ADDIS ABABA UNIVERSITY

SCHOOL OF GRADUATE STUDIES

SCHOOL OF INFORMATION SCIENCE

KNOWLEDGE DISCOVERY FOR EFFECTIVE

CUSTOMER SEGMENTATION: THE CASE OF
ETHIOPIAN REVENUE AND CUSTOMS AUTHORITY

BELETE BIAZEN

JUNE 2011
 ADDIS ABABA UNIVERSITY
SCHOOL OF GRADUATE STUDIES
SCHOOL OF INFORMATION SCIENCE

KNOWLEDGE DISCOVERY FOR EFFECTIVE

CUSTOMER SEGMENTATION: THE CASE OF
ETHIOPIAN REVENUE AND CUSTOMS AUTHORITY

A Thesis Submitted to the School of Graduate Studies of Addis

Ababa University in Partial Fulfillment of the Requirements for the
Degree of Master of Science in Information Science

By

BELETE BIAZEN

JUNE 2011
 ADDIS ABABA UNIVERSITY
SCHOOL OF GRADUATE STUDIES
SCHOOL OF INFORMATION SCIENCE

KNOWLEDGE DISCOVERY FOR EFFECTIVE

CUSTOMER SEGMENTATION: THE CASE OF
ETHIOPIAN REVENUE AND CUSTOMS AUTHORITY

By

BELETE BIAZEN

Name and signature of Members of the Examining Board

Name Title Signature Date

Ato Mahder Alemayehu Chairperson ____________ ________________

Ato Getachew Jemaneh Advisor ____________ ________________

Gashaw Kebede (PhD) Examiner ____________ ________________

 DECLARATION
I declare that the thesis is my original work and has not been presented for a degree in
any other university.

_________________

Date

This thesis has been submitted for examination with my approval as university advisor.

_________________

Advisor
 DEDICATION
I would like to dedicate this paper to my family, who have always been there, and
supported me when I face any difficulty.

I
 ACKNOWLEDGEMENT
First and foremost, I would like to thank the almighty GOD for giving me the ability to
do this research work.

My pleasure and deepest gratitude goes to my advisor Ato Getachew Jemaneh for his
patience, valuable ideas, and supportive advice. His interest and encouragement has
always stimulated me to accelerate to the completion of the work.

I would like to forward my special thanks to the Ethiopian Revenue and Customs
Authority employees. Many thanks go particularly to the Database Administrator of the
Authority, Ato Getacher Alemu, and the customer service officers for their kind
cooperation and consistent assistance in explaining the work in the problem domain.

I would also like to extend my very sincere and special thanks to Dr. Million Meshesha
who provided me with the necessary assistance during the proposal writing.

I am also very thankful to my brother Endegena Biazen and my sister Zegeju Biazen for
their support and encouragement in all my studies, starting from my early school age.

Lastly, my special gratitude and respects go to my friends Hilemariam Abebe, Kendie

Alebachew, Shumet Taddese, Tadele Asitatikie and Tariku Adane, who supplied me with
all the necessary information and valuable comments.

II
 LIST OF ACRONYMS
AI: Artificial Intelligence

ANN: Artificial Neural Network

ARFF: Attribute Relationship File Format

ASYCUDA: Automation System Customs Data

CRISP-DM: CRoss Industry Standard Process of Data Mining

CRM: Customer Relationship Management

CSV: Comma Separated Value

DBMS: Database Management Systems

ERCA: Ethiopian Revenue and Customs Authority

IT: Information Technology

KDD: Knowledge Discovery in Databases

MOLAP: Multidimensional Online Analytical Processing

OLAP: Online Analytical Processing

ROLAP: Relational Online Analytical Processing

SAD: Single Administration Document

SOM: Self Organized Maps

SQL: Structured Query Language

WEKA: Waikato Environment for Knowledge Analysis

III
 TABLE OF CONTENTS

DEDICATION.................................................................................................................... I
ACKNOWLEDGEMENT ............................................................................................... II
LIST OF ACRONYMS .................................................................................................. III
LIST OF TABLES ........................................................................................................ VII
LIST OF FIGURES ........................................................................................................ IX
LIST OF APPENDICES ................................................................................................. X
ABSTRACT ..................................................................................................................... XI
CHAPTER ONE ............................................................................................................... 1
BACKGROUND ............................................................................................................... 1
1.1 Introduction ........................................................................................................ 1
1.2 Statement of the Problem and Justification ..................................................... 3
1.3 Objectives ............................................................................................................ 5
1.3.1 General Objective ......................................................................................... 5
1.3.2 Specific Objectives ........................................................................................ 5
1.4 Scope of the Study .............................................................................................. 6
1.5 Methodology ....................................................................................................... 6
1.6 Significance of the Study.................................................................................... 9
1.7 Organization of the Thesis ................................................................................. 9
CHAPTER TWO ............................................................................................................ 11
DATA MINING .............................................................................................................. 11
2.1 Overview ........................................................................................................... 11
2.2 Data Mining Process ........................................................................................ 12
2.2.1 Defining the Data Mining Problem ............................................................ 12
2.2.2 Collecting the Data Mining Data ............................................................... 13
2.2.3 Detecting and Correcting the Data............................................................. 13
2.2.4 Estimating and Building the Model ............................................................ 13
2.2.5 Model Description and Validation.............................................................. 14
2.3 Data Mining as KDD Process .......................................................................... 15
2.4 Data Mining and Related Fields ..................................................................... 16
2.4.1 Data Mining and Data Warehousing .......................................................... 16
2.4.2 Data Mining and DBMS ............................................................................. 17
2.4.3 Data Mining and OLAP .............................................................................. 17

IV
 2.4.4 Data Mining, Artificial Intelligence and Statistics ..................................... 18
2.5 Data Mining Functionality .............................................................................. 19
2.5.1 Concept/Class Description: Characterization and Discrimination ........... 20
2.5.2 Association Analysis ................................................................................... 20
2.5.3 Classification and Prediction ..................................................................... 20
2.5.4 Cluster Analysis .......................................................................................... 21
2.5.5 Outlier Analysis .......................................................................................... 22
2.5.6 Evolution and Deviation Analysis ............................................................... 22
2.6 Data Mining Techniques and Algorithms ...................................................... 23
2.6.1 Clustering Technique .................................................................................. 23
2.6.2 Classification .............................................................................................. 25
2.7 Applications of Data Mining ........................................................................... 30
2.7.1 Application of Data Mining for CRM ......................................................... 34
CHAPTER THREE ........................................................................................................ 37
CUSTOMER RELATIONSHIP MANAGEMENT AND CUSTOMER
SEGMENTATION ......................................................................................................... 37
3.1 Loyalty and Customer Relationship Management ....................................... 37
3.1.1 Overview ..................................................................................................... 37
3.1.2 Customer Loyalty ........................................................................................ 39
3.1.3 Principles and Tasks of CRM ..................................................................... 40
3.2 Customer Segmentation ................................................................................... 42
3.2.1 Overview ..................................................................................................... 42
3.2.2 Applications of Customer Segmentation ..................................................... 43
3.2.3 Difficulties in Making Good Segmentation ................................................. 44
3.3 CRM in Ethiopian Revenue and Customs Authority ................................... 45
3.3.1 Overview ..................................................................................................... 45
3.3.2 Powers and Duties of the ERCA ................................................................. 46
3.3.3 Organization of the ERCA .......................................................................... 47
3.3.4 Complaints Handling in ERCA ................................................................... 48
CHAPTER FOUR ........................................................................................................... 50
EXPERIMENTATION .................................................................................................. 50
4.1 Overview ........................................................................................................... 50
4.2 Understanding the Problem Domain .............................................................. 51
4.2.1 Data Mining Goals ..................................................................................... 51
4.2.2 Data mining Tool Selection ........................................................................ 52
V
 4.3 Understanding the Data ................................................................................... 52
4.3.1 Collection of Initial Data ............................................................................ 53
4.3.2 Description of the Data Collected .............................................................. 53
4.3.3 Data Quality Verification ........................................................................... 57
4.4 Preparation of the Data ................................................................................... 57
4.4.1 Data Cleaning ............................................................................................. 57
4.4.2 Data Integration and Transformation ........................................................ 58
4.4.3 Data Formatting ......................................................................................... 59
4.4.4 Attribute Selection ....................................................................................... 59
4.5 Data Mining ...................................................................................................... 61
4.5.1 Selection of Modeling Techniques .............................................................. 61
4.5.2 Test Design.................................................................................................. 62
4.5.3 Model Building............................................................................................ 63
4.6 Evaluation of the Discovered Knowledge ...................................................... 87
4.7 Use of the Discovered Knowledge ................................................................... 90
CHAPTER FIVE ............................................................................................................ 91
CONCLUSION AND RECOMMENDATIONS .......................................................... 91
5.1 Conclusion ......................................................................................................... 91
5.2 Recommendations ............................................................................................ 92
REFERENCES ................................................................................................................ 95
APPENDICES ............................................................................................................... 101

VI
 LIST OF TABLES
Table 1.1 Comparison of the five KDD process models.................................................... 7

Table 4.1 Attributes and description of the SAD general segment table.......................... 54

Table 4.2 Attributes and description of the SAD_ITEM table......................................... 55

Table 4.3 Attributes and description of the company table ............................................. 55

Table 4.4 Attributes and description of the cash declaration payment in table................ 56

Table 4.5 Attributes and description of the country table................................................. 56

Table 4.6 Selected attributes with their description.......................................................... 60

Table 4.7 List of range of conditions by which a cluster result is assessed...................... 64

Table 4.8 Cluster distributions with number of iterations 4 and sum of squared errors
5199.169............................................................................................................................ 65

Table 4.9 List of abbreviated words and attributes of the dataset along with their
description......................................................................................................................... 66

Table 4.10 Clustering result of the first experiment......................................................... 67

Table 4.11 Cluster rank for the first experiment............................................................... 68

Table 4.12 Clustering result of the second experiment……………………………..…... 70

Table 4.13 Cluster rank for the second experiment…...................................................... 71

Table 4.14 Clustering result of the third experiment........................................................ 73

Table 4.15 Cluster rank for the third experiment…….……………………...………….. 75

Table 4.16 Clustering result of the fourth experiment...................................................... 76

Table 4.17 Cluster rank for the forth experiment………..……………………………... 78

VII
Table 4.18 Output from J48 decision tree algorithm with 10-fold cross-validation default
parameter value……………………………………………………………...….............. 80

Table 4.19 Output from J48 decision tree algorithm with minNumObj=25 and
confidenceFactor=0.25 .................................................................................................... 81

Table 4.20 Summary of the confusion matrix with default parameters value and 70 % for
training and 30 % for testing dataset................................................................................. 82

Table 4.21 Parameters and their default values of the neural network classifier............. 84

Table 4.22 10-fold cross-validation output from MultilayerPerceptron ANN algorithm

with hiddenLayers=8, learningRate=0.5 and momentum=0.4......................................... 84

Table 4.23 Split output from MultilayerPerceptron ANN algorithm with hiddenLayers=8
learningRate=0.6 and momentum=0.4.............................................................................. 85

Table 4.24 Summary of the accuracy level of the decision tree and neural net
classification models……………………..………………………………………………86

VIII
 LIST OF FIGURES
Figure 2.1 Knowledge discovery steps………………………………………..…………16

Figure 2.2 Diagram of a typical neural network…………………………………………30

Figure 3.1 Organizational charts of the Ethiopian Revenue and Customs Authority…....48

Figure 4.1 The six-step Cios et al. (2000) KDD process model …...…….…...…........…50

IX
 LIST OF APPENDICES
Appendix 1. Partial view of the initial collected sample data………………….....……101

Appendix 2. Sample of the decision tree generated with 10-fold cross-validation

technique…………………………………………………………………..…………....102

Appendix 3. Output of the K-means cluster modeling with different K and seed
values………………………………………………………………………………….. 104

X
 ABSTRACT
CRM is a process by which an organization maximizes customer satisfaction in an effort
to increase loyalty and retain customers‟ business over their lifetimes. On the other hand,
customer segmentation is the grouping of customers into different groups based on their
common attributes and it is the main part of CRM. In order to analyze CRM data, one
needs to explore the data from different angles and look at its different aspects. This
should require application of different types of data mining techniques. Data mining finds
and extracts knowledge hidden in corporate data warehouses.

The aim of this study is to test the applicability of clustering and classification data
mining techniques to support CRM activities for ERCA using the Cios et al. (2000) KDD
process model. In this study, different characteristics of the ERCA customers‟ data were
collected from the customs ASYCUDA database. Once the customers‟ data were
collected, the necessary data preparation steps were conducted on it and finally a dataset
consisting of 46748 records was attained.

To segment customers, the K-means clustering algorithm was used. During the cluster
modeling different experiments have been conducted using different cluster numbers
(K=3, 4, 5, 6) and seed values. From the different experiments, the one which had better
performance has been selected. Hence, the cluster model at K=5 had better performance
and its output was used for the next classification modeling.

The classification modeling was built by using J48 decision tree and multilayerperceptron
ANN algorithms with 10-fold cross-validation and splitting (70% training and 30%
testing) techniques. Among these models, a model which was built using J48 decision
tree algorithm with default 10-fold cross-validation shows better performance which is
99.95% of overall accuracy rate; hence this model was selected.

The results of this research were encouraging as very high classification accuracy has
been obtained.

XI
 CHAPTER ONE

BACKGROUND
1.1 Introduction
Customer relationship management (CRM) has become one of the strategies of an
organization for sustained competitive advantage. CRM in its broadest sense simply
means managing all customer interaction (Trappey et al. 2009). The new millennium is in
the middle of explosive change witnessing rapidly changing market conditions, volatile
equity markets, reconstructed value chains and new global competitors (Kumar and
Solanki 2010). Customers themselves are changing, and consider natural customer
loyalty which is a thing of the past (Suresh 2002). CRM includes all measures for
understanding the customers and for exploiting this knowledge to design and implement
marketing activities, align production and coordinate the supply-chain (Srivastava 2002).
It is an integration of technology and business to satisfy the need of the customer. CRM
gains in its importance for companies that serve multiple groups of customers and exploit
different interaction channels for them. CRM is a process by which a company
maximizes customer satisfaction in an effort to increase loyalty and retain customers‟
business over their lifetimes. The primary goals of CRM are to build long term and
profitable relationships with chosen customers and get closer to those customers at every
point of contact (Verhoef 2003).

Segmentation can be defined as aggregating customers into groups with similar

characteristics such as demographic, geographic or behavioral traits and marketing to
them as a group (Parvatiyar and Sheth 2001). Consequently, each member of the segment
has similar needs and wants; however, they are not completely uniform. The result was
that customers often received most of what they wanted but still had to compromise on
many desires (Bose 2002).

Customer segmentation is the grouping of customers into different groups based on their
common attributes and it is the main part of CRM (Verhoef 2003). Segmentation requires

1
the collection, organization and analysis of customer data. With proper segmentations of
a customer‟s data it is possible to identify the reliability/loyalty of customers so as to
increase the revenue of the organization. CRM creates interaction of customers with the
organization by using information technology (IT). Moreover, identifying customer‟s
need/interest better and treating them accordingly can increase their life time (Verhoef
2003).

In order to analyze CRM data, one needs to explore the data from different angles and
look at its different aspects. This should require application of different types of data
mining techniques. There is a strong requirement for data integration before data mining.
Data mining involves specialized software tools that allow users to filter through large
amounts of data to uncover data content relationships and build models to predict
customer behavior; data mining uses well-established statistical and machine learning
techniques to build models that predict customer behavior (Suresh 2002).

Rygielski et al. (2002) define data mining as a sophisticated data search capability that
uses machine learning algorithms to discover patterns and correlations in data. Data
mining finds and extracts knowledge hidden in corporate data warehouses, or information
that users have dropped on a website, most of which can lead to improvements in the
understanding and use of the data (Rygielski et al. 2002). In simple terms, data mining is
another way to find meaning in data. Data mining discovers patterns and relationships
hidden in data, and is actually part of a larger process called knowledge discovery in
database (KDD), which describes the steps that must be taken to ensure meaningful
results (Edelstein 2002). Data mining software does not, however, eliminate the need to
know the business, understand the data, or be aware of general statistical methods. Data
mining helps business analysts to generate hypotheses, but it does not validate/confirm
the hypotheses.

Fayyad et al. (1996) also describe data mining as the application of specific algorithms
for extracting patterns from data. In addition to data mining, the additional steps in the
KDD process, such as data preparation, data selection, data cleaning, are also essential to
ensure that useful knowledge is derived from the data. So, to effectively segment the

2
customers of a certain organization it is possible to apply different kinds of data mining
techniques

Classification is a data mining technique which maps the target data into the predefined
groups or classes. Classification is a supervised learning technique because the classes
are predefined before the examination of the target data (Deshpande and Thakare 2010).
Classification is mainly used for predictive model which makes prediction about
unknown data values by using the known values.

Similar to classification, clustering is also the organization of data in classes. However,

unlike classification, in clustering, class labels are unknown (Han and Kamber 2006).
Clustering is also referred to as unsupervised learning or segmentation and it is the
partitioning or segmentation of the data into groups or clusters (Deshpande and Thakare
2010).

The present research is on the Ethiopian Revenue and Customs Authority (ERCA).
ERCA has the responsibility to collect revenue for the federal government and prevent
contraband. In addition to raising revenue, ERCA is responsible to facilitate the
legitimate movement of people and goods across the border. Simultaneously, ERCA
focuses on those people and vehicles that may involve in the act of smuggling i.e. the act
of bringing into or taking out of the country goods on which customs duty and taxes are
not paid and goods the importation or exportation of which are prohibited by law. ERCA
conducts investigation, audits and prosecutes offenders/ criminals. In order to discharge
this responsibility the ERCA works by holding the objective of using established modern
revenue assessment and collection system; and provide customers with equitable,
efficient and quality service, and cause taxpayers voluntarily discharge their tax
obligations.

1.2 Statement of the Problem and Justification

Currently, the ERCA is using statistical analysis and assessment techniques to identify
potential and low valued customers. The Authority check whether the customers
discharge their responsibility or not in the revenues database and at the same time they

3
cross check the customs database whether the items are imported/exported with paying
the required tax by using assessment and statically analysis techniques. The Authority is
also using these techniques to check the credibility of the customers, i.e., whether or not
they import/export goods in accordance with what they had specified in their declaration
form. However, these techniques are not effective and efficient and took a considerable
amount of time to treat customers according to their characteristics. For instance,
international trade participants (importers, exporters) were facing some difficulties to
deliver their goods to domestic and international market on time. Owing to it, importers
or exporters viewed ERCA procedure with disfavor or as an obstacle that blocked the
movement of international trade. These techniques are also less efficient to control
taxpayers who fail to declare their actual income in order to reduce their tax bill and the
federal government‟s revenue.

Chen et al. (2005) define CRM as the strategy integrating sales, marketing and services,
which unites operating procedures and technology to better understand customers from
different perspectives. ERCA has well organized customers database. Even though the
organization has well-organized customers database, there is no such an integrated
system or model being applied to segment customers. In addition, the ERCA has no any
well-organized set of rules or procedures that are used for segmenting customers
according to their culture, behavior, and characteristics to say whether the customers are
potential or low valued customers. As a result, the organization is unable to collect the
required revenue and control the act of smuggling.

In order to solve the problems that are described earlier, conducting a research using the
data mining technology is appropriate. In the context of Ethiopia, there are some attempts
made by Fekadu (2004) and Melaku (2009) in Ethiopian Telecommunications
Corporation, Kumnegere (2006) in Ethiopian Shipping Lines, and Henok (2002) and
Denekew (2003) in Ethiopian Airlines to solve CRM problems using the data mining
technology. Moreover, different substantial research works have also been done on
revenue and customs authority CRM abroad, such as Mahler and Hennessey (1996) on
“Taking Internal Customer Satisfaction Seriously at the U.S. Customs Service”, Doye
(2010) a study on “Collaborative Border Management” and Boulding et al. (2005)

4
investigates on “A Customer Relationship Management Roadmap: What is Known,
Potential Pitfalls, and Where to Go”. Finally, they come up with a conclusion that
achieving customer satisfaction has been a keystone in the total quality management
movement and in the national performance review efforts. However, to the best of the
knowledge of the present researcher, there is no work done to solve CRM problems on
ERCA.

Since there has been an increase in demand and supply from time to time, there is also an
increase in transactions of goods in the borders of the country. So, the organization has to
revisit its CRM strategy and should support it with new technology so as to satisfy its
customers and reduce smuggling.

This study is attempted to address CRM related problems according to customers‟

behavior, culture and characteristics. The study also focuses on grouping of customers
based on the hidden information, which is found in the customers‟ database of ERCA and
to generate rules for identifying potential and low valued customers.

1.3 Objectives
1.3.1 General Objective

The general objective of this study is to design a model using data mining techniques for
customer segmentation that helps the organization to maintain the potential CRM. This
can be done through transforming customer data into meaningful categorization of
customer that has been useful for designing appropriate CRM strategies for the purpose
of maximizing revenue, preventing smuggling and increase customer satisfaction.

1.3.2 Specific Objectives

The specific objectives of this study are the following:

 To review the literature, to have a conceptual understanding about the study area
 To identify and preprocess the type of customers data which is found in the
organizations database

5
  To select the data mining tool and algorithms to be used based on the objective of
the study
 To prepare the data for analysis (data selection, data cleaning and organizing the
training dataset)
 To build data mining model using training dataset for effective customer
segmentation
 To evaluate the performance of the model

1.4 Scope of the Study

The present study focuses only on customers who perform import/export transactions
through ERCA; that means the study is only restricted on the ERCA customs database.
The study is also restricted on building a data mining model for segmenting the
customers, interpreting the resulting segments, and developing a classification rules for
each segment. The researcher uses clustering for describing the dataset and classification
techniques for developing a predictive model.

1.5 Methodology

This study generally follows both quantitative and qualitative methods. The quantitative
method is used to collect and analyze customers‟ data. On the other hand, the qualitative
aspect is used to understand the business operation by making a close relationship with a
domain experts and the responsible body such as the Database Administrator of the
organization. In data mining or KDD process, there are different kinds of standard
methodologies or models, such as Cross-Industry Standard Process for Data Mining
(CRISP-DM), Anand & Buchner, Fayyad et al., Cios et al. (2000) and Cabena et al.
(Kurgan and Musilek 2006). All process models consist of multiple steps executed in a
sequence, which often includes loops and iterations. As Kurgan and Musilek (2006)
explained the main differences between the models lie in the number and scope of their
specific steps.

6
The following table shows the different steps of the KDD process models.

Model Fayyad et al. Anand & Buchner Cios et al. (2000) Cabena et al. CRISP-DM
Area Academic Academic Hybrid Industry Industry
# of 9 8 6 5 6
steps
Steps 1. Developing & 1. Human 1. Understanding of 1. Business 1. Business
Understanding the Resource the Problem Domain Objective Understanding
Application Domain Identification Determination

2. Creating a Target 2. Problem 2. Understanding of 2. Data 2. Data

Data Set Specification the Data Preparation Understanding
3. Data Cleaning & 3. Data 3. Preparation of the 3. Data Mining 3. Data
Preprocessing Prospecting Data Preparation
4. Data Reduction & 4. Domain 4. Data Mining 4. Analysis of 4. Modeling
projection Knowledge Results
Elicitation
5. Choosing the 5. Methodology 5. Evaluation of the 5. Assimilation 5. Evaluation
Data Mining Task Identification Discovered of Knowledge
Knowledge
6. Choosing the 6. Data 6. Use of the 6. Deployment
Data Mining Preprocessing Discovered
Algorithm Knowledge
7. Data Mining 7. Pattern
Discovery
8. Interpreting 8. Knowledge
Mined Patterns Post-Processing
9. Consolidating
Discovered
Knowledge

Table 1.1 Comparison of the five KDD process models Source: Kurgan and
Musilek (2006)

As it is indicated in Table 1.1, most process models come under the categories of either
industrial or academic process models. However, there are some models which combine
both aspects; this kind of models are called hybrid models. An example of such kind of
model is a six step KDD process model developed by Cios et al. (2000). It was developed
based on the CRISP-DM model by adopting it as an academic research model. As
Kurgan and Musilek (2006) explained the main difference of the Cios et al. (2000) model
from other data mining model is that the Cios et al. (2000) model has the advantages of:

 providing more general, research-oriented description of the steps

 introducing a data mining step instead of the modeling step

7
  introducing several new explicit feedback mechanisms, (the CRISP-DM model
has only three major feedback sources, while the hybrid model has more detailed
feedback mechanisms) and
 modification of the last step, since in the hybrid model, the knowledge
discovered for a particular domain may be applied in other domains

So, this study employed Cios et al. (2000) KDD process model. The Cios et al. (2000)
KDD process model focuses on the process of data mining projects‟ life cycle.

 Understanding of the Problem Domain

Interviews, observation and document analysis have been made to thoroughly assess the
need of users and analyze the business problems. Interviews with domain experts and
other officers are preferred to other methods, such as questionnaires, because interviews
are used to understand the business problem very well by asking further questions or
interviews is more flexible than the other techniques. Observation of the existing system
of the organization is also important to understand in what manner they treat or handle
their customers. Any relevant document related to customer relationship/handling would
also be analyzed by the researcher to get the required information.

 Understanding of the Data

The data source has been the database of ERCA which contain the data of the customers
engaged in import and export of different goods.

The dataset contains customer types with respect to their attributes and selection of
suitable data has been made according to the organizations need and problem.

 Preparation of the Data

Before using the customers‟ data it must be converted into an appropriate form for
analysis. Data preparation techniques such as analysis, editing, cleaning, integration and
transformation have been preformed.

8
  Data Mining

After the data has been prepared for analysis it is used to build clustering and/or
classification models using data mining techniques. This has enabled to design a better
strategy for effective customer segmentation.

 Evaluating of the Discovered Knowledge

In order to check the output of the model‟s performance, counterchecking with domain
experts and responsible persons (other officers) of the organization is made. Based on the
confusion matrix accuracy of the system is also evaluated to see the performance of the
model in segmenting customers.

 Use of the Discovered Knowledge

At the end of this study, the researcher has generated the final written report of the data
mining engagement, including all of the deliverables, summarizing and organizing the
results.

1.6 Significance of the Study

This study is important for the organization to properly segment the customers and treat
them accordingly. If the organization implements proper CRM with the help of IT, the
satisfaction of the customers is expected to increase, which in turn would result in an
increase of revenue and reduction of smuggling. The results of this research would
support the routine and strategic decision making processes of the ERCA.

1.7 Organization of the Thesis

This research paper is organized in five chapters. The first chapter deals with the
background of the study which mainly introduces the problem area, states the problem,
the general and specific objectives of the study, the research methodology, the scope of
the study, and significance of the study. The second chapter reviews the KDD or data

9
mining technology. In this chapter, different data mining techniques, such as clustering
and classification with their respective algorithms, are reviewed. Chapter Three discusses
CRM, customer segmentation and introduces the ERCA (the organization which this
research is on). Chapter Four presents the experimentation phase of the study, which
mainly discusses the different stages of the experiment to build the KDD model and
interprets the results of the clustering and classification experiments. The final chapter,
Chapter Five, presents the conclusion of the result of the study and provides
recommendation based on the investigation of the research.

10
 CHAPTER TWO

DATA MINING
2.1 Overview
The amount of data kept in computer files and databases is growing at a phenomenal rate.
At the same time, the users of these data are expecting more sophisticated information
from them. A marketing manager or customer service officer is no longer satisfied with a
simple listing of marketing contexts, but wants detailed information about customers past
purchases as well as predictions of future ones (Dunham 2000). Simple SQL queries are
not adequate to support these increased demands on information. Information retrieval
(IR) is also not adequate anymore for decision-making. With huge collections of data,
now there would be created new needs to help make better managerial choices. These
needs are automatic summarization of data, extraction of the core information stored, and
the discovery of patterns in raw data (Han and Kamber 2006). Data mining, also known
as knowledge discovery in database (KDD), is able to solve these needs.

As explained by Trappey et al. (2009), data mining enables the extraction of hidden
predictive information from large databases. As a result, organizations identify valuable
customers, predict their future behaviors, which enable them to make positive,
knowledgeable decisions. Data mining involves the use of sophisticated data analysis
tools to discover previously unknown, valid patterns and relationships in large dataset.
These tools can include statistical models, mathematical algorithms, and machine
learning methods (algorithms that improve their performance automatically through
experience, such as neural networks or decision trees). Consequently, data mining
consists of more than collecting and managing data; it also includes analysis and
prediction (Trappey et al. 2009).

As indicated earlier, Rygielski et al. (2002) describe data mining as a powerful data
search capability that uses statistical algorithms to discover patterns and correlations in
data. It is possible to apply data mining techniques on data represented in quantitative,

11
textual, or multimedia forms (Rygielski et al. 2002). Data mining applications can use a
variety of parameters to examine the data. They include association (patterns where one
event is connected to another event), sequence or path analysis (patterns where one event
leads to another event), classification (identification of new patterns), clustering (finding
and visually documenting groups of previously unknown facts), and forecasting
(discovering patterns from which one can make reasonable predictions regarding future
activities) (Singh and Chauhan 2009).

2.2 Data Mining Process

Data mining process is a step in KDD process which consists of methods that produce
useful patterns or models from the data (Nasereddin 2009). In data mining process there
are two possibilities; in some cases when the problem is known and correct data is
available as well, there is an attempt to find the models or tools which will be used. On
the other hand, some problems might occur because of duplicate, missing, incorrect,
outlier values and there is a need to make some statistical methods. According to
Nasereddin (2009), to solve these kinds of problems the data mining process passes
through the following five steps:

 Defining the data mining problem

 Collecting the data mining data
 Detecting and correcting the data
 Estimating and building the model
 Model description, and validation

2.2.1 Defining the Data Mining Problem

Understanding of the projects‟ objectives and requirements are the first step in data
mining process. Once the project have specified from a business perspective, it is
possible to formulate it as a data mining problem and develop a preliminary
implementation plan. Most data-based modeling studies are performed for a particular
application domain. Hence, domain-specific knowledge and experience are usually
necessary in order to come up with a meaningful problem statement (Nasereddin 2009).

12
 2.2.2 Collecting the Data Mining Data

This process mainly focuses on the collection of data from different sources and
locations. Currently, there are two methods which are used to collect the data mining
data. These are:

 Internal data: Data are usually collected from existing databases, data
warehouses, and OLAP. Actual transactions recorded by individuals are the
richest source of information.
 External data: In addition to data shared within a company, data items can also be
collected from demographics, psychographics and web graphics.

2.2.3 Detecting and Correcting the Data

Real-world databases are usually confronted with noise, missing, and inconsistent data
due to their typically huge size. Data preprocessing is commonly used as a preliminary
data mining practice. It transforms the data into a format that has been easily and
effectively processed by data mining algorithms. As indicated by Nasereddin (2009),
there are a number of data preprocessing techniques which include:

 Data cleaning: This can be applied to remove noise and correct inconsistencies,
outliers and missing values.
 Data integration: Merges data from multiple sources into a coherent data store,
such as a data warehouse or a data cube.
 Data transformations: It is the process which improves the accuracy and
efficiency of mining algorithms involving distance measurements.
 Data reduction: It is the reduction of the data size by aggregating, eliminating
redundant features.

2.2.4 Estimating and Building the Model

According to Nasereddin (2009), the process of estimating and building the model
includes the following four parts:

13
  select data mining task
 select data mining method
 select suitable algorithm and
 extract knowledge

Select data mining task (s): Selecting which task to use depends on the model whether it
is predictive or descriptive (Two Crows Corporation 1999). Predictive models predict the
values of data using known results and/or information found in large dataset.
Classification, regression, time series analysis, prediction, and estimation are tasks for
predictive model (Zaiane 1999). On the other hand, descriptive model identifies patterns
or relationships in data and serves as a way to explore the properties of the data
examined. Clustering, summarization, association rules and sequence discovery are
usually viewed as descriptive (Zaiane 1999). The importance of prediction and
description for particular data mining applications can be varying. That means selecting
which task to use depends on the model whether it is predictive or descriptive.

Select data mining method (s): After selecting the task the next step is choosing the
method of the data mining. There are a number of methods for model estimation
including neural networks, decision trees, association rules, genetic algorithms, cluster
detection, fuzzy logic and so on.

Select suitable algorithm: The next step is to select a suitable specific algorithm that
implements the general methods.

Extracting knowledge: This is the last step in building the model which is the result (the
answers for the problem solved in data mining) after making the simulation for the
algorithm.

2.2.5 Model Description and Validation

In all cases, the function of the data mining models is to assist users in decision making.
However, the model by itself doesn‟t help users to give decisions; hence, there is a need

14
to interpret such models because humans are unable to make good decisions on these
complex models.

2.3 Data Mining as KDD Process

Data mining discovers patterns and relationships hidden in data, and is actually part of a
larger process called KDD process which describes the steps that must be taken to ensure
meaningful results (Rygielski et al. 2002). KDD refers to the serious extraction of
implicit, previously unknown and potentially useful information from data in databases.
While data mining and KDD are frequently treated as synonyms, data mining is actually
part of the KDD process.

As described by Dunham (2000), the KDD process is an iterative process which consists
of the following steps:

Data cleaning: Also known as data cleansing, it is a phase in which noise data and
irrelevant data are removed from the collection. The process of data cleansing is normally
expensive, hence it was not possible to do with old technologies. Nowadays, faster
computers allow data cleansing to be performed in an acceptable amount of time on a
large amount of data (Huang 2003). In addition to KDD, data cleansing is also applied in
data warehouse and total data quality management

Data integration: At this stage, multiple heterogeneous data sources are combined in a
common source because the data may come from several resources. The combination of
different sources of data is needed so that data is integrated. Issues of data integration
include identifying similar entities, removing redundancy, detecting and removing
conflicts and errors.

Data selection: At this step, the data relevant to the analysis is decided on and retrieved
from the data collection.

Data transformation: Also known as data consolidation, it is a phase in which the

selected data is transformed into an appropriate forms for the mining process.

15
Data mining: It is an essential process where intelligent techniques are applied to extract
potentially useful patterns.

Pattern evaluation: Is a step in which strictly interesting patterns representing

knowledge are evaluated based on given measures.

Knowledge representation: Is the final phase in which the discovered knowledge is

visually represented to the user. This essential step uses visualization techniques to help
users understand and interpret the data mining results.

Figure 2.1 Knowledge discovery steps (Source: Fayyad et al. 1996)

2.4 Data Mining and Related Fields

2.4.1 Data Mining and Data Warehousing

Before starting any kinds of work on data mining it is better to bring all the data together
because in real business environment the data is found in different departments.
However, integrating data from different sources is not an easy task because different
departments have been used different styles of record keeping, different conventions,
different time period, different degree of data aggregation, and there have been different
kinds of errors. A data warehouse is a repository/storage area of data collected from
multiple heterogeneous data sources and is intended to be used as a whole under the same
unified schema (Han and Kamber 2006). A data warehouse is also defined as a subject-
oriented, integrated, time-variant, and nonvolatile collection of data in support of
management‟s decision making process (Two Crows Corporation 1999). Data warehouse

16
construction can be viewed as an important preprocessing step for data mining because
data warehouse involves data cleaning, data integration and data transformation. A data
warehouse gives the option to analyze data from different sources under the same roof.

A data warehouse is not a requirement for data mining. However, setting up a large data
warehouse that consolidates data from multiple sources resolves data integrity problems
(Two Crows Corporation 1999).

As it is described by Two Crows Corporation (1999), frequently the data to be mined is

first extracted from an enterprise data warehouse into a data mining database or data
mart. There is some real benefit if data is already part of a data warehouse. For example,
the problems of cleansing data for a data warehouse and for data mining are very similar.
So, if the data has already been cleansed for a data warehouse, then it most likely will not
need further cleaning in order to be mined. The data mining database may be a logical
rather than a physical subset of data warehouse, provided that the data warehouse DBMS
can support the additional resource demands of data mining. If it cannot, then it will be
better off with a separate data mining database.

2.4.2 Data Mining and DBMS

Database management system (DBMS) provides well ground to data mining to discover
knowledge from large databases. DBMSs provide a number of essential capabilities to
data mining, for example as a persistent storage, a consistent data model, and a high-level
query language, that allows users to request the required information. DBMS also provide
the integrity of data in the database by constraint enforcement and transaction
management. In addition, databases also provide a metadata description that can be used
to help understand the data which is to be mined.

2.4.3 Data Mining and OLAP

There exists confusion with the difference between data mining and OLAP (On Line
Analytical Processing). According to Han and Kamber (2006), OLAP is defined as the
dynamic synthesis, analysis and consolidation of large volumes of multidimensional data.

17
OLAP is a field which is part of the decision support tools. Traditional query and report
tools describe what is in a database but OLAP is more than this kind of activity; it is used
to answer why certain things are true. The user of OLAP forms a hypothesis about a
relationship and verifies it with a series of queries against the data.

According to Dunham (2000), OLAP tools can be classified as relational OLAP

(ROLAP), multidimensional OLAP (MOLAP) or hybrid OLAP (HOLAP). During
MOLAP data is modeled, viewed, and physically stored in a multidimensional database.
MOLAP tools are supported by specialized DBMS and software systems capable of
supporting the multidimensional data directly. With MOLAP, data is stored as n-
dimensional array, thus the cube view is stored directly. Although MOLAP has extremely
high storage requirements, indexes are used to speed up processing. On the other hand,
with ROLAP data is stored in a relational database and ROLAP server (middleware)
creates the multidimensional view for the user. Even though the ROLAP tools are less
complex, it is also less efficient at the same time. Hybrid OLAP combines the best
features of ROLAP and MOLAP together. Queries are stated in multidimensional terms.
Data which is not updated frequently will be stored as multidimensional database, while
data which is updated frequently will be stored as relational database.

Data mining is different from OLAP, because in OLAP the user should first verify
hypothetical patterns and use OLAP and use the data for support where as in data mining
rather than verify hypothetical patterns, it uses the data itself to discover such patterns. In
addition, OLAP systems focus on providing access to multidimensional data while data
mining systems deal with influence analysis of data along a single dimension (Two
Crows Corporation 1999).

2.4.4 Data Mining, Artificial Intelligence and Statistics

Data mining takes advantage of advances in the fields of artificial intelligence (AI) and
statistics. Both disciplines have been working on problems of pattern recognition and
classification and they have made great contributions to the understanding and
application of neural nets and decision trees (Two Crows Corporation 1999).

18
Statistics is very useful in providing a language and framework for quantifying the
uncertainty, which results when one tries to conclude general patterns from a particular
sample of an overall population. So, data mining does not replace these traditional
statistical techniques. Rather, it is an extension of statistical methods that is in part the
result of a major change in the statistics community. The development of most statistical
techniques was, until recently, based on elegant theory and analytical methods that
worked quite well on the modest amounts of data being analyzed. The increased power of
computers and their lower cost, together with the need to analyze enormous/ huge data
sets with millions of rows have allowed the development of new techniques. Traditional
statistical techniques rely on the modeler to specify the functional form and interactions.

The key point is that data mining is the application of these and other AI and statistical
techniques to common business problems in a way that makes these techniques available
to the skilled knowledge worker as well as the trained statistics professional (Two Crows
Corporation 1999).

2.5 Data Mining Functionality

The kinds of patterns which are found in data mining tasks are specified by data mining
functionalities. Generally, data mining tasks can be classified into two categories:
descriptive and predictive. Descriptive mining tasks such as association analysis,
clustering analysis, outlier analysis and so on, characterize the general properties of the
data in the database whereas predictive mining tasks like classification and prediction
perform inference on the current data in order to make predictions (Han and Kamber
2006).

Data mining functionalities can be applied to customers at all stages of the life cycle;
including customer identification and customer analysis, access to new customers and
upgrade existing customers value, and to maintain the valuable clients (Huang 2003).

19
 2.5.1 Concept/Class Description: Characterization and Discrimination

There is a difference between class and customer segmentation; class is the input to find
customer segment. Class, or concept description, is describing individual classes or
concepts in summarized, concise, and precise terms. These descriptions can be derived
via characterization and discrimination:

Characterization: Data characterization is a summarization of general features of

objects in a target class, and produces what is called characteristic rules (Han and
Kamber 2006). The data relevant to a user-specified class are normally retrieved by a
database query and run through a summarization module to extract the essence of the data
at different levels of abstractions.

Discrimination: Data discrimination produces what are called discriminate rules and is
basically the comparison of the general features of objects between two classes referred
to as the target class and the contrasting class. The techniques used for data
discrimination are very similar to the techniques used for data characterization with the
exception that data discrimination results include comparative measures.

2.5.2 Association Analysis

Association analysis is the discovery of what are commonly called association rules
(Witten and Frank 2005). It studies the frequency of items occurring together in
transactional databases, and based on a threshold called support, identifies the frequent
item sets. Another threshold, confidence, which is the conditional probability that an item
appears in a transaction when another item appears, is used to locate association rules.
Association rules mining try to find interesting association or correlation relationship
among a large set of data items. Association analysis is commonly used for market basket
analysis (Bounsaythip and Rinta-Runsala 2001).

2.5.3 Classification and Prediction

Classification analysis: - Also known as supervised classification, classification analysis

is the process of finding a model (or function) that describes and distinguishes data
20
classes or concepts, for the purpose of being able to use the model to predict the class of
objects whose class label is unknown (Han and Kamber 2006). By using classification it
is possible to organize data in a given classes. The classification uses given class labels to
order the objects in the data collection.

Classification approaches normally use a training set where all objects are already
associated with known class labels. Then the classification algorithm learns from the
training set and builds a model. The model is used to classify new objects. In other words
classification is a two-step process first, a classification model is built based on training
data set and then the model is applied to new data for classification (Huang 2003).

Prediction: - There are two major types of predictions: numeric prediction and class
label prediction. The first type of prediction that is numeric prediction predicts some
unavailable data values or pending trends, and the second type of prediction (the one
which is tied to classification) predicts a class label for some data.

Prediction has attracted considerable attentions by giving the potential implications of

successful forecasting in a business context. Once a classification model is built based on
a training set, the class label of an object can be forecasted based on the attribute values
of the object and the attribute values of the classes (Two Crows Corporation 1999).
Prediction is more often referred to the forecast of missing numerical values, or increase/
decrease trends in time related data. The major idea in prediction is to use a large number
of past values to consider probable future values.

2.5.4 Cluster Analysis

Clustering can be defined as the process of grouping a set of physical or abstract objects
into classes of similar objects (Han and Kamber 2006). Clustering is also called
unsupervised classification, because the classification is not dictated/ordered by given
class labels. There are many clustering approaches, all based on the principle of
maximizing the similarity between objects in a same class (intra-class similarity) and
minimizing the similarity between objects of different classes (inter-class similarity).

21
Clustering is similar to classification, but classes are not predefined and it is up to the
clustering algorithm to discover acceptable classes.

Often it is necessary to modify the clustering by excluding variables that have been
employed to group instances, because upon examination the user identifies them as
irrelevant or not meaningful. After clusters are found that reasonably segment the
database, these clusters then used to classify new data. Some of the common algorithms
used to perform clustering include Kohonen feature maps and K-means.

Some peoples are confused with the difference between clustering and segmentation;
however, clustering is different from segmentation. Segmentation refers to the general
problem of identifying groups that have common characteristics where as clustering is a
way to segment data into groups that are not previously defined (Two Crows Corporation
1999). Clustering is useful to find natural groups of data which is called clusters. A
cluster is a collection of data that are similar to one another. Clustering can be used to
group customers with similar behavior and to make business decisions in industry (Huang
2003).

2.5.5 Outlier Analysis

Outliers are data elements in the database that cannot be grouped in a given class or
cluster. In other words, outliers are data objects that don‟t fulfill the general behavior or
model of the data. Outliers are also known as exceptions or surprises, they are often very
important to identify. Although outliers can be considered as noise and discarded in some
applications, they can produce important knowledge in other domains, so their analysis is
very significant.

2.5.6 Evolution and Deviation Analysis

Evolution and deviation analysis is relevant to the study of time related data that changes
in time. Evolution analysis models evolutionary trends in data, which permit to
characterizing, comparing, classifying or clustering of time related data. Deviation

22
analysis, on the other hand, considers differences between measured values and expected
values, and attempts to find the cause of the deviations from the anticipated values.

2.6 Data Mining Techniques and Algorithms

2.6.1 Clustering Technique

Clustering is a tool for data analysis, which solves classification problems. Its objective is
to distribute cases (people, objects, events etc.) into groups, so that the degree of
similarity can be strong between members of the same cluster and weak between
members of different clusters (Hajizadeh et al. 2010). In clustering, there is no pre-
classified data and no distinction between independent and dependent variables. Instead,
clustering algorithms search for groups of records (the clusters composed of records
similar to each other). The algorithms discover these similarities.

According to Faber (1994), clustering involves dividing a set of data points into non-
overlapping groups, or clusters, of points, where points in a cluster are more similar to
one another than to points in other clusters. When a dataset is clustered, every point is
assigned to some cluster, and every cluster can be characterized by a single reference
point, usually an average of the points in the cluster. Any particular division of all points
in a dataset into clusters is called a partitioning.

Clustering algorithms are used when there have been a large complex dataset with many
variables and lots of internal structures. Clustering algorithms are often used to find
outliers or records that don‟t fit the predictive model. Many clustering algorithms have
been developed including K-means (K-nearest neighbors), hierarchical, fuzzy C-means
and a special type of neural network called Self Organized Maps (SOM) or kohonen net
maps (Bounsaythip and Rinta-Runsala 2001).

2.6.1.1 The K-Means Algorithm

K-means clustering, which was developed by MacQueen in 1967, is the most frequently
cited algorithm for database applications (Trappey et al. 2009). Since K-means algorithm
is simple it is used in various fields.

23
K-means is a partition clustering method that separates data into K mutually excessive
groups. By iterating such partitioning, K-means minimizes the sum of distance from each
data to its clusters. K-means method is very popular because of its ability to cluster huge
data, and also outliers, quickly and efficiently. Even though, K-means algorithm is very
sensitive in initial starting points K-means generates initial cluster randomly. When
random initial starting points close to the final solution, K-means has high possibility to
find out the cluster center. Otherwise, it will lead to incorrect clustering results. Because
of initial starting points generated randomly, K-means does not guarantee the unique
clustering results (Arai and Barakbah 2007).

Arai and Barakbah (2007) explain that the K-means method is a well known geometric
clustering algorithm. Given a set of n data points, the algorithm uses a local search
approach to partition the points into k clusters. A set of k initial cluster centers is chosen
arbitrarily. Each point is then assigned to the center closest to it, and the centers are
recomputed as centers of mass of their assigned points. This is repeated until the process
stabilizes. In simple K-means clustering, there is no partitioning a single point twice, so
the algorithm is guaranteed to terminate.

K-means clustering generates a specific number of disjoint, flat (non-hierarchical)

clusters. It is well suited to generating globular clusters especially for large or huge
datasets. Since the K-means clusters are non-hierarchical they do not overlap. With a
large number of variables, K-means may be computationally faster than hierarchical
clustering (if K is small). K-means may produce tighter/rigid clusters than hierarchical
clustering, especially if the clusters are globular.

2.6.1.2 Self Organized Maps (SOM)

SOM or kohonen feature map, is a special kind of network architecture that provides a
mapping from the multi-dimensional input space to a lower order regular lattice of cells
(typically two dimensional grids). SOM is important in multi-dimensional inputs, because
when the set of input is multi-dimensional, traditional clustering algorithms do not offer
an easy way to visualize the closeness of other clusters (Bounsaythip and Rinta-Runsala

24
2001). Such a mapping is used to identify clusters of elements that are similar in the
original space.

Unlike other neural network approaches, the SOM network performs unsupervised
training. The most common approach to neural networks requires supervised training of
the networks, i.e, the network is fed with a set of training cases and the generated output
is compared with the known correct output. The SOM network, on the other hand, does
not require the knowledge of the corresponding outputs. The nodes in the network
converge to form clusters to represent groups of entities with similar properties. The
number and composition of clusters can be visually determined based on the output
distribution generated by the training process.

The SOM network is typically a feed-forward neural network with no hidden layer; it has
two layers of nodes, the input layer and the Kohonen layer. The input layer is fully
connected to a two or one or multi-dimensional Kohonen layer. In effect multiple outputs
are created and the best one is chosen (Dunham 2000).

2.6.2 Classification

Classification is one of the most common learning models in data mining techniques and
it aims at building a model to predict future customer behaviors through classifying
database records into a number of predefined classes based on certain criteria (Ngai et al.
2009). According to Dunham (2000), classification is viewed as a mapping from the
database to the set of classes. Classes which produced by classification are predefined,
none overlapping and partition the entire database. There are several classification
techniques including decision trees and neural networks. Classification approaches
normally use a training set where all objects are already associated with known class
labels. The classification algorithm learns from the training set and builds a model. The
model is used to classify new objects. This pattern can be used both to understand the
existing data and to predict how new instances will behave (Two Crows Corporation
1999).

25
Issues in Classification

According to Dunham (2000), there are different issues when using classification
techniques. The most common ones are the following:

Missing data: One of the most common classification issues is missing data values,
which cause problems during both the training phase and the classification process itself.
Missing values in the training data has to be handled and may produce an inaccurate
result. There are many approaches to handle missing data such as, ignore the missing
data, assume a missing value for the data (this may be determined by using some method
to predict what the value could be) and assume a special value for the missing data (this
means that the missing data is taken to be a specific value all of its own).

Measuring performance: The other issue in classification techniques is performance

measurement. The performance of classification algorithms is usually examined by
evaluating the accuracy of the classification. However, since classification is often a
fuzzy problem, the correct answer may depend on the user. Classification accuracy is
usually calculated by determining the percentage of tuples placed in the correct class.

2.6.2.1 Decision Tree

Decision tree algorithm is a data mining induction technique that recursively partitions a
dataset of records using either depth-first greedy approach or breadth-first approach until
all the data items belong to a particular class (Anyanwu and Shiva n.d). Decision trees are
a way of representing a series of rules that lead to a class or value (Two Crows
Corporation 1999) and it is a powerful and popular tool for classification and prediction.
In data mining, a decision tree is a predictive model which can be used to represent both
classifiers and regression models. Decision trees are also useful for exploring data to gain
sight into the relationships of a large number of candidate input variable to the target
variable. Since decision trees combine both data exploration and modeling, they are a
powerful first step in modeling process even when building the final model using some
other techniques. When a decision tree is used for classification tasks, it is more

26
appropriately referred to as a classification tree and when it is used for regression tasks, it
is called regression tree.

The attractiveness of decision trees is due to the fact that, in contrast to neural networks,
decision trees represent rules. Rules can readily be expressed so that humans can
understand them or even directly used in a database access language like SQL so records
falling into a particular category may be retrieved.

Hajizadeh et al. (2010) also described decision tree as it is a model which consists of a set
of rules for dividing a large heterogeneous population into smaller, more homogeneous
groups with respect to a particular target variable. The target variable is usually
categorical and the decision tree model is used either to calculate the probability that a
given record belongs to each of the categories, or to classify the record by assigning it to
the most likely class.

According to Anyanwu & Shiva (n.d), decision tree classification technique has tree
building and tree pruning phases. Tree building is a top-down approach in which the tree
is recursively partitioned until all the data items belong to the same class label. The tree
growing or building phase is an iterative process which involves splitting the data into
progressively smaller subsets (Bounsaythip and Rinta-Runsala 2001). Tree-building
algorithms usually have several stopping rules. These rules are usually based on several
factors including maximum tree depth, minimum number of elements in a node
considered for splitting, or it‟s near equivalent, the minimum number of elements that
must be in a new node (Bounsaythip and Rinta-Runsala 2001). Tree pruning, on the other
hand, is a bottom-up fashion. It is used to improve the prediction and classification
accuracy of the algorithm by minimizing over-fitting (noise or much detail in the training
data set); over-fitting in decision tree algorithm results in misclassification error
(Dunham 2000). There are two methods of tree pruning; the first type of pruning is pre-
pruning which involves try to decide during the tree-building process when to stop
developing sub trees; this technique involves look ahead during tree generation. The
second type of pruning is post-pruning which involve heuristics to determine which
branch to prune or using cross-validation to get a better estimate on good prunes. Sub tree

27
replacement and sub tree raising are examples of post-pruning methods (Witten and
Frank 2005).

The disadvantage of many classification techniques like Bayesian classifier is that the
classification process is difficult to understand. However, humans do easily understand
and accept decision rules. So, by using decision tree method it is possible to make good
decisions especially decisions that involve high costs and risks.

2.6.2.1.1 Decision Tree Algorithms

According to Bounsaythip and Rinta-Runsala (2001), the most commonly implemented

decision tree algorithms which are well suited for classification techniques include Chi-
squared Automatic Interaction Detection (CHAID), Classification and Regression Trees
(CART), C4.5 and C5.0.

CART: - This is a technique which generates a binary decision tree. CART uses binary
split based on GINI (recursive partitioning motivated by statistical prediction) techniques
and in CART algorithm there exists exactly two branches from each non-terminal node.
In CART algorithm pruning is performed based on measure of complexity of the tree.

C4.5 and C5.0: - The decision tree algorithm C4.5 is an improvement to ID3 and it
produce tree with multiple branches per node (Dunham 2000). The number of branches is
equal to the number of categories of predictor. C4.5 combines multiple decision trees into
a single classifier. In C4.5 pruning is performed based on error rate at each leaf. C4.5
algorithm has also the ability to handles missing and continuous data. C5.0 is a
commercial version of C4.5. Unlike C4.5 the precision algorithm used for C5.0 have not
be disclosed.

CHAID: - This is a multi-way splitting algorithm using chi-square tests (detection of

complex statistical relationship). The number of branches varies from two to the number
of predictor categories.

28
 2.6.2.2 Neural Network Algorithm

According to Singh and Chauhan (2009), neural networks are defined as a non-linear
statistical data modeling tools that can be used to model complex relationships between
inputs and outputs or to find patterns in data. The neural networks approach, like decision
trees, requires that a graphical structure be built to represent the model and then that the
structure be applied to the data (Dunham 2000). According to Dunham (2000), a neural
network can be viewed as a directed graph with source (input), sink (output), and internal
(hidden) nodes. The hidden nodes may exist over one or more hidden layers. To perform
the data mining tasks, a tuple is input through the input nodes and the output node
determines what the prediction is. Unlike decision tree which have only one input node
(the root of the tree), a neural net has one input node for each attribute value to be
examined to solve the data mining function. Unlike decision trees, after a tuple is
classified, the neural network may be changed to improve future classification
applications. Although the structure of the graph does not change, the labeling of the
nodes and edges may change.

It is possible to collect important information from data warehouse in the process of data
mining by using neural networks as a tool. Neural networks essentially consist of three
pieces: the architecture or model; the learning algorithm; and the activation functions.
Neural networks are programmed or trained to store, recognize, and associatively retrieve
patterns or database entries; to solve combinatorial optimization problems; to filter noise
from measurement data; to control ill-defined problems; in summary, to estimate sampled
functions when it is difficult to know the form of the functions (Singh and Chauhan
2009).

It is possible to successfully applied neural networks in both supervised and unsupervised

learning applications. In data mining tasks we can commonly use neural network
methods, because neural network often produce clear and understandable models. A
neural network is a computational technique which is similar to what the human brain
works. It is designed to imitate or copy the ability of the human brain to process data and
information and understand or realized patterns. It imitates the structure and operations of

29
the three dimensional web of network among brain cells (nodes or neurons, and hence the
term „„neural‟‟) (Hajizadeh et al. 2010).

The major drawback to the use of neural networks is the fact that they are difficult to
explain to the end users (unlike decision trees which are easy to understand), it does not
generate rules like decision trees; thus, no rules are derived for the model and unlike
decision trees, neural networks usually work only with numerical data (Dunham 2000).

Neural networks except SOM can work well with noisy data and data with missing some
values. Neural networks are good tool for predictions and classification. When there are
lots of input features select only important features for the training phase by using other
methods (e.g associations) (Bounsaythip and Rinta-Runsala 2001).

Figure 2.2 Diagram of a typical neural network (Source: Rygielski et al. (2002))

2.7 Applications of Data Mining

Data mining has so many applications in reduced costs of doing business, improved
profitability, or enhanced quality of service for different fields. Areas in which such
benefits have been demonstrated includes Insurance, Direct Mail Marketing,

30
Telecommunications, Retail, Medicine, CRM and others (Huang 2003). Data mining
applications have been shown to be highly effective in addressing many important
business problems. Each business is interested in predicting the behavior of its customers
through the knowledge gained in data mining

Application of data mining in medical science: In healthcare data mining is applicable

in many different ways. These applications can generally group as the evaluation of
treatment effectiveness, management of healthcare, CRM, and detection of fraud and
abuse (Koh and Tan n.d). Diagnosis of disease, health care, patient profiling and history
generation etc. are the few examples of application of data mining in medical science.

Mammography is the method used in breast cancer detection. Radiologists face lot of
difficulties in detection of tumors/swelling. Computer-aided methods could assist medical
staff and improve the accuracy of detection (Deshpande and Thakare 2010). The neural
networks with back-propagation and association rule mining, used for tumor
classification in mammograms.

Koh and Tan (n.d) describe the applications of data mining in healthcare as follows:

 Treatment effectiveness: Data mining applications can be used to assess the

effectiveness of medical treatments. By comparing and contrasting causes,
symptoms, and courses of treatments, data mining can deliver an analysis, of
which courses of action verify effective (Koh and Tan n.d). For example, the
outcomes of patient groups treated with different drug regimens for the same
disease or condition can be compared to determine which treatments work best
and are most cost-effective.
 Healthcare management: Identifying chronic disease states and high-risk
patients, design appropriate interventions, and reduce the number of patients and
claims are an example of data mining applications that assist healthcare
management.
 CRM: While CRM is a core approach in managing interactions between business
organizations such as banks and retailers and their customers, it is also important

31
 in a healthcare context. Customer interactions may occur through reservation,
physicians‟ offices, inpatient settings, and ambulatory care settings.
 Fraud and abuse: Data mining applications that attempt to detect fraud and
abuse often establish norms and then identify unusual or abnormal patterns of
claims by physicians, laboratories, clinics, or others (Koh and Tan n.d).

Application of data mining in distance learning: Data mining is useful in distance

learning to enhance the learning process based on the vast amount of data generated by
the tutors and student‟s interactions with web based distance-learning environment
(Deshpande and Thakare 2010). The data mining applications transfers the data into
information and feedback to the e-learning environment. This solution transforms large
amounts of useless data into an intelligent monitoring and recommendation system
applied to the learning process.

Application of data mining in anomaly detection: Anomaly detection in the network is

very difficult and needs a very close watch on the data traffic. The intrusion detection
plays an essential role in computer security. The classification method of data mining is
used to classify the network traffic in to normal traffic or abnormal traffic (Deshpande
and Thakare 2010). The classification is done by if any TCP header does not belong to
any of the existing TCP header clusters, then it can be considered as anomaly.

Application of data mining in retail: Through the use of store-branded credit cards and
point-of-sale systems, retailers can keep detailed records of every shopping transaction.
This enables them to better understand their various customer segments. According to
Rygielski et al. (2002), some of the data mining application for retail includes:

 Performing basket analysis: Basket analysis also known as affinity analysis,

reveals which items customers tends to purchase together. This analysis can
improve stocking, store layout strategies, and promotions.
 Sales forecasting: Examining time-based patterns helps retailers make stocking
decisions. It used to analyze if a customer purchases an item today, when are they
most likely to purchase a complementary item.

32
  Database marketing: Retailers can develop profiles of customers with certain
behaviors, for example, those who purchase designer labels clothing or those who
attend sales. This information can be used to focus cost-effective promotions.
 Merchandise planning and allocation: When retailers add new stores, they can
improve merchandise planning and allocation by examining patterns in stores
with similar demographic characteristics. Retailers can also use data mining to
determine the ideal layout for a specific store.

Application of data mining in telecommunications: The telecommunications industry

was an early adopter of data mining technology and therefore many data mining
applications exist. Generally there are three data mining application areas in
telecommunication industries these are fraud detection, marketing/customer profiling and
network fault isolation.

Telecommunications companies around the world face increasing competition which is

forcing them to aggressively market special pricing programs aimed at retaining existing
customers and attracting new ones. Knowledge discovery in telecommunications mainly
include the following (Rygielski et al. 2002):

 Call detail record analysis: Telecommunications companies accumulate detailed

call records. By identifying customer segments with similar use patterns, the
companies can develop attractive pricing and feature promotions.
 Customer loyalty: Some customers repeatedly switch providers, or churn, to take
advantage of attractive incentives by competing companies. The companies can
use data mining to identify the characteristics of customers who are likely to
remain loyal once and those who switch their providers‟ frequently, thus enabling
the companies to target their expenditure on customers who will produce the most
profit.

Application of data mining in banking: According to Rygielski et al. (2002), banks can
utilize knowledge discovery for various applications, including:

33
  Card marketing: By identifying customer segments, card issuers and acquirers
can improve profitability with more effective acquisition and retention programs,
targeted product development, and customized pricing.
 Card holder pricing and profitability: Card issuers can take advantage of data
mining technology to price their products so as to maximize profit and minimize
loss of customers. Includes risk-based pricing.
 Fraud detection: Fraud is enormously costly. By analyzing past transactions that
were later determined to be fraudulent, banks can identify patterns.
 Predictive life-cycle management: Data mining helps banks predict each
customer‟s lifetime value and to service each segment appropriately (for example,
offering special deals and discounts).

2.7.1 Application of Data Mining for CRM

The application of data mining tools in CRM is an emerging trend in the global economy.
Analyzing and understanding customer behaviors and characteristics is the foundation of
the development of a competitive CRM strategy, so as to acquire and retain potential
customers and maximize customer value. Appropriate data mining tools, which are good
at extracting and identifying useful information and knowledge from enormous customer
databases, are one of the best supporting tools for making different CRM decisions (Ngai
et al. 2009). Many organizations are using data mining to help manage all phases of the
customer life cycle, including acquiring new customers, increasing revenue from existing
customers, and retaining good customers (Two Crows Corporation 1999).

Customer identification: Customer identification which is also known as customer

acquisition is the beginning of CRM. The main target of this phase is to identify people
who are most likely to become customers or most profitable to the company. Elements
for customer identification include target customer analysis and customer segmentation.
Target customer analysis involves seeking the profitable segments of customers through
analysis of customers‟ underlying characteristics, whereas customer segmentation
involves the subdivision of an entire customer base into smaller customer groups or

34
segments, consisting of customers who are relatively similar within each specific
segment.

Customer attraction: This is the phase following customer identification. After

identifying the segments of potential customers, organizations can direct effort and
resources into attracting the target customer segments. An element of customer attraction
is direct marketing. Direct marketing is a promotion process which motivates customers
to place orders through various channels.

Customer retention: This is the central concern for CRM. Customer satisfaction, which
refers to the comparison of customers‟ expectations with his or her perception of being
satisfied, is the essential condition for retaining customers. As such, elements of customer
retention include one-to-one marketing, loyalty programs and complaints management.

One-to-one marketing refers to personalized marketing campaigns which are supported

by analyzing, detecting and predicting changes in customer behaviors. Loyalty programs
involve supporting activities which aim at maintaining a long term relationship with
customers. Specifically churn analysis, credit scoring, service quality or satisfaction
forms are part of loyalty programs.

Customer development: This involves consistent expansion of transaction intensity,

transaction value and individual customer profitability. Elements of customer
development include customer lifetime value analysis, up/cross selling and market basket
analysis. Customer lifetime value analysis is defined as the prediction of the total net
income a company can expect from a customer.

Application of data mining in customer segmentation: All industries can take

advantage of data mining to discover discrete segments in their customer bases by
considering additional variables beyond traditional analysis. That means by incorporating
behavioral components into traditional segmentation techniques, organizations can target
and market customers far more effectively. Behavioral attributes which used to segment
customers includes product usage, buying cycle, purchase history, online activity,
responsiveness to marketing materials and others.

35
Substantial research works in many other data mining applications has been done abroad.
In the context of Ethiopia, data mining applications for CRM has been done by Henok
(2002) and Denekew (2003) in Ethiopian Airlines. They performed their study by using
clustering and classification techniques; for the case of clustering they use K-means
algorithm, on the other hand, for classification purpose they prefer decision tree
algorithm. The other research in CRM was conducted by Fekadue (2004) and Melaku
(2009) in Ethiopian Telecommunications Corporation; they also conducted their research
by using clustering and classification data mining techniques. Still another research was
done by Kumnegere (2006) in Ethiopian Shipping Lines. She has also used clustering and
classification techniques in her study, at the same time she selected K-means and decision
tree algorithms. Tilahun (2009) has also conducted a research on application of data
mining in the area of Banking Industry by using the same techniques.

There are also research projects done abroad on revenue and customs authority CRM,
such as Mahler and Hennessey (1996), Doye (2010) and Boulding et al. (2005).

36
 CHAPTER THREE

CUSTOMER RELATIONSHIP MANAGEMENT AND

CUSTOMER SEGMENTATION

3.1 Loyalty and Customer Relationship Management

3.1.1 Overview

Customer relationship management (CRM) has become one of the leading business
strategies in the new millennium. Nowadays, companies are changing their business
process models and build information technology (IT) solutions that enable the
acquisition of new customers, retain existing ones, and maximize the customers‟ lifetime.

It is difficult to find out a totally approved definition of CRM. However, there are two
approaches to define CRM, i.e. from the perspective of management and IT approaches
(Bose 2002, Wahab and Ali 2010). In management approach, CRM refers to an
integrated approach to identifying, acquiring, and retaining customers (Bull 2003, Bose
2002). On the other hand, CRM on IT approach refers to the tools or system design to
support the relationship strategy activities such as identifying, acquiring, and retaining
customer (Farn and Huang 2009).

CRM is an enterprise approach to understand and influence customer behavior through

meaningful communications in order to improve customer acquisition, customer
retention, customer loyalty, and customer profitability (Kim et al. 2003). By using
customer satisfaction as a principle of CRM, the modern marketing promotes a customer
orientation (Parvatiyar and Sheth 2001). As Trappey et al. (2009) explained, business
enterprises understand the value of customers, target the most profitable customers, and
build high-quality relationships that increase loyalty and profits by the help of CRM;
because the most crucial elements for the success of CRM are targeting the most
profitable customers and evaluating customer profitability. In addition, organizations are
realizing the need for in-depth and integrated customer knowledge in order to build close

37
cooperative and partnering relationships with their customers (Parvatiyar and Sheth
2001).

According to Parvatiyar and Sheth (2001), CRM is the creation of superior value for both
the company and the customer by the comprehensive strategy and process of acquiring,
retaining, and partnering with selective customers. To achieve a better efficiencies and
effectiveness in delivering customer value, CRM involves the integration of marketing,
sales, customer service, and the supply-chain functions of the organization.

CRM is also viewed as the strategy of transforming enterprises to become customer

centric while still expanding revenue and profit is one of the main strategies in business
today (Kim et al. 2003). Many business enterprises, therefore, realize the importance of
CRM and the potential of these techniques to achieve and sustain a competitive
advantage. To realize CRM success, business and IT executives should implement CRM
processes and technologies and promote employee behavior that supports coordinated
and more effective customer interactions throughout all customer channels (Kim et al.
2003). Therefore, one of the most important processes of CRM is extracting valid,
previously unknown, and comprehensible information from a large database and using it
for profit.

CRM is illustrated as a combination of people, processes, and technology which provide

understanding of customer needs, business strategy support, and build long-term
relationships with customers (Shang and Chen n.d). As a result, companies make every
effort to deliver the highest value to customers through better communication,
customized promotions, faster delivery, and personalized products and services (Trappey
et al. 2009). To effectively address human behavioral elements appropriate business
processes and organizational culture are also required for successful utilization of the
integrated technology (Shang and Chen n.d).

CRM is also described as a customer-focused business strategy that dynamically

integrates sales, marketing and customer care service in order to create and add value for

38
the company and its customers (Chalmeta 2006). As Chalmeta (2006) explained, CRM
systems basically make the following things possible:
 CRM enables to have an integrated, single view of customers by using analytical
tools
 Managing customer relationships in a single way, regardless of the
communication channel such as telephone, website, personal visit, and so on,
 Improving the effectiveness and efficiency of the processes involved in customer
relationships.

Generally, nowadays companies are becoming more and more aware of the many
potential benefits provided by CRM. According to Kim et al. (2003), some of the
potential benefits of CRM are the following:

 Increased customer retention and loyalty

 Higher customer profitability
 Creation of value for the customer
 Customization of products and services
 Lower process, higher quality products and services

3.1.2 Customer Loyalty

The importance of creating and maintaining customer loyalty is not new. It is a common
belief among businesses that it costs more to find a new customer than to keep and grow
an existing one. However, recent studies indicate that in spite of heavy investments in
customer satisfaction efforts and rewards programs, loyalty remains an intangible goal in
almost every industry (Mc Kinsey & Company 2001).

According to Hwang et al. (2004), customer loyalty can be defined as the index that
customers would like to remain as customers of a company. That is:

Customer Loyalty = 1 – Churn Rate.

As Hwang et al. (2004) explained, churn describes the number or percentage of regular
customers who discard relationship with a service provider. So, by using customer loyalty
39
it is possible to measure customer retention. As Farn and Huang (2009) explained, there
are two fundamental methods to retain customers. One is to improve customers‟
satisfaction and then motivate their true attitudinal and behavioral loyalty, so that
customers do not be diverted by competitive offering. The other is to increase switching
barriers to discourage customers from using other suppliers, such as switching cost and
changing to alternatives as disincentives. To retain customers, understanding the value of
customers and the most profitable customers are also essential. Level of customer
retention can be derived from churn rate. It is significant for customer cultivation and
retention to consider the churn rates (Hwang et al. 2004).

3.1.3 Principles and Tasks of CRM

3.1.3.1 Principles of CRM

According to Gray and Byun (2001), personalization, loyalty and lifetime value are the
three main principles of CRM implementation. These principles have the benefit of
improving the organization‟s ability to retain and acquire customers, maximize the
lifetime value of each customer and improve services without increasing cost of service.

Personalization (treat customer individually): Personalization deals about treating

customers individually so that the content and services to customer should be designed
based on customer preferences and behavior (Adomavicius and Tuzhilin 2001).

Acquire and retain customer loyalty through personal relationship: Once

personalization is takes place, then a company needs to maintain relationships with the
customer. Continuous contacts with the customer intern can create customer loyalty.

Select good customer instead of bad customer based on lifetime value: This principle
mainly focuses on finding and keeping the right customers who generate the most profits
for the organization. According to this principle the best customers earn the most
customer care and the worst customers should be dropped.

40
 3.1.3.2 CRM Tasks

The database marketing approach is highly company centric; however, customers were
not kept loyal by the discount programs and the one-time promotions that were used in
the database-marketing programs (Gray and Byun 2001). So, there is a need to follow/use
the CRM approach which is customer-centric; because it is better to gain customer
loyalty in customer-centric approach than in database marketing approach (Gray and
Byun 2001). Rather than based on what the company wants to sell, the customer centric
approach of CRM focuses on the long-term relationship with the customers by providing
the customer benefits and values from the customer‟s point of view. So, to achieve these
basic goals CRM requires the following four basic tasks (Gray and Byun 2001,
Kumnegere 2006). These are:

Customer identification: It refers to the selection and or knowing of customers through

marketing channels, transactions, and interactions over time for the purpose of serving or
providing value to the customer (Berndt et al. 2005).

Customer differentiation: This step refers to segmenting customers into different

perspective from the company‟s point of view; because, from the company's point of
view each customer has their own lifetime value.

Customer interaction: The main purpose of this step is to analyze the customer‟s
behavior over a long period of time in order to provide the right goods and services at the
right time because customer demands are sensitive and it change over time. From a CRM
perspective, the customer‟s long-term profitability and relationship to the company is
important. Therefore, the company needs to learn about the customer continually.

Customization / Personalization: The other goal of CRM is to treat each customer

uniquely so as to increase customer loyalty. Through this personalization process, the
company can also increase customer. Customization is carried out by the organization in
order to ensure that customer needs are met. It requires that the organization adapts its
product, service or communication in such a way that have something unique for each
customer. Communication can be customized to address the specific needs and profile the

41
customer, and organization also makes use of personalization as part of this process.
Products can be customized as to the specific desires that the customer has of the
organization (Berndt et al. 2005).

3.2 Customer Segmentation

3.2.1 Overview

Segmentation is the process of developing meaningful customer groups that are similar
based on individual explanations characteristics and behaviors (Trappey et al. 2009).
According to Bounsaythip and Rinta-Runsala (2001), segmentation also viewed as a way
to have more targeted communication with the customers; and the process of
segmentation describes the characteristics of the customers groups (called segments or
clusters) within the data.

Greengrove (2002) explained that there are two main segmentation approaches: the first
type of segmentation is the process of segmenting the customers based on understanding
the needs of the end user which is called needs-based segmentation. The second type of
segmentation, characteristics-based segmentation, is the process of segmenting customers
based on their characteristics, attitudes or behaviors.

Customer segmentation is defined as the practice of classifying customer base into

distinct groups (Farn and Huang 2009). In other words, customer segmentation is also
described as the process of dividing customers into homogeneous groups on the basis of
shared or common attributes (Bounsaythip and Rinta-Runsala 2001). The goal of
segmentation is to know the customer better and to apply that knowledge to increase
profitability, reduce operational cost, and enhance customer service. Segmentation can
provide a multidimensional view of the customer for better treatment strategy.

One of the challenging tasks during segmentation is to choose and derives appropriate
segmentation bases and variables in which the segmentation process is performed and
resulting segments are interpreted. These variables may be demographic, geographic,
psychographic or behavioral variables (Bounsaythip and Rinta-Runsala 2001).

42
Several segments may be formed by using customer profitability. For instance, the most
profitable segment consisting of the highest-profit customers should be retained through
loyalty and retention program. Another possible segment is the most unprofitable
customer group who generate more costs than profit (Kim et al. 2006).

3.2.2 Applications of Customer Segmentation

According to McGuirk (2007), customer segmentation has different applications for a

certain organization which is described as follows.

Making customer investment decisions: Segmentation provides a framework to help

identify the optimal customer investment strategy for each unique segment. For some
segments, the investment may be directed towards further developing customer
relationships, while for other segments the investment is made to introduce new products
and services that address unsatisfactory customer needs. Ultimately, the key factor
driving customer investment decisions has been the expected return on that investment.
Segmentation not only helps to determine how much to invest in a customer segment, but
how to spend it.

Managing customer relationships: Segmentation also provides an excellent framework

to manage the varied needs of customers. Customized customer management and
development strategies can be developed for each unique segment. The development
plans should include a set of objectives, goals and performance metrics that are derived
from the unique opportunities and challenges present within each customer group. The
segment-level plans function as a strategic roadmap, supporting business growth and
attempting to maximize the potential of each customer relationship.

Tailoring marketing programs: It is true that successful data-driven marketers

understand how to communicate with their customers at the right time, right place and
with the right message. Distinctive customer preferences and needs represent unique
opportunities and challenges that can be pursued by introducing tailored or modified
programs for each segment. The make-up of each customer segment and their past and
projected behaviors and needs should guide the tailored use of key marketing levers to

43
maximize program effectiveness. Segmentation becomes the focus, supporting program
development and ongoing test and learns activities.

Guiding product development and research: A comprehensive segmentation solution

has been emphasized the fact that individuals have different product needs and usage
patterns. Customers in one segment may use a company‟s full portfolio of products quite
frequently, while customers in another segment may only have a need for a single
product that is used at irregular intervals (infrequently). McGuirk (2007) further
explained, segments that contain less active customers often exposes opportunities to
strengthen and broaden these customer relationships by introducing new or re-packaged
products that meet a specific customer need. Segmentation provides the means to target
research and product development activities with the goal of further stimulating customer
demand.

3.2.3 Difficulties in Making Good Segmentation

As Bounsaythip and Rinta-Runsala (2001) explained, there are different factors which
make the segmentation process more complex such as:

Relevance and quality of data: Only relevance and good quality of data is needed
because if the company has inadequate or too much customer data, this can lead to
complex and time-consuming analysis. If the organizations data are also poorly organized
(different formats, different source systems) then it is also difficult to extract relevance
information. Furthermore, if the company‟s customer data is insufficient, too much or
poorly organized, the resulting segmentation can be very complicated for the organization
to implement it effectively.

Intuition: Although data can be highly informative, marketers need to be continuously

developing segmentation hypotheses in order to identify the right data for analysis.

Continuous process: Segmentation needs continuous development and updating as new

customer data is acquired. In addition, effective segmentation strategies have been
influenced the behavior of the customers affected by them; thereby necessitating revision
and reclassification of customers.
44
Over-segmentation: Segmentation can become too small and/or insufficiently distinct
to justify treatment as separate segments.

3.3 CRM in Ethiopian Revenue and Customs Authority

3.3.1 Overview

The Ethiopian Revenue and Customs Authority (ERCA) came into existence on 14 July
2008, by the merger of the Ministry of Revenue, Ethiopian Customs Authority and The
Federal Inland Revenue Authority, formerly responsible to raise revenue for the Federal
Government and to prevent contraband. According to Article 3 of the proclamation No
.587/2008 of the Ethiopia Federal Negarit Gazeta, the ERCA is looked upon as "an
autonomous federal agency having its own legal personality". The reasons for the merge
of the previous separately administrations into a single autonomous authority are, to
provide the basis for modern tax and customs administrations; to reduce or avoid
unnecessary and redundant procedures which causes delay and are considered cost-
inefficient; to become more effective and efficient in keeping and utilizing information,
promoting law and order, resource utilization and service delivery; and to transform the
efficiency of the revenue sector to a high level (ERCA official website ).

According to the experts, the newly designed tax and customs procedures have brought
about a positive result and it has produced beneficial effects to both the Authority and its
customers or stakeholders. However, to further improve the satisfaction of customers and
the beneficial effects of the Authority, there is a need to implement modern CRM in the
organization.

Objective of the ERCA

The ERCA has the following objectives:

 To establish or set up modern revenue assessment and collection system; and

provide customers with fair, efficient and quality service
 To cause taxpayers voluntarily discharge their tax obligations

45
  To enforce tax and customs laws by preventing and controlling contraband as well
as tax fraud and evasion
 To collect timely and effectively tax revenues generated by the economy
 To provide the necessary support to regions with a view to harmonizing federal
and regional tax administration systems.

Vision

The ERCA‟s vision is to see “fair and modern taxes and customs administration system
that enhances proper and effective revenue collection”.

Mission

“The ERCA shall promote the voluntary compliance of taxpayers, ensure integrity and
develop the skill of the employees, support the modernization and harmonization of the
taxes and customs administration system, contribute to economic development and social
welfare through effective revenue collection”.

3.3.2 Powers and Duties of the ERCA

According to the proclamation No. 587/2008 of the Ethiopia Federal Negarit Gazeta the
ERCA has the powers and duties to establish and implement modern revenue assessment
and collection system; the ERCA has also the powers and duties of provide efficient,
equitable and quality service within the sector; properly enforce incentives of tax
exemptions given to investors and ensure that such incentives are used for the intended
purposes.

In addition, ERCA has the responsibility to collect and analyze information necessary for
the control of import and export goods and the assessment and determination of taxes;
compile statistical data on criminal offences relating to the sector, and disseminate the
information to others respected sectors. Furthermore, the Authority provide information
and appropriate support to the Federal Police in the control of illegal trafficking of goods
and struggle contraband; and cause appropriate measures be taken in accordance with the

46
law. The powers of deciding the place where import and export goods are to be deposited
and establish are also given to the ERCA.

3.3.3 Organization of the ERCA

The head office of ERCA is found in Addis Ababa. The ERCA is led by a Director
General with the assistance of four Deputy Director Generals, namely Deputy Director
General for Corporate Functions Sector, Operations Sector, Change Management and
Support Sector, and Enforcement Sector. The appointment of both the Director General
and the Deputies are made by the Prime Minister of the country. At the headquarters, the
Authority has 19 business processes and 2 directorates namely Women's Affairs
Directorate and Ethics Directorate. According to the proclamation No. 587/2008 of the
Ethiopia Federal Negarit Gazeta, “business processes” is defined as “a class or some
class of work flow in which a serious and succeeding group of activities are performed
step by step by a case worker or a case team in a single cycle of performance to bring
about a result from the external or internal inputs and information utilized in the general
course of performance of the Authority”. The Authority‟s directorates and business
process are led by a Director. Among the directorates, two (Audit and Inspection
Business Process and Public Relations and Image Building Business Process) are
accountable to the Director General while the remaining are accountable to the four
Deputy Director Generals.

Apart from the 19 business processes and the 2 directorates at the headquarters level, the
Authority has 17 branch offices which can be divided as follows: 15 branch offices
located in a regional state or city administrations and 2 coordination offices located
outside of Ethiopia at the port of Djibouti and at the port of Burbera, Somalia. Each
branch office is directed by a manager who is accountable to the Director General of the
Authority.

47
 Figure 3.1 Organizational charts of the Ethiopian Revenue and Customs Authority

3.3.4 Complaints Handling in ERCA

Customer complaints are firstly a private communication between a company and a

customer who communicates his/her dissatisfaction on a defined matter. Customer
complaint is defined as the expression of a customer‟s dissatisfaction through various
possible channels such as letter, email, phone call, physical claim and others (ERGEG
2010). Complaints handling and managing it accordingly, i.e dealing with customers after
a service failure should be the corner stone of an organization‟s customer-satisfaction
strategy (Johnston and Mehra 2002). Customer complaint handling is a key instrument of
identifying the problem in the route of service providing and appropriate response to
complaints. The most important tasks in the customer services are designing and

48
implementing of customer‟s compliant handling system and identify problems that may
have been encountered and propose solutions to those problems. It is agreed that
customers make complaints now and then should be considered as loyal to the
organization rather than the sources of chaos (ERGEG 2010). Hence an organization
should pay due attention of these parties and implement systematic compliant handling
mechanism.

According to Johnston and Mehra (2002), the objectives of effective compliant handling
are:

 To meet customer‟s satisfaction

 To build trust and confidence
 To avoid unfavorable publicity
 To provide fair, transparent and timely service
 To ease burden in the day-to-day work of legal system and
 To inform management

There are different techniques of customer complaints handling systems such as, handling
complaints face-to face, handling complaints on the telephone and handling complaints in
writing. When contacts in customers are difficult let customers make contact by
telephone. During telephone customer complaints handling, make easy for customers to
find the number; established telephone procedure to ensure all calls are dealt with
punctually and train staff to be positive and friendly (ERGEG 2010).

In order to improve customer satisfaction, the ERCA tries to established customer

compliant handling systems. However, to make the current strategic thinking more
transparent and practical ERCA should implement modern CRM techniques.

49
 CHAPTER FOUR
EXPERIMENTATION

4.1 Overview

This chapter is the main part of the study, which used to successfully meet the objectives
of the research. To effectively achieve these research objectives, it is necessary to follow
appropriate research methodology (model). This research follows all the fundamental
steps of the Cios et al. (2000) KDD process model. Figure 4.1 indicates that Cios et al.
(2000) model consists of a cycle that comprises six stages: understanding of the problem
domain, understanding of data, preparation of the data, data mining, evaluation of the
discovered knowledge, and use of the discovered knowledge. The main reason the
researcher prefers the Cios et al. (2000) model to other data mining model is that the Cios
et al. (2000) model draws from both academic and industrial models and emphasizes
iterative aspects; the Cios et al. (2000) model has also an advantage to identify and
describe several explicit feedback loops.

Understanding of the
problem
Input Data (database,
Understanding of data images, video, semi
structured data, etc.)

Preparation of the data

Data mining

Knowledge (patterns, rules,

Evaluation of the clusters, classification,
discovered knowledge associations, etc.)

Use of the discovered Extend knowledge to

knowledge other domains

Figure 4.1 The six-step Cios et al. (2000) KDD process model

50
 4.2 Understanding the Problem Domain

As it is explained in Section 3.2.2, which mainly deals about application of customer

segmentation, there is a need to implement effective customer segmentation in the
Authority, in order to increase the profitability of the Authority and to raise the
satisfaction of the customers. As stated in Section 1.5, the researcher used different
techniques to understand the problem domain of this investigation.

Based on the experts explanation and evaluation made, there are different kinds of
customers in the Authority. Some customers are high value customers; these types of
customers are those customers who generate substantial revenue for the organization by
importing smaller amount of items and spending low invoice. On the other hand, there
are also customers who generate lower revenue for the Authority. These types of
customers are called low value customers. Low value customers import large amount of
items and spend higher invoice but they generate smaller amount of revenue for the
Authority; the type of goods which are imported by these low value customers were
cheap. The intermediate between high and low value customers are medium value
customers. Thus, the main aim of this research is to differentiate customers whether they
are high, medium and low value customers based on the revenue they generate and the
transaction of the data.

4.2.1 Data Mining Goals

The first data mining goal for this study was segmenting or clustering customers into
different groups based on their characteristics to treat or handle them accordingly. To
effectively accomplish this task, identifying important variables from the total data
collected was necessary. The identified variables were in turn used to build a model by
using clustering techniques. In addition, the identified variables were also used to better
understand the customers. The data preparation and data analysis phases help the
identification of important attributes that could serve as an input for the model building.

The other data mining goal for this research was building a classification model which
used to assign new records into the identified clusters. The classification model was done
51
by using the most appropriate data mining classification techniques which is decision tree
and neural network classification techniques and the one which performs better accuracy
is selected.

4.2.2 Data mining Tool Selection

Finding and selecting the most appropriate data mining tool with better capability is one
of the challenging tasks in data mining process. For this study, selection of an appropriate
data mining tool was done by first listing certain criteria. Among the criteria used for the
selection, the most important ones were:

 The speed and quality of the tool (performance)

 The data mining tasks that the tool is intended for (clustering, classification)
 The algorithm supported by the tool (K-means, decision tree, neural network)
 The number of records the tool can handle
 User friendliness
 Availability of the tool on the internet (freely downloaded )

By considering the above criteria, the researcher found different tools in the Internet,
such as Weka Version 3.6.4, Tanagra 1.4 and Sipina research Version 3.6. After though
evaluation of the tool as per the criteria the researcher selected Weka Version 3.6.4 and
more or less Weka fulfilled the above criteria. Weka Version 3.6.4 is better than the other
two in that it had better performance than both Tanagra and Sipina; in addition, Weka has
more facility for preprocessing of the data and number of algorithms than the remaining
ones.

4.3 Understanding the Data

Having defined the data mining goal, or after understanding of the problem to be
addressed and selecting an appropriate tool, the next step was to analyze and understand
the available customers‟ data. Therefore, to fulfill the data requirement, data was initially
collected regarding customers‟ behavior from ERCA‟s ASYCUDA database. Owning to
this, analysis of the data and its structure was done together with the Database

52
Administrator and other domain experts by evaluating the relationship of the data with
the problem at hand and the particular data mining task were dealt with the experts in the
Authority. In Cios et al. (2000) methodology during the data understanding step there are
subtasks which are considered by the users of the model. These are collection of initial
data, description of data and verification of data quality.

4.3.1 Collection of Initial Data

The primary step of data understanding is identifying the initial source of data that is used
for the data mining process. As indicated earlier, the major and initial data source for this
research was the ERCA‟s ASYCUDA database. The ERCA‟s ASYCUDA database is
very huge and rich data source because every declaration of import/export items and
other detail information of the customers are found in the ERCA‟s ASYCUDA database.
For instance, the customer/company name, the item type, total revenue paid, the price of
the item, total number of items imported, total amount of invoice spent to import the
items etc. of every customer are found in this database. The researcher collected all the
two year declaration of these customers‟ data which was around 110740 records. The
responsibility of administering or controlling the ERCA‟s ASYCUDA database is given
to the IT department of the Authority.

4.3.2 Description of the Data Collected

The researcher took a two-year transaction data for this study. Since every transaction of
goods within the country is registered in the ASYCUDA database, the amount of data
found is very huge but the researcher took only selected data. To select the data, the
researcher used the Weka preprocessing facilities which is stratified remove folds by
setting number of folds with two; and only around 46,748 records were taken with around
8 detail selected attributes. However, in the ERCA‟s ASYCUDA database there are a lot
of tables and many attributes. A sample data which is found from the initial data
collection is attached into this research as Appendix 1.

Among the different tables of the ASYCUDA database the researcher together with the
domain experts, selected the customer related tables and used the following once.

53
SAD_GEN TABLE (SAD General Segment Table): The SAD_GEN table contains all
information related to the SAD general segments. SAD means an abbreviation for single
administration document.

Attribute Data Type Description

SAD_REG_DATE Date Date that the SAD was register. (registration
date)
SAD_ITEM_TOTAL Small integer Total number of items imported
SAD_PACK_TOTAL Char Total number of packet
SAD_CONSIGNEE Char Company code
SAD_TOT_INVOICED Decimal Total amount of invoiced for entire
declaration
SAD_TYP_DEC Char Type of declaration (import or export)
SAD_CTY_IDLP Char If the SAD is an import SAD this field
indicates the country of last consignment. If
the SAD is for an export declaration, then this
is the field showing the country of first
destination.
SAD_CUR_COD Char Currency code indicates that used on the
invoice.
SAD_MOT_BORD Char Mode of transport at border.
SAD_WHS_COD Char If the goods are being warehoused (for re-
export, later home use etc) the warehouse
code is noted here.

Table 4.1 Attributes and description of the SAD general segment table

54
SAD_ITM TABLE: - The SAD_ITM TABLE contains all data about the declaration of
items.

Attribute Data Type Description

SADITM_ITM_PRICE Decimal Declaration per item price
SADITM_GOODS_DESC Char Description of the goods
SADITM_PACK_KNDCOD Char The code used to identify the kind of
package
SADITM_CTY_ORIGCOD Char Country of origin
SADITM_HS_COD Char Goods / commodity code
SADITM_GROSS_MASS Decimal Total weight of the imported item
SADITM_SUPP_UNITS Decimal Supplementary units
SADITM_PREV_DOC Char Previous document reference (for
temporary import for instance)
SADITM_ITM_TOTAMT Char Total amount of duties and taxes for
the item
SADITM_NUM Integer Declaration version (if the
declaration is amend)

Table 4.2 Attributes and description of the SAD_ITEM table

UNCMPTAB (Companies Table): - This table records all companies‟/customers‟

related data. As this table is not historized, ASYCUDA doesn‟t record all modifications
done on a company code but only keep the last updated value
Attributes Data type Description
CMP_COD Char Company code
CMP_NAM Char Company name
CMP_ADR Char Company address
CMP_TEL Char Telephone number
CMP_FAX Char Fax number
CMP_TLX Char Telex number
LST_OPE Char Last operation
SER_STA Integer Server status number

Table 4.3 Attributes and description of the company table

55
CSH_GEN_CDI (Cash Declaration Payment In Table)

This table contains information such as declarant reference and number of declaration
paid regarding cash payment in declarations. This is the master table for the multiple
rows to be inserted in CSH_DEC (declaration), in CSH_TRA (other payment), in
CSH_MEA (means of payment) and in CSH_OCC (if occasional exporter / importer).

Attributes Data type Description

RCP_YEAR Char Year of receipt
RCP_CUO Char Customs office code
RCP_SERIAL Char Number of the serial in which the receipt is
issued
RCP_NBER Integer Number of the receipt within the previous
serial
CDI_DEC_STM Integer Cash declaration statement
CDI_DECLARANT Char Code of SAD declarant
CDI_COMPANY Char Code of SAD consignee or exporter
CDI_DECL_TOT SmallInt Number of declaration on this receipt
CDI_AMOUNT_TOTAL Decimal Total amount paid on this receipt
CDI_RCPT_TIME Char Time when the receipt was issued
CDI_TOTAL_DECL Decimal Total amount paid for declarations on this
receipt

Table 4.4 Attributes and description of the cash declaration payment in table

Unctytab (country table): - This table contains information‟s about the country where
the item is imported.

Attributes Data type Descriptions

CTY_COD Char City code (ET)
CTY_DSC Char City description(Ethiopia)
LST_OPE Char City status (deleted or active)

Table 4.5 Attributes and description of the country table

56
 4.3.3 Data Quality Verification

Generally, the researcher was satisfied with the reliability of data and completeness of the
records. However, the collected data contains missing, incomplete and irrelevant data; for
instance, item price and total revenue columns contained missing or incomplete values.
Hence, revenue and item price column details which were missing or contain zero are
either handled the missing value or discarded them from the dataset.

4.4 Preparation of the Data

The main purpose of the data preparation phase is to make the data more suitable for the
next step which is the modeling (data mining in Cios et al. (2000) model). Starting from
the initial data collection there were a number of transformations performed until the
final dataset was found which is used for model building. In KDD process model, there
are different methods for data preprocessing such as data cleaning, data construction, data
integration and data formatting.

4.4.1 Data Cleaning

During the data cleaning phase, the researcher tried to handle missing values of the
attribute and remove some incomplete data which does not affect the entire dataset. From
the total dataset, only two attributes have missing values; these are ITEM_PRICE and
TOTAL_REVENUE. The researcher replaced the missing ITEM_PRICE attribute value
with the Weka mean/mode missing value replacement preprocess facility. When the total
revenue is missed or became zero the researcher learns from the Database Administrator
that these types of items are imported freely (without any taxation), as a result the
researcher discarded theses instances from the dataset; because, these types of instances
were not important for the data mining step. The researcher also discretizes one attribute
called SADITM_CTY_ORIGCOD; this attribute has around 80 values which is the name
of the country where the item is imported or originated. However, the researcher
discretizes this value into 6 which is in their continent of origin. MS-Excel is selected for
cleaning the data because the researcher is more familiar with it.

57
 4.4.2 Data Integration and Transformation

Data integration method for retrieving important fields from different tables and data
transformation method, which is attributes construction, are done in the effort to prepare
the data ready for the data mining techniques to be undertaken in this research.

After selecting different fields from different tables together with the domain experts, the
integration part which is bringing different attributes from different tables into a single
dataset was performed by together with the Database Administrator by using SQL
commands in the original ASCUDA database. The attributes were retrieved from the
above tables which were stated in the description of the data collected parts.

Most of the attributes which were used for clustering and classification model building
have been derived from the SAD_GEN and SAD_ITEM tables; because, these tables
contains the main attributes, which describe the characteristics of the items which
transfers around the border. Accordingly, many of the attributes such as registration date,
total number of items, total amount of invoice, item price, descriptions of the goods, and
total number of the pack, have been derived from these tables. In addition, the researcher
has also derived additional attributes which help the modeling process more efficient.
According to Saarenvirta (1998), data creation involves the creation of new variables by
combining existing variables to form ratios, difference and so forth. REVENUE_ITEM,
REVENUE_PRICE and INVOICED_REVENNUE are attributes which were derived
from the other original attributes. Accordingly, the derived attributes are then defined
using the existing attributes as follows:

TOTAL_REVENUE
REVENUE_ITEM =
TOTAL_ITEM

TOTAL_INVOICED
INVOICED_REVENUE =
TOTAL_REVENUE

TOTAL_REVENUE
REVENUE_PRICE =
ITEM_PRICE

58
 4.4.3 Data Formatting

At this step the researcher changes the data into a format which was suitable for the data
mining tool or algorithms. Since the preprocessing of the data was performed in MS-
Excel the final dataset was also in MS-Excel format; however, the selected tool Weka
3.6.4 doesn‟t accept the data in Excel format. So, the researcher first tried to convert the
data in comma delimited (CSV) text file in ARFF (Attribute Relation File Format)
format. Comma delimited applied for a list of records where the items are separated by
commas, whereas ARFF is an extension of a file format that the Weka software can read.
In addition, outliers and extreme values which may mislead the K-means algorithm were
also totally discarded from the dataset.

4.4.4 Attribute Selection

After the preparation of the final dataset the next step was to identify the best attributes
which is used for cluster modeling. To distinguish attributes which have high information
content, the researcher input all the selected attributes in the selected data mining tool
which is Weka 3.6.4. The following table shows attributes with their description after the
final preprocessed dataset. The dataset containing these attributes is fed into the data
mining software.

59
Attribute Data Type Description
CMP_COD Char Company code
CMP_ADR Char Company address
CMP_NAM Char Company name
SAD_REG_DATE Date Date that the SAD was register.
(registration date)
SAD_ITEM_TOTAL Integer Total number of items imported
SAD_TOT_INVOICED Decimal Total amount of invoiced for entire
declaration
SADITM_ITM_PRICE Decimal Declaration per item price
SADITM_GOODS_DESC Char Description of the goods
SADITM_PACK_KNDCOD Char The code used to identify the kind of
package
SADITM_CTY_ORIGCOD Char Country of origin
CDI_RCPT_DATE Date Date that the declaration was paid and the
original receipt issued.
SAD_PACK_TOTAL Char Total number of packet
SAD_CONSIGNEE Char Company code
SADITM_ITM_TOTAMT Char Total amount of duties and taxes for the
item
RCP_YEAR Char Year of receipt
REVENUE_ITEM Decimal Ratio of revenue to total number of items
REVENUE_PRICE Decimal Ratio of revenue to item price
INVOICE_REVENUE Decimal Ratio of invoice to revenue

Table 4.6 Selected attributes with their description

Once the attributes were fed in the data mining tool, the researcher evaluated the
information content of the attributes using the select attribute technique with
GainRationAttributeEval attribute evaluator and Ranker search method. So, the tool
arranged the attributes in order of their gain ratio; hence, the researcher together with the
domain experts removed those attributes which had less gain ratio.

In addition to evaluating the gain ratio of the attributes, Weka 3.6.4 explorer also enables
to know the minimum, maximum, mean, standard deviation, data type, number of
missing values, number of distinct values, and number of unique values of each currently
selected attribute. It is possible to see these values for all other attributes by selecting

60
each of them one-by-one. These values in turn enable the researcher together with the
domain experts to determine the threshold values of each variable for the analysis of the
result.

4.5 Data Mining

There are different tasks performed during the data mining phase the major once are
selection of modeling techniques and building a model.

4.5.1 Selection of Modeling Techniques

The main objective of this research being to effectively segment the customers of ERCA
into various groups in which there is less similar between segments /clusters however
whose members being very similar. To successful meet this research objectives there was
a need of selecting appropriate modeling techniques.

Customer segmentation of this research is performed by using clustering and

classification data mining techniques. These techniques are selected because
segmentation problems are most of the time performed by using these techniques.
Furthermore, the techniques are also well implemented in the selected data mining tool.

For the case of clustering the selected tool supports various algorithms, such as simple K-
means, Cobweb, DBScan, Expectation Maximization (EM), FarthestFirst,
FilteredClusterer, HierarchicalClusterer, MakeDensityBasedClusterer and OPTICS. From
all these clusters, the researcher selected simple K-means algorithm; simple K-means
algorithm is selected as it is better in handling discrete and numeric attributes. Moreover,
K-means method is very popular because of its ability to cluster huge data, and outliers,
quickly and efficiently (Kanungo et al. 2002). When using simple K-means algorithm
there are different activities which are performed by the users of the algorithm, such as
determining the value of K which is the number of segments/clusters. In order to decide
the value of K there is a need to consult the domain experts because the number of
clusters chosen should be driven by how many clusters the business can manage. The

61
experts suggested that the K value to be in the range of 3-6, this four cluster models were
evaluated against their performance of creating dissimilar segments.

For classification, the most common classification techniques, decision tree and neural
network classifier, are tested. Regarding decision tree, the selected tool holds various
decision tree algorithms, such as BFTree, J48, ID3, LADTree, NBTree, RandomForest,
RandomTreee and so on. From these options J48 decision tree algorithm has been
selected because it handles a large amount of variables with either a continuous or
discrete variable. Moreover, The J48 algorithm gives several options related to tree
pruning (Han and Kamber 2006). As described in Section 2.6.2 tree pruning produces
fewer, more easily interpreted results and it used as a tool to correct for potential over-
fitting.

J48 decision tree algorithm employs two pruning methods. The first method of pruning is
known as sub-tree replacement and the second type of pruning used in J48 is termed sub-
tree raising (Han and Kamber 2006, Witten and Frank 2005).

 Sub-tree replacement: - in this method of pruning, nodes in a decision tree are

replaced with a leaf. This process starts from the leaves of the fully formed tree,
and works backwards toward the root.
 Sub-tree raising: - In this type of pruning, a node may be moved upwards towards
the root of the tree, replacing other nodes along the way. This type of pruning is
computationally complex

4.5.2 Test Design

Before passing to the actual model building, it is better to set a guide for the training and
testing process. For the case of clustering the total instance of the dataset is used to train
the clustering model. On the other hand, for classification model the researcher used both
10-Fold Cross-validation and percentage split. In 10-fold cross-validation, the initial data
are randomly partitioned into 10 mutually exclusive subsets or “folds,” 1, 2, 3, …, 10,
each approximately equal size. Training and testing is performed 10 times. In the first
iteration, the first fold is reserved as a test set, and the remaining 9 folds are collectively

62
used to train the classifier; the classifier of the second iteration is trained on folds 1, 3, 4,
…, 10 and tested on the second fold; and so on (Han and Kamber 2006). For the case of
percentage split 70% was used for train the model, whereas 30% of the dataset was used
for test data.

The following variables are used for model building, while the rest attributes are not used
because they are less important for the model building.

 Total number of items imported (SAD_ITEM_TOTAL)

 Total amount of invoice spent to import the item (SAD_TOTAL_INVOICED)
 Total amount of revenue generated (TOTAL_REVENUE)
 The price of the item (ITEM_PRICE)
 Origin (continent) of the item where it is imported (CTY_DSC)
 Ratio of total revenue to total number of items (REVENUE_ITEM)
 Ratio of total revenue to item price (REVENUE_PRICE)
 Ratio of total invoiced to total amount of revenue generated
(INVOICED_REVENUE)

There is one attribute called risk color, it is important for model building but it doesn‟t
included during the attribute selection, because this attribute has three values (green,
yellow, and red) out of the total data more than 90% has red value and only 0.9% of the
total data has green value.

4.5.3 Model Building

The model building data mining phase mainly consists of the cluster modeling and
classification modeling subsections.

4.5.3.1 Cluster Modeling

Once the final dataset has been selected the next step was to build the clustering models
by using the selected data mining tools, Weka 3.6.4 Version. In clustering of data using
simple K-means there is a need to set the number of clusters or the value of K, the
number of data tuples the cluster must start with (seed size) and also there is a need to

63
specify the variables to be used for building the cluster. In this cluster modeling the entire
population of the dataset (100%) is used to train the clustering model.

Before starting the experimentation it is important to set the threshold values of each
numeric attributes of the dataset, because the threshold values are used to interpret the
output of the clustered model. The threshold value has been determined with the domain
experts and with the aid of the Weka‟s minimum, maximum, and mean values display for
each attribute. So the threshold value for each attribute has been as follow.

TOTAL_ITEM Very High High Average Low Very Low

TOTAL_INVOICE Very High High Average Low Very Low

ITEM_PRICE Very High High Average Low Very Low

TOTAL_REVENUE Very High High Average Low Very Low

REVENUE_ITEM Very High High Average Low Very Low

INVOICE_REVENUE Very High High Average Low Very Low

REVENUE_PRICE Very High High Avarage Low Very Low

Threshold Values
List of Attributes
Very High High Average Low Very Low
1. TOTAL_ITEM (TIT) >50 26-50 15-25.9 10-14.9 <10
2. TOTAL_INVOICE (TIV) >200,000 151,000- 110,000- 45,000- <45,000
200,000 150,999 109,999
3. TOTAL_REVENUE (TRN) >1,000,000 750,000- 500,000- 300,000- <300,000
1,000,000 749,999 499,999
4. REVENUE_ITEM (RIT) >60,000 39,000- 15,000- 10,000- <10,000
60,000 38,999 14,999
5. INVOICE_REVENUE (IVR) >5 4-5 2-3.9 1-1.9 <1
6. ITEM_PRICE (IPR) >50,000 25,000- 15,000- 2,000- <2,000
50,000 24,999 14,999
7. REVENUE_PRICE (RPR) >12,000 10,000- 5,000-9,999 2,000- <2,000
12,000 4,999

Table 4.7 List of range of conditions by which a cluster result is assessed

64
 4.5.3.1.1 Experiment 1

The first experiment is performed by adjusting the number of cluster or the value of K= 3
and using the seed size100. During this experiment all of the selected variables were
input into the Weka cluster run. In clustering all the variables were set as independent
variables because clustering is unsupervised data mining techniques. The output of the
first experiment is summarized in the following table

Cluster distributions
Cluster 1 Cluster 2 Cluster 3
6544(14%) 38992(83%) 1212(3%)

Table 4.8 Cluster distributions with number of iterations 4 and sum of squared
errors 5199.169

During the default seed size (seed=10) the total number of instances were not fairly
distributed; which means the sum of the whole cluster was 99%, but when the seed
number was set to 100 the distribution of instances were significantly improved. In
addition, the number of iterations and sum of squared errors were also minimized. When
the seed size is 10, the number of iterations and sum of squared errors were 22 and
9409.92, respectively. However, if the seed size is set to 100 the number of iterations was
4 and sum of squared errors was 5199.169.

As describe in Table 4.8, majority of the customers were grouped in the second cluster.
From the total (46748) customers 38992 (83%) of them are grouped in this cluster. The
first cluster contains the second largest number of customers that is it consists of 6544
(14%) number of customers and the third cluster contains the list number of customers
which is only 1212(3%) number of customers.

In an attempt to improved the data distribution of the segments, different seed values
have been tried. But, the seed value at 100 gives a better data distribution in the
segments. The optimal seed value has been obtained after a number of experimentations.

65
The following abbreviations are used in the cluster modeling interpretation

Attributes of the dataset with their Abbreviated terms used in the cluster
description interpretation
Abbreviated Description Abbreviated Description
attributes words
TIT Total number of items VH Very High
imported
TIV Total amount of H High
invoice spent
TRN Total amount of A Average
revenue generated
RIT Ratio of revenue per L Low
total number of items
IVR Ratio of invoiced per VL Very Low
revenue generated
RPR Ratio of revenue AS Asia
generated per item
price
IPR The price of the item EU Europe
AM America
LA Latin America
AF Africa
AU Australia

Table 4.9 List of abbreviated words and attributes of the dataset along with
their description

66
Segmentati
on (K=3,
Seed=100)
Cluster No. Distributi TIT TIV TRN RIT IVR RPR IPR City
on of
instances
(in %)
1 6544 13.33 74041.98 639766.33 69301.05 1.81 5025.53 20555.43 EU
(14%)
2 38992 12.58 105916.33 879519.78 35285.70 2.53 6414.65 32368.00 AS
(83%)
3 1212 (3%) 15.83 152172.10 555970.81 33285.87 7.02 5414.98 27767.38 AM
Total 46748
(100%)
Cluster No. Distributi
on of
instances
(in %)
1 6544 L L A VH L A A
(14%)
2 38992 L L H A A A H
(83%)
3 1212 (3%) A H A A VH A H
Total 46748
(100%)

Table 4.10 Clustering result of the first experiment

Table 4.10 shows the mapping of the attributes average value of each segment with the
corresponding discrete value. This mapping is used to compare the segments which are
generated from the clustering algorithm.

Once the average value of each attribute in each cluster has been replaced with the
corresponding discrete value, the next step is to describe each cluster according to the
above table.

Cluster 1:- Cluster One of the first experiment contains customers who import small
amount of total number of items, and customers who spend low amount of invoice.
Customers that grouped in this cluster has generated average amount of revenue; and the
ratio of total revenue per total number of items was very high. However, the ratio of total
invoice with total revenue generated was low; on the other hand, the ratio of revenue
generated with the item price and the price of the item which is imported by customers

67
that grouped in this cluster has average value. Items of this cluster have been imported
from Europe countries.

Cluster 2:- The second cluster also holds those customers who import lower amount of
total number of items and those customers who spend low amount of invoice. The cluster
consists of high amount of revenue generated customers. The ratio of total revenue per
total amount of items, the ratio of total invoice per total revenue generated, and the ratio
of total revenue generated to the item price were average. The customers that grouped in
this cluster have been imported high price item. This cluster customer has been imported
their items from Asian countries.

Cluster 3:- The last cluster of the first experiment embraces customers that import
average number of items and those spend high amount of invoice. The revenue generated
by the customers of this cluster was average. The ratio of total revenue generated to total
number of items and the ratio of total revenue generated to the item price were average.
On the other hand, the ratio of total invoice to total revenue generated was very high. The
price of the item which is grouped in this cluster was high. The cluster holds items which
were imported from North America.

As stated in the business understandings of the Authority high value customers were
those customers who generate high amount of revenue by importing small number of
items and spending lower amount of invoice. On the other hand, low value customers
were that customers who generate lower amount of revenue by importing large amount of
items and spending higher amount of invoice.

By considering the above facts of the business, each cluster has been assigned in the
following ranking order.

Cluster No. Possible Rank

Cluster 1 2
Cluster 2 1
Cluster 3 3

Table 4.11 Cluster rank for the first experiment

68
As it is clearly indicated in Table 4.11, the second cluster is ranked first. This is due to
the fact that customers of this cluster import smaller amount of items, spend lower
invoice and generate high amount of revenue for the Authority. In addition, ratio of total
revenue to total number of items, ratio of invoice to total revenue generated and ratio of
revenue generated to item price were medium. For this segment the price of the item was
high.

The first cluster is ranked second; customers in this segment generate average amount of
revenue for the Authority by importing small number of items. In addition, the invoices
spend by this segment customer were also low. The ratio of total revenue to total number
of items was very high, whereas the ratio of total revenue to item price was medium.
Again, the ratio of total invoice to total revenue generated was low and the price of the
item was average.

Cluster Three is ranked last, because, although customers of this cluster generate average
amount of revenue, the total number of items imported was also average. Customers of
this segment also spent high amount of invoice. Moreover, the ratio of total revenue to
total number of items and the ratio of total invoice to total revenue generated were
average and very high, respectively.

Therefore, the second cluster customers were high value customers, because these
customers generate high revenue by importing small number of items and spending lower
invoice. However, customers categorize under Cluster Three were low value customers
because this cluster customers spent high amount of invoice and they generated average
amount of revenue from average number of items.

69
 4.5.3.1.2 Experiment 2

The second clustering model experiment is conducted by setting the value of K=4 and
seed size=1000. When the default seed size (seed=10) took the total number of instances
were not properly distributed. The output of the second experiment is displayed in the
following table.

Segmentation
(K=4,
Seed=1000)
Cluster No. Distribution TIT TIV TRN RIT IVR RPR IPR City
of instances
(in %)
1 6651 (14%) 13.40 73379.33 640087.27 68448.49 1.78 5031.43 20372.4 EU
2 9263 (20%) 10.35 138122.13 1800030.56 75083.10 0.08 12172.93 58714.05 AS
3 18790 14.13 128316.08 327627.05 14652.31 5.64 3465.22 27325.13 AS
(40%)
4 12044 12.15 51504.68 1001966.74 36835.44 0.05 6495.89 19715.84 AS
(26%)
Total 46748
(100%)
Cluster No. Distribution
of instances
(in %)
1 6651 (14%) L L A VH L A A
2 9263 (20%) L A VH VH VL VH VH
3 18790 L A L L VH L H
(40%)
4 12044 L L VH A VL A A
(26%)
Total 46748
(100%)

Table 4.12 Clustering result of the second experiment

In the second experiment four segments were formed, but these four segments didn‟t
contain equal number of customers. The majority of the customers are grouped under
Cluster 3 that is 18790 customers out of 46748. Cluster 4 consists of the second larger
group (12044 out of 46748), Cluster 2, the third (9263 out of 46748) and Cluster 1
consists the least number of customers (6651 out of 46748).

70
Cluster 1: - This cluster consists of customers that imported low number of items, spent
low amount of invoice, generated average amount of revenue, imported items which have
average price, the ratio of revenue to total number of item was very high, the ratio of
invoice to total number of item was low, the ratio of total revenue to item price was
average and items originated from European countries.

Cluster 2: - The second cluster consists of customers that imported small number of
items, spent average total invoice, generated very high revenue, had very high ratio of
total revenue to total number of items, had very low ratio of total invoice to total revenue,
had very high ratio of total revenue to item price, had very high item price and items
originated from Asian countries.

Cluster 3: - The third segment holds customers that imported small number of items,
spent average invoice, generated low revenue, had low ratio of total revenue to total
number of items, had very high ratio of total invoice to total revenue, had low ratio of
total revenue to item price, had high item price and that items which were imported from
Asian countries.

Cluster 4: - The last cluster consists of customers that imported small number of items,
spent low number of invoice, generated very high amount of revenue, had average ratio
of total amount of revenue to total number of items, had very low ratio of total invoice to
total amount of revenue generated, had average ratio of total revenue to item price, had
average item price and items imported from Asian countries.

Rank of the second experiment is arranged in the following table.

Cluster No. Possible Rank

1 3
2 2
3 4
4 1

Table 4.13 Cluster rank for the second experiment

71
As it is shown in Table 4.13, Cluster 4 is ranked first, as customers in this cluster
generated very high amount of revenue by importing small number of items and spending
low invoice. This segment consists of average ratio of total revenue to total number of
item, very low ratio of invoice to total revenue, average ratio of total revenue to item
price and average item price. The items for the fourth segment are imported from Asian
countries. Customers in the second segment is ranked second as they generated very high
revenue by importing small number of items and spending average amount of invoice.
The ratio of total revenue to total number of items and the ratio of total invoice to total
revenue were very high and very low, respectively. The ratio of total revenue to item
price and the price of the item were very high. The second cluster also consists of items
which imported from Asian countries. The first cluster is ranked third because this cluster
included customers who generated average amount of revenue by importing small
number of items and spending low amount of revenue. The ratio of total revenue to total
number of items was very high. On the other hand, the ratio of total invoice to total
revenue generated was low. The ratio of total revenue to item price and the price of the
item for this cluster was average. The third segment is ranked last, because customers of
this cluster generated low amount of revenue by providing average amount of invoice and
bringing in small number of items. The ratio of total revenue to total number of items was
low and the ratio of total invoice to total amount of revenue generated was very high. The
ratio of total revenue to item price and the price of the item were low and high,
respectively. Items of this cluster originated from Asian countries.

72
 4.5.3.1.3 Experiment 3

To get the best clustering the researcher tried different experiments by varying the
number of clusters (the value of K) and the seed size. So, the researcher continued with
the third experiment by setting the value of K=5 and using the default seed size, which is
10. The Weka output of the third experiment is presented in the following table.

Segmentation
(K=5,
Seed=10)
Cluster No. Distribution TIT TIV TRN RIT IVR RPR IPR City
of instances
(in %)
1 17640 11.23 132532.47 319067.19 15308.78 5.97 2985.01 28946.29 AS
(38%)
2 9210 (20%) 9.52 138128.41 1797414.48 75515.58 0.08 12241.55 59053.13 AS
3 11746 9.97 50533.09 994265.18 38099.33 0.05 6205.96 20326.46 AS
(25%)
4 1486 70.51 71479.32 694342.74 852.17 0.36 11953.49 1096.22 AS
(3%)
5 6666 (14%) 13.37 73362.92 639324.67 68314.16 1.73 5027.24 20310.03 EU
Total 46748
(100%)
Cluster No. Distribution
of instances
(in %)
1 17640 L A L A VH L H
(38%)
2 9210 (20%) VL A VH VH VL VH VH
3 11746 VL L H A VL A A
(25%)
4 1486 VH L A VL VL H L
(3%)
5 6666 (14%) L L A VH L A A
Total 46748
(100%)

Table 4.14 Clustering result of the third experiment

Cluster 1: - Customers that grouped under this segment generated low amount of
revenue by importing small number of items and spending average amount of invoice.
The ratio of total revenue to total number of items was average, the ratio of total invoice
to total amount of revenue was very high, the ratio of total revenue to item price was low,

73
the price of the item was high and this cluster consists of items which imported from
Asian countries.

Cluster 2: - The second cluster consists of customers that generated very high amount of
revenue by importing very low number of items and spending average amount of invoice.
The ratio of total revenue to total number of items and the ratio of total revenue to item
price were very high. The ratio of total invoice to total revenue was very low; the price of
the item which is grouped under this cluster was very high. This cluster holds items
originated from Asian countries.

Cluster 3: - This segment holds customers that generated high amount of revenue by
importing very small number of items and spending low amount of invoice. The ratio of
revenue to total number of items was average; the ratio of total invoice to total amount of
revenue generated was very low; the ratio of total revenue to item price and the price of
items were average. This cluster also holds items which imported from Asian countries.

Cluster 4: - The fourth segment of this experiment consists of customers who generated
average amount of revenue by importing very high number of items and spending low
amount of invoice. The ratio of total revenue to total number of items and the ratio of
invoice to revenue were very low; the ratio of revenue to item price was high; the price of
item which grouped under this cluster was low and this cluster holds items which
imported from Asian countries.

Cluster 5: - The last cluster of this experiment consists of customers who generated
average amount of revenue by importing lower number of items and spending lower
amount of invoice. The ratio of total revenue to total number of items was very high; the
ratio of total invoice to total amount of revenue generated was low; the ratio of total
revenue to item price and the price of the item were average. This segment holds items
which were imported from European countries.

Based on the facts of the business problem of the Authority the above clusters are ranked
in the following order.

74
 Cluster No. Possible Rank
1 5
2 1
3 2
4 4
5 3

Table 4.15 Cluster rank for the third experiment

Cluster 2 customers are ranked first because this group of customers generated very high
amount of revenue by importing very small number of items and spending medium
amount of invoice. In addition, both the ratio of revenue to total number of items and the
ratio of revenue to item price were very high. For this segment, the ratio of total invoice
to total revenue generated was also very low. So, Cluster Two customers were high value
customers. The third segment customers are ranked second; this segment customers
generated high amount of revenue by importing very low amount of items and spending
low amount of invoice. The ratio of revenue to total number of items and the ratio of
revenue to item price were average. The ratio of total invoice to total amount of revenue
generated was very low. This segment consists of customers who imported medium price
of item. Cluster 5 holds the third rank; this segment embraces those customers who
generated average amount of revenue by importing small number of items and spending
low amount of invoice. The ratio of total revenue to total number of items and the ratio of
total invoice to total amount of revenue generated were very high and low, respectively.
The ratio of total revenue to item price and the price of the item were average. Cluster 4
holds the fourth rank; this cluster consists of customers that generated average amount of
revenue by importing very large number of items and spending low amount of invoice.
For this cluster, the ratio of revenue to total number of items and the ratio of invoice to
total amount of revenue generated were very low but the ratio of total revenue to item
price is high that means the types of items which is imported by this group of customers
were low price items.

Cluster 1 is ranked last because the customers of this segment generated low amount of
revenue. Even thought the customers of this cluster imported low amount of items they
spent average amount of invoice. So, according to the business problem the first segment

75
customers are low-value customers, whereas Cluster 2 customers are considered high-
value customers.

4.5.3.1.4 Experiment 4

The fourth experiment is continued by adjusting the number of clusters or value of K=6
and using the default seed size which is 10. The Weka output of the fourth experiment
looks like as follows.

segmentation
(K=6,
Seed=10)
Cluster No. Distribution TIT TIV TRN RIT IVR RPR IPR City
of instances
(in %)
1 13854 12.10 157070.73 243884.97 13300.12 7.59 2276.08 33073.11 AS
(30%)
2 6662 (14%) 8.84 104125.23 1908117.51 82442.08 0.06 13668.66 72714.90 AS
3 10740 10.84 47558.55 729263.28 30509.01 0.07 5426.38 18969.74 AS
(23%)
4 1120 (2%) 78.36 75492.02 795137.70 929.30 0.36 12229.10 994.73 AS
5 6610 (14%) 13.45 73674.61 638412.01 68930.34 1.79 5036.62 20336.02 EU
6 7762 (17%) 9.94 109094.76 1303954.71 45333.99 0.08 7950.09 19114.95 AS
Total 46748
(100%)
Cluster No. Distribution
of instances
(in %)
1 13854 L H VL L VH L H
(30%)
2 6662 (14%) VL L VH VH VL VH VH
3 10740 L L A A VL A A
(23%)
4 1120 (2%) VH L H VL VL VH VL
5 6610 (14%) L L A VH L A A
6 7762 (17%) VL L VH H VL A A
Total 46748
(100%)

Table 4.16 Clustering result of the fourth experiment

Cluster 1: - The first cluster of the last experiment consists of small number of items,
high amount of invoice, very low amount of revenue, low amount of ratio of revenue to

76
total number of items, very high amount of ratio of invoice to revenue, low amount of
ratio of revenue to item price, high amount of item price and items which originated from
Asian countries.

Cluster 2: - The second segment of this experiment embraces very low number of items,
low amount of invoice, very high amount of revenue, very high amount of ratio of
revenue to total number of items, very low amount of ratio of invoice to revenue, very
high amount of ratio of revenue to item price, very high item price and items which
imported from Asian countries.

Cluster 3: - This cluster consists of small number of items, low amount of invoice,
average amount of revenue, average amount of ratio of revenue to total number of items,
very low amount of ratio of invoice to revenue, average amount of ratio of revenue to
item price, average amount of item price and items which originated from Asian
countries.

Cluster 4: - The fourth segment holds very high number of items, low amount of invoice,
high amount of revenue, very low amount of ratio of revenue to total number of items,
very low amount of ratio of invoice to revenue, very high amount of ratio of revenue to
item price, very low item price and items which originated from Asian countries.

Cluster 5: - The fifth cluster consists of low number of items, low amount of invoice,
average amount of revenue, very high amount of ratio of revenue to total number of
items, low amount of ratio of invoice to revenue, medium amount of ratio of revenue to
item price, average amount of item price and items originated from European countries.

Cluster 6: - The last cluster of this experiment consists of very low number of items, low
amount of invoice, very high amount of revenue, high ratio of revenue to total number of
items, very low ratio of invoice to revenue, average ratio of revenue to item price,
average item price and items which imported from Asian countries.

Based on the above description and the business problems of the Authority the above
clusters are ranked in the following order.

77
 Cluster No. Possible Rank
1 6
2 1
3 3
4 5
5 4
6 2

Table 4.17 Cluster rank for the forth experiment

The second cluster generated very high amount of revenue, very high amount of revenue
per total number of items and very high amount of revenue per item price. Since the price
of items was very high Cluster Two customers are imported very low number of items by
spending low amount of invoice; therefore, this cluster is ranked first. The sixth cluster
held the second rank, because this cluster also generated very high amount of revenue,
high amount of revenue per total number of items and average amount of revenue per
item price. This cluster customers imported very low amount of items by spending low
amount of invoices. The price of item which is grouped under this segment was average.
So, Cluster Two customers are high value customers. Cluster 6, 3, 5 and 4 customers held
the next rank this cluster generally consist of medium-level value customers. The first
segment held the last rank, because this cluster generated very low amount of revenue by
spending high amount of invoice and importing small number of items. The ratio of
revenue per total number of items was also low.

4.5.3.1.5 Choosing the Best Clustering Model

To get the best segmentation of customers, four cluster modeling experiments were
conducted by varying the number of clusters or the value of K (3, 4, 5 and 6) and using
different seed size. Though various experiments with different seed size are tried, only
the seed size which resulted in fair distribution of segments is reported.

From the four experiments, the best clustering are selected by considering the output of
clusters, which contains less inter-cluster similarity and high intra-cluster similarity.
Moreover, within cluster sum of squared errors and the domain experts‟ judgment were
also taking into consideration. From all of the above experiments, Experiment 3 (K=5)

78
and Experiment 4 (k=6) have relatively small number of within cluster sum of squared
errors. Moreover, the clusters have good distribution of instances. From these two
clusters, the domain experts chase the third experiment (K=5); this is because the experts
justified that the fourth experiment did not create dissimilar clusters; the experts further
clarified that, in the fourth experiment Cluster 2 and Cluster 6 as well as Cluster 3 and
Cluster 5 customers were similar customers. So, according to the experts, the third
experiment (K=5) is the best cluster model from the other models experimented in this
research.

4.5.3.2 Classification Modeling

As described in Section 2.6.2, classification is a learning model in data mining techniques

which aims at building a model to predict future customer behaviors through classifying
database records into a number of predefined classes based on certain criteria. In this
study, most common classification techniques, such as decision tree and neural network
classification techniques were tested; for decision tree J48 algorithm and for neural
network MultilayerPerceptron algorithm were investigated.

4.5.3.2.1 Decision Tree Model Building

The output of the selected clustering model was fed to the J48 decision tree algorithm.
Here the cluster index is used as the dependant variable, whereas the remaining all
attributes which are selected for the cluster model building, are fed as independent
variables. The J48 decision tree provided a descriptive classification model of the
clusters, thus enabling exploration and detection of the characteristic of each cluster.

During the generation of a classification model, both options, the 10-fold cross-validation
and percentage split (with 70% train and 30% test), were investigated.

When using the J48 decision tree algorithm with 10-fold cross-validation default
parameter value, the output of the tree consisted of 91 nodes and 54 leaves. The output of
the confusion matrix for this learning algorithm looks as follows.

79
 Predicted Accuracy
Actual Total
Cluster1 Cluster 2 Cluster 3 Cluster 4 Cluster 5 Rate
Cluster 1 17633 0 1 0 6 17640 99.96%
Cluster 2 0 9209 1 0 0 9210 99.98%
Cluster 3 0 0 11742 0 4 11746 99.96%
Cluster 4 0 2 0 1484 0 1486 99.96%
Cluster 5 3 0 3 0 6660 6666 99.90%
Total 17636 9211 11747 1484 6670 46748 99.95%

Table 4.18 Output from J48 decision tree algorithm with 10-fold cross-validation
default parameter value

The output of the confusion matrix shows that, in this experiment from the total 46748
amounts of data, 46728 (99.95%) of the records were correctly classified, while the
remaining 20 (0.05%) of the records were incorrectly classified.

Table 4.18 clearly shows that, from 17640 records, 17633 (99.96%) of the records were
correctly classified as Cluster One (low-value customers), while the remaining 1 record
was incorrectly classified as Cluster Three, and the other 6 were incorrectly classified as
Cluster Five. From 9210 records, 9209 (99.98%) of the records were correctly classified
as Cluster Two (high-value customers), whereas the remaining 1 record was incorrectly
classified as Cluster Three. Out of 11746 records, 11742 (99.96%) records were correctly
classified as Cluster Three, while the remaining 4 records were incorrectly classified as
Cluster Five. From 1486 records, 1484 (99.96%) records were correctly classified as
Cluster Four, whereas the remaining 2 records were misclassified as Cluster Two (high-
value customer). Out of 6666 records, 6660 (99.90%) records were correctly classified as
Cluster Five; while the remaining 3 records were misclassified as Cluster One (low-value
customers) and the other 3 records were also incorrectly classified as Cluster Three.

Although the above experiment had good accuracy, the researcher tried to find the best
classification with small tree size; because small tree size decision tree classifications are
easier to generate rules. So, the researcher tried various experiments by changing the
default value of the J48 decision tree 10-fold cross-validation parameter values. The

80
default values of 10-fold cross-validation are the minimum number of instances per leaf
(minNumObj) with 2 and the confidence factor used for pruning (confidenceFactor) with
0.25. The different values of minNumObj with 5, 10, 15, 20, 25 were tested and the
output of the confusion matrix for minNumObj=25 displayed in the following table.

Predicted Accuracy
Actual Total
Cluster1 Cluster 2 Cluster 3 Cluster 4 Cluster 5 Rate
Cluster 1 17583 0 29 20 8 17640 99.67%
Cluster 2 0 9209 1 0 0 9210 99.98%
Cluster 3 0 6 11727 0 13 11746 99.83%
Cluster 4 5 2 0 1479 0 1486 99.52%
Cluster 5 12 0 16 14 6624 6666 99.36%
Total 17600 9217 11773 1513 6645 46748 99.73%

Table 4.19 Output from J48 decision tree algorithm with minNumObj=25 and
confidenceFactor=0.25

Although increasing the value of the minimum number of instance per leaf (minNumObj)
as 25 reduces the accuracy rate, the tree was pruned significantly. At this experiment the
tree is generated by 34 leaves and 55 nodes. The overall accuracy rate of the algorithm
was 99.73%, which means, from the total dataset (46748 records), the algorithm correctly
classified 46622 number of records, whereas the remaining 126 number of records were
misclassified. The accuracy was minimized when compare with the pervious experiment,
which had 99.95% of accuracy.

Moreover, out of 17640 numbers of low value customers in Cluster One, only 17583
(99.67%) of them were correctly classified, while the remaining 57 (0.33%) customers
were incorrectly classified. Compared with the first experiment, which had (99.96%) of
accuracy, the second algorithm‟s accuracy rate was reduced by 0.29%. From 9210
numbers of high value Cluster Two customers, 9209 (99.98%) of them were correctly
classified; only one customer was misclassified and its accuracy rate was similar to the
pervious experiment. Out of 11746 number of medium value Cluster Three customers,
11727 (99.83%) of them were correctly classified, whereas the remaining 19 (0.17%)

81
customers were incorrectly classified, so its accuracy was reduced by 0.13% from the
previous experiment. From 1486 average value customers of Cluster Four, 1479
(99.52%) of them were correctly classified, while the rest 7 (0.48%) were misclassified;
when it was compared with the previous experiment, the accuracy rate was reduced by
(0.44%). Out of 6666 average value customers of Cluster Five, 6624 (99.36%) of them
were correctly classified, while the remaining 42 (0.64%) customers were misclassified;
hence, the accuracy rate was reduced by 0.54% from the previous experiment.

The experimentation of the J48 decision tree model building is continued by using
percentage split with 70% of the records (dataset) as training and the remaining 30% of
the record as testing. From the total 30% (14024) testing data, the algorithm correctly
classified 13975 (99.65%) records, whereas the remaining 49 (0.35%) records were
incorrectly classified. The following table shows the result of the percentage split
experimentation.

Predicted Accuracy
Actual Total
Cluster1 Cluster 2 Cluster 3 Cluster 4 Cluster 5 Rate
Cluster 1 5264 0 10 7 0 5281 99.67%
Cluster 2 0 2748 0 0 0 2748 100%
Cluster 3 0 6 3559 0 8 3573 99.60%
Cluster 4 2 0 0 437 0 439 99.54%
Cluster 5 6 0 4 6 1967 1983 99.19%
Total 5272 2754 3573 450 1975 14024 99.65%

Table 4.20 Summary of the confusion matrix with default parameters value and 70
% for training and 30 % for testing dataset

Compared with the previous experiments, both 10-fold cross-validation experiments, this
experiment had the least overall accuracy rate. Even in the individual cluster
classification, except for the second cluster, the 10-fold cross-validation had better
classification accuracy than the percentage split experimentation.

82
Hence, from all of the above experimentations, the first model, 10-fold cross-validation
with the default parameter values, was selected because this model had better accuracy
both in the overall and individual cluster classification.

4.5.3.2.2 ANN Classification Model

The other common classification technique is the Artificial Neural Network (ANN)
classification model. According to Han and Kamber (2006), ANN classification model
learn very fast when the attributes‟ values fall in the range [-1, 1]. So, to put the attribute
value in the range of [-1, 1], the researcher used the Weka normalizing preprocessing
facilities. Consequently, all numeric attributes were normalized; that is their value fell in
between the range of [-1, 1]. However, CTY_DSC (the origin of items) is a nominal
attribute but as it is stated in Section 2.6.2.2, ANN classification model usually works
only with numerical data. So, there are six distinct values in this attribute and each of
them is assigned a numeric value from 1-6. After mapping the nominal value into the
numeric value, the Weka preprocessing facility normalizes all values to fall in the range
of [-1, 1]. Hence, the same attributes that were used to build the decision tree models,
were also used in the neural net modeling.

Like the J48 decision tree algorithm in this experiment also the researcher used both the
10-fold cross-validation and percentage split tests. During the experimentation various
runs are performed by using the default values and changing the hidden layers, learning
rate and momentum parameter values. The default hidden layers, learning rate and
momentum parameters are shown in the following table.

83
 Parameter Description Default value
hiddenLayers This defines the hidden layers of the 'a' = (attributes + classes) / 2
neural network. This is a list of a=(8 (number of attributes +5
positive whole numbers. (number of clusters))/2=6.5 (7)
learningRate The amount the weights are updated 0.3
momentum Momentum applied to the weights 0.2
during updating
10-folds cross-validation

Table 4.21 Parameters and their default values of the neural network classifier

From the different runs of 10-fold cross-validation better accuracy rate are found by
using 0.5 learning rate, 8 hidden layers, and 0.4 momentum. At this specific run the
output of the confusion matrix looks as follows.

Predicted Accuracy
Actual Total
Cluster1 Cluster 2 Cluster 3 Cluster 4 Cluster 5 Rate
Cluster 1 16744 1 49 0 0 16794 99.70%
Cluster 2 1 8834 8 0 38 8881 99.47%
Cluster 3 13 12 12305 0 0 12330 99.79%
Cluster 4 0 0 0 1168 0 1168 100%
Cluster 5 0 13 0 0 7562 7575 99.83%
Total 16758 8860 12362 1168 7600 46748 99.71%

Table 4.22 10-fold cross-validation output from MultilayerPerceptron ANN

algorithm with hiddenLayers=8, learningRate=0.5 and momentum=0.4

As it is shown in Table 4.22, the multilayerperceptron ANN algorithm had good overall
and individual cluster classification accuracy rate. From the total dataset (46748) records,
46613 (99.71%) records were correctly classified by this algorithm; only 135 (0.29%)
records were misclassified.

84
For the percentage split run also better performance accuracy rate was found when
hiddenLayers=8, learningRate=0.5 and momentum=0.4. For this split run like in the
above J48 decision tree experiment, 70% of the records were used for training and 30%
of the records were used for testing. The output of the confusion matrix for this split run
is presented in the following table.

Predicted Accuracy
Actual Total
Cluster1 Cluster 2 Cluster 3 Cluster 4 Cluster 5 Rate
Cluster 1 5048 0 44 0 0 5092 99.04%
Cluster 2 0 2659 0 0 9 2668 99.66%
Cluster 3 0 2 3641 0 0 3643 99.95%
Cluster 4 0 0 0 345 0 345 100%
Cluster 5 0 0 0 0 2276 2276 100%
Total 5048 2661 3685 345 2285 14024 99.60%

Table 4.23 Split output from MultilayerPerceptron ANN algorithm with

hiddenLayers=8 learningRate=0.6 and momentum=0.4

The output of the split run shows that from the total 14024 testing dataset, 13969
(99.60%) of the records were correctly classified, while the remaining 55 (0.40%) records
were misclassified. For individual cluster level classification the split run better correctly
classified the medium-value (Cluster 4 and 5) customers and high-value (Cluster 2)
customers. However, this run was still less efficient on classifying low-value (Cluster 1)
customers. Moreover, the split run had less overall accuracy rate than 10-fold cross-
validation.

From the above two MultilayerPerceptron ANN classification models, the first model
built using 10-fold cross-validation with hiddenLayers=8 learningRate=0.5 and
momentum=0.4 was selected; because this model had better accuracy rate than the
second split model.

85
 4.5.3.3 Comparison of Decision Tree and Neural Network Models

So far two classification models were tested; the next step was comparing the above two
classification models, which were the J48 decision tree model and MultilayerPerceptron
ANN classification model. The purpose of the comparison was to choose the best from
these two algorithms, which was appropriate for the problem domain of this research,
CRM.

From the above two classification models the best algorithm was selected based on the
following three parameters.

 The overall classification accuracy rate

 The model accuracy in classifying high value customers
 The model accuracy in classifying low value customers

So, based on the above criteria the algorithm, which had high overall accuracy rate and
high accuracy in correctly classifying high value and low value customers in their
clusters were selected. Consequently, the comparison of the decision tree and ANN
models are described as follows.

High-value customers Low-value customers

10-fold Overall accuracy
(Cluster 2) accuracy (Cluster 1) accuracy
classification (46748 records)
model
Correctly Incorrectly Correctly Incorrectly Correctly Incorrectly
classified classified classified classified classified classified
Decision 46728 20 9209 1 17633 7
tree model 99.95% 0.05% 99.98% 0.02% 99.96% 0.04%
Neural 46613 135 8834 47 16744 48
network 99.71% 0.29% 99.47% 0.53% 99.70% 0.30%
model

Table 4.24 Summary of the accuracy level of the decision tree and neural net
classification models

Table 4.24 shows that the decision tree model had better overall accuracy rate than the
neural net model. From the total (46748) dataset, the decision tree model correctly

86
classified 46728 (99.95%) records. Only 20 (0.05%) records were misclassified, while
the neural net algorithm correctly classified 46613 (99.71%) records and the remaining
135 (0.33%) records were misclassified. Moreover, decision tree had also better accuracy
in classifying high value customers; the model had an accuracy of 99.98% for classifying
high value customer only one customer was classified as other cluster. However, the
neural network model had relatively less accurate in classifying high-value customers;
this model had an accuracy of 99.47% for classifying high-value customer as Cluster
Two. Furthermore, the decision tree model had better classification accuracy for
classifying low-value customers; it had an accuracy of 99.96%, whereas the neural
network model had only 99.70% accuracy.

From the above comparison of the two classification models, it is possible to conclude
that the decision tree classification model is the best classifier for CRM applications,
because the result shows that from the three parameters the decision tree model had better
accuracy rate than the neural network classification model. Moreover, the decision tree
classification model had also advantages in generation rules easily. Sample decision tree
rules are attached in Appendix 3

Finally, the model developed with the J48 decision tree algorithm using the 10-fold cross
validation with the default parameter values is tested with 17012 testing dataset to
evaluate the predicting performance of the model. This developed model scored a
prediction accuracy of 99.87 on the given testing dataset.

4.6 Evaluation of the Discovered Knowledge

Unfortunately, the real life data, which was collected from the business organization,
consisted of missing value and outliers. Moreover, the collected data were also stored in
different tables and in the format to which the selected data mining tool was not accepted.
Accordingly, the researcher had taken considerable time for the data preparation task.
During the preprocessing phase the researcher handled those records, which contained
missing values and detected and removed outliers. Furthermore, data integration and
transformation tasks were carried out during the data preparation phase. Finally, the

87
preprocessed data were converted in the format (ARFF) to which the selected tool was
accepted.

Once the data was preprocessed, the next step was to build a model, which could
effectively segment the customers of ERCA. The model-building phase was divided into
two. The first phase embraced building different clustering model using the K-means
algorithm. The cluster model, which best differentiated high, medium and low value
customers, was selected. The next phase was to build a classification model with the J48
decision tree and multilayerperceptron neural net algorithm using the cluster index as the
dependent variable. The classification task was carried out using the 10-fold cross-
validation and percentage split (70% for training and 30% for testing) test options.
Among the various models of J48 decision tree and multilayerperceptron neural net
algorithms, the best one was selected with the criteria of overall classification accuracy
rate, accuracy in classifying high value customers and accuracy in classifying low value
customers. In all of these criterions the decision tree scored better classification accuracy.
As a result, the decision tree with 10-fold cross-validation, which scores an overall
accuracy of 99.95 %, was selected as the best classification model. Then the J48 decision
tree model was tested with 17012 separate testing dataset and scored an accuracy of
99.87%.

Different rules are generated from the decision tree developed by the J48 algorithm.
Some of the sample rules are the following:

Rule # 1: If TOTAL_REVENUE <= 652333.11 AND

CTY_DSC = AS AND
TOTAL_ITM <= 37 AND
REVENUE_ITEM <= 162139.61, then a customer is classified into cluster 0
Rule # 2: If 630615<TOTAL_REVENUE < 652333.11 AND
CTY_DSC = AS AND
TOTAL_ITM <= 37 AND
REVENUE_ITEM >162139.61, then a customer is classified into cluster 2
Rule # 3: If TOTAL_REVENUE <= 652333.11 AND
CTY_DSC = AS AND

88
 37<TOTAL_ITM <= 43 AND
TOTAL_INVOICED > 87408.66, then a customer is classified into cluster 3
Rule # 4: If TOTAL_REVENUE<=478053.49
CTY_DSC = AM AND
TOTAL_ITM <=42, then a customer is classified into cluster 0
Rule # 5: If 652333.11<TOTAL_REVENUE <= 1393587.35 AND
TOTAL_ITM <= 39 AND
CTY_DSC = EU, then a customer is classified into cluster 4
Rule # 6: If TOTAL_REVENUE > 1393587.35 AND
CTY_DSC = AS AND
TOTAL_ITM > 75, then a customer is classified into cluster 3
Rule # 7: If TOTAL_REVENUE >1393587.35 AND
CTY_DSC = AM AND
TOTAL_ITM < 9, then a customer is classified into cluster 1
Rule # 8: If TOTAL_REVENUE <= 498887.78 AND
TOTAL_ITM > 9, then a customer is classified into cluster 4
The above-generated rules are consistent with the business rules of the company, which
helps to identify potential and low valued customers. For instance, Rule # 7 shows those
customers who generated higher revenue and import small number of items. And this is
one of the characteristics of potential customers. On the other hand, Rule # 1 shows
those customers who generated small amount of revenue and import large amount of
items, which is one of the characteristics of low value customers. In addition to this the
generated rules showed that the items origin, item price, and revenue per number of item
attributes are important for identifying potential and low value customers.

Those items imported from North America, Europe, and Latin America, item price and
revenue per number of item is high showed potential customers. On the other hand, those
items imported from Asia, item price low, and revenue per number of item low showed
low valued customers.

89
 4.7 Use of the Discovered Knowledge
The above-generated rules in this research are encouraging. These discovered
knowledge/rules could be used for identifying high, medium, and low value customers
and treating them accordingly. For the company to use these generated rules effectively
and efficiently the company should have to first design a knowledge base system, which
can provide advice for the domain experts. The company can also incorporate the newly
generated rules with those current rules used for identifying potential and low value
customers in order to improve the decision making process.

90
 CHAPTER FIVE

CONCLUSION AND RECOMMENDATIONS

5.1 Conclusion
The main objective of this research was to effectively segment the customers of ERCA
based on their revenue generated, invoice spent, total number of items imported, the price
of items imported, and the items originated behaviors. This, in turn, could enable the
Authority to identify high, medium, and low value customers. The identification of these
different groups of customers would enable the ERCA to create good CRM with high
value customers.

This investigation is carried out using Cios et al. (2000) hybrid data mining process
model. The Cios et al. (2000) data mining model consists of understanding of the
problem domain, understanding of the data, preparation of the data, data mining,
evaluation of the discovered knowledge, and use of the discovered knowledge.

The collection and preparation of data, to make the data suitable for the data mining task,
were the major tasks which took considerable time in this research. Next, K-means
clustering algorithm was applied to segment the customer‟s data into meaningful groups.
As a result, four different clustering models were built by changing the value of K (3, 4, 5
and 6) and the seed size value. Among the four models, a model which best segment
high, medium and low value customers were selected together with the domain experts.
Consequently, from the four experiments, the third (K=5) experiment was selected.

The classification models were built with J48 decision tree and multilayerperceptron
neural net algorithms. From these two classification models the best classification model
was selected by comparing the overall accuracy, accuracy in classifying high value
customers and accuracy in classifying low value customers. The model which was
developed with J48 decision tree algorithm had 99.95% overall accuracy rate, whereas
the multilayerperceptron neural net algorithm had 99.71% of overall accuracy. For
classifying high value customers, the decision tree algorithm had 99.98% of accuracy,

91
while the neural net algorithm had 99.47% of accuracy. Moreover, in classifying low
value customers the decision tree algorithm had 99.96% of accuracy, while neural net
algorithm had 99.70% of accuracy. Since, the decision tree model had scored better
performance in all these evaluation parameters, it is the researcher‟s belief that decision
tree classification model has an appropriate technique for this research on CRM.

In general, the results from this study were encouraging. It was possible to segment
customers‟ data using data mining techniques that made business sense. To this effect,
related literature on data mining techniques, CRM and customer segmentation was
reviewed.

5.2 Recommendations
The researcher makes the following recommendations based on the findings of this study.

Build a Customer Data Warehouse

This study has taken considerable time for the data preparation phase. This time can be
shortened if the data warehouse is established beforehand. So, the researcher strongly
recommends that the ERCA to develop data warehouse which contains customer
behavior/profiles. The data warehouse could be used not only for data mining but also for
other statistical analysis (OLAP) and report generation.

Model Performance Improvements

The model building process in this investigation was carried out in two sub-phases. For
clustering, the researcher used the simple K-means algorithm, whereas for classification
J48 decision tree and multilayerperceptron neural net learner algorithms were tested.
Though the results found are encouraging, refinement to both the segmentation and
multilayerperceptron neural net classification were needed. Based on this, the researcher
recommends the following for further data mining research:

 Inclusion of additional customer attributes: Though the attribute selection was

done together with the domain experts, only limited numbers of all the possible

92
 member of attributes were used. However, the researcher believes that better
results were found through the use of as much attributes as possible.
 Integration of customs and revenue database: The present investigation only
focuses on the customs ASYCUDA databases. This database only contains
revenue which is generated when customers import an item. However, further
investigations by integrating the customs and revenue databases will bring about
better customer segmentation in ERCA.
 Comparison of the K-means clustering algorithm with other clustering algorithms:
The simple K-means algorithm which was used in this investigation has good
performance to segment high, medium, and low value customers. However,
further investigation by using other clustering algorithms such as
HierarchicalClusterer, DBScan, FarthestFirst, FillteredClusterer and SOM, would
have either confirmed or adjusted the result of this investigation.
 Additional trial by changing different parameter value of multilayerperceptron
neural net algorithm: As Melaku (2009) recommended that by increasing the
number of records and adjusting the default parameter values of
multilayerperceptron neural net algorithm, it is possible to increase the accuracy
of the classification model. As a result, in this research, by increasing the dataset
size from 10,090 to 46,748 and adjusting the default momentum parameter value
from 0.2 to 0.4, the overall accuracy rate increases from 98.62% to 99.71%.
However, in this investigation only momentum, learning rate and hidden nodes of
multilayerperceptron neural network algorithm‟s parameters were tried at
different values to improve the classification accuracy of the resulting model. So,
the researcher further recommends testing the remaining parameter values, such
as value of number of iteration (training time), validation set size, and validation
threshold.
 Reducing within cluster sum of squared error: During the cluster experimentation,
the cluster sum of squared error was relatively large, which was around 3101. So,
it is the researcher‟s believe that by further adjusting some parameter values, the
sum of squared error may be reduced.

93
Enhancing CRM and Data Mining Understanding

To increase the satisfaction of customers so as to generate higher revenue, the ERCA

should expose its employees to the importance of CRM. Moreover, data mining
techniques for customer segmentation should enable the Authority to identify potential
customers. Therefore, to use this capability, awareness on the advantages of CRM among
employees at all levels should be created.

94
 REFERENCES
Adomavicius, Gediminas and Alexander Tuzhilin. 2001. Using data mining methods to
build customer profiles. New York: IEEE 1: 74-82

Anyanwu, N. Matthew and Sajjan G. Shiva. n.d. Comparative analysis of serial decision
tree classification algorithms. International Journal of Computer Science and
Security, (IJCSS) 3: 230-240

Apte, Chidanand, Bing Liu, Edwin Pednault and Padhraic Smyth. 2002. Business
applications of data mining. Communications of the ACM 8:49-53

Arai, Kohei and Ali Ridho Barakbah. 2007. Hierarchical K-means: an algorithm for
centroids initialization for K-means. Saga University: 1:25-31

Arthur, David and Sergei Vassilvitskii. 2006. How slow is the K-means method.
http://www.cs.duke.edu/courses/spring07/cps296.2/papers/kMeans-socg.pdf (access
date February 17, 2011)

Berndt, Adele, Frikkie Herbst, and Lindie Roux. 2005. Implementing a customer
relationship management program in an emerging market. Journal of Global
Business and Technology 1: 81-89

Bose, Ranjit. 2002. Customer relationship management key components for IT success.
Industrial Management and Data System 2:89-97

Boulding, William, Richard Staelin, Michael Ehret, & Wesley J. Johnston. 2005. A
customer relationship management roadmap: what is known, potential pitfalls, and
where to go. Journal of Marketing 1:155–166

Bounsaythip, Catherine and Esa Rinta-Runsala. 2001. Overview of data mining for
customer behavior modeling. VTT Information Technology 18: 1-53

Bull, Christopher. 2003. Strategic issues in customer relationship management (CRM)

implementation. Business Process Management Journal 9: 592-602

Chalmeta, Ricardo. 2006. Methodology for customer relationship management. The

Journal of Systems and Software 79:1015–1024

95
Chen, Ruey-Shun, Ruey-Chyi Wu and J. Y. Chen. 2005. Data mining application in
customer relationship management of credit card business. IEEE Annual
International Computer Software and Applications Conference 2:730-3157

Cross, Glendon and Wayne Thompson. 2008. Understanding your customer:

segmentation techniques for gaining customer insight and predicting risk in the
telecom industry. SAS Global forum data mining and predictive model
http://www2.sas.com/proceedings/forum2008/154-2008.pdf (access date February
10, 2011)

Denekew, Abera. 2003. The application of data mining to support customer relationship
management at Ethiopian Airlines. Unpublished Master‟s Thesis, Department of
Information Science, Faculty of Informatics, Addis Ababa University, Addis Ababa

Deshpande, S. P. and V. M. Thakare. 2010. Data mining system and application: a

review. International Journal of Distributed and Parallel systems (IJDPS) 1:32-44

Doye, Tom. 2010. Collaborative border management. World Customs Journal 4:333-340

Dunham, H. Margaret. 2000. Data mining techniques and algorithms partial draft of
forthcoming book from prentice hall. http://www.cba.ua.edu/~mhardin/dunham.pdf
(access date January 5, 2011)

Edelstein, H. 2002. Data mining: exploiting the hidden trends in your data. Technology
in Society 4:483-502

European Regulators Group for Electricity and Gas (ERGEG). 2010. GGP on customer
complaint handling, reporting and classification. Ref: E10-CEM-33-05

Faber, Vance. 1994. Clustering and the continuous K-means algorithm. Los Alamos
Science 22: 138-144

Farn, Cheng-Kiang and Li Ting Huang. 2009. A study on industrial customer’s loyalty to
application service providers: the case of logistics information services.
International Journal of Computers 3: 151-160

Fayyad, Usama, Gregory Piatetsky-Shapiro, and Padhraic Smyth. 1996. From data
mining to knowledge discovery in databases. American Association for Artificial
Intelligence 37-54

96
Fayyad, Usama, Gregory Piatetsky-Shapiro, and Padhraic Smyth. 1996. Knowledge
discovery and data mining: towards a unifying framework.
http://www-aig.jpl.nasa.gov/kdd96 (access date March 2, 2011)

Federal Negarit Gazeta of the Federal Democratic Republic of Ethiopia. 2008.

Proclamation No. 587/2008, Proclamation Page 4123

Fekadu, Mekonnen. 2004. Application of data mining to support customer relationship

management at Ethiopian Telecommunications Corporation. Unpublished Master‟s
Thesis, Department of Information Science, Faculty of Informatics, Addis Ababa
University, Addis Ababa

Gray, Paul and Jongbok Byun. 2001. Customer relationship management.

http://www.wcfia.harvard.edu/us-japan/research/pdf/06-13.ueno.pdf (access date
February 15, 2011)

Gray, Paul. 2001. Customer relationship management.

http://www.crito.uci.edu/papers/2001/crm.pdf (access date February 16, 2011)

Greengrove, Kathryn. 2002. Needs-based segmentation: principles and practice. USA:

International Journal of Market Research 44: 405-421

Hajizadeh, Ehsan, Hamed Davari Ardakani and Jamal Shahrabi. 2010. Application of
data mining techniques in stock markets. Journal of Economics and International
Finance 7:109-118

Han, Jiawei and Micheline Kamber. 2006. Data mining: concepts and techniques. 2nd
ed.USA: Morgan Kaufmann

Henock, Woubishet. 2002. Application of data mining techniques to support customer

relationship management at Ethiopian Airlines. Unpublished Master‟s Thesis,
Deparment of Information Science, Faculty of Informatics, Addis Ababa University,
Addis Ababa

http://www. Ethiopian Revenues and Customs Authority (access date January 7, 2011)

Huang, Yingping. 2003. Infrastructure, data cleansing and mining for support of
scientific simulation. Department of Computer Science and Engineering, University
of Notre Dame, Indiana

97
Hwang, Hyunseok, Taesoo Jung and Euiho Suh. 2004. An LTV model and customer
segmentation based on customer value: a case study on the wireless
telecommunication industry. Expert Systems with Applications 26: 181–188

Johnston, Robert and Sandy Mehra. 2002. Best-practice complaint management.

Academy of Management Executive 4:145-154

Kanungo, Tapas, David M. Mount, Nathan S. Netanyahu, Christine D. Piatko, Ruth

Silverman, and Angela Y. Wu. 2002. An efficient k-means clustering algorithm:
analysis and implementation. IEEE Transactions on Pattern Analysis and Machine
Intelligence 24:881-892

Kim, Jonghyeok, Euiho Suh, and Hyunseok Hwang. 2003. A model for evaluating the
effectiveness of CRM using the balanced scorecard. Journal of Interactive Marketing
2: 5-19

Kim, Su-Yeon, Tae-Soo Jung, Eui-Ho Suh, and Hyun-Seok Hwang. 2006. Customer
segmentation and strategy development based on customer lifetime value: a case
study. Expert Systems with Applications 31: 101–107

King, F. Stephen and Thomas F. Burgess. 2008. Understanding success and failure in
customer relationship management. Industrial Marketing Management 37: 421–431

Koh, Chye Hian and Gerald Tan. n.d. Data mining applications in healthcare. Journal of
Healthcare Information Management 2:64-72

Kumar, Ela and Arun Solanki .2010. A combined mining approach and application in tax
administration. International Journal of Engineering and Technology 2:38-44

Kumneger, Fekrie. 2006. Application of data mining techniques to support customer

relationship management for Ethiopian Shipping Lines (ESL). Unpublished Master‟s
Thesis, Department of Information Science, Faculty of Informatics, Addis Ababa
University, Addis Ababa

Kurgan, A. Lukasz, and Petr Musilek. 2006. A survey of knowledge discovery and data
mining process models. United Kingdom: Cambridge University Press 21: 1-24

Mahler, J. Juliannem and Thomas Hennessey. 1996. Taking internal customer

satisfaction seriously at the U.S. customs service. Jstor‟s M.E Sharpe 4:487-497

98
McGuirk, Mike. 2007. Customer segmentation and predictive modeling.
http://www.iknowtion.com/downloads/Segmentation.pdf (access date March 2,
2011)

McKinsey and Company. 2001. The new era of customer loyalty management.
htttp://www.marketing.mckinsey.com (access date February 4, 2011)

Melaku, Girma. 2009. Applicability of data mining techniques to customer relationship

management (CRM): the case of Ethiopian Telecommunications Corporation’s
(ETC) code division multiple access (CDMA) telephone service. Unpublished
Master‟s Thesis, Department of Information Science, Faculty of Informatics, Addis
Ababa University, Addis Ababa

Nasereddin, H. O. Hebah. 2009. Stream data mining. International Journal of Web

Applications 4:183-190

Ngai, E.W.T. 2005. Customer relationship management research: an academic literature

review and classification. Marketing Intelligence and Planning 6: 582-605

Ngai, E.W.T., Li Xiu , D.C.K. Chau. 2009. Application of data mining techniques in
customer relationship management: a literature review and classification. Expert
Systems with Applications 36:2592-2602

Parvatiyar, Atul and Jagdish N. Sheth. 2001. Customer relationship management:

emerging practice, process, and discipline. Journal of Economic and Social
Research 2:1-34

Rygielski, A. Chris, Jyun-Cheng Wang B, David C. Yen. 2002. Data mining techniques
for customer relationship management. Technology in Society 24:483–502

Saarenvirta, G.. 1998. Mining customer data.

http://www.db2mag.com/db_area/archives/1998/q3/98fsaar.html (access date
January 6, 2011)

Shang, S. C. Shari and Chih-Hsiang Chen. n.d. Human processes in customer

relationship management. 11th Pacific-Asia Conference on Information Systems
http://www.pacis-net.org/file/2007/1259.pdf (access date January 6, 2011)

Singh, Yashpal and Alok Singh Chauhan. 2009. Neural networks in data mining. India:
Journal of Theoretical and Applied Information Technology 37-42

99
Srivastava, Jaideep. 2002. Data mining for customer relationship management (CRM).
Advances in Knowledge Discovery and Data Mining 2336:14-27

Suresh, Hemamalini. 2002. Customer relationship management: an opportunity for

competitive advantage. India: PSG Institute of Management
http://www.realmarket.com/required/psginst1.pdf (access date February 6, 2011)

Tilahun, Muluneh. 2009. Possible application of data mining techniques to target

potential visa card users in direct marketing: the case of Dashen Bank s.c.
Unpublished Master‟s Thesis, Department of Information Science, Faculty of
Informatics, Addis Ababa University, Addis Ababa

Trappey, V .Charles, Amy J.C. Trappey, Ai-Che Chang, and Ashley Y.L. Huang. 2009.
The analysis of customer service choices and promotion preferences using
hierarchical clustering. China: Journal of the Chinese Institute of Industrial
Engineers 5:367-376

Two Crows Corporation. 1999. Introduction to data mining and knowledge discovery.
3rd ed. ISBN: 1-892095-02-5, Potomac, MD 20854 (U.S.A.)

Verhoef, C. Peter. 2003. Understanding the effect of customer relationship management

efforts on customer retention and customer share development. Jstor the Journal of
Marketing 4: 30-45

Wahab, Samsudin and Juhary Ali. 2010. The evolution of relationship marketing (RM)
towards customer relationship management (CRM): a step towards company
sustainability. Information Management and Business Review 1: 88-96

WeiWang and Shidong Fan. 2010. Application of data mining technique in

customer segmentation of shipping enterprises. China: IEEE 1:4

Witten, H. Ian and Eibe Frank. 2005. Data mining practical machine learning tools and
techniques. 2nd ed. USA: Morgan Kaufmann

Zaiane, R. Osmar .1999. Principles of knowledge discovery in databases. University of

Alberta http://webdocs.cs.ualberta.ca/~zaiane/courses/cmput690/slides/ch0s.pdf
(access date February 5, 2011)

100
 APPENDICES
Appendix 1 Partial View of the Initial Collected Sample Data

101
Appendix 2 Sample of the Decision Tree Generated With 10-Fold Cross-Validation
Technique
J48 pruned tree
------------------

TOTAL_REVENUE <= 652333.11

| CTY_DSC = AS

| | TOTAL_ITM <= 43

| | | TOTAL_ITM <= 37

| | | | REVENUE_ITEM <= 162139.61: cluster0 (16044.0)

| | | | REVENUE_ITEM > 162139.61

| | | | | TOTAL_REVENUE <= 630615.86: cluster0 (432.0)

| | | | | TOTAL_REVENUE > 630615.86: cluster2 (21.0)

| | | TOTAL_ITM > 37

| | | | TOT_INVOICED <= 87408.66: cluster0 (257.0)

| | | | TOT_INVOICED > 87408.66: cluster3 (71.0)

| | TOTAL_ITM > 43

| | | TOTAL_REVENUE <= 33962.11: cluster0 (26.0)

| | | TOTAL_REVENUE > 33962.11: cluster3 (628.0)

| CTY_DSC = EU: cluster4 (4210.0)

| CTY_DSC = AM

| | TOTAL_REVENUE <= 478053.49

| | | TOTAL_ITM <= 42: cluster0 (353.0)

| | | TOTAL_ITM > 42

| | | | TOTAL_ITM <= 49: cluster0 (8.0)

| | | | TOTAL_ITM > 49: cluster3 (3.0)

| | TOTAL_REVENUE > 478053.49: cluster4 (61.0)

| CTY_DSC = AF

| | TOTAL_REVENUE <= 448994.77: cluster0 (348.0)

| | TOTAL_REVENUE > 448994.77

102
| | | TOTAL_REVENUE <= 498887.78

| | | | TOTAL_ITM <= 9: cluster0 (7.0/1.0)

| | | | TOTAL_ITM > 9: cluster4 (8.0/1.0)

| | | TOTAL_REVENUE > 498887.78: cluster4 (22.0)

| CTY_DSC = LA

| | TOTAL_REVENUE <= 490154.39: cluster0 (50.0)

| | TOTAL_REVENUE > 490154.39: cluster4 (10.0)

| CTY_DSC = AU

| | TOTAL_REVENUE <= 435824.59: cluster0 (39.0)

| | TOTAL_REVENUE > 435824.59: cluster4 (2.0)

TOTAL_REVENUE > 652333.11

| TOTAL_REVENUE <= 1393587.35

| | TOTAL_ITM <= 39

| | | CTY_DSC = AS

| | | | REVENUE_ITEM <= 1327510.21

| | | | | TOTAL_REVENUE <= 665252.37

| | | | | | TOTAL_ITM <= 14

| | | | | | | TOTAL_REVENUE <= 653509.24

| | | | | | | | TOTAL_ITM <= 5: cluster2 (7.0)

| | | | | | | | TOTAL_ITM > 5: cluster0 (6.0)

| | | | | | | TOTAL_REVENUE > 653509.24: cluster2 (129.0)

| | | | | | TOTAL_ITM > 14

| | | | | | | TOT_INVOICED <= 27476.95

| | | | | | | | TOTAL_ITM <= 21: cluster0 (4.0)

| | | | | | | | TOTAL_ITM > 21: cluster2 (21.0)

| | | | | | | TOT_INVOICED > 27476.95: cluster0 (66.0)

| | | | | TOTAL_REVENUE > 665252.37: cluster2 (11304.0/1.0)

| | | | REVENUE_ITEM > 1327510.21

| | | | | TOTAL_REVENUE <= 1337529.49: cluster2 (2.0)

103
Appendix 3 Output of the K-means Cluster Modeling with Different K and Seed
Values

Output of cluster run with K=3 and seed=100

Output of cluster run with K=4 and seed=1000

Output of cluster run with K=5 and seed=10

104