6
big cloud of data points), only 500 values (representing 500 customers) from
this 2-dimensional dataset will be randomly picked. For the K-means and the
K-medoid algorithm, the Sammon’s mapping gives the best visualization of
the results. For the other cluster algorithms, the Fuzzy Sammon’s mapping
visualization gives the best projection with respect to the partitions of the data
set. These visualization methods are used for the following plots. Figures 4.x-
4.x show the different clustering results for c = 4 and c = 6 on the data set.
Figure 4.6 and 4.7 show that hard clustering methods can find a solution
Figure 4.6: Result of K-means algorithm
for the clustering problem. None of the clusters contain sufficient more or less
customers than other clusters. The plot of the Fuzzy C-means algorithm, in
Figure 4.8, shows unexpected results. For the situation with 4 clusters, there
are only 2 clusters clearly visible. By a detailed look at the plot, one can see that
there are actually 4 cluster centers, but the cluster centers are almost situated
on the same location. In the situation with 6 clusters, one can see three big
cluster, with one small cluster in one of the big clusters. The other two cluster
centers are nearly invisible. This implies that the Fuzzy C-means algorithm is
not able to find good clusters for this data set. In Figure 4.9, the results of the
Gustafon-Kessel algorithm are plotted. For both situations, the clusters are well
separated. Note that the cluster in the left bottom corner and the cluster in the
43
Figure 4.7: Result of K-medoid algorithm
Figure 4.8: Result of Fuzzy C-means algorithm
Figure 4.9: Result of Gustafson-Kessel algorithm
44
top right corner in Figure 4.9 are also maintained in the situation with 6 clusters.
This may indicate that the data points in these clusters represents customers
that differ on multiple fields with the other customers of Vodafone. The results
Figure 4.10: Result of Gath-Geva algorithm
of the Gath-Geva algorithm, visualized in Figure 4.10, for the situation c = 4
look similar to the result of the Gustafson-Kessel algorithm. The result for the
c = 6 situation is remarkable. Here are also appearing clusters in other clusters.
In the real high-dimensional situation, the clusters are not a subset of each
other, but are separated. The fact that this is the case in the two-dimensional
plot, indicates that a clustering with six clusters with the Gustafson-Kessel
algorithms not a good solution. With the results of the validation methods and
the visualization of the clustering, one can conclude that there are two possible
best solutions: The Gath-Geva algorithm for c = 4 and the Gustafson-Kessel
algorithm for c = 6. To determine which partitioning will be used to define
the segments, a closer look to the meaning of the clusters will be needed. In
the next section, the two different partitions will be closely compared with each
other.
4.3 Designing the segments
To define which clustering method will be used for the segmentation, one can
look at the distances from the points to each cluster. In Figure 4.11 and 4.12, two
box plots of the distances from the data points to the cluster are plotted. The
box indicates the upper and lower quartiles. In both situations, the results show
that the clusters are homogeneous. This indicates that, based on the distances
to the cluster, one can not distinguish between the two cluster algorithms.
Another way to view the differences between the cluster methods is to profile
the clusters. For each cluster, a profile can be made by drawing a line between
all normalized feature values (each feature value is represented at the x-as)
of the customers within this cluster. The result is visible for the Gath-Geva
algoithm for c = 4 and for the Gustafson-Kessel algorithm with six clusters.
45
Figure 4.11: Distribution of distances from cluster centers within clusters for
the Gath-Geva algorithm with c = 4
Figure 4.12: Distribution of distances from cluster centers within clusters for
the Gustafson-Kessel algorithm with c = 6
46
The profiles of the different clusters do not differ much in shape. However, in
each cluster, at least one value differs sufficient from the values of the other
cluster. This confirms the assumption that customers of different clusters have
indeed a different usage behavior. Most of the lines in one profile are drawn
closely together. This means that the customers in one profile contain similar
values of the feature values.
Figure 4.13: Cluster profiles for c = 4
47
Figure 4.14: Cluster profiles for c = 6
48
More relevant plots are shown in Figure 4.15 and ??. The mean of all the
lines (equivalent to the cluster center) was calculated and a line between all the
(normalized) feature vales was drawn. The difference between the clusters are
visible by some feature values. For instance, in the situation with four clusters,
Cluster 1 has customers, compared with other cluster, have a high feature value
at feature 8. Cluster 2 has high values at position 6 and 9, while Cluster 3
contains peaks at features 2 and 12. The 4th and final cluster has high values
at feature 8 and 9.
Figure 4.15: Cluster profiles of centers for c = 4
49
Figure 4.16: Cluster profiles of centers for c = 6
50
With the previous clustering results, validation measures and plots, it is not
possible to decide which of the two clustering methods gives a better result.
Therefor, both results will be used as a solution for the customer segmentation.
For the Gath-Geva algorithm with c = 4 and the Gustafson-Kessel algorithm
with c = 6, table 4.5 shows the result of the customer segmentation. The feature
Feature 1 2 3 4 5 6
Average 119.5 1.7 3.9 65.8 87.0 75.7
Segment 1 (27.2%) 91.3 0.9 2.9 54.8 86.6 58.2
c = 4 Segment 2 (28.7%) 120.1 1.8 3.6 73.6 87.1 93.7
Segment 3 (23.9%) 132.8 2.4 4.4 60.1 86.7 72.1
Segment 4 (20.2%) 133.8 1.7 4.7 74.7 87.6 78.8
Segment 1 (18.1%) 94.7 1.2 2.8 66.3 88.0 72.6
Segment 2 (14.4%) 121.8 1.7 4.1 65.9 86.4 73.0
c = 6 Segment 3 (18.3%) 121.6 2.5 4.9 66.0 84.3 71.5
Segment 4 (17.6%) 126.8 1.6 4.0 65.7 87.3 71.2
Segment 5 (14.8%) 96.8 1.1 3.5 65.2 88.6 92.9
Segment 6 (16.8%) 155.3 2.1 4.1 65.7 87.4 73.0
Feature 7 8 9 10 11 12
Average 1.6 3.7 2.2 14.4 6.9 25.1
Segment 1 (27.2%) 1.7 4.0 1.6 12.3 6.2 12.2
c = 4 Segment 2 (28.7%) 1.2 3.1 2.1 12.8 6.6 30.6
Segment 3 (23.9%) 1.4 3.4 2.1 22.4 9.4 39.7
Segment 4 (20.2%) 2.1 4.3 3.0 10.1 5.4 17.9
Segment 1 (18.1%) 2.3 4.5 1.8 11.3 6.1 13.5
Segment 2 (14.4%) 1.6 3.7 1.9 17.8 9.5 40.4
c = 6 Segment 3 (18.3%) 1.0 2.9 2.9 15.1 6.6 26.9
Segment 4 (17.6%) 1.5 3.6 1.9 15.0 6.2 24.0
Segment 5 (14.8%) 0.8 2.9 1.8 12.4 6.1 23.1
Segment 6 (16.8%) 2.4 4.6 2.9 14.8 6.9 22.7
Table 4.5: Segmentation results
numbers correspond to the feature numbers of Section 2.1.2. (Feature 1 is the
call duration, feature 2 the received voices calls and feature 3 the originated
calls, feature 4 the daytime calls, feature 5 the weekday calls, 6 are calls to
mobile phones, 7 received sms, 8 originated sms, feature 9 the international
calls, feature 10 the calls to Vodafone mobiles, 11 the unique are codes and
feature 12 the number of different numbers called). In words, the segments can
be described as follows: For the situation with 4 segments:
• Segment 1: In this segment are customers with a relative low number of
voice calls. This customers call more in the evening (in proportion) and to
fixed lines then other customers. Their sms usage is higher then normal.
The number of international calls is low.
• Segment 2: This segment contains customers with an average voice call
51
usage. They call often to mobile phones during day time. They do not
send and receive many sms messages.
• Segment 3: The customers in this segment make relative many voice
calls. These customers call to many different numbers and have a lot of
contacts which are Vodafone customers.
• Segment 4: These customers originate many voice calls. They also send
and receive many sms messages. They call often during daytime and call
more then average to international numbers. Their call duration is high.
Remarkable is the fact that they don not have so many contacts as the
number of calls do suspect. They have a relative small number of contacts.
For the situation with 6 segments, the customers in this segments can be de-
scribed as follows:
• Segment 1: In this segment are customers with a relative low number
of voice calls. Their average call duration is also lower than average.
However, their sms usage is relative high. These customers do not call to
many different numbers.
• Segment 2: This segment contains customers with a relative high number
of contacts. They also call to many different areas. They have also more
contacts with a Vodafone mobile.
• Segment 3: The customers in this segment make relative many voice
calls. Their sms usage is low. In proportion, they make more international
phone calls than other customers.
• Segment 4: These customers are the average customers. None of the
feature values is high or low.
• Segment 5: These customers do not receive many voice calls. The aver-
age call duration is low. They also receive and originate a low number of
sms messages.
• Segment 6: These customers originate and receive many voice calls.
They also send and receive many sms messages. The duration of their
voice calls is longer than average. The percentage of international calls is
high.
In the next session the classification method Support Vector Machine will be
explained. This technique will be used to classify/estimate the segment of a
customer by personal information as age, gender and lifestyle (the customer
data of Section 2.1.3).
52
Chapter 5
Support Vector Machines
A Support Vector Machine is a algorithm that learns by example to assign
labels to objects [16]. In this research a Support Vector machine will be used
to recognize the segment of a customer by examining thousands of customers
(e.g. the customer data features of Section 2.1.3) of each segment. In general, a
Support Vector Machine is a mathematical entity; an algorithm for maximizing
a particular mathematical function with respect to a given collection of data.
However, the basic ideas of Support Vector Machines can be explained without
any equations. The next few sections will describe the four basic concepts:
• The separating hyper plane
• The maximum-margin hyperplane
• The soft margin
• The kernel function
For now, to allow an easy, geometric interpretation of the data, imagine that
there exists only two segments. In this case the customer data consist of 2
feature values, age and income, which can be easily plotted. The green dots
represent the customers that are in segment 1 and the red dots are customers
that are in segment 2. The goal of the SVM is learn to tell the difference between
the groups and, given an unlabeled customer, such as the one labeled ’Unknown’
in Figure 5.1, predict whether it corresponds to segment 1 or segment 2.
5.1 The separating hyperplane
A human being is very good at pattern recognition. Even a quick glance at Fig-
ure 5.1a shows that the green dots form a group and the reds dots form another
group that can easily be separated by drawing a line between the two groups
(Figure 5.1b). Subsequently, predicting the label of an unknown customer is
simple: one simply needs to ask whether the new customer falls on the segment
53
(a) Two-dimensional representation of the
customers
(b) A separating hyperplane
Figure 5.1: Two-dimensional customer data of segment 1 and segment 2
1 or the segment 2 side of the separating line. Now, to define the notion of
a separating hyperplane, consider the situation where there are not just two
feature values to describe the customer. For example, if there was just 1 feature
value to describe the customer, then the space in which the corresponding one-
dimensional feature resides is a one-dimensional line. This line can be divided
in half by using a single point (see Figure 5.2a). In two dimensions, a straight
line divides the space in half (remember Figure 5.1b) In a three-dimensional
space, a plane is needed to divide the space, illustrated in Figure 5.2b. This
procedure can be extrapolated mathematically in higher dimensions. The term
for a straight line in a high-dimensional space is a hyperplane. So the term
separating hyperplane is, essentially, the line that separates the segments.
(a) One dimension (b) Three dimensions
Figure 5.2: Separating hyperplanes in different dimensions
54
5.2 The maximum-margin hyperplane
The concept of treating objects as points in a high-dimensional space and finding
a line that separates them, is a common way of classification, and therefore not
unique to the SVM. However, the SVM differs from all other classifier methods
by virtue of how the hyperplane should be selected. Consider again the classifi-
cation problem of Figure 5.1a The goal of SVM is to find a line that separates
the segment 1 customers from the segment 2 customers. However, there are
an infinite number of possible lines, as portrayed in Figure 5.2 The question is
which line should be chosen as the optimal classifier and how should the optimal
line be defined. A logical way of selecting the optimal line, is selecting the line
that is, roughly speaking, ’in the middle’. In other words, the line that sepa-
rates the two segments and adopts the maximal distance from any of the given
customers (see Figure 5.2). It is not surprising that a theorem of the statistical
learning theory is supporting this choice [6]. By defining the distance from the
hyperplane to the nearest customer (in general an expression vector) as the mar-
gin of the hyperplane, the SVM selects the maximum separating hyperplane.
By selecting this hyper plane, the SVM is able to predict the unknown segment
of the customer in Figure 5.1a. The vectors (points) that constrain the width
of the margin are the support vectors. This theorem, is in many ways, the key
(a) Many possibilities (b) The maximum-margin hyperplane
Figure 5.3: Demonstration of the maximum-margin hyperplane
to the success of Support Vector Machines. However, there are a some remarks
and caveats to deal with. First at all, the theorem is based on the fact that the
data on which the SVM is trained are drawn from the same distribution as the
data it has to classify. This is of course logical, since it is not reasonable that
a Support Vector machine trained on customer data is able to classify different
car types. More relevantly, it is not reasonable to expect that the SVM can
classify well if the training data set is prepared with a different protocol then
the test data set. On the other hand, the theorem of a SVM indicates that the
two data sets has to be drawn from the same distribution. For example, a SVM
55
does not assume that the data is drawn from a normal distribution.
5.3 The soft margin
So far, the theory assumed that the data can be separated by a straight line.
However, many real data sets are not cleanly separable by a straight line, for
example the data of Figure 5.4a. In this figure, the data contains an error
object. A intuitively way to deal with the problems of errors is designing the
SVM in such a way that it allows a few anomalous customers to fall on the
’wrong side’ of separation line. This can be achieved by adding a ’soft margin’
to the SVM. The soft margin allows a small percentage of the data points
to push their way through the margin of the separating hyperplane without
affecting the final result. With the soft margin, the data set of Figure 5.4a will
be separated in the way it is illustrated in Figure 5.3 The customer can be seen
as an outlier and resides on the same side of the line with customers of segment
1. Of course, a SVM should not allow too many misclassification. Note, that
(a) Data set containing one error (b) Separating with soft margin
Figure 5.4: Demonstration of the soft margin
with the introduction of the soft margin, a user-specified parameter is involved
that controls the soft margin and, roughly, controls the number of customers
that is allowed to violate the separation line and determines how far across the
line they are allowed. Setting this parameter is a complicated process, by the
fact that a large margin will be achieved with respect to the number of correct
classifications. In other words, the soft margin specifies a trade-off between
hyper plane violations and the size of the margin.
5.4 The kernel functions
To understand the notion of a kernel function, the example data will be sim-
plified even further. Assume that, instead of a two-dimensional data set, there
56
is a one-dimensional data set, as seen before in Figure 5.1. In that case, the
separating hyperplane was a single point. Now, consider the situation of Figure
5.4, which illustrates an non separable data set. No single point can separate
the two segments and introducing a soft margin would not help. A kernel func-
tion provides a solution to this problem. The kernel function adds an extra
dimension to the data, in this case by squaring the one dimensional data set.
The result is plotted in Figure 5.4. Within the new higher dimensional space,
as shown in the figure, the SVM can separate the data in two segments by one
straight line. In general, the kernel function can be seen as a mathematical
trick for the SVM to project data from a low-dimensional space to a space of
higher dimensions. If one chooses a good kernel function, the data will become
separable in the corresponding higher dimension. To understand kernels better,
(a) None separable dataset (b) Separating previously non separable
dataset
Figure 5.5: Demonstration of kernels
some extra examples will be given. In Figure 5.4 is plotted a two-dimensional
data set. With a relative simple kernel function, this data can be projected to a
four-dimensional space. It is not possible to draw the data in the 4 dimensional
space, but with a projection of the SVM hyperplane in the four-dimensional
space back down to the original two-dimensional space, the result is shown as
the curved line in Figure 5.4. it is possible to prove that for any data set exists
a kernel function that allows the SVM to separate the data linearly in a higher
dimension. Of course, the data set must contain consistent labels, which means
that two identical data points may not have different labels. So, in theory,
the SVM should be a perfect classifier. However, there are some drawbacks of
projecting data in a very high-dimensional space to find the separating hyper-
plane. the first problem is the so called curse of dimensionality: as the numbers
of variables under consideration increases, the number of possible solutions also
increases, but exponentially. Consequently, it becomes harder for any algorithm
to find a correct solution. In Figure 5.4 the situation is drawn when the data is
project into a space with too many dimensions. The figure contains the same
data as Figure 5.4, but the projected hyperplane is found by a very high dimen-
57
sional kernel. This results in boundaries which are to specific to the examples
of the data set. This phenomenon is called over fitting. The SVM will not
function well on new unseen unlabeled data. There exists another large practi-
(a) Linearly separable in four dimensions (b) A SVM that has over fit the data
Figure 5.6: Examples of separation with kernels
cal difficulty when applying new unseen data to the SVM. This problems relies
on the question how to choose a kernel function that separates the data, but
without introducing too many irrelevant dimensions. Unfortunately, the answer
too this question is, in most cases, trial and error. In this research a SVM will
be experimented with a variety of ’standard’ kernel functions. By using the
cross-validation method, the optimal kernel will be selected on a statistical way.
However, this is a time-consuming process and it is not guaranteed that the
best kernel function that was found during cross-validation, is actually the best
kernel function that exists. It is more likely that there exists a kernel function
that was not tested and performs better than the selected kernel function. Prac-
tically, the method described above, mainly gives sufficient results. In general
the kernel function is defined by:
K(x
i
, x
j
) = Φ(x
i
)
T
Φ(x
j
), (5.1)
where x
i
are the training vectors. The vectors are mapped into a higher dimen-
sional space by the function Φ. Many kernel mapping functions can be used,
probably an infinite number, but a few kernel functions have been found to work
well in for a wide variety of applications [16]. The default and recommended
kernel functions were used during this research and will be discussed now.
• Linear: which function is defined by:
K(x
i
, x
j
) = x
T
i
x
j
. (5.2)
• Polynomial: the polynomial kernel of degree d is of the form
K(x
i
, x
j
) = (γx
T
i
x
j
+c
0
)
d
. (5.3)
58
• Radial basis function: also known as the Gaussian kernel is of the form
K(x
i
, x
j
) = exp(−γ||x
i
−x
j
||
2
). (5.4)
• Sigmoid: the sigmoid function, which is also used in neural networks, is
defined by
K(x
i
, x
j
) = tanh(γx
T
i
x
j
+c
0
). (5.5)
When the sigmoid function is used, one can regard it with a as a two-layer
neural network.
In this research the constant c
0
is set to 1. The concept of a kernel mapping
function is very powerful. It allows a SVM to perform separations even with
very complex boundaries as shown in Figure 5.7
Figure 5.7: A separation of classes with complex boundaries
5.5 Multi class classification
So far, the idea of using a hyperplane to separate the feature vectors into two
groups was described, but only for two target categories. How does a SVM
discriminate between a large variety of classes, as in our case 4 or 6 segments?
There are several approaches proposed, but two methods are the most popu-
lar and most used [16]. The first approach is to train multiple, one-versus-all
classifiers. For example, if the SVM has to recognize three classes, A, B and C,
one can simply train three separate SVM to answer the binary questions, ”Is it
A?”, ”Is it B?” and ”Is it C?”. Another simple approach is the one-versus-one
where k(k −1)/2 models are constructed, where k is the number of classes. In
this research the one-verses-one technique will be used.
59
Chapter 6
Experiments and results of
classifying the customer
segments
6.1 K-fold cross validation
To avoid over fitting, cross-validation is used to evaluate the fitting provided by
each parameter value set tried during the experiments. Figure 6.1 demonstrates
how important the training process is. Different parameter values may cause
under or over fitting. By K-fold cross validation the training dataset will be
Figure 6.1: Under fitting and over fitting
divided into two groups, the training set, the test set and the validation set.
The training set will be used to train the SVM. The test set will be used to
estimate the error during the training of the SVM. With the validation set,
the actual performance of the SVM will be measured after the SVM is trained.
The training of the SVM will be stopped when the test error reached a local
60
minimum, see Figure 6.2. By K-fold cross validation, a k-fold partition of the
Figure 6.2: Determining the stopping point of training the SVM
data set is created. For each of K experiments, K-1 folds will be used for training
and the remaining one for testing. Figure 6.3 illustrates this process. In this
Figure 6.3: A K-fold partition of the dataset
research, K is set to 10. The advantage of K-fold cross validation is that all the
examples in the dataset are eventually used for both training and testing. The
error is calculated by taking the average off all K experiments.
6.2 Parameter setting
In this section, the optimal parameters for the Support Vector Machine will
be researched and examined. Each kernel function with its parameters will be
tested on their performance. The linear Kernel function itself has no parameters.
The only parameter that can be researched is the soft margin value of the
Support Vector Machin, denoted by C. In table 6.1 and table 6.2 the results for
the different C-values are summarized. For the situation with 4 clusters, the
C 1 2 5 10 20 50 100 200 500
42.1% 42.6% 43.0% 43.2% 43.0% 42.4% 41.7% 40.8% 36.1%
Table 6.1: Linear Kernel, 4 segments
61
C 1 2 5 10 20 50 100 200 500
28.9% 29.4% 30.9% 31.3% 31.4% 32.0% 27.6% 27.6% 21.8%
Table 6.2: Linear Kernel, 6 segments
optimal value for the soft margin is C = 10 and by using the 6 segments C = 50.
The correct number of classifications are respectively, 43.2% and 32.0%. For the
polynomial kernel function, there are two parameters. The number of degrees,
denoted by d and the width γ. Therefor, the optimal number for the maximal
margin will be determined. This is done by multiple test runs with random
values for d and γ. The average value for each soft margin C can be found in the
tables 6.3 and 6.4. These C-values are used to find out which d and γ give the
C 1 2 5 10 20 50 100 200 500
73.8% 77.4% 76.6% 74.6% 73.5% 72.8% 70.6% 63.2% 53.7%
Table 6.3: Average C-value for polynomial kernel, 4 segments
C 1 2 5 10 20 50 100 200 500
70.1% 74.4% 75.3% 75.1% 75.0% 75.1% 50.6% 42.7% 26.0%
Table 6.4: Average C-value for polynomial kernel, 6 segments
best results. The results are shown in tables 6.5 and 6.6. For the situation with
d 1 2 3 4 5 6 7
γ = 0.4 76.1% 76.3% 78.1% 73.2% 74.8% 76.0% 75.0%
γ = 0.6 76.0% 76.3% 77.6% 74.1% 74.5% 75.4% 75.8%
γ = 0.8 75.8% 76.3% 77.2% 74.0% 74.4% 77.1% 75.2%
γ = 1.0 76.2% 76.4% 78.0% 75.0% 75.2% 75.6% 75.8%
γ = 1.2 76.0% 76.2% 78.1% 74.6% 75.1% 76.0% 75.8%
γ = 1.4 75.2% 76.2% 78.1% 74.9% 75.5% 76.3% 74.9%
Table 6.5: Polynomial kernel, 4 segments
d 1 2 3 4 5 6 7
γ = 0.4 75.0% 74.6% 75.9% 76.0% 75.8% 74.3% 73.9%
γ = 0.6 74.2% 75.1% 74.9% 76.2% 75.0% 74.5% 74.0%
γ = 0.8 73.8% 74.7% 74.3% 76.2% 75.9% 74.8% 73.1%
γ = 1.0 74.1% 75.0% 73.6% 76.1% 75.3% 74.2% 72.8%
γ = 1.2 72.1% 74.1% 75.5% 75.4% 75.4% 74.1% 73.0%
γ = 1.4 73.6% 74.3% 72.2% 76.0% 74.4% 74.3% 72.9%
Table 6.6: Polynomial kernel, 6 segments
62
4 segments, the optimal score is 78.1% and for 6 segments 76.2%. The following
kernel function, the radial basis function has only one variable, namely γ. The
results of the Radial Basis function are given in table 6.7 and table 6.8. The
C 1 2 5 10 20 50 100 200 500
γ = 0.4 80.0 79.0 76.6 78.3 76.4 73.3 60.2 52.4 37.5
γ = 0.6 80.1 80.3 77.7 79.0 79.9 72.8 63.6 59.6 27.5
γ = 0.8 79.3 79.5 78.2 80.2 78.4 69.3 59.3 51.4 29.7
γ = 1.0 78.4 78.2 80.3 78.5 76.9 66.2 61.7 47.9 30.6
γ = 1.2 79.6 79.9 79.8 80.2 80.1 69.0 61.3 45.5 26.3
γ = 1.4 77.4 76.9 76.5 79.4 77.7 71.4 61.3 41.2 26.0
Table 6.7: Radial basis function, 4 segments
C 1 2 5 10 20 50 100 200 500
γ = 0.4 73.6 77.4 72.6 70.9 68.0 65.1 52.7 51.8 40.0
γ = 0.6 72.5 74.8 74.8 72.7 73.0 70.4 54.0 49.3 39.1
γ = 0.8 74.1 76.6 80.3 80.0 68.4 60.5 55.5 54.1 40.9
γ = 1.0 70.7 72.9 73.8 70.9 66.1 64.7 52.2 48.5 34.2
γ = 1.2 72.6 73.5 73.4 73.1 71.9 74.6 64.8 60.0 38.3
γ = 1.4 69.4 68.5 70.7 69.1 68.0 68.5 54.4 52.4 31.0
Table 6.8: Radial basis function, 6 segments
best result with 4 segments is 80.3%, with 6 segments the best score is 78.5%.
The sigmoid function has also only 1 variable. The results are given in table
6.9 and 6.10 The results show that 66.1% and 44.6% of the data is classified
C 1 2 5 10 20 50 100 200 500
γ = 0.4 58.2 53.0 57.7 58.2 56.1 57.9 30.3 47.5 38.9
γ = 0.6 47.6 56.1 55.5 46.0 58.3 44.1 30.6 30.7 34.5
γ = 0.8 52.1 60.5 54.6 57.9 58.6 44.7 43.2 44.3 38.7
γ = 1.0 51.4 57.3 52.0 50.7 50.2 48.6 44.7 42.2 40.0
γ = 1.2 66.1 64.8 61.3 62.8 59.6 57.1 46.5 44.0 42.0
γ = 1.4 63.2 61.4 59.7 65.0 53.8 51.1 52.2 47.6 41.4
Table 6.9: Sigmoid function, 4 segments
correct, with respectively 4 and 6 segemtents, by the Sigmoid function. This
means that the Radial basis function has the best score for both situations, with
80.3% and 78.5%. Remarkable is the fact that the difference is small between
the two situations, while there are two extra clusters. The confusion matrix for
both situations, table 6.11 and 6.12, show that there are two clusters which can
easily be classified with the customer profile. This corresponds to the cluster in
the top right corner and the cluster in the bottom of Figures 4.9 and 4.10.
63
C 1 2 5 10 20 50 100 200 500
γ = 0.4 33.8 34.0 34.7 33.1 34.6 30.0 32.6 28.8 28.8
γ = 0.6 29.6 27.4 28.5 29.7 21.4 20.8 20.0 18.8 18.1
γ = 0.8 39.1 36.4 33.6 35.7 38.9 32.0 26.4 24.6 22.9
γ = 1.0 40.0 42.5 39.8 40.7 39.9 39.8 30.4 31.1 28.0
γ = 1.2 41.9 40.6 43.6 43.2 44.1 43.2 44.6 40.6 41.7
γ = 1.4 38.6 34.5 32.1 30.6 30.2 27.5 24.3 26.3 27.9
Table 6.10: Sigmoid function, 6 segments
Predicted → Segment 1 Segment 2 Segment 3 Segment 4
Actual ↓
Segment 1 97.1% 0.5% 1.9% 0.5%
Segment 2 3.6% 76.6% 7.8% 12.0%
Segment 3 2.2% 0.8% 96.3% 0.7%
Segment 4 7.1% 13.0% 6.9% 73.0%
Table 6.11: Confusion matrix, 4 segments
Predicted → Segm. 1 Segm. 2 Segm. 3 Segm. 4 Segm. 5 Segm. 6
Actual ↓
Segment 1 74.1% 1.1% 10.1% 8.4% 0.6% 5.7%
Segment 2 0.2% 94.5% 0.6% 1.4% 1.2% 2.1%
Segment 3 5.6% 4.7% 71.2% 9.1% 2.1% 7.3%
Segment 4 12.3% 4.1% 3.9% 68.9% 6.8% 4.0%
Segment 5 2.0% 0.6% 0.7% 1.3% 92.6% 2.8%
Segment 6 12.5% 2.4% 3.7% 10.4% 1.3% 69.7%
Table 6.12: Confusion matrix, 6 segments
64
6.3 Feature Validation
In this section, the features will be validated. The importance of each feature
will be measured. This will be done, by leaving one feature out of the feature
vector and train the SVM without this feature. The results of both situations,
are shown in Figure 6.4 and 6.5. The result show that Age is an important
Figure 6.4: Results while leaving out one of the features with 4 segments
Figure 6.5: Results while leaving out one of the features with 6 segments
feature for classifying the right segment. This is in contrast with the type of
telephone, which increase the result with only tenths of percents. Each feature
increases the result and therefore each feature is useful for the classification.
65
Chapter 7
Conclusions and discussion
This section concludes the research and the corresponding results and will give
some recommendations for future work.
7.1 Conclusions
The first objective of our research was to perform automatic customer segmen-
tation based on usage behavior, without the direct intervention of a human
specialist. The second part of the research was focused on profiling customers
and finding a relation between the profile and the segments. The customer
segments were constructed by applying several clustering algorithms. The clus-
tering algorithms used selected and preprocessed data from the Vodafone data
warehouse. This led to solutions for the customer segmentation with respec-
tively four segments and six segments. The customer’s profile was based on
personal information of the customers. A novel data mining technique, called
Support Vector Machines was used to estimate the segment of a customer based
on his profile.
There are various ways for selecting suitable feature values for the clustering
algorithms. This selection is vital for the resulting quality of the clustering.
One different feature value will result in different segments. The result of the
clustering can therefore not be regarded as universally valid, but merely as one
possible outcome. In this research, the feature values were selected in such a
way that it would describe the customer’s behavior as complete as possible.
However, it is not possible to include all possible combinations of usage behav-
ior characteristics within the scope of this research. To find the optimal number
of clusters, the so-called elbow criterion was applied. Unfortunately, this crite-
rion could not always be unambiguously identified. An other problem was that
the location of the elbow could differ between the validation measures for the
same algorithm. For some algorithms, the elbow was located at c = 4 and for
other algorithms, the location was c = 6. To identify the best algorithm, several
validation measures were used. Not every validation method marked the same
66
algorithm as the best algorithm. Therefore, some widely established validation
measures were employed to determine the most optimal algorithm. It was how-
ever not possible to determine one algorithm that was optimal for c = 4 and
c = 6. For the situation with four clusters, the Gath-Geva algorithm appears to
be the best algorithm and the Gustafson-Kessel algorithm gives the best results
by six clusters. To determine which customer segmentation algorithm is best
suited for a particular data set and a specific parameter setting, the clustering
results were interpreted in a profiling format. The results show, that in both
situations the clusters were well separated and clearly distinguished from each
other. It is hard to compare the two clustering results, because of the different
number of clusters. Therefore, both clustering results were used as a starting
point for the segmentation algorithm. The corresponding segments differ on
features as number of voice calls, sms usage, call duration, international calls,
different numbers called and percentage of weekday and daytime calls. A short
characterization of each cluster was made.
A Support Vector Machine algorithm was used to classify the segment of a
customer, based on the customer’s profile. The profile exists of the age, gen-
der, telephone type, subscription type, company size, and residential area of
the customer. As a comparison, four different kernel functions with different
parameters were tested on their performance. It was found that the radial basis
function gives the best result with a classification of 80.3% for the situation
with four segments and 78.5% for the situation with six segments. It appeared
that the resulting percentage of correctly classified segments was not as high
as expected. A possible explanation could be that the features of the customer
are not adequate for making a customer’s profile. This is caused by the fre-
quently missing data in the Vodafone data warehouse about lifestyle, habits
and income of the customers. A second reason for the low number of correct
classification is the fact that the usage behavior in the database corresponds
to a telephone number and this telephone number corresponds to a person. In
real life, however, this telephone is maybe not used exclusively by the person
(and the corresponding customer’s profile) as stored in the database. Customers
may lend their telephone to relatives, and companies may exchange telephones
among their employees. In such cases, the usage behavior does not correspond
to a single customer’s profile and this impairs the classification process.
The last part of the research involves the relative importance of each individ-
ual feature of the customer’s profile. By leaving out one feature value during
classification, the effect of each feature value became visible. It was found that
without the concept of ’customer age’, the resulting quality of the classifica-
tion was significantly decreased. On the other hand, leaving out a feature such
as the ’telephone type’ barely decreased the classification result. However, this
and some other features did well increase the performance of classification. This
implies, that this feature bears some importance for the customer profiling and
the classification of the customer’s segment.
67
7.2 Recommendations for future work
Based on our research and experiments, it is possible to formulate some recom-
mendations for obtaining more suitable customer profiling and segmentation.
The first recommendation is to use different feature values for the customer
segmentation. This can lead to different clusters and thus different segments.
To know the influence of the feature values on the outcome of the clustering, a
complete data analysis research is required. Also, a detailed data analysis of the
meaning of the cluster is recommended. In this research, the results are given
by a short description of each segment. Extrapolating this approach, a more
detailed view of the clusters and their boundaries can be obtained. Another
way to validate the resulting clusters is to offer them to a human expert, and
use his feed-back for improving the clustering criteria.
To improve on determining the actual number of clusters present in the data
set, the application of more specialized methods than the elbow criterion could
be applied. An interesting alternative is, for instance, the application of evolu-
tionary algorithms, as proposed by Wei Lu [21]. Another way of improving this
research is to extent the number of cluster algorithms like main shift cluster-
ing, hierarchical clustering or mixture of Gaussians. To estimate the segment
of the customer, also, other classification methods can be used. For instance,
neural networks, genetic algorithms or Bayesian algorithms. Of specific interest
is, within the framework of Support Vector Machines, cluster analysis of the
application of miscellaneous (non-linear) kernel functions.
Furthermore, it should be noted that the most obvious and best way to improve
the classification is to come to a more accurate and precise definition of the
customer profiles. The customer profile used in this research is not sufficient
detailed enough to describe the wide spectrum of customers. One reason for this
is the missing data in the Vodafone data warehouse. Consequently, an enhanced
and more precise analysis of the data ware house will lead to improved features
and, thus, to an improved classification.
Finally, we note that the study would improve noticeably by involving multiple
criteria to evaluate the user behavior, rather than mere phone usage as em-
ployed here. Similarly, it is challenging to classify the profile of the customer
based on the corresponding segment alone. However, this is a complex course
and it essentially requires the availability of high-quality features.
68
Bibliography
[1] Ahola, J. and Rinta-Runsala E., Data mining case studies in customer profiling.
Research report TTE1-2001-29, VTT Information Technology (2001).
[2] Amat, J.L., Using reporting and data mining techniques to improve knowledge of
subscribers; applications to customer profiling and fraud management. J. Telecom-
mun. Inform. Technol., no. 3 (2002), pp. 11-16.
[3] Balasko, B., Abonyi, J. and Balazs, F., Fuzzy Clustering and Data Analysis Tool-
box For Use with Matlab. (2006).
[4] Bounsaythip, C. and Rinta-Runsala, E., Overview of Data Mining for Customer
Behavior Modeling. Research report TTE1-2001-18, VTT Information Technol-
ogy (2001).
[5] Bezdek, J.C. and Dunn, J.C., Optimal fuzzy partition: A heuristic for estimating
the parameters in a mixture of normal distributions. IEEE Trans. Comput., vol.
C-24 (1975), pp. 835-838.
[6] Dibike, Y.B., Velickov, S., Solomatine D. and Abbott, M.B., Model Induction
with Support Vector Machines: Introduction and Applications. J. Comp. in Civ.
Engrg., vol. 15 iss. 3 (2001), pp. 208-216.
[7] Feldman, R. and Dagan, I., Knowledge discovery in textual databases (KDT). In
Proc. 1st Int. Conf. Knowledge Discovery and Data Mining, (2005), pp. 112-117.
[8] Frawley, W.J., Piatetsky-Shapiro, G. and Matheus, C.J., Knowledge discovery in
databases, AAAI/MIT Press (1991), pp. 1-27.
[9] Gath, I. and Geva, A.B., Unsupervised optimal fuzzy clustering. IEEE Trans
Pattern and Machine Intell, vol. 11 no. 7 (1989), pp. 773-781.
[10] Giha, F.E., Singh, Y.P. and Ewe, H.T., Customer Profiling and Segmentation
based on Association Rule Mining Technique. Proc. Softw. Engin. and Appl., no.
397 (2003).
[11] Gustafson, D.E. and Kessel, W.E., Fuzzy clustering with a fuzzy covariance ma-
trix. In Proc. IEEE CDC, (1979), pp. 761766.
[12] Janusz, G., Data mining and complex telecommunications problems modeling. J.
Telecommun. Inform. Technol., no. 3 (2003), pp. 115-120.
69
[13] Mali, K., Clustering and its validation in a symbolic framework. Patt. Recogn.
Lett., vol. 24 (2003), pp. 2367-2376.
[14] Mattison, R., Data Warehousing and Data Mining for Telecommunications.
Boston, London: Artech House, (1997).
[15] McDonald, M. and Dunbar, I., Market segmentation. How to do it, how to profit
from it. Palgrave Publ., (1998).
[16] Noble, W.S., What is a support vector machine? Nature Biotechnology, vol. 24
no. 12 (2006), pp. 1565-1567.
[17] Shaw, M.J., Subramaniam, C., Tan, G.W. and Welge, M.E., Knowledge manage-
ment and data mining for marketing. Decision Support Systems, vol. 31 (2001),
pp. 127137.
[18] Verhoef, P., Spring, P., Hoekstra, J. and Lee, P., The commercial use of segmenta-
tion and predictive modeling techniques for database marketing in the Netherlands.
Decis. Supp. Syst., vol. 34 (2002), pp. 471-481.
[19] Virvou, M., Savvopoulos, A. Tsihrintzis, G.A. and Sotiropoulos, D.N., Construct-
ing Stereotypes for an Adaptive e-Shop Using AIN-Based Clustering. ICANNGA
(2007), pp. 837-845.
[20] Wei, C.P. and Chiu, I.T., Turning telecommunications call detail to churn pre-
diction: a data mining approach. Expert Syst. Appl., vol. 23 (2002), pp. 103112.
[21] Wei Lu, I.T., A New Evolutionary Algorithm for Determining the Optimal Num-
ber of Clusters. CIMCA/IAWTIC (2005), pp. 648-653.
[22] Weiss, G.M., Data Mining in Telecommunications. The Data Mining and Knowl-
edge Discovery Handbook (2005), pp. 1189-1201.
70
Appendix A
Model of data warehouse
In this Appendix a simplified model of the data ware house can be found. The
white rectangles correspond to the tables that were used for this research. The
most important data fields of these tables are written in the table. The colored
boxes group the tables in a category. To connected the tables with each other,
the relation tables (the red tables in the middle) are needed.
71
Figure A.1: Model of the Vodafone data warehouse
72
Appendix B
Extra results for optimal
number of clusters
In this Appendix, the plots of the validation measures, for the algorithms that
not were discussed in Section 4.1, are given.
The K-medoid algorithm:
Figure B.1: Partition index and Separation index of K-medoid
73
Figure B.2: Dunn’s index and Alternative Dunn’s index of K-medoid
The Fuzzy-C-means algorithm:
Figure B.3: Partition coefficient and Classification Entropy of Fuzzy C-means
74
Figure B.4: Partition index, Separation index and Xie Beni index of Fuzzy
C-means
Figure B.5: Dunn’s index and Alternative Dunn’s index of Fuzzy C-means
75

Acknowledgments
This Master thesis was written to complete the study Operations Research at the University of Maastricht (UM). The research took place at the Department of Mathematics of UM and at the Department of Information Management of Vodafone Maastricht. During this research, I had the privilege to work together with several people. I would like to express my gratitude to all those people for giving me the support to complete this thesis. I want to thank the Department of Information Management for giving me permission to commence this thesis in the first instance, to do the necessary research work and to use departmental data. I am deeply indebted to my supervisor Dr. Ronald Westra, whose help, stimulating suggestions and encouragement helped me in all the time of research for and writing of this thesis. Furthermore, I would like to give my special thanks to my second supervisor Dr. Ralf Peeters, whose patience and enthusiasm enabled me to complete this work. I have also to thank my thesis instructor, Drs. Annette Schade, for her stimulating support and encouraging me to go ahead with my thesis. My former colleagues from the Department of Information Management supported me in my research work. I want to thank them for all their help, support, interest and valuable hints. Especially I am obliged to Drs. Philippe Theunen and Laurens Alberts, MSc. Finally, I would like to thank the people, who looked closely at the final version of the thesis for English style and grammar, correcting both and offering suggestions for improvement.

1

Contents
1 Introduction 1.1 Customer segmentation and customer profiling 1.1.1 Customer segmentation . . . . . . . . . 1.1.2 Customer profiling . . . . . . . . . . . . 1.2 Data mining . . . . . . . . . . . . . . . . . . . . 1.3 Structure of the report . . . . . . . . . . . . . . 2 Data collection and preparation 2.1 Data warehouse . . . . . . . . . 2.1.1 Selecting the customers 2.1.2 Call detail data . . . . . 2.1.3 Customer data . . . . . 2.2 Data preparation . . . . . . . . 8 9 9 10 11 13 14 14 14 15 19 20 22 22 23 23 24 27 27 28 28 29 30 31 33 33 34 35

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

3 Clustering 3.1 Cluster analysis . . . . . . . . . . . . . . 3.1.1 The data . . . . . . . . . . . . . 3.1.2 The clusters . . . . . . . . . . . . 3.1.3 Cluster partition . . . . . . . . . 3.2 Cluster algorithms . . . . . . . . . . . . 3.2.1 K-means . . . . . . . . . . . . . . 3.2.2 K-medoid . . . . . . . . . . . . . 3.2.3 Fuzzy C-means . . . . . . . . . . 3.2.4 The Gustafson-Kessel algorithm 3.2.5 The Gath Geva algorithm . . . . 3.3 Validation . . . . . . . . . . . . . . . . . 3.4 Visualization . . . . . . . . . . . . . . . 3.4.1 Principal Component Analysis . 3.4.2 Sammon mapping . . . . . . . . 3.4.3 Fuzzy Sammon mapping . . . . .

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

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

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

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

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

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

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

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

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

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

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

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

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

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

4 Experiments and results of customer segmentation 37 4.1 Determining the optimal number of clusters . . . . . . . . . . . . 37 4.2 Comparing the clustering algorithms . . . . . . . . . . . . . . . . 42

2

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68 Bibliography A Model of data warehouse B Extra results for optimal number of clusters 68 71 73 3 .1 K-fold cross validation 6.4 The kernel functions . . . . . . . . . . . . . . . 6 Experiments and results 6. . . . . . . . . .2 Recommendations for future work . . . . . . . . . . . . . . . . .2 Parameter setting . . . . . . .2 The maximum-margin hyperplane 5. . 5. . . . . . . . . . .5 Multi class classification . . . . . . . . . . . . . . . 61 . . . . . . . . . . . . . . . . . . . . . . . 65 7 Conclusions and discussion 66 7. . . . . . . . 6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 53 53 55 56 56 59 5 Support Vector Machines 5. . . . . . 5. . . . .1 Conclusions . . . . . . of classifying the customer segments 60 . . . . . . . .1 The separating hyperplane . . . . . .3 Designing the segments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 . . 66 7. .4. .3 The soft margin . . .3 Feature Validation . . . . . . . 5. . .

. .7 4. . . . . .2 3. . . . . . . . . . . . . . .4 4. . . . . . . . . . . . . Relation between originated and received calls . . . . . . . . . . . . . . . . . . . . . .1 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Result of Gustafson-Kessel algorithm . . . . . Relation between daytime and weekday calls . . . . . Cluster profiles of centers for c = 6 . . . . . . .3 4.2 4. . . . . . Values of Partition Index. . . . . . .9 4. . . . . . . Separation Index and the Xie Beni Index Values of Dunn’s Index and the Alternative Dunn Index . . . Result of Gath-Geva algorithm . . . . . Result of K-medoid algorithm . . Cluster profiles for c = 4 . . . . . . . . . . . . . . . . . . Distribution of distances from cluster centers within clusters for the Gath-Geva algorithm with c = 4 . . . . . . . . . . . . . . . . . . . Values of Partition coefficient and Classification Entropy with Gustafson-Kessel clustering . . . . . . . . . . . . . . . . . .4 2. . . . . . .11 4. . . . . . . . . Values of Partition Index. . .14 4. . . . . . . .1 4. . . . . . . . . Hard and fuzzy clustering . . . . . . . . . . . . Values of Dunn’s Index and Alternative Dunn Index with GustafsonKessel clustering . . . . . . . . . . . . . .15 4. . . . . . . . . . .3 2. . . . . . . . . .3 4. Cluster profiles for c = 6 . . . . . . . . 4 . .1 3. . . . . . . . . . . . . . . . . . . . . . .5 3. . . . . . . . . . .16 5. . . . . . Structure of customers by Vodafone . . Visualization of phone calls per hour . . . . . . . .5 4. . . . . . . . . . . Distribution of distances from cluster centers within clusters for the Gustafson-Kessel algorithm with c = 6 . . . . . . . . . . .List of Figures 1. .1 A taxonomy of data mining tasks . . . . . . . . . . Two-dimensional customer data of segment 1 and segment 2 . . . . . Histograms of feature values . . . .8 4. . . . . Separation Index and the Xie Beni Index with Gustafson-Kessel clustering .13 4. . . . . . . . Result of Fuzzy C-means algorithm . . . . . . . . . . . . . . . . . . . . . . . .12 4.6 4. . Cluster profiles of centers for c = 4 .1 2. . . . . . . . . Result of K-means algorithm .2 2. .10 4. . 12 15 17 18 18 19 22 24 25 38 39 40 41 41 43 44 44 44 45 46 46 47 48 49 50 54 Example of clustering data . . Different cluster shapes in R2 . . . . . . . . .

. . . . . . .4 5. . . . .5 5.2 5. . . . . Results while leaving out one of the features with 4 segments Results while leaving out one of the features with 6 segments A. . . . .4 Partition index and Separation index of K-medoid . . .6 5.2 B. . . . . . . . .4 6.3 B. . . . . . Demonstration of kernels . . . . . Demonstration of the maximum-margin hyperplane Demonstration of the soft margin . . . .7 6. . . . . . . . . . . . . . . . B. . . . Determining the stopping point of training the SVM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Separation index and Xie Beni index of Fuzzy C-means . A separation of classes with complex boundaries . . . . . . . . . 54 55 56 57 58 59 60 61 61 65 65 72 73 74 74 75 75 Under fitting and over fitting . . . . . . . . . . . . B. . . . . . . . . . . . . . .3 5. . . Partition coefficient and Classification Entropy of Fuzzy C-means Partition index. . . . . . . .2 6. .5 Separating hyperplanes in different dimensions . . . . . . . . . 5 . . . . . . . Dunn’s index and Alternative Dunn’s index of K-medoid . . . . . . . . . Examples of separation with kernels . . . . . .1 Model of the Vodafone data warehouse . . . . . . . . . . .3 6. . .5. . . A K-fold partition of the dataset . . . .5 Dunn’s index and Alternative Dunn’s index of Fuzzy C-means . . . .1 B. .1 6. . . . .

.4 6. . 4 6 . . . . .1 4. . . . 6 segments . The values of all the validation measures with K-means clustering The values of all the validation measures with Gustafson-Kessel clustering . 6 segments . .List of Tables 2. . . . . . . . . . . . .11 6. . . . . . . . . . . . . . . .5 6. . . . . . . . . . . . . . . .1 4. . . . . . . . . . Linear Kernel. .1 6. . . . . . . Average C-value for polynomial kernel.6 6.4 4. 20 39 42 42 43 51 61 62 62 62 62 62 63 63 63 64 64 64 6 . . . . . Average C-value for polynomial kernel. . . . . .12 Proportions within the different classification groups . . . . . Linear Kernel. . . . .5 6. . . Polynomial kernel. .8 6. segments segments . . .10 6. . . . . .2 6. . Segmentation results . Sigmoid function. . . Confusion matrix. .9 6. . . . . . Polynomial kernel. . . . . . . . . . 6 segments . . . . . . . . . . . . . 4 segments . . . . . .2 4.3 6. . . . . . . . . . . . . . . 4 segments . . . . . . . . . . . 6 segments . . . Radial basis function. . 4 segments . .3 4. . . . . Confusion matrix. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Radial basis function. . . . . . . . . . . . . . . . . . . .7 6. . . . . The numerical values of validation measures for c = 4 . . . . . . . . . . . . . . . . . . . . 4 segments . Sigmoid function. . . . . . . . . . . . . . . . . . . . . The numerical values of validation measures for c = 6 . . . . . . . . . . 6 segments . . . . . . . . . 4 segments . . .

This research will address the question how to perform customer segmentation and customer profiling with data mining techniques. A number of advanced and state-of-the-art clustering algorithms are modified and applied for creating customer segments. this data holds valuable information that can be applied for operational and strategical purposes. automatic analysis is essential. The magnitude of this data is so huge that manual analysis of data is not feasible. the segment of a customer will be estimated based on the customers profile. the customer segmentation is based on usage call behavior. ’Customer profiling’ is describing customers by their attributes. managers can decide which marketing actions to take for each segment. such as age. Having these two components. gender. With six segments. with a recent data mining technique. This thesis describes the process of selecting and preparing the accurate data from the data warehouse. Different kernel functions with different parameters will be examined and analyzed. tastes. gender and residential area information.e. The best i. i. income and lifestyles. in order to extract such information from this data. Customer profiling can be accomplished with information from the data warehouse. In our context. An optimality criterion is constructed in order to measure their performance. Each segment will be described and analyzed. called Support Vector Machines. Therefore.3% of the cases to classify the segment of a customer based on its profile for the situation with four segments. a correct classification of 78. an International mobile telecommunications company. etc). ’customer segmentation’ is a term used to describe the process of dividing customers into homogeneous groups on the basis of shared or common attributes (habits. In this research.5% is obtained.e. the behavior of a customer measured in the amounts of incoming or outgoing communication of whichever form. These data mining techniques search and analyze the data in order to find implicit and useful information.Abstract Vodafone. by means of advanced data mining techniques. most optimal in the sense of the optimality criterion clustering technique will be used to perform customer segmentation. such as age. One solution with four segments and one solution with six segments. With the Support Vector Machine approach it is possible in 80. The customer segmentation will lead to two solutions. However. without direct knowledge of human experts. in order to perform customer segmentation and to profile the customer. 7 . Finally. has accumulated vast amounts of data on consumer mobile phone behavior in a data warehouse.

These automated systems perform important functions such as identifying network faults and detecting fraudulent phone calls. The need to handle such large volumes of data led to the development of knowledge-based expert systems [17. Having these two components. network data and customer data. A disadvantage of this approach is that it is based on knowledge from human experts. Customer profiling is describing customers by their attributes. information [12]. 10]. with approximately 4.1 million customers in The Netherlands. Vodafone is interested in a complete different issue. if not impossible [22]. the experts do not have the requisite knowledge [2]. etc) [10]. From all these customers a tremendous amount of data is stored. tastes. many data mining tasks can be distinguished. but potentially useful. while the network data gives a description of the state of the hardware and software components in the network. 22]. gender. among others. The customer data contains information of the telecommunication customers. Examples of main problems for marketing and sales departments of telecommunication operators are churn prediction. fraud detection. identifying trends in customer behavior and cross selling and up-selling.Chapter 1 Introduction Vodafone is world’s leading mobile telecommunications company. such as age. Data mining is the process of searching and analyzing data in order to find implicit. Solutions to these problems were promised by data mining techniques. These data include. The amount of data is so great that manual analysis of data is difficult. income and lifestyles [1. namely customer segmentation and customer profiling and the relation between them. A basic way to perform customer segmentation is to define segmentations in 8 . marketers can decide which marketing actions to take for each segment and then allocate scarce resources to segments in order to meet specific business objectives. Within the telecommunication branch. call detail data. Call detail data gives a description of the calls that traverse the telecommunication networks. and in many cases. Obtaining knowledge from human experts is a time consuming process. Customer segmentation is a term used to describe the process of dividing customers into homogeneous groups on the basis of shared or common attributes (habits.

it is needed to divide customers in segments and to profile the customers. Once the segmentations are obtained. different settings of the Support Vector Machines will be examined and the best working estimation model will be used. Customer profiling is a way of applying external data to a population of possible customers. In this research. called clustering techniques.advance with knowledge of an expert. The construction of user 9 . In this report. Profiling is performed after customer segmentation. Customer segmentation is a preparation step for classifying each customer according to the customer groups that have been defined. The goal is to predict behavior based on the information we have on each customer [18]. Segmenting means putting the population in to segments according to their affinity or similar characteristics. a data mining technique called Support Vector Machines (SVM) will be used. The process of segmentation describes the characteristics of the customer groups (called segments or clusters) within the data. To find a relation between the profile and the segments. for each customer a profile will be determined with the customer data. Customer profiling is done by building a customer’s behavior model and estimating its parameters. tested. such as age.1 Customer segmentation Segmentation is a way to have more targeted communication with the customers. the segment can be estimated and the usage behavior of the customer profile can be determined. Based on the combination of the personal information (the customer profile). will be developed. and dividing the customers over these segmentations by their best fits. Another key benefit of utilizing the customer profile is making effective marketing strategies. 1. 1. Segmentation is essential to cope with today’s dynamically fragmenting consumer marketplace. validated and compared to each other.1 Customer segmentation and customer profiling To compete with other providers of mobile telecommunications it is important to know enough about your customers and to know the wants and needs of your customers [15]. A Support Vector machine is able to estimate the segment of a customer by personal information. different data mining techniques. By using segmentation. This research will deal with the problem of making customer segmentations without knowledge of an expert and without defining the segmentations in advance. the principals of the clustering techniques will be described and the process of determining the best technique will be discussed. gender and lifestyle. Depending on data available.1. To realize this. marketers are more effective in channeling resources and discovering opportunities. The segmentations will be determined based on (call) usage behavior. it can be used to prospect new customers or to recognize existing bad customers. To realize this.

This makes it possible to estimate for each profile (the combination of demographic and personal information) the related segment and visa versa. In addition. Difficulties in making good segmentation are [18]: • Relevance and quality of data are essential to develop meaningful segments. A simple customer profile is a file that contains at least age and 10 . • Intuition: Although data can be highly informative. If the company has insufficient customer data. apparently effective variables may not be identifiable. thereby necessitating revision and reclassification of customers. Alternatively. Depending on the goal. the meaning of a customer segmentation in unreliable and almost worthless. effective segmentation strategies will influence the behavior of the customers affected by them. Many of these problems are due to an inadequate customer database. such as demographic data purchased from various sources.segmentations is not an easy task. More directly. for each profile. Poorly organize data (different formats. the use of too many segmentation variables can be confusing and result in segments which are unfit for management decision making. In this report. the resulting segmentation can be too complicated for the organization to implement effectively. several clustering algorithms will be discussed and compared to each other. Furthermore. Moreover.2 Customer profiling Customer profiling provides a basis for marketers to ’communicate’ with existing customers in order to offer them better services and retaining them. In particular. One solution to construct segments can be provided by data mining methods that belong to the category of clustering algorithms. • Over-segmentation: A segment can become too small and/or insufficiently distinct to justify treatment as separate segments.1. On the other hand. 1. data analysts need to be continuously developing segmentation hypotheses in order to identify the ’right’ data for analysis. too much data can lead to complex and time-consuming analysis. one has to select what is the profile that will be relevant to the project. This is done by assembling collected information on the customer such as demographic and personal data. an estimation of the usage behavior can be obtained. in an e-commerce environment where feedback is almost immediate. segmentation would require almost a daily update. This data is used to find a relation with the customer segmentations that were constructed before. Customer profiling is also used to prospect new customers using external sources. • Continuous process: Segmentation demands continuous development and updating as new customer data is acquired. different source systems) makes it also difficult to extract interesting information.

What languages do they speak? Does ethnicity affect their tastes or buying behaviors? • Economic conditions. 10. exploring and modeling large amounts of data to uncover previously unknown patterns. This can be realized by a data mining method called Support Vector Machines (SVM). income and/or purchasing power. the file would contain product information and/or volume of money spent. data mining tools have been available for a long time. Data mining uses a broad family of computational methods that include statistical analysis. the term data mining was used. attitudes and beliefs. How long has the customer been regularly purchasing products? • Knowledge and awareness. but potentially useful. information [12]. How many lifestyle characteristics about purchasers are useful? • Recruitment method. nationally or globally • Cultural and ethnic. How much knowledge do customers have about a product. It involves selecting. What is the average household incom or power of the customers? Do they have any payment difficulty? How much or how often does a customer spend on each product? • Age and gender. and ultimately comprehensible information. or industry? How much education is needed? How much brand building advertising is needed to make a pool of customers aware of offer? • Lifestyle. from large databases. How was the customer recruited? The choice of the features depends also on the availability of the data. and graphic visualization. are described in [2. 19]: • Geographic. Although. neural networks.2 Data mining In section 1. particularly exploratory tools like data visualization and neural 11 . an estimation model can be made. Data mining is the process of searching and analyzing data in order to find implicit. 1. If one needs profiles for specific products. What is the customers’ attitude toward your kind of product or service? • Life cycle. This report gives an description of SVM’s and it will be researched under which circumstances and parameters a SVM works best in this case. the advances in computer hardware and software. What is the predominant age group of your target buyers? How many children and what age are in the family? Are more female or males using a certain service or product? • Values. Customer features one can use for profiling.service. With these features.gender.1. decision trees. Are they grouped regionally. rule induction and refinement.

Classification algorithms groups customers in predefined classes. For example. a pattern is defined as [4]: A statement S in L that describes relationships among a subsets of facts Fs of a given set of facts F. The typical data mining process consist of the following steps [4]: • problem formulation • data preparation • model building • interpretation and evaluation of the results Pattern extraction is an important component of any data mining activity and it deals with relationships between subsets of data. have made data mining more attractive and practical. Data mining tasks are used to extract patterns from large data sets. Validation of the results is also a data mining task. this task was not mentioned as a separate one. The taxonomy reflects the emerging role of data visualization as Figure 1. The specific tasks to be used in this research are Clustering (for the customer segmentation). A drawback of this method is that the number of clusters has to be given in advance. The data mining tasks generate an assortment of customer and market knowledge which form the core of knowledge management process. The advantage of clustering is that expert knowledge is not required. Clustering algorithms produce classes that maximize similarity within clusters but minimize similarity between classes. based on user behavior data. Different data mining tasks are grouped into categories depending on the type of knowledge extracted by the tasks. By the fact that the validation supports the other data mining tasks and is always necessary within a research.networks. Formally. clustering algorithms can classify the Vodafone customers into ”call only” users. such that S is simpler than the enumeration of all facts in Fs . Classification (for estimating the segment) and Data visualization. with some certainty C. The identification of patterns in a large data set is the first step to gaining useful marketing insights and marking critical marketing decisions.1: A taxonomy of data mining tasks a separate data mining task. The various data mining tasks can be broadly divided into six categories as summarized in Figure 1. even as it is used to support other data mining tasks.1. ”international callers”. For example. 12 . ”SMS only” users etc.

Chapter 5 delves into a data mining technique called Support Vector Machines. The chapter ends with a description of visualization methods. a profile can be made. data miners use applications that provide advanced manipulation capabilities to slice. 1. It provides information about the structure of the data and the data ware house. gender and type of subscription and then target its user behavior.4) can be used. In Chapter 3 the process of clustering is discussed. It ends with an explanation of the preprocessing techniques that were used to prepare the data for further usage. Then. This will be tested with the prepared call detail data as described in Chapter 2 For each algorithm. algorithms as Principal Component Analysis and Sammon’s Mapping (discussed in Section 3. Different cluster algorithms will be studied. Finally. Once the segments are determined. Conclusions and recommendations are given and future work is proposed. To realize this. Chapter 4 analyzes the different cluster algorithms of Chapter 3. It also focuses on validation methods. Multiple plots and figures will show the working of the different cluster methods and the meaning of each segment will be described.Vodafone can classify its customers based on their age.3 Structure of the report The report comprises 6 chapters and several appendices. In some cases it is needed to reduce high dimensional data into three or two dimensions. that in this research is used to determine the customer segmentations. with the customer data of Chapter 2. This technique will be used to classify the right segment for each customer profile. the optimal numbers of cluster will be determined. in Chapter 7. the cluster algorithms will be compared to each other and the best algorithm will be chosen to determine the segments. These methods are used to analyze the results of the clustering. rotate or zoom the objects. Data visualization allow data miners to view complex patterns in their customer data as visual objects complete in three or two dimensions and colors. the research will be discussed. it gives an overview of the data that is used to perform customer segmentation and customer profiling. Different parameter settings of the Support Vector Machines will be researched and examined in Chapter 6 to find the best working model. In addition to to this introductory chapter. To provide varying levels of details of observed patterns. The chapter starts with explaining the general process of clustering. Furthermore. 13 . Clustering is a data mining technique. Chapter 2 describes the process of selecting the right data from the data ware house. which can be used to determine the optimal number of clusters and to measure the performance of the different cluster algorithms.

Debitel and InterCity Mobile Communications (ICMC). It is clear to see. that their customers can use the Vodafone network. In this chapter. that prepaid users are always consumers. A more precisely view can be found in Figure 2. Vodafone has made an accomplishment with two other telecommunications companies. All data of Vodafone is stored in a data warehouse. Debitel customers are always consumers and ICMC customers are always business customers. Furthermore. A captive customer has a business account if his telephone or subscription is bought in relation with the business 14 . In the postpaid group. Without such an understanding. A non-captive customer is using the Vodafone network but has not a Vodafone subscription or prepaid (called roaming). 2. useful applications cannot be developed. The ICMC customers will also be involved in this research. will be described.1. This data warehouse exists off more than 200 tables. the process of preparing the data for customer segmentation and customer profiling will be explained. there are captive and non captive users.1 Data warehouse Vodafone has stored vast amounts of data in a Teradata data warehouse.1. In general.1 Selecting the customers Vodafone Maastricht is interested in customer segmentation and customer profiling for (postpaid) business customers. the process of collecting the right data from this data ware house. business customers can be seen as employees of a business that have a subscription by Vodafone in relation with that business. 2.Chapter 2 Data collection and preparation The first step (after the problem formulation) in the data mining process is to understand the data. A simplified model of the data warehouse can be found in Appendix A.

2 Call detail data Every time a call is placed on the telecommunications network of Vodafone. These customers also count as business users.000. each call detail record will include the originating and terminating phone numbers. The total number of (postpaid) business users at Vodafone is more than 800. the date and the time of the call and the duration of the call. To define the features. The number of call detail records that are generated and stored is huge. who. In some cases. Call detail records include sufficient information to describe the important characteristics of each call. etc. can have a subscription that is under normal circumstances only available for business users. Vodafone customers generate over 20 million call detail records per day. the call detail records associated with a customer must be summarized into a single record that describes the customer’s calling behavior. not at the level of individual phone calls [7. when. descriptive information about the call is saved as a call detail record. The next sections describe which data of these customers is needed for customer segmentation and profiling.Figure 2. how often. These customers are called business users. 8]. can help with this process: 15 . Call detail records can not be used directly for data mining. Keywords like what. Given that 12 months of call detail data is typically kept on line. Call detail records are generated in two or three days after the day the calls were made.1. Thus. which is typically made available only once per month. 2. one can think of the smallest set of variables that describe the complete behavior of a customer. At a minimum. where.1: Structure of customers by Vodafone he works. customers with a consumer account. This is in contrast with billing data. this means that hundreds of millions of call detail data will need to be stored at any time. The choice of summary variables (features) is critical in order to obtain a useful description of the customer []. For example. and will be available almost immediately for data mining. since the goal of data applications is to extract knowledge at the customer level.

% daytime calls (9am . The customer can also receive an SMS or voice call. 19. # unique area codes called during P • 12. % of calls to mobile phones • 7. average # calls received per day • 3. • Where? : Where is the customer calling? Is he calling abroad? • How long? : How long is the customer calling? • How often? : How often does a customer call or receive a call? Based on these keywords and based on proposed features in the literature [1. average # calls originated per day • 4. • Who? : Who is the customer calling? Does he call to fixed lines? Does he call to Vodafone mobiles? • What? : What is the location of the customer and the recipient? They can make international phone calls.6pm) • 5. % of weekday calls (Monday . but their appearances are so rare that they were not used during this research). 15. a list of features that can be used as a summary description of a customer based on the calls they originate and receive over some time period P is obtained: • 1.Friday) • 6. average call duration • 2. % international calls • 10. • When? : When does a customer call? A business customer can call during office daytime. average # sms received per day • 8. or in private time in the evening or at night and during the weekend. 20] . average # sms originated per day • 9. # different numbers called during P 16 .• How? : How can a customer cause a call detail record? By making a voice call. % of outgoing calls within the same operator • 11. or sending an SMS (there are more possibilities.

4 is also visible that the customers that originated more calls. Note that the histograms resemble well known distributions. but some features require a little more creativity and operations on the data. including summary features.4 demonstrates this. data mining will not be successful. In Figure 2. customers who use their telephone only at their office could be in a different segment then users that use their telephone also for private purposes. More detailed exploratory data analysis. the number of weekday and daytime calls and the originated calls have sufficient variance. For examples. Interesting to see is the relation between the number of calls originated and received.2: Visualization of phone calls per hour variance within the data. it should include exploratory data analysis. For example. On the other hand. values above the blue line represent customers with more originating calls than receiving calls. Furthermore. Another aspect that is simple to figure out is the fact that customer 17 .2 indicates that the period from 9am to 6pm is actually more appropriate for this purpose. is a critical step within the data mining process. to much variance hampers the process of segmentation. Such a segment describes a certain behavior of group of customers. This also indicates that the chosen features are suited for the customer segmentation. shown in Figure 2. receive also more calls in proportion. in general. First of all. Figure 2. For some features values. In that case. customers originating more calls than receiving. Figure 2. there should be sufficient Figure 2. It may be clear that generating useful features. Most of the twelve features listed above can be generated in a straightforward manner from the underlying data of the data ware house.3 shows that the average call duration. for each summary feature. the segmentation was based on the percentage weekday and daytime calls. otherwise distinguish between customers is not possible and the feature is not useful. the variance is visible in the following histograms. Should poor features be generated.These twelve features can be used to build customer segments. Although the construction of these features may be guided by common sense. the use of the time period 9am-6pm in the fourth feature is not based on the commonsense knowledge that the typical workday on a office is from 9am to 5pm.

3: Histograms of feature values Figure 2.4: Relation between originated and received calls 18 .(a) Call duration (b) Weekday calls (c) Daytime calls (d) Originated calls Figure 2.

3 Customer data To profile the customer. advance.5. The chosen features appear to be well chosen and useful for customer segmentation. 25-40 40-55 >55 • Gender: male. with some creativity. small city /town 19 . the following variables can be used to define a customers profile: • Age group: <25.5: Relation between daytime and weekday calls 2. The information that Vodafone stored in the data ware house include name and address information and also include other information such as service plan. intermediate.that make more weekday calls also call more at daytime (in proportion). basic. However.1. The proposed data in Section 1. some information can be subtracted from the data ware house. With this information. Information about lifestyles and income is missing. It is clear to see that the chosen features contain sufficient variance and that certain relations and different customer behavior are already visible.2 is not completely available. big • Living area: (big) city. female • Type telephone: simple. customer data is needed.1. This is plotted in Figure 2. expanded • Company size: small. Figure 2. advanced • Type subscription: basic. contract information and telephone equipment information.

Chapter 5 and Chapter 6 contain information and results of this method. a Support Vector Machine will be used to estimate the segment of the customer. This is caused by the fact that from each segment a relative high number of customers is represented in this group.8% expanded 29.9% small 31.7% advanced 36.8% basic 38. • Deleting unwanted data fields. If there is one group with a sufficient higher amount of customers than other groups. the age of the customers has to be grouped.2 Data preparation Before the data can be used for the actual data mining process.5% simple 34. In general.3% small city/town 58. this feature will not increase the performance of the classification. 2. Table 2. spelling errors.With this profile.0% 25-40 29.5% (big) city 42.2% Male 60. Otherwise. The composition of the groups should be chosen with care. it need to cleaned and prepared in a required format.9% >55 21.1 shows the percentages of customers within the chosen groups. It is of high importance that the sizes of the groups are almost equal (if this is possible) [22].1: Proportions within the different classification groups and the values can be used for defining the customers profile. Data may contain many meaningless fields from an analysis point of view.4% advanced 27. These tasks are [7]: • Discovering and repairing inconsistent data formats and inconsistent data encoding. the segment of a customer can not be determined.Because a relative small difference in age between customers should show close relationships. abbreviations and punctuation.5% Female 39. It is clear to see that sizes of the groups were chosen with care Age: Gender: Telephone type: Type of subscription: Company size: Living area: <25 21. • Interpreting codes into text or replacing text into meaningful numbers. the result of the classification algorithm is too specific to the trainings data [14].0% 40-55 27.2% Table 2.2% simple 33. the goal of grouping variables is to reduce the number of variables to a more manageable size and to remove the correlations between each variable. 20 .1% big 34.0% intermediate 34. such as production keys and version numbers. Based on this feature.

correspondence analysis and conjoint analysis [14]. • Combining data. The second type is to normalize the variance to one. • Converting from textual to numeral or numeric data.g. • Checking missing data fields or fields that have been replaced by a default value. Dimension reduction means that one has to select relevant feature to a minimum set of attributes such that the resulting probability distribution of data classes is a close as possible to the original distribution given the values of all features. for instance the customer data. clustering. • Mapping continuous values into ranges. e. it is also useful to apply data reduction techniques (data cube aggregation. decision trees or associations rules. discretization and concept hierarchy generation). New fields can be generated through combinations of e.g. out of bounds or ambiguous values. frequencies. The following data preparations were needed during this research: • Checking abnormal. • Finding multiple used fields.Data may contain cryptic codes. exhaustive. These codes has to be augmented and replaced by recognizable and equivalent text. Some of these outliers may be correct but this is highly unusual. There are two types of normalization. For this additional tools may be needed. from multiple tables into one common variable. decision trees. • Converting nominal data (for example yes/no answers) to metric scales.1]. random or heuristic search. • Normalization of the variables. The first type is to normalize the values between [0. dimension and numerosity reduction. When there is a large amount of data. e.g. averages and minimum/maximum values. • Adding computed fields as inputs or targets. The goal of this approach is to reduce the number of variables to a more manageable size while also the correlations between each variable will be removed. 21 . A possible way to determine is to count or list all the distinct variables of a field. thus almost impossible to explain. Techniques used for this purpose are often referred to as factor analysis.

3. Another way of clustering is conceptual clustering. Figure 3. This is called distance-based clustering. In this case the 3 Figure 3. Clustering can be considered the most important unsupervised learning method.Chapter 3 Clustering In this chapter. The similarity criterion that was used in this case is distance: two or more objects belong to the same cluster if they are ”close” according to a given distance (in this case geometrical distance). Within this method. A cluster can be defined as a collection of objects which are ”similar” between them and ”dissimilar” to the objects belonging to other clusters. As every other unsupervised method.1 shows this with a simple graphical example.1: Example of clustering data clusters into which the data can be divided were easily identified.1 Cluster analysis The objective of cluster analysis is the organization of objects into groups. according to similarities among them [13]. it does not use prior class identifiers to detect the underlying structure in a collection of data. two or more objects 22 . the used techniques for the cluster segmentation will be explained.

1. In other words. (3. qualitative (categoric) data. the purpose of clustering is to find relationships between independent system variables. In metric spaces. And therefore. In this research. Distance can be measured in different ways. 2.1. one should realize. similarity is often defined by means of a distance norm. and the columns are the feature variables of their behavior as described in Section 2. 3. xk2 . and will be calculated by the clustering algorithms simultaneously with the partitioning of the data. To obtain such a model.1) . .2. . grouped into an n-dimensional row vector xk = [xk1 . As mentioned before. and is represented as an N x n matrix:   x11 x12 · · · x1n  x21 x22 · · · x2n    X= . . called the regressands. However.. where xk ∈ Rn . . In general.. or distance measure. 23 . measured in some well-defined sense. Each observation of the customers calling behavior consists of n measured values. as described in Section 2... N }. The rows of X represent the customers. they will not automatically constitute a prediction model of the given system. The cluster centers are usually (and also in this research) not known a priori. xN 1 xN 2 · · · xN n In pattern recognition terminology. and X is called the pattern matrix. additional steps are needed. or a mixture of both.belong to the same cluster if this one defines a concept common to all that objects. xkn ]T .2 The clusters The definition of a cluster can be formulated in various ways. the rows of X are called patterns or objects. The data.1 The data One can apply clustering techniques to quantitative (numerical) data. the columns are called the features or attributes. . X will be referred to the data matrix. . .. The term ”similarity” can be interpreted as mathematical similarity.2. In this research. the clustering of quantitative data is considered. 3. .   . . are typically summarized observations of a physical process (call behavior of a customer). called the regressors. A set of N observations is denoted by X = {xk |k = 1. objects are grouped according to their fit to descriptive concepts.. . depending on the objective of the clustering. A second way is to measure the distance form the data vector to some prototypical object of the cluster. one can accept the definition that a cluster is a group of objects that are more similar to another than to members of other clusters.1. and future values of dependent variables. not according to simple similarity measures. The first possibility is to measure among the data vectors themselves. In this research.. that the relations revealed by clustering are not more than associations among the data vectors. only distance-based clustering algorithms were used.1. .

2: Different cluster shapes in R2 hollow.1. Clusters can be well-separated. Clustering algorithms are able to detect subspaces of the data space. and therefore reliable for identification. elongated and also be (a) Elongated (b) Spherical (c) Hollow (d) Hollow Figure 3.c and d can be characterized as linear and non linear subspaces of the data space (R2 in this case). but can also be defined as ”higher-level” geometrical objects.The cluster centers may be vectors of the same dimensions as the data objects. Clusters a. such as linear or nonlinear subspaces or functions. Cluster can be found in any n-dimensional space.3 Cluster partition Clusters can formally be seen as subsets of the data set. Data can reveal clusters of different geometrical shapes. 3. One can distinguish two possible outcomes of the classification of clustering methods. or overlapping each other. Subsets can 24 .2 Clusters can be spherical. continuously connected to each other. The performance of most clustering algorithms is influenced not only by the geometrical shapes and densities of the individual clusters. sizes and densities as demonstrated in Figure 3. but also by the spatial relations and distances among the clusters.

3: Hard and fuzzy clustering causes analytical and algorithmic intractability of algorithms based on analytic functionals.   . . 1 ≤ i ≤ c. 25 . .c    (3.g. which requires that an object either does or does not belong to a cluster. .. µN. Hard clustering methods are based on the classical set theory. . The data set X is thus partitioned into c fuzzy subsets. Using classical sets.3) (3.2 · · · µN. as objects on the boundaries between several classes are not forced to fully belong to one of the classes. since these functionals are not differentiable. but rather are assigned membership degrees between 0 and 1 indicating their partial memberships (illustrated by Figure 3. .c Hard partition The objective of clustering is to partition the data set X into c clusters. with different degrees of membership. Ø ⊂ Ai ⊂ X. .4) (3. 1 ≤ i = j ≤ c. Assume that c is known.2 · · · µ2. .3 The discrete nature of hard partitioning also Figure 3.2 · · · µ1.5) Ai ∩ Aj . Hard clustering in a data set X means partitioning the data into a specified number of exclusive subsets of X.1 µ1. based on prior knowledge. The structure of the partition matrix U = [µik ]:   µ1. e.either be fuzzy or crisp (hard).1 µN. or it is a trial value. of witch partition results must be validated. fuzzy clustering is more natural than hard clustering. . its properties can be defined as follows: c Ai = X.1 µ2. . a hard partition can be seen as a family of subsets {Ai |1 ≤ i ≤ c ⊂ P (X)}. The number of subsets (clusters) is denoted by c.c  µ2. In many real situations. i=1 (3.2) U= . Fuzzy clustering methods allow objects to belong to several clusters simultaneously. .

6) (3. Then. in this case µik is allowed to acquire all real values between zero and 1. 1 ≤ k ≤ c.15) Note that there is only one difference with the conditions of the hard partitioning. (3. partitions can be represented in a matrix notation. they must be disjoint and none of them is empty nor contains all the data in X. c (3.8) µAi ∨ µAi . 1 ≤ i ≤ N. 1}.9) (3. 1 ≤ i ≤ N. 1]. c (3.13) (3. 0 ≤ µAi < 1.11) A definition of a hard partitioning space can be defined as follows: Let X be a finite data set and the number of clusters 2 ≤ c < N ∈ N.These conditions imply that the subsets Ai contain all the data in X. To simplify these notations. 1 ≤ i ≤ N. k=1 N 0< i=1 µik < N. ∀i. 1 ≤ i ≤ N. 0 < i=1 µik < N . 1 ≤ k ≤ c.7) (3. its conditions are given by: µij ∈ [0. containing the fuzzy partitions. 1 ≤ k ≤ c. (3. (3. k=1 µik = 1. Where µAi represents the characteristic function of the subset Ai which value is zero or one. 1 ≤ k ≤ c. k=1 N 0< i=1 µik < N. a N xc matrix. 1}. i=1 (3. Also the definition of the fuzzy partitioning space will not much differ with 26 . U = [µik ]. Expressed in the terms of membership functions: c µAi = 1.14) µik = 1. 1 ≤ i = j ≤ c. the hard partitioning space for X can be seen as the set: c N Mhc = {U ∈ RN xc |µik ∈ {0. and denoting µi (xk ) by µik . Consider the matrix U = [µik ]. ∀k}.10) µik = 1. is a representation of the hard partition if and only if its elements satisfy: µij ∈ {0. µi will be used instead of µAi . ∀i. 1 ≤ i ≤ c.12) Fuzzy partition Fuzzy partition can be defined as a generalization of hard partitioning. k.

1]. This method will result into hard partitioned clusters. ∀i. 0 < i=1 µik < N . (3. The possibilistic partition will not be used in this researched and will not be discussed here. xk ∈ Ai . Then.17) Ai represents a set of data points in the i-th cluster and vi is the average of the data points in cluster i. The K-means algorithm allocates each data point to one of the c clusters to minimize the within sum of squares: c sumk∈Ai ||xk − vi ||2 . 3. k=1 µik = 1. 27 . ∀k}. the fuzzy partitioning space for X can be seen as the set: c N Mf c = {U ∈ RN xc |µik ∈ [0. ∀i.16) The i-th column of U contains values of the membership functions of the i-th fuzzy subset of X.14) implies that the sum of each column should be 1. This research will focus on hard partitioning. It can be defined as follows: Let X be a finite data set and the number of clusters 2 ≤ c < N ∈ N. The procedure follows an easy way to classify a given N x n data set through a certain numbers of c clusters defined in advance. To deal with the problem of fuzzy memberships. the results of this hard partitioning method are not always reliable and this algorithm has numerical problems as well. 3.18) vi = k=1 . fuzzy cluster algorithms will be applied as well. However. Equation (1. which means that the total membership of each xk in X equals one.2 Cluster algorithms This section gives an overview of the clustering algorithms that were used during the research. k.2. vi is the cluster center (also called prototype) of cluster i: Ni xk (3. Ni where Ni is the number of data points in Ai .1 K-means K-means is one of the simplest unsupervised learning algorithms that solves the clustering problem. There are no constraints on the distribution of memberships among the fuzzy clusters. Within the cluster algorithms. Note that ||xk −vi ||2 is actually a chosen distance norm. i=1 (3. However.the definition of the hard partitioning space. the cluster with the highest degree of membership will be chosen as the cluster were the object belongs to.

14 to J by means of Lagrange multipliers: c N 2 (µik )m DikA + i=1 k=1 k=1 N c ¯ J(X. V and λ.20) V denotes the vector with the cluster centers that has to be determined.19) with V = [v1 . to zero. one can adjoint the constrained in 3.22) ˆ and by setting the gradients of (J). vi ∈ Rn . λ) = λk i=1 µik − 1 .3 Fuzzy C-means The Fuzzy C-means algorithm (FCM) minimizes an objective function. equation 3.2 K-medoid K-medoid clustering. 1 ≤ i ≤ c. This method is called the fuzzy cmeans algorithm. then the C-means functional may only be minimized by (U.23) 28 .. there is no continuity in the data space. (DikA /DjkA )2/(m−1) (3. V ) ∈ Mf c xRnxc if µik = c j=1 1 . uses the same equations as the K-means algorithm. ∀i.19 measures the total number of variance of xk from vi . (3. k and m > 1. This can be useful when. U. with respect to U. The distance norm ||xk − vi ||2 is called a squared inner-product distance norm and A is defined by: 2 DikA = ||kk − vi ||2 = (xk − vi )T A(xk − vi ). V ) = i=1 k=1 (µik )m ||xk − vi ||2 . The C-means functional.21) On a statistical point of view. (3. V. This implies that a mean of the points in one cluster does actually not exist. 3. vc ]. To find the stationary points of the c-means functional. Examples of methods that can solve non linear optimization problems are grouped coordinate minimization and genetic algorithms. . A (3.. to define the clusters.2.2. When 2 DikA > 0. The minimization of the C-means functional can be seen as a non linear optimization problem.19. The only difference is that in K-medoid the cluster centers are the nearest data points to the mean of the data in one cluster V = {vi ∈ X|1 ≤ i ≤ c}. v2 . also a hard partitioning algorithm. invented by Dunn.3. 1 ≤ k ≤ N. that can be solved by a variety of methods. The simplest method to solve this problem is utilizing the Picard iteration through the first-order conditions for the stationary points of equation 3. is defined as follows: c N J(X. A (3. for example. U. called C-means functional..

matrix A is based ˆ on the Mahalanobis distance norm. (3. The objective functional of the GK algorithm can be calculated by: c N 2 (uik )m DikAi .   . Note that it can only detect clusters with the same shape. . where F = 1 N N (xk − x)(xk − x)T ˆ ˆ k=1 (3.25) AD =  . 3. This implies that each cluster is allowed to adapt the distance norm to the local topological structure of the data.24).23) and (3.2. Each cluster will have its own norm-inducing matrix Ai . Remark that the vi of equation (3.24) The solution of these equations are satisfying the constraints that were given in equation (3. it is able to define hyper spherical clusters. 0 0 · · · (1/σn )2 This matrix accounts for different variances in the directions of the coordinate axes of X.. Another possibility is to choose A as the inverse of the nxn covariance matrix A = F −1 . V.k ≤ i ≤ c.24) is the weighted average of the data points that belong to a cluster and the weights represents the membership degrees. (3. satisfying the following inner-product norm: 2 DikA = (xk − vi )T · Ai (xk − vi ). . The FCM algorithm uses the standard Euclidean distance for its computations. Therefor. . in this case.26) and x denotes the mean of the data. The norm inducing matrix can also be chosen as an nxn diagonal matrix of the form:   (1/σ1 )2 0 ··· 0   0 (1/σ2 )2 · · · 0   (3. i=1 k=1 J(X.15). The Fuzzy C-means algorithm is actually an iteration between the equations (3. . Hence that.. . where A = (A1 .1 N m k=1 µi.and vi = N m k=1 µik xk .13) and (3.. This explains why the name of the algorithm is c-means. A2 . A) = (3.28) 29 . . where 1 ≤ i ≤ c and 1 ≤ k ≤ N.4 The Gustafson-Kessel algorithm The Gustafson and Kessel (GK) algorithm is a variation on the Fuzzy c-means algorithm [11]. . . A c-tuple of the norm-inducing matrices is defined by A. Ac ). U.27) The matrices Ai are used as optimization variables in the c-means functional.. caused by the common choice of the norm inducing matrix A = I. . It employs a different and adaptive distance norm to recognize geometrical shapes in the data. . .

If A is fixed.13). w will be set to 2. Unfortunately. the fuzzy covariance matrix F i is defined by: Fwi = N k=1 (µik )w (xk − vi )(xk − vi )T N k=1 (µik )w . the conditions under (3.28) can not be minimized in a straight forward manner.30) . (3. The outcome of the inner-product norm of (3.5 The Gath Geva algorithm Bezdek and Dunn [5] proposed a fuzzy maximum likelihood estimation (FMLE) algorithm with a corresponding distance norm: Dik (xk .14) and (3.34) N k=1 30 .27).15) can be applied without any problems.27) is a generalized squared Mahalanobis norm between the data points and the cluster center. Ai can be expressed in the following way: Ai = [ρi det(Fi )]1/n Fi−1 .33) The reason for using the w variable is to generalize this expression. A varying Ai with a fixed determinant relates to the optimization of the cluster whit a fixed volume: ||Ai || = ρi . 1 ≤ i ≤ c. In combination with the Lagrange multiplier. equation (3. the distance norm includes an exponentional term. (3.2. (3. ρ > 0. αi can be defines as follows: N 1 αi = µik . Hence that this equation in combination with equation (3. This implies that this distance norm will decrease faster than the inner-product norm. N k=1 (µik )m (xk − vi )(xk − vi )T 3. with Fi = N k=1 (3. (3.30) can be substituted into equation (3. vi ) = det(Fwi ) αi (l) (l) T −1 1 2 (xk −vi ) Fwi (xk −vi ) . Ai has to be constrained to obtain a feasible solution. In this case. The variable αi in equation (3. A general way to this is by constraining the determinant of the matrix. This implies that J can be made as small as desired by making Ai less positive definite.32) Comparing this with the Gustafson-Kessel algorithm. To avoid this. In the original FMLE algorithm.32) is the prior probability of selecting cluster i. to compensate the exponential term and obtain clusters that are more fuzzy. (3. The covariance is weighted by the membership degrees of U . Because of the generalization.29) Here ρ is a remaining constant for each cluster. two weighted covariance matrices arise. since it is linear in Ai . In this research.31) (µik )m Fi is also called the fuzzy covariance matrix. w = 1. (3.

One can distinguish two main approaches to determine the correct number of clusters in the data: • Start with a sufficiently large number of clusters. none of them is perfect by oneself. CE(c) = − 1 N c N uij log(uij ) i=1 j=1 (3. However. the data can not be grouped in a meaningful way at all. 3. Different validation methods have been proposed in the literature. it is not know how reliable the results of this algorithm are. this does not apply that the best fit is meaningful at all. in this research are used several indexes. and successively reducing this number by combining clusters that have the same properties. To be able to perform the second approach. which is a slightly variation on the Partition Coefficient. In the worst case.Gath and Geva [9] discovered that the FMLE algorithm is able to detect clusters of different shapes. • Classification Entropy (CE): measures only the fuzziness of the cluster. The number of clusters might not be correct or the cluster shapes do not correspond to the actual groups in the data.3 Validation Cluster validation refers to the problem whether a found partition is correct and how to measure the correctness of a partition. since the exponential distance norm can converge to a local optimum. Furthermore. The main drawback of this validity measure is the lack of direct connection to the data itself. which are described below: • Partition Coefficient (PC): measures the amount of ”overlapping” between clusters. It is defined by Bezdek [5] as follows: 1 P C(c) = N c N (uij )2 i=1 j=1 (3. The main drawback of this algorithm is the robustness. sizes and densities and that the clusters are not constrained in volume. however. Therefore. • Cluster the data for different values of c and validate the correctness of the obtained clusters with validation measures.35) where uij is the membership of data point j in cluster i. The optimal number of clusters can be found by the maximum value.36) 31 . A clustering algorithm is designed to parameterize clusters in a way that it gives the best fit. validation measures has to be designed.

is rated in under bound by the triangle-inequality: d(x. y).y∈Cj d(x. the separation index uses a minimum-distance separation to validate the partitioning.41) minxi ∈Ci . DI(c) = min{ min { i∈c j∈c.• Partition Index (PI): expresses the ratio of the sum of compactness and separation of the clusters. the sum of the value for each individual cluster is used.40) The main disadvantage of the Dunn’s index is the very expansive computational complexity as c and N increase. y) }} maxk∈c {maxx. XB(c) = c i=1 N j=1 (uij )m ||xj − vi ||2 N mini. y)} 32 (3.39) The lowest value of the XB index should indicate the optimal number of clusters. • Alternative Dunn Index (ADI):To simplify the calculation of the Dunn index. • Dunn’s Index (DI): this index was originally designed for the identification of hard partitioning clustering. vj )| were vj represents the cluster center of the j-th cluster.37) P I is mainly used for the comparing of different partitions with the same number of clusters.38) N mini. vj ) − d(x.42) . y) ≥ |d(y. c P I(c) = i=1 N j=1 (uij )m ||xj − vi ||2 c k=1 Ni ||vk − vi ||2 (3. vj )| }} maxk∈c {maxx.j ||xj − vi ||2 (3. ADI(c) = min{ min { i∈c j∈c. the result of the clustering has to be recalculated.y∈C d(x.y∈C d(x. the Alternative Dunn Index was designed.i=j (3.i=j minx∈Ci .k ||vk − vi ||2 • Xie and Beni’s Index (XB): is a method to quantify the ratio of the total variation within the clusters and the separations of the clusters [3]. This value is normalized by dividing it by the fuzzy cardinality of the cluster. c N 2 2 i=1 j=1 (uij ) ||xj − vi || SI(c) = (3.xj ∈Cj |d(y.y∈Cj d(x. Therefor. Each individual cluster is measured with the cluster validation method. measured with minx∈Ci . To receive the Partition index. This will be the case when the dissimilarity between two clusters. y)} (3. A minor value of a SC means a better partitioning. • Separation Index (SI): in contrast with the partition index (PI). vj ) − d(xi .

the three mapping methods will be used for the visualization of the clustering results. • The Sammon mapping method aims to find in a high n-dimensional space N points in a lower q-dimensional subspace. However. the used data set is a highdimensional data set. To achieve this. such in a way the inter point distances correspond to the distances measured in the n-dimensional space. To avoid these problems of the Sammon mapping method. that the Partition Coefficient and the Classification Entropy are only useful for fuzzy partitioned clustering. This implies that the Sammon mapping only can be applied on clustering algorithms that use the Euclidean distance norm during the calculations of the clusters. it is useful to visualize the data and the results. The three visualisation methods will be explained in more detail in the following subsections. This is caused by the repartitioning of the results with the hard partition method. this report will focus on the Sammon mapping method. In case of fuzzy clusters the values of the Dunn’s Index and the Alternative Dunn Index are not reliable. a standard and a most widely method to map high-dimensional data into a lower dimensional space. A draw back of this Fuzzy Sammon mapping is the loose of precision in distance. the Sammon mapping application has two main drawbacks: • Sammon mapping is a projection method. which can not be plotted and visualized directly. The first principal component represents 33 . is used during this research. a computational expensive algorithm is needed. This section describes three methods that can map the data points into a lower dimensional space. which is based on the preservation of the Euclidean inter point distance norm. called the Fuzzy Sammon mapping. In this research. 3.Note.1 Principal Component Analysis Principal component analysis (PCA) include a mathematical procedure that maps a number of correlated variables into a smaller set of uncorrelated variables. The first method is the Principal Component Analysis (PCA). However. This kind of mapping of distances is much closer related to the proposition of clustering than saving the variances (which will be done by PCA). Then.4 Visualization To understand the data and the results of the clustering methods. because in every iteration step a computation of N (N − 1)/2 distances is required. a modified algorithm. 3. called the principal components. since only the distance between the data points and the cluster centers considered to be important. The advantage of the Sammon mapping is the ability to preserve inter pattern distances.4.

a minimization criterion of the error: E= where λ is a constant: N −1 N 1 λ N −1 i=1 (dij − d∗ )2 ij . The inter point distance measure of the n-dimensional space. (3. The eigenvalue associated with the second largest eigenvalue correspond to the second principal component. In this research.47) λ= i<j dij = i=1 j=i+1 dij . The direction of the first principal component is diverted from the eigenvector with the largest eigenvalue. Principal Component Analysis is based on the projection of ¯ correlated high-dimensional data onto a hyperplane [3]. (3.46) 3.q .q Λi. • Discovering and/or reducing the dimensionality of a data set.2 Sammon mapping As mentioned before. which can be defined as follows: yi.as much of the variability in the data as possible.4. given by d∗ = d∗ (yi . Furthermore.the Sammon mapping uses inter point distance measures to find N points in a q-dimensional data space. The main goals of the PCA method are: • Identifying new meaningful underlying variables. defined by dij = d(xi . xj ) correspond to the inter point distances in the q-dimensional space. the second objective is used.k = Wi−1 (xk ) = WiT (xk ).43) where v = xk . N (3. which are representative for a higher n-dimensional data set. etc. This methods uses only the first q nonzero eigenvalues and the corresponding eigenvectors of the covariance matrix: Fi = Ui Λi UiT . dij j=i+1 N (3.44) With Λi as a matrix that contains the eigenvalues λi. In a mathematical way.45) The weight matrix Wi contains the q principal orthonormal axes in its column: 1 2 Wi = Ui. yj ). The succeeding components describe the remaining variability. the covariance matrix of the data set can be described by: F = 1 (xk − v)(xk − v)T .48) 34 .j of Fi in its diagonal in decreasing order and Ui is a matrix containing the eigenvectors corresponding to the eigenvalues in its columns. (3. there is a q-dimensional reduced vector that represents the vector xk of X. the principal components will be achieved by analyzing the eigenvectors and eigenvalues. This is achieved by Sammon’s ij stress. (3. In this case.

52) with d(xk . 2. it is not possible to reach a local minimum in the error surface. The Fuzzy Sammon mapping algorithm is similar to the original Sammon mapping. with i ∈ {1..k=i ∂ 2 E(t) 2 =− 2 ∂yil (t) λ N k=1. q} which implies that yi = [yi1 . independently to the shape of the original cluster. in a projected two dimensional space every cluster is represented by a single point. but in this case the projected cluster 35 . Sammon’s mapping has several drawbacks. 2. since a constant does not change the result of the optimization process. N } and l ∈ {1. vi ) representing the distance between data point xk and the cluster center vi in the original n-dimensional space. yiq ]T . ki (3. weighted by the membership values similarly to equation (3. The minimization of the error E is an optimization problem in the N xq variables yil .19): c N Ef uzz = i=1 k=1 (µki )m (d(xk .. . called Fuzzy Sammon mapping.. with a recommended value α 0..Note that there is no need to maintain λ. 2 ∂yil (t) ∂E(t) ∂y (t) (3. .50) ∂yil (t) λ dki d∗ ki k=1. ..49) where α is a nonnegative scalar constant.. it is possible to estimate the correct initialization based on the information which is obtained from the data. zi ). 3. This scalar constant represents the step size for gradient search in the direction of N 2 dki − d∗ ∂E(t) ki =− (yil − ykl ) (3. The modified algorithm.. vi ) − d∗ )2 .. To avoid this drawbacks. while searching for the minimum of E. This is a disadvantage. a modified mapping method is designed which takes into account the basic properties of fuzzy clustering algorithms where only the distance between the data points and the clustering centers are considered to be important [3].3 − 0. The rating of yil at the t-th iteration can defined by:   il yil (t + 1) = yil (t) − α  ∂ 2 E(t)  .3 Fuzzy Sammon mapping As mentioned in the introduction of this section. According to this information. However. because multiple experiments with different random initializations are necessary to find the minimum.k=i 1 (dki − d∗ ) − ki dki d∗ ki (yil − ykl )2 d∗ ki 1+ dki − d∗ ki dki (3.51) With this gradient-descent method. The Euclidean distance between the cluster center zi and the data yk of the projected q-dimensional space is represented by d∗ (yk .4.. uses only N ∗c distances.4.

center will be recalculated in every iteration after the adaption of the projected data points. The recalculation will be based on the weighted mean formula of the fuzzy clustering algorithms, described in Section 3.2.3 (equation 3.19). The membership values of the projected data can be plotted based on the standard equation for the calculation of the membership values: µ∗ = ki
c j=1

1
d∗ (xk ,ηi) d∗ (xk ,vj )
2 m−1

,

(3.53)

where U ∗ = [µ∗ ] is the partition matrix with the recalculated memberships. ki The plot will only give an approximation of the high dimensional clustering in a two dimensional space. To measure the quality of this rating, an evaluation function that determines the mean square error between the original and the recalculated error can be defined as follows: P = ||U − U ∗ ||. (3.54)

In the next chapter, the cluster algorithms will be tested and evaluated. The PCA and the (Fuzzy) Sammon mapping methods will be used to visualize the data and the clusters.

36

Chapter 4

Experiments and results of customer segmentation
In this chapter, the cluster algorithms will be tested and their performance will be measured with the proposed validation methods of the previous chapter. The best working cluster method will be used to determine the segments. The chapter ends with an evaluation of the segments.

4.1

Determining the optimal number of clusters

The disadvantage of the proposed cluster algorithms is the number of clusters that has to be given in advance. In this research the number of clusters is not known. Therefor, the optimal number of clusters has to be searched with the given validation methods of Section 3.3. For each algorithm, calculations for each cluster, c ∈ [215], were executed. To find the optimal number of clusters, a process called Elbow Criterion is used. The elbow criterion is a common rule of thumb to determine what number of clusters should be chosen. The elbow criterion says that one should choose a number of clusters so that adding another cluster does not add sufficient information. More precisely, by graphing a validation measure explained by the clusters against the number of clusters, the first clusters will add much information (explain a lot of variance), but at some point the marginal gain will drop, giving an angle in the graph (the elbow). Unfortunately, this elbow can not always be unambiguously identified. To demonstrate the working of the elbow criterion, the feature values that represent the call behavior of the customers, as described in Section 2.1.2, are used as input for the cluster algorithms. From the 800,000 business customers of Vodafone, 25,000 customers were randomly selected for the experiments. More customers would lead to computational problems. First, the K-means algorithm will be evaluated. The values of the validation methods depending on the number of clusters will be plotted. The value of the Partition Coefficient is for all 37

clusters 1, and the classification entropy is always ’NaN’. This is caused by the fact that these 2 measures were designed for fuzzy partitioning methods, and in this case the hard partitioning algorithm K-means is used. In Figure 4.1, the values of the Partition Index, Separation Index and Xie and Beni’s Index are shown. Mention again, that no validation index is reliable only by itself.

Figure 4.1: Values of Partition Index, Separation Index and the Xie Beni Index Therefor, all the validation indexes are shown. The optimum could differ by using different validation methods. This means that the optimum only could be detected by the comparison of all the results. To find the optimal number of cluster, partitions with less clusters are considered better, when the difference between the values of the validation measure are small. Figure 4.1 shows that for the PI and SI, the number of clusters easily could be rated to 4. For the Xie and Beni index, this is much harder. The elbow can be found at c = 3, c = 6, c = 9 or c = 13, depending on the definition and parameters of an elbow. In Figure 4.2 there are more informative plots shown. The Dunn’s index and the Alternative Dunn’s index confirm that the optimal number of clusters for the K-means algorithm should be chosen to 4. The values of all the validation measures for the K-means algorithm, are embraced in table 4.1

38

0000 5 1.0001 3.0071 0.9519 0.0443 0.0005 5.8758 0.0063 0.0000 N aN 0.7557 0.2907 0.2: Values of Dunn’s Index and the Alternative Dunn Index c PC CE PI SI XBI DI ADI c PC CE PI SI XBI DI ADI 2 1.0000 N aN 3.0000 N aN 0.9386 0.8318 0.0001 3.3998 0.0002 4.0002 3.5737 0.4379 0.7489 0.0000 4 1.0000 7 1.8620 0.0072 0.0002 4.0000 N aN 1.0002 3.0082 0.0002 3.8384 0.1: The values of all the validation measures with K-means clustering 39 .0002 11 1.8362 0.0000 N aN 1.8080 0.0002 3.7225 0.0052 0.0041 0.Figure 4.0061 0.0034 0.0000 N aN 0.0003 4.9253 0.0071 0.0000 N aN 1.0000 N aN 0.0061 0.0000 N aN 0.0013 10 1.0001 13 1.9079 0.9109 0.0034 0.7696 0.0001 14 1.8828 0.3353 0.0001 12 1.0002 3.0000 N aN 0.0061 0.8261 0.0000 6 1.0000 N aN 0.0000 8 1.4626 0.0018 9 1.7783 0.0000 N aN 1.0000 Table 4.0001 3.0000 N aN 0.0065 0.2214 0.0000 15 1.0000 N aN 0.0000 3 1.0070 0.1571 0.0061 0.0002 5.0001 3.

has an elbow at the point c = 3.the optimal number of clusters is chosen at c = 4. On the other hand. In Figure 4. Compared to the hard clustering methods. for the XBI.3: Values of Partition coefficient and Classification Entropy with Gustafson-Kessel clustering ters. caused by the lack of direct connection to the data. the local minimum is reached at c = 6. the main drawback of PC is the monotonic decreasing with c. The same problem holds for CE: monotonic increasing. The points at c = 3.2. for the Alternative Dunn Index is not known how reliable its results are. the Alternative Dunn index. Again.It is also possible to define the optimal numbers of clusters for fuzzy clustering algorithms with this method. The optimal number of cluster can not be rated based on those two validation methods. This process can be repeated for all other cluster algorithms. the Dunn index also indicates that the optimal number of clusters should be at c = 6. the optimal number of clusters is located at c = 6. However. the validation methods can be used now for the fuzzy clustering. so the optimal number of clusters for the Gustafson-Kessel algorithm will be six. which makes it hardly to detect the optimal number of cluster. For the other algorithms. can be seen as an elbow. The results of the validation measures for the Gustafson-Kessel algorithm are written in table 4. The results can be found in Appendix B.5. 40 . Figure 4. For the PI and the SI. K-medoid and the Gath-Geva. it is difficult to find the optimal number of clusters. the results of the GustafsonKessel algorithm will be shown.3 the results of the Partition Index and the Classification Entropy are plotted. To illustrate this. For the K-means. In Figure 4.4 gives more information about the optimal number of clus- Figure 4. c = 6 and c = 11. However.

4: Values of Partition Index.5: Values of Dunn’s Index and Alternative Dunn Index with GustafsonKessel clustering 41 . Separation Index and the Xie Beni Index with Gustafson-Kessel clustering Figure 4.Figure 4.

3046 0. the optimal number of clusters was found at c = 4 or c = 6.0002 2.0007 13 0.0009 15 0.0000 7 0.7293 0.4183 0.5512 0.5675 0.6882 0.2: The values of all the validation measures with Gustafson-Kessel clustering 4.0063 10 0.0039 00030 K-means K-medoid FCM GK GG Table 4.0001 0.8128 0. The validation measures can also be used to compare the different cluster methods.0063 0.0102 0.0017 0.2366 0.3863 1.3 PC 1 1 0.0083 0.0853 0.0084 0. To avoid visibility problems (plotting too much values will cause one 42 .0041 0.0007 1. With these visualization methods.4183 1. as mentioned in the previous section.0062 0.0007 0.0027 0.9019 0.0002 4 0.0001 0.3550 0.2 Comparing the clustering algorithms The optimal number of cluster can be determined with the validation methods.4.2489 1.7813 0.2800 0.1149 2. To visualize the clustering results.0009 1.6620 0.1611 2.1479 2.0001 SI 0.4293 0.2057 0.9364 0.2024 1.0002 0.1469 2.0029 ADI 0.5603 0.0006 0.0039 0.0644 DI 0. depending on the clustering algorithm.9205 0.0003 3 0. the dataset can be reduced to a 2-dimensional space.0028 0. On the score of the values of the three most used indexes. The validation measures for c = 4 and c = 6 of all the clustering methods are collected in the tables 4.5930 0.0039 11 0.0004 1. Xie and Beni’s index and Dunn’s index.2066 1.9305 0.0002 6 0.9203 0.0092 0.0002 0.2189 0.0001 0.7575 0.0082 0.5303 0. As examined in the previous section.5085 0.0003 1.2737 0.0015 0.7797 0.0000 Table 4.0004 5 0.5547 0.4982 CE NaN NaN 1.0034 0.3: The numerical values of validation measures for c = 4 and 4.0001 0.3500 0.7688 0.6462 0.0030 0.0001 0.0263 9 0.7447 0.0002 0. Separation index.c PC CE PI SI XBI DI ADI c PC CE PI SI XBI DI ADI 2 0.5930 0.0001 0.0002 0.0046 0.4684 0. one can conclude that for c = 4 the Gath-Geva algorithm has the best results and for c = 6 the Gustafson-Kessel algorithm.5131 0.0009 1.8536 0.8903 0.5034 PI 1.4 can be used. the validation methods that are described in Section 3.1573 0.3983 0.3044 1.8218 1.4 show that the PC and CE are useless for the hard clustering methods K-means and K-medoid.0001 8 0.0001 14 0.1571 0.0029 0.0018 12 0.3 and 4.0852 0.0001 XBI 5.0002 1.1410 2.1702 2.0001 42.0034 Inf 1.0012 0.7233 0.3209 1.0083 0.5978 0.0002 0.0002 0.0002 0.2741 1.5819 0.3983 1. Table 4.0867 1.7149 0.

0457 DI 0. only 500 values (representing 500 customers) from this 2-dimensional dataset will be randomly picked.2907 0. By a detailed look at the plot. Note that the cluster in the left bottom corner and the cluster in the 43 . This implies that the Fuzzy C-means algorithm is not able to find good clusters for this data set.0007 0. there are only 2 clusters clearly visible.1238 0. the clusters are well separated. For the situation with 4 clusters. For the other cluster algorithms.0029 0.9. one can see that there are actually 4 cluster centers. but the cluster centers are almost situated on the same location.0063 0.6: Result of K-means algorithm for the clustering problem.3044 0.x show the different clustering results for c = 4 and c = 6 on the data set. For both situations.7918 1. Figures 4.4293 1.K-means K-medoid FCM GK GG PC 1 1 0.x4.0008 0.0001 0.9203 1. The plot of the Fuzzy C-means algorithm.6 and 4.1667 0.0001 0.3773 CE NaN NaN 1.0001 19. None of the clusters contain sufficient more or less customers than other clusters. Figure 4. For the K-means and the K-medoid algorithm.0008 XBI 3.4: The numerical values of validation measures for c = 6 big cloud of data points).0099 ADI 0.0009 Table 4. the Sammon’s mapping gives the best visualization of the results. in Figure 4. the Fuzzy Sammon’s mapping visualization gives the best projection with respect to the partitions of the data set. with one small cluster in one of the big clusters.8903 0. In Figure 4.7 show that hard clustering methods can find a solution Figure 4.1043 SI 0. shows unexpected results.0008 0.0102 0. the results of the Gustafon-Kessel algorithm are plotted.6490 PI 1. These visualization methods are used for the following plots. one can see three big cluster. In the situation with 6 clusters.0002 0.9253 Inf 0.0070 0.8.4613 0. The other two cluster centers are nearly invisible.0001 0.9245 0.

9: Result of Gustafson-Kessel algorithm 44 .8: Result of Fuzzy C-means algorithm Figure 4.Figure 4.7: Result of K-medoid algorithm Figure 4.

The fact that this is the case in the two-dimensional plot. for the situation c = 4 look similar to the result of the Gustafson-Kessel algorithm. one can not distinguish between the two cluster algorithms. the results show that the clusters are homogeneous. In the next section. 45 .10. In both situations. one can look at the distances from the points to each cluster.11 and 4.top right corner in Figure 4. In the real high-dimensional situation. The result for the c = 6 situation is remarkable.10: Result of Gath-Geva algorithm of the Gath-Geva algorithm. indicates that a clustering with six clusters with the Gustafson-Kessel algorithms not a good solution.12. visualized in Figure 4. The box indicates the upper and lower quartiles. The results Figure 4. based on the distances to the cluster. To determine which partitioning will be used to define the segments. two box plots of the distances from the data points to the cluster are plotted. This may indicate that the data points in these clusters represents customers that differ on multiple fields with the other customers of Vodafone.9 are also maintained in the situation with 6 clusters. but are separated. the two different partitions will be closely compared with each other.3 Designing the segments To define which clustering method will be used for the segmentation. The result is visible for the Gath-Geva algoithm for c = 4 and for the Gustafson-Kessel algorithm with six clusters. Another way to view the differences between the cluster methods is to profile the clusters. a closer look to the meaning of the clusters will be needed. In Figure 4. With the results of the validation methods and the visualization of the clustering. This indicates that. a profile can be made by drawing a line between all normalized feature values (each feature value is represented at the x-as) of the customers within this cluster. Here are also appearing clusters in other clusters. 4. one can conclude that there are two possible best solutions: The Gath-Geva algorithm for c = 4 and the Gustafson-Kessel algorithm for c = 6. the clusters are not a subset of each other. For each cluster.

11: Distribution of distances from cluster centers within clusters for the Gath-Geva algorithm with c = 4 Figure 4.Figure 4.12: Distribution of distances from cluster centers within clusters for the Gustafson-Kessel algorithm with c = 6 46 .

Figure 4.The profiles of the different clusters do not differ much in shape. However. at least one value differs sufficient from the values of the other cluster. This means that the customers in one profile contain similar values of the feature values. This confirms the assumption that customers of different clusters have indeed a different usage behavior. Most of the lines in one profile are drawn closely together. in each cluster.13: Cluster profiles for c = 4 47 .

Figure 4.14: Cluster profiles for c = 6 48 .

The mean of all the lines (equivalent to the cluster center) was calculated and a line between all the (normalized) feature vales was drawn. Cluster 2 has high values at position 6 and 9.More relevant plots are shown in Figure 4.15: Cluster profiles of centers for c = 4 49 . For instance. have a high feature value at feature 8. Figure 4. The difference between the clusters are visible by some feature values. compared with other cluster. while Cluster 3 contains peaks at features 2 and 12. Cluster 1 has customers. The 4th and final cluster has high values at feature 8 and 9.15 and ??. in the situation with four clusters.

Figure 4.16: Cluster profiles of centers for c = 6 50 .

0 71.2%) Segment 1 (18.7 66.2 65.1 78.8 2.0 1.4 4.9 1.4 6.1 4.8%) Segment 6 (16.4 2.3 87.9 4.0 86.6 88.3 17.9 66.5 4. • Segment 2: This segment contains customers with an average voice call 51 .9%) Segment 4 (20.4%) Segment 3 (18.9 13.3 12.8 96.1 22. The feature Feature Average Segment 1 (27.1 8 3.0 1.6 1.4 26.1%) Segment 2 (14.7 17.3 65. Their sms usage is higher then normal. For the Gath-Geva algorithm with c = 4 and the Gustafson-Kessel algorithm with c = 6.8%) Feature Average Segment 1 (27.8 121.4 1.1 132.1 2.6 2. the segments can be described as follows: For the situation with 4 segments: • Segment 1: In this segment are customers with a relative low number of voice calls.3 7 1.7 4. 11 the unique are codes and feature 12 the number of different numbers called).8 72.1 11.0 3.6 126.8%) 1 119.9 4.2%) Segment 2 (28.1 86.7 1.6 4.2 1.2 93.8 4.8 15.0 12.4%) Segment 3 (18.7 58.7%) Segment 3 (23.4 5.5 91.5 0.9 6.4 2 1.1 3.7%) Segment 3 (23.1 2. validation measures and plots.4 84.4 12.0 65. In words.1 2.6 87.7 2.9 4 65. table 4.6 1.3%) Segment 4 (17.With the previous clustering results.1.0 12 25.8 155.6 6.2 6.5 3.9 3. feature 9 the international calls.2%) Segment 1 (18.3 4.9 24.0 86.6 1.8 1.7 87.5 6.4 14.3%) Segment 4 (17.5 71.7 121.6 39.2 92.0 3. feature 10 the calls to Vodafone mobiles.1%) Segment 2 (14.6 3 3.9 2.1 9.5 40.0 23. both results will be used as a solution for the customer segmentation. feature 4 the daytime calls. Therefor. 7 received sms.8%) Segment 6 (16.3 1.1 74.9 1.7 10 14. The number of international calls is low. feature 5 the weekday calls.5: Segmentation results numbers correspond to the feature numbers of Section 2. This customers call more in the evening (in proportion) and to fixed lines then other customers.9 1.5 shows the result of the customer segmentation. it is not possible to decide which of the two clustering methods gives a better result.8 94.6 2.7 72.6%) Segment 5 (14.9 2.9%) Segment 4 (20.1 3.3 120.2%) Segment 2 (28.1 6. (Feature 1 is the call duration.8 73.7 2. feature 2 the received voices calls and feature 3 the originated calls.7 1.4 10.2 1.2 6.6 73.7 c=4 c=6 c=4 c=6 Table 4.1 12.2.8 2.9 6 75.7 0.1 15.1 9 2.3 88.7 65.7 2.4 11 6.9 73.8 2.8 5 87.8 133.9 3.8 54. 8 originated sms.8 22.6 9.2 1.6 87.6 60.2 30.6%) Segment 5 (14.5 1. 6 are calls to mobile phones.4 4.

• Segment 4: These customers originate many voice calls. Their average call duration is also lower than average.3). For the situation with 6 segments. • Segment 6: These customers originate and receive many voice calls. Their call duration is high. They have a relative small number of contacts. they make more international phone calls than other customers. In proportion. 52 . The percentage of international calls is high. • Segment 5: These customers do not receive many voice calls. They also send and receive many sms messages. the customers in this segments can be described as follows: • Segment 1: In this segment are customers with a relative low number of voice calls. The duration of their voice calls is longer than average. In the next session the classification method Support Vector Machine will be explained. • Segment 3: The customers in this segment make relative many voice calls. They also receive and originate a low number of sms messages. The average call duration is low. However.1. These customers do not call to many different numbers. They also call to many different areas. None of the feature values is high or low. their sms usage is relative high. They have also more contacts with a Vodafone mobile. They call often during daytime and call more then average to international numbers. Their sms usage is low. This technique will be used to classify/estimate the segment of a customer by personal information as age. Remarkable is the fact that they don not have so many contacts as the number of calls do suspect. gender and lifestyle (the customer data of Section 2. • Segment 3: The customers in this segment make relative many voice calls. These customers call to many different numbers and have a lot of contacts which are Vodafone customers. • Segment 2: This segment contains customers with a relative high number of contacts. They do not send and receive many sms messages.usage. • Segment 4: These customers are the average customers. They also send and receive many sms messages. They call often to mobile phones during day time.

age and income. 5. The goal of the SVM is learn to tell the difference between the groups and.1 The separating hyperplane A human being is very good at pattern recognition. geometric interpretation of the data. Subsequently. such as the one labeled ’Unknown’ in Figure 5. which can be easily plotted. In this case the customer data consist of 2 feature values.1. given an unlabeled customer. the basic ideas of Support Vector Machines can be explained without any equations. a Support Vector Machine is a mathematical entity.1a shows that the green dots form a group and the reds dots form another group that can easily be separated by drawing a line between the two groups (Figure 5.1b).Chapter 5 Support Vector Machines A Support Vector Machine is a algorithm that learns by example to assign labels to objects [16]. In this research a Support Vector machine will be used to recognize the segment of a customer by examining thousands of customers (e. imagine that there exists only two segments. The next few sections will describe the four basic concepts: • The separating hyper plane • The maximum-margin hyperplane • The soft margin • The kernel function For now. the customer data features of Section 2. In general. However. an algorithm for maximizing a particular mathematical function with respect to a given collection of data. Even a quick glance at Figure 5.1. predicting the label of an unknown customer is simple: one simply needs to ask whether the new customer falls on the segment 53 . The green dots represent the customers that are in segment 1 and the red dots are customers that are in segment 2. predict whether it corresponds to segment 1 or segment 2.3) of each segment. to allow an easy.g.

2: Separating hyperplanes in different dimensions 54 . So the term separating hyperplane is. consider the situation where there are not just two feature values to describe the customer. a straight line divides the space in half (remember Figure 5. The term for a straight line in a high-dimensional space is a hyperplane.2a). essentially. For example. a plane is needed to divide the space. the line that separates the segments. illustrated in Figure 5.2b. This procedure can be extrapolated mathematically in higher dimensions.1b) In a three-dimensional space. This line can be divided in half by using a single point (see Figure 5. then the space in which the corresponding onedimensional feature resides is a one-dimensional line. Now.1: Two-dimensional customer data of segment 1 and segment 2 1 or the segment 2 side of the separating line. In two dimensions.(a) Two-dimensional representation of the customers (b) A separating hyperplane Figure 5. to define the notion of a separating hyperplane. if there was just 1 feature value to describe the customer. (a) One dimension (b) Three dimensions Figure 5.

3: Demonstration of the maximum-margin hyperplane to the success of Support Vector Machines. However. However. since it is not reasonable that a Support Vector machine trained on customer data is able to classify different car types.2 The maximum-margin hyperplane The concept of treating objects as points in a high-dimensional space and finding a line that separates them.1a The goal of SVM is to find a line that separates the segment 1 customers from the segment 2 customers. More relevantly. the line that separates the two segments and adopts the maximal distance from any of the given customers (see Figure 5. For example. In other words. a SVM 55 . By selecting this hyper plane. is in many ways. there are an infinite number of possible lines.5. the SVM selects the maximum separating hyperplane. the key (a) Many possibilities (b) The maximum-margin hyperplane Figure 5. This is of course logical. Consider again the classification problem of Figure 5. the theorem of a SVM indicates that the two data sets has to be drawn from the same distribution.2 The question is which line should be chosen as the optimal classifier and how should the optimal line be defined. as portrayed in Figure 5. roughly speaking. there are a some remarks and caveats to deal with.1a. is selecting the line that is. A logical way of selecting the optimal line.2). It is not surprising that a theorem of the statistical learning theory is supporting this choice [6]. This theorem. is a common way of classification. the SVM is able to predict the unknown segment of the customer in Figure 5. the theorem is based on the fact that the data on which the SVM is trained are drawn from the same distribution as the data it has to classify. First at all. By defining the distance from the hyperplane to the nearest customer (in general an expression vector) as the margin of the hyperplane. However. the SVM differs from all other classifier methods by virtue of how the hyperplane should be selected. On the other hand. The vectors (points) that constrain the width of the margin are the support vectors. and therefore not unique to the SVM. it is not reasonable to expect that the SVM can classify well if the training data set is prepared with a different protocol then the test data set. ’in the middle’.

However. the theory assumed that the data can be separated by a straight line. instead of a two-dimensional data set. a SVM should not allow too many misclassification.4a. In this figure. many real data sets are not cleanly separable by a straight line. the example data will be simplified even further. roughly.does not assume that the data is drawn from a normal distribution. 5. 5.4 The kernel functions To understand the notion of a kernel function. A intuitively way to deal with the problems of errors is designing the SVM in such a way that it allows a few anomalous customers to fall on the ’wrong side’ of separation line.3 The customer can be seen as an outlier and resides on the same side of the line with customers of segment 1. In other words. Of course.4: Demonstration of the soft margin with the introduction of the soft margin. that (a) Data set containing one error (b) Separating with soft margin Figure 5.4a will be separated in the way it is illustrated in Figure 5. Note. The soft margin allows a small percentage of the data points to push their way through the margin of the separating hyperplane without affecting the final result. by the fact that a large margin will be achieved with respect to the number of correct classifications. With the soft margin. controls the number of customers that is allowed to violate the separation line and determines how far across the line they are allowed. for example the data of Figure 5. the soft margin specifies a trade-off between hyper plane violations and the size of the margin.3 The soft margin So far. a user-specified parameter is involved that controls the soft margin and. there 56 . This can be achieved by adding a ’soft margin’ to the SVM. Assume that. the data contains an error object. the data set of Figure 5. Setting this parameter is a complicated process.

consider the situation of Figure 5.4.4. In general.1. So. In Figure 5. In that case. (a) None separable dataset (b) Separating previously non separable dataset Figure 5.4 is plotted a two-dimensional data set. To understand kernels better. In Figure 5. Now. the data set must contain consistent labels. in this case by squaring the one dimensional data set. the number of possible solutions also increases. The figure contains the same data as Figure 5. the SVM should be a perfect classifier. which means that two identical data points may not have different labels. but with a projection of the SVM hyperplane in the four-dimensional space back down to the original two-dimensional space. but the projected hyperplane is found by a very high dimen57 . If one chooses a good kernel function. it is possible to prove that for any data set exists a kernel function that allows the SVM to separate the data linearly in a higher dimension. the SVM can separate the data in two segments by one straight line. However. Consequently.4 the situation is drawn when the data is project into a space with too many dimensions. which illustrates an non separable data set. the kernel function can be seen as a mathematical trick for the SVM to project data from a low-dimensional space to a space of higher dimensions. Of course. but exponentially. the result is shown as the curved line in Figure 5.4. The result is plotted in Figure 5. No single point can separate the two segments and introducing a soft margin would not help. this data can be projected to a four-dimensional space. the separating hyperplane was a single point. as shown in the figure. the first problem is the so called curse of dimensionality: as the numbers of variables under consideration increases. in theory. the data will become separable in the corresponding higher dimension. A kernel function provides a solution to this problem. The kernel function adds an extra dimension to the data. as seen before in Figure 5. it becomes harder for any algorithm to find a correct solution. there are some drawbacks of projecting data in a very high-dimensional space to find the separating hyperplane.5: Demonstration of kernels some extra examples will be given. It is not possible to draw the data in the 4 dimensional space. Within the new higher dimensional space. With a relative simple kernel function.4.is a one-dimensional data set.

(5. trial and error. In general the kernel function is defined by: K(xi . It is more likely that there exists a kernel function that was not tested and performs better than the selected kernel function. However. i 58 (5. In this research a SVM will be experimented with a variety of ’standard’ kernel functions. The SVM will not function well on new unseen unlabeled data. is actually the best kernel function that exists.2) . This results in boundaries which are to specific to the examples of the data set. xj ) = (γxT xj + c0 )d . • Linear: which function is defined by: K(xi . probably an infinite number. i • Polynomial: the polynomial kernel of degree d is of the form K(xi . This phenomenon is called over fitting. The vectors are mapped into a higher dimensional space by the function Φ. Unfortunately.1) where xi are the training vectors. xj ) = Φ(xi )T Φ(xj ). There exists another large practi- (a) Linearly separable in four dimensions (b) A SVM that has over fit the data Figure 5. Many kernel mapping functions can be used.6: Examples of separation with kernels cal difficulty when applying new unseen data to the SVM. in most cases. the method described above. the answer too this question is. mainly gives sufficient results. this is a time-consuming process and it is not guaranteed that the best kernel function that was found during cross-validation. but without introducing too many irrelevant dimensions. Practically.sional kernel. The default and recommended kernel functions were used during this research and will be discussed now. This problems relies on the question how to choose a kernel function that separates the data. By using the cross-validation method. xj ) = xT xj .3) (5. but a few kernel functions have been found to work well in for a wide variety of applications [16]. the optimal kernel will be selected on a statistical way.

if the SVM has to recognize three classes. A. xj ) = tanh(γxT xj + c0 ). In this research the constant c0 is set to 1. (5. (5. one-versus-all classifiers. ”Is it B?” and ”Is it C?”. In this research the one-verses-one technique will be used. Another simple approach is the one-versus-one where k(k − 1)/2 models are constructed. ”Is it A?”.5 Multi class classification So far. which is also used in neural networks.5) i When the sigmoid function is used. but two methods are the most popular and most used [16].7: A separation of classes with complex boundaries 5. one can regard it with a as a two-layer neural network. is defined by K(xi . B and C.7 Figure 5. How does a SVM discriminate between a large variety of classes. the idea of using a hyperplane to separate the feature vectors into two groups was described.4) • Sigmoid: the sigmoid function. one can simply train three separate SVM to answer the binary questions. but only for two target categories. where k is the number of classes. The first approach is to train multiple. xj ) = exp(−γ||xi − xj ||2 ). The concept of a kernel mapping function is very powerful.• Radial basis function: also known as the Gaussian kernel is of the form K(xi . 59 . For example. It allows a SVM to perform separations even with very complex boundaries as shown in Figure 5. as in our case 4 or 6 segments? There are several approaches proposed.

cross-validation is used to evaluate the fitting provided by each parameter value set tried during the experiments.Chapter 6 Experiments and results of classifying the customer segments 6. With the validation set. the actual performance of the SVM will be measured after the SVM is trained. The training of the SVM will be stopped when the test error reached a local 60 . the training set. The test set will be used to estimate the error during the training of the SVM.1: Under fitting and over fitting divided into two groups. By K-fold cross validation the training dataset will be Figure 6. Figure 6. the test set and the validation set.1 K-fold cross validation To avoid over fitting. Different parameter values may cause under or over fitting. The training set will be used to train the SVM.1 demonstrates how important the training process is.

The linear Kernel function itself has no parameters. a k-fold partition of the Figure 6.1% 2 42. In table 6.2 Parameter setting In this section.6% 5 43. denoted by C.0% 10 43.2% 20 43.2 the results for the different C-values are summarized. In this Figure 6. see Figure 6. 4 segments 61 . Figure 6.0% 50 42.3 illustrates this process. By K-fold cross validation. For the situation with 4 clusters. the optimal parameters for the Support Vector Machine will be researched and examined.2. the C 1 42.minimum. K is set to 10. The only parameter that can be researched is the soft margin value of the Support Vector Machin. Each kernel function with its parameters will be tested on their performance.1 and table 6. K-1 folds will be used for training and the remaining one for testing.2: Determining the stopping point of training the SVM data set is created.7% 200 40. For each of K experiments. 6.1: Linear Kernel. The advantage of K-fold cross validation is that all the examples in the dataset are eventually used for both training and testing.3: A K-fold partition of the dataset research. The error is calculated by taking the average off all K experiments.1% Table 6.4% 100 41.8% 500 36.

9% 10 31.0% Table 6.6% 2 74. For the situation with γ γ γ γ γ γ d = 0.6% 200 42. 4 segments γ γ γ γ γ γ d = 0.8% 73.8% 75. These C-values are used to find out which d and γ give the C 1 73.8% 74.9% Table 6.1% 75.3% 10 75.8% 75.4% 77.2% 2 76.6% 74.4% 74.0% 73.8% 74.2% 75.2: Linear Kernel.4 1 76.2% 76.0 = 1.1% 2 74.3% 20 31.4 1 75.0% 75.6.2% 76.5 and 6.3% 76.1% 76.0% 75.8% 74.0% 76. 6 segments 62 .0 = 1.1% 100 50.2% 73.8% 100 70. the optimal number for the maximal margin will be determined.5% 50 72.7% 500 26.3% 7 73.4% 75.2% 3 78.7% Table 6.6: Polynomial kernel.1% 20 75.1% 75.1% 74.8% Table 6.2% 78. The correct number of classifications are respectively. For the polynomial kernel function. denoted by d and the width γ.6% 500 21.2% 500 53.6% 75.4% 6 74.5% 6 76.1% 77. 6 segments optimal value for the soft margin is C = 10 and by using the 6 segments C = 50.1% 4 73.3% 76.6% 10 74.6% 20 73.4% 5 30.9% 5 74.1% 74.6% 77. The average value for each soft margin C can be found in the tables 6.0% 72.5% 74.2% 76.3% 3 75.8% 74. The results are shown in tables 6.3% 74.4 = 0.0% 75.2% 76.0% 75.1% 78. This is done by multiple test runs with random values for d and γ.1% 74.0% 74.8% 75. Therefor.1% 72. 4 segments C 1 70.0%.2% 74.5% 72.9% Table 6.5% 74.2% 74.1% 73.4% 5 75.9% 74.0% 50 75.3% 76.6% 75.3 and 6.4% 5 76.6 = 0.8 = 1.9% 2 29.9% 75.0% 74.8 = 1.5: Polynomial kernel.6% 200 27.9% 74.4% 50 32.4 = 0.4% 76.0% 78. 43.1% 75.0% 100 27. The number of degrees.C 1 28.2% 75.8% 2 77.6% 76.4.3% 7 75. 6 segments best results.0% 74.1% 74.4: Average C-value for polynomial kernel.2% and 32.6 = 0.2 = 1.6% 200 63.2% 4 76.2 = 1.7% 75.8% 76.3: Average C-value for polynomial kernel. there are two parameters.9% 74.4% 76.0% 75.3% 73.0% 76.0% 5 75.0% 75.1% 72.3% 75.

8 = 1. with respectively 4 and 6 segemtents.7 80.9 73.8.4 59.3 30.4 2 79.8 = 1.0 61.6 72.4 200 51.3 200 52.8 76.0 55.6 74.4 74.2 500 37.4 = 0.4 segments.0 Table 6.0 39.9 76. with 6 segments the best score is 78.5 52.2 63.7 54. The results of the Radial Basis function are given in table 6.7 30.3 79.4 79.2 69.2 38.2 2 53.9 78.2 44.6 69.5%.0 42.4 5 57.2 78.2 200 47.4 60. The sigmoid function has also only 1 variable.3 72.6 = 0.9: Sigmoid function. show that there are two clusters which can easily be classified with the customer profile. The C = 0.6 26.5 74.7 74.3% and 78.8 = 1.7 50 73.4 = 0.9 5 76.6 43.4 Table 6.1 71.4 66.9 80.1 63.8 76.10.4 1 80.6 53. The results are given in table 6.7 62. the radial basis function has only one variable. This means that the Radial basis function has the best score for both situations.7 55.2 47. 4 segments correct.7 44.8 80.6 72.5 100 52.9 72.3 73.4 79. table 6.0 41.2 80.5 30.7 61.1 58.0 = 1.6 52.6 52.9 34.5 64.1 51.7 40.9 and 4.11 and 6.2 46.2 = 1.9 50.9 73.3 42.9 44.2 59.1 100 30.7 72.6 = 0.4 500 40.1 48.5%.1 40.5 68.1% and for 6 segments 76.5 60.3 79.2 = 1.6 = 0.8 54.0 71.3 54.3 64. The confusion matrix for both situations.0 73.0 68.6 59.5 78.9 and 6.5 80. by the Sigmoid function. 63 .3 79.1 20 68.0 = 1.2%.2 79.5 52.2 = 1.5 10 78.5 54.0 52. while there are two extra clusters.7 48.7 10 58.10 The results show that 66.4 20 76.0 50 65.1 51.0 80.0 70.1 70.3 66.3 61.1% and 44.5 57.5 38.3 61.7 10 70.0 20 56.4 70.1 77. The following kernel function.9 68.0 γ γ γ γ γ γ Table 6.12.4 = 0.8 73.0 57.1 70.3 59.6 500 38.0 56.7: Radial basis function.4 2 77.0 80.6 68.4 1 58.7 78.6 77.8 61.3 26.0 47. the optimal score is 78.9 45. Remarkable is the fact that the difference is small between the two situations.7 46.8: Radial basis function. namely γ.4 47.0 80.4 76.9 34.1 60.6 57.5 41.1 79.7 and table 6.8 50 57.1 69.5 5 72. 6 segments best result with 4 segments is 80.5 29.3 31.6 51.6 77.6 50.8 65.1 44.4 100 60.3 78.6% of the data is classified γ γ γ γ γ γ C = 0.3 58.0 = 1.8 49.5 27.3%.2 44. 4 segments γ γ γ γ γ γ C = 0.4 66. This corresponds to the cluster in the top right corner and the cluster in the bottom of Figures 4.8 69.2 64.2 79. with 80.4 1 73.

0 41.8 24.1% Segment 2 0.1% 0.7% 4.1 40.0% 12.0 26.3% 6.5% 76.9% 7.8 43.7 43.8% 69. 5 0.7 28.9% Segment 4 0.8 32.6 2 34.0% Segment 3 1.1% 68.8 43. 6 segments 64 .8 29.0% Table 6.0 27.5 5 34.6 34.8 18.8 = 1.5% Segm.8% 13.4% 1.4% Segm.3% 10.5 33.1% 3.1 40.2% 7.8 18.7% Segm. 1 74.4 1 33.7 27.2% 5.6% 71.3 200 28.2 30.1% 0.6 21.6 39.4% Segm.3 500 28.9% 1.7 35.11: Confusion matrix.0% 0.4 30.6 39.7% 73.3% 2.5 40.8% 92.1 30. 6 segments Predicted → Actual ↓ Segment 1 Segment 2 Segment 3 Segment 4 Segment 1 97.6 24.6 = 0. 4 segments Predicted → Actual ↓ Segment 1 Segment 2 Segment 3 Segment 4 Segment 5 Segment 6 Segm.6% 12.1% 94.0% 2.1% 7.5 100 32.1 10 33.7% 2.7% Table 6.4 44.1% 0.9 39.5% 12.9% 0.5% 4.2% 2.8% 96. 6 5.2 50 30.4 42.10: Sigmoid function.6 20 34.4 38.1 22.4 36.6% 1.γ γ γ γ γ γ C = 0.2% 3. 2 1.6% 2.0 41.9 44.4% 9.6 32.6 26.3% Segm.0 20.9 28.6% 0.0 = 1.0 39.3% 4. 4 8.4 = 0.2 27.6 20.7% 3.12: Confusion matrix. 3 10.6 31.1 29.6% 2.7 40.9 Table 6.6% 1.9 38.1% 6.2 = 1.

The result show that Age is an important Figure 6. are shown in Figure 6.3 Feature Validation In this section. This will be done. by leaving one feature out of the feature vector and train the SVM without this feature. The results of both situations.4 and 6. which increase the result with only tenths of percents. the features will be validated.5.6. This is in contrast with the type of telephone.5: Results while leaving out one of the features with 6 segments feature for classifying the right segment. 65 . The importance of each feature will be measured. Each feature increases the result and therefore each feature is useful for the classification.4: Results while leaving out one of the features with 4 segments Figure 6.

the so-called elbow criterion was applied. several validation measures were used. There are various ways for selecting suitable feature values for the clustering algorithms. called Support Vector Machines was used to estimate the segment of a customer based on his profile. The second part of the research was focused on profiling customers and finding a relation between the profile and the segments. An other problem was that the location of the elbow could differ between the validation measures for the same algorithm. this criterion could not always be unambiguously identified. This selection is vital for the resulting quality of the clustering. 7. but merely as one possible outcome. However. The customer’s profile was based on personal information of the customers. This led to solutions for the customer segmentation with respectively four segments and six segments. For some algorithms. Not every validation method marked the same 66 . the location was c = 6. the elbow was located at c = 4 and for other algorithms. In this research. A novel data mining technique. The customer segments were constructed by applying several clustering algorithms. Unfortunately. To identify the best algorithm. The clustering algorithms used selected and preprocessed data from the Vodafone data warehouse. The result of the clustering can therefore not be regarded as universally valid. One different feature value will result in different segments. without the direct intervention of a human specialist. it is not possible to include all possible combinations of usage behavior characteristics within the scope of this research.1 Conclusions The first objective of our research was to perform automatic customer segmentation based on usage behavior.Chapter 7 Conclusions and discussion This section concludes the research and the corresponding results and will give some recommendations for future work. To find the optimal number of clusters. the feature values were selected in such a way that it would describe the customer’s behavior as complete as possible.

the usage behavior does not correspond to a single customer’s profile and this impairs the classification process. It was however not possible to determine one algorithm that was optimal for c = 4 and c = 6. the resulting quality of the classification was significantly decreased. call duration. Therefore. The last part of the research involves the relative importance of each individual feature of the customer’s profile. 67 . the clustering results were interpreted in a profiling format. By leaving out one feature value during classification. some widely established validation measures were employed to determine the most optimal algorithm. The profile exists of the age. On the other hand. international calls.5% for the situation with six segments. It was found that the radial basis function gives the best result with a classification of 80.algorithm as the best algorithm. The results show. This implies. the Gath-Geva algorithm appears to be the best algorithm and the Gustafson-Kessel algorithm gives the best results by six clusters. this telephone is maybe not used exclusively by the person (and the corresponding customer’s profile) as stored in the database. and companies may exchange telephones among their employees. the effect of each feature value became visible. However. For the situation with four clusters. Therefore. A Support Vector Machine algorithm was used to classify the segment of a customer. sms usage. It was found that without the concept of ’customer age’. that this feature bears some importance for the customer profiling and the classification of the customer’s segment. leaving out a feature such as the ’telephone type’ barely decreased the classification result. habits and income of the customers. Customers may lend their telephone to relatives. and residential area of the customer. different numbers called and percentage of weekday and daytime calls. It is hard to compare the two clustering results. To determine which customer segmentation algorithm is best suited for a particular data set and a specific parameter setting. telephone type. In real life.3% for the situation with four segments and 78. In such cases. A short characterization of each cluster was made. based on the customer’s profile. gender. this and some other features did well increase the performance of classification. both clustering results were used as a starting point for the segmentation algorithm. This is caused by the frequently missing data in the Vodafone data warehouse about lifestyle. It appeared that the resulting percentage of correctly classified segments was not as high as expected. company size. subscription type. because of the different number of clusters. A possible explanation could be that the features of the customer are not adequate for making a customer’s profile. As a comparison. however. that in both situations the clusters were well separated and clearly distinguished from each other. The corresponding segments differ on features as number of voice calls. four different kernel functions with different parameters were tested on their performance. A second reason for the low number of correct classification is the fact that the usage behavior in the database corresponds to a telephone number and this telephone number corresponds to a person.

7. genetic algorithms or Bayesian algorithms. it should be noted that the most obvious and best way to improve the classification is to come to a more accurate and precise definition of the customer profiles. thus. for instance. cluster analysis of the application of miscellaneous (non-linear) kernel functions. The first recommendation is to use different feature values for the customer segmentation. For instance. hierarchical clustering or mixture of Gaussians. a detailed data analysis of the meaning of the cluster is recommended. This can lead to different clusters and thus different segments. an enhanced and more precise analysis of the data ware house will lead to improved features and. this is a complex course and it essentially requires the availability of high-quality features. In this research. Extrapolating this approach. a complete data analysis research is required. An interesting alternative is. However. it is possible to formulate some recommendations for obtaining more suitable customer profiling and segmentation. The customer profile used in this research is not sufficient detailed enough to describe the wide spectrum of customers. Also. 68 . the application of evolutionary algorithms. to an improved classification. a more detailed view of the clusters and their boundaries can be obtained.2 Recommendations for future work Based on our research and experiments. also. One reason for this is the missing data in the Vodafone data warehouse. Consequently. neural networks. Of specific interest is. it is challenging to classify the profile of the customer based on the corresponding segment alone. Finally. and use his feed-back for improving the clustering criteria. the results are given by a short description of each segment. the application of more specialized methods than the elbow criterion could be applied. To estimate the segment of the customer. To know the influence of the feature values on the outcome of the clustering. within the framework of Support Vector Machines. Furthermore. To improve on determining the actual number of clusters present in the data set. Another way to validate the resulting clusters is to offer them to a human expert. we note that the study would improve noticeably by involving multiple criteria to evaluate the user behavior. other classification methods can be used. Similarly. as proposed by Wei Lu [21]. rather than mere phone usage as employed here. Another way of improving this research is to extent the number of cluster algorithms like main shift clustering.

1-27. Fuzzy clustering with a fuzzy covariance matrix. no. Y. Piatetsky-Shapiro. Data mining and complex telecommunications problems modeling. In Proc. F. pp.. Inform. 3 (2001). Engin. [6] Dibike. Research report TTE1-2001-29. Comp. IEEE Trans.J. Knowledge Discovery and Data Mining.E. vol. and Appl. H. J. W.. Unsupervised optimal fuzzy clustering..L. J. J.. Softw. W. VTT Information Technology (2001).. (2005). pp. and Matheus. [5] Bezdek. D. G. Comput. Technol.. Engrg. pp. AAAI/MIT Press (1991). 761766. [2] Amat. Y. C. Proc... pp. 7 (1989).Bibliography [1] Ahola. [4] Bounsaythip. VTT Information Technology (2001). G. Data mining case studies in customer profiling. 3 (2002). [10] Giha. (1979). Singh. and Rinta-Runsala. 1st Int. In Proc. in Civ. and Dagan.. and Ewe.. pp. C. 112-117.E.C.J. [8] Frawley. 773-781.. Fuzzy Clustering and Data Analysis Toolbox For Use with Matlab. Technol. Knowledge discovery in textual databases (KDT). Conf.B. and Kessel. no.. Overview of Data Mining for Customer Behavior Modeling. pp. [3] Balasko. Abonyi.. M. R. [11] Gustafson. Knowledge discovery in databases. Velickov. J. Telecommun. B. J. [9] Gath.C... 3 (2003). 11 no. applications to customer profiling and fraud management. Using reporting and data mining techniques to improve knowledge of subscribers.. I. Solomatine D.B. C-24 (1975). and Abbott. pp. 69 . Research report TTE1-2001-18. [7] Feldman. Inform. S. 208-216. J.B. F. (2006). J. Telecommun... pp. J. and Dunn. and Rinta-Runsala E. 11-16. and Balazs.. and Geva.T.E. vol.. A. I. 15 iss. no. 397 (2003).P. Model Induction with Support Vector Machines: Introduction and Applications. 835-838.. 115-120. IEEE Trans Pattern and Machine Intell. Optimal fuzzy partition: A heuristic for estimating the parameters in a mixture of normal distributions. Customer Profiling and Segmentation based on Association Rule Mining Technique. E. [12] Janusz. vol. IEEE CDC..

. Subramaniam. W. ICANNGA (2007). pp. G. Recogn. J. and Dunbar. How to do it.. how to profit from it.E. G. 24 no. and Chiu. P. 70 . 2367-2376. Turning telecommunications call detail to churn prediction: a data mining approach. 24 (2003).. Lett.. Constructing Stereotypes for an Adaptive e-Shop Using AIN-Based Clustering.. M.. C.. [20] Wei. 127137. M. I. Spring. D. vol. Decision Support Systems. Data Warehousing and Data Mining for Telecommunications.N.T. [16] Noble. P. [21] Wei Lu. and Lee. A. P. 837-845. Syst. London: Artech House. [22] Weiss.. Supp. 103112. Tan. Knowledge management and data mining for marketing. 23 (2002). Market segmentation.W. The Data Mining and Knowledge Discovery Handbook (2005). (1998). I. Tsihrintzis. [15] McDonald. Savvopoulos. C.. 12 (2006). and Sotiropoulos. pp. A New Evolutionary Algorithm for Determining the Optimal Number of Clusters. Expert Syst. vol. vol. pp. vol.. pp.T. [18] Verhoef.. Palgrave Publ. Data Mining in Telecommunications.P.. [14] Mattison. 471-481.. and Welge. The commercial use of segmentation and predictive modeling techniques for database marketing in the Netherlands.. Hoekstra. CIMCA/IAWTIC (2005). [19] Virvou. pp.. K.J. pp. What is a support vector machine? Nature Biotechnology. pp...S. Patt. pp. R. M. Clustering and its validation in a symbolic framework. 648-653. vol. Appl.. 31 (2001). Boston. 1189-1201.. [17] Shaw. G. (1997). I.[13] Mali.A.M. 1565-1567. 34 (2002). Decis. M.

The colored boxes group the tables in a category. the relation tables (the red tables in the middle) are needed. The white rectangles correspond to the tables that were used for this research.Appendix A Model of data warehouse In this Appendix a simplified model of the data ware house can be found. The most important data fields of these tables are written in the table. To connected the tables with each other. 71 .

Figure A.1: Model of the Vodafone data warehouse 72 .

Appendix B Extra results for optimal number of clusters In this Appendix. for the algorithms that not were discussed in Section 4.1.1: Partition index and Separation index of K-medoid 73 . are given. The K-medoid algorithm: Figure B. the plots of the validation measures.

3: Partition coefficient and Classification Entropy of Fuzzy C-means 74 .2: Dunn’s index and Alternative Dunn’s index of K-medoid The Fuzzy-C-means algorithm: Figure B.Figure B.

Separation index and Xie Beni index of Fuzzy C-means Figure B.4: Partition index.5: Dunn’s index and Alternative Dunn’s index of Fuzzy C-means 75 .Figure B.

You're Reading a Free Preview

Descarga
scribd
/*********** DO NOT ALTER ANYTHING BELOW THIS LINE ! ************/ var s_code=s.t();if(s_code)document.write(s_code)//-->