Title Page

CUSTOMER SEGMENTATION ON CLUSTERING ALGORITHMS

BY
TOH WEI XUAN

A REPORT
SUBMITTED TO
Universiti Tunku Abdul Rahman
in partial fulfillment of the requirements
for the degree of
BACHELOR OF COMPUTER SCIENCE (HONOURS)
Faculty of Information and Communication Technology
(Kampar Campus)

JAN 2023
 UNIVERSITI TUNKU ABDUL RAHMAN

REPORT STATUS DECLARATION FORM

Title: CUSTOMER SEGMENTATION ON CLUSTERING ALGORITHMS_

__________________________________________________________
__________________________________________________________

Academic Session: January 2023

I ____________ TOH WEI XUAN_________________

(CAPITAL LETTER)

declare that I allow this Final Year Project Report to be kept in

Universiti Tunku Abdul Rahman Library subject to the regulations as follows:
1. The dissertation is a property of the Library.
2. The Library is allowed to make copies of this dissertation for academic purposes.

Verified by,

_________________________ _________________________
(Author’s signature) (Supervisor’s signature)

Address:
7, Laluan Keledang Utara 2,
Taman Wang, _______Lim Jia Qi_________
31450 Menglembu, Perak Supervisor’s name

Date: 25/4/2023 Date: ___28/04/2023________

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
ii
 Universiti Tunku Abdul Rahman
Form Title : Sample of Submission Sheet for FYP/Dissertation/Thesis
Form Number: FM-IAD-004 Rev No.: 0 Effective Date: 21 JUNE 2011 Page No.: 1 of 1

FACULTY OF INFORMATION AND COMMUNICATION TECHNOLOGY

UNIVERSITI TUNKU ABDUL RAHMAN

Date: __1/April/2023____

SUBMISSION OF FINAL YEAR PROJECT /DISSERTATION/THESIS

It is hereby certified that Toh Wei Xuan (ID No: 19ACB03568) has completed this final year
project/ dissertation/ thesis* entitled “_Customer Segmentation on Clustering Algorithms_” under
the supervision of _Dr Lim Jia Qi_ (Supervisor) from the Department of Computer Science, Faculty
of Information and Communication Technology, and Dr Kh'ng Xin Yi (Co-Supervisor)* from the
Department of Computer Science, Faculty of Information and Communication Technology.

I understand that University will upload softcopy of my final year project / dissertation/ thesis* in
pdf format into UTAR Institutional Repository, which may be made accessible to UTAR
community and public.

Yours truly,

____________________
Toh Wei Xuan

*Delete whichever not applicable

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
iii
 DECLARATION OF ORIGINALITY
I declare that this report entitled “Customer Segmentation on Clustering Algorithms” is my
own work except as cited in the references. The report has not been accepted for any degree
and is not being submitted concurrently in candidature for any degree or other award.

Signature : _________________________

Name : Toh Wei Xuan

Date : 1 April 2023

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
iv
 ACKNOWLEDGEMENTS
I would like to express my sincere gratitude and appreciation to my supervisor, Dr. Lim Jia
Qi and my moderator, Dr Kh'ng Xin Yi, for their invaluable guidance, support, and
encouragement throughout my journey in the field of Machine Learning. Without their expert
knowledge and unwavering commitment to my development, I would not have been able to
complete this project successfully.

Their continuous support and advice have been instrumental in overcoming the challenges I
faced during this project. Their unwavering belief in my abilities and willingness to share
their knowledge and expertise have been invaluable in shaping my understanding of the
subject matter.

Once again, I would like to express my heartfelt thanks to my supervisor and moderator for
their golden opportunity, and for the immense impact they have had on my academic and
professional growth. I will forever cherish the experience and knowledge gained under their
mentorship.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
v
 ABSTRACT
This report presents an analysis of customer segmentation using various clustering
algorithms, including k-means, DBSCAN, GMM, and RFM. The aim of the study is to
identify customer groups based on their buying behaviour and demographic characteristics.
The study utilizes a dataset consisting of transactional and demographic data of customers
from an e-commerce company.

Firstly, descriptive analysis is performed to explore the characteristics of the dataset. Then, k-
means, DBSCAN, and GMM clustering algorithms are applied to segment customers based
on their buying behaviour. Finally, RFM (Recency, Frequency, Monetary) analysis is used to
segment customers based on their purchasing history.

The results show that all clustering algorithms were able to identify distinct customer groups
with varying characteristics. Furthermore, the RFM analysis was able to segment customers
based on their buying patterns, and provide insights into their behaviour.

Overall, the study demonstrates the effectiveness of different clustering algorithms and RFM
analysis in identifying customer segments. The insights gained from this study could
potentially be used by the e-commerce company to improve their marketing strategies and
customer engagement.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
vi
 TABLE OF CONTENTS

Title Page _________________________________________________________________ i

REPORT STATUS DECLARATION FORM_____________________________________ii

DECLARATION OF ORIGINALITY__________________________________________ iv

ACKNOWLEDGEMENTS ___________________________________________________ v

ABSTRACT ______________________________________________________________ vi

TABLE OF CONTENTS ____________________________________________________ vii

LIST OF FIGURES _________________________________________________________ x

LIST OF TABLES _________________________________________________________xii

LIST OF SYMBOLS ______________________________________________________ xiii

LIST OF ABBREVIATIONS _______________________________________________ xiv

1 Introduction___________________________________________________________ 1

1.1 Background Information ____________________________________________ 1

1.1.1 Customer Segmentation ____________________________________________ 2
1.1.2 Clustering _______________________________________________________ 4

1.2 Problem Statement _________________________________________________ 6

1.2.1 Selection of sample size, features, and variable __________________________ 6
1.2.2 Data cleaning and preprocessing _____________________________________ 6
1.2.3 Use appropriate clustering algorithms _________________________________ 6
1.2.4 Calculate optimal number of clusters. _________________________________ 7
1.2.5 Validation test for clustering_________________________________________ 7
1.2.6 Cluster interpretation and description __________________________________ 7

1.3 Motivation ________________________________________________________ 8

1.4 Objective _________________________________________________________ 8

1.5 Project Scope and Direction __________________________________________ 9

1.6 Contributions_____________________________________________________ 10

1.7 Report Organization _______________________________________________ 11

Bachelor of Computer Science (Honours)
Faculty of Information and Communication Technology (Kampar Campus), UTAR
vii
2 Literature Review ______________________________________________________ 12

2.1 Previous works on customer segmentation_____________________________ 12

2.1.1 Clustering techniques for customer segmentation _______________________ 13

2.2 Limitation of previous study ________________________________________ 14

3 System Methodology ___________________________________________________ 16

3.1 Data Pre-processing _______________________________________________ 16

3.1.1 Data Import _____________________________________________________ 16
3.1.2 Data Cleaning ___________________________________________________ 18
3.1.3 Feature Engineering ______________________________________________ 20
3.1.4 Data Scaling and Encoding _________________________________________ 21
3.1.5 Dimensionality Reduction _________________________________________ 22

3.2 Performance Metric (clarify internal and why not external) ______________ 24

3.3 Selection of optimal number of clusters, k _____________________________ 25

3.4 Clustering Algorithms _____________________________________________ 26

3.4.1 K-means Clustering ______________________________________________ 27
3.4.2 GMM Clustering _________________________________________________ 29
3.4.3 Dbscan Clustering ________________________________________________ 32
3.4.4 RFM+k-means Clustering__________________________________________ 34

4 Experiment/Simulation _________________________________________________ 38

4.1 Hardware ________________________________________________________ 38

4.2 Software _________________________________________________________ 38

4.2.1 Jupyter Notebook (Python3) ________________________________________ 38
4.2.2 Anaconda Prompt ________________________________________________ 39

4.3 Kmeans extension _________________________________________________ 40

4.4 GUI _____________________________________________________________ 41

5 Model Evaluation and Discussion_________________________________________ 43

5.1 Determine optimal number of clusters ________________________________ 43

5.1.1 k-means, GMM, DBSCAN _________________________________________ 43
5.1.2 RFM+kmeans ___________________________________________________ 44
Bachelor of Computer Science (Honours)
Faculty of Information and Communication Technology (Kampar Campus), UTAR
viii
 5.2 Clustering Evaluation ______________________________________________ 45
5.2.1 k-means and GMM _______________________________________________ 45
5.2.2 DBSCAN ______________________________________________________ 46
5.2.3 3D Plot Performance ______________________________________________ 47

5.3 Cluster Analysis (k-means. GMM and DBSCAN) _______________________ 49

5.3.1 Cluster Analysis (RFM) ___________________________________________ 53

6 Conclusion and Recommendations ________________________________________ 61

6.1 Summary ________________________________________________________ 61

6.2 Further Study ____________________________________________________ 63

REFERENCES ___________________________________________________________ 64

Appendix _________________________________________________________________ 67

FINAL YEAR PROJECT WEEKLY REPORT __________________________________ 86

POSTER _________________________________________________________________ 96

PLAGIARISM CHECK RESULT _____________________________________________ 97

Checklist FYP2______________________________________ Error! Bookmark not defined.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
ix
 LIST OF FIGURES

Figure 1.1 Targeted market segmentation phrases 2

Figure 1.2 Unlabeled examples grouped into three cluster. 4
Figure 1.3 An Overview of clustering taxonomy 5
Figure 3.1 Feature name with Datatype 17
Figure 3.2 Schematic diagram of Data cleaning 18
Figure 3.3 Detect Outliers 20
Figure 3.4 Drop Outliers 21
Figure 3.5 PCA with 3 components 23
Figure 3.6 k-means output 29
Figure 3.7 GMM output 31
Figure 3.8 DBSCAN output 34
Figure 3.9 RFM Mapping 35
Figure 3.10 RFM output 37
Figure 5.1 Silhouette score and Calinski-Harabaz 43
Figure 5.2 Comparison of all metric prediction 44
Figure 5.3 Elbow Method for RFM 44
Figure 5.4 Silhouette Score for RFM 45
Figure 5.5 Clustering Metrics Summary for k-means and GMM 45
Figure 5.6 GridSearch for eps 46
Figure 5.7 Clustering Metrics Summary for DBSCAN 47
Figure 5.8 Comparison between kmeans, GMM and DBSCAN 48
Figure 5.9 3D plot for RFM 49
Figure 5.10 Tabulate total number of clusters 49
Figure 5.11 Age vs spending 50
Figure 5.12 Education vs Spending 50
Figure 5.13 NumwebvisitsMonth vs Spending 51
Figure 5.14 Marital Status vs Spending 51
Figure 5.15 Familysize vs Spending 52
Figure 5.16 Tabulate number of customers in each cluster 53
Figure 5.17 Summary boxplot 53
Bachelor of Computer Science (Honours)
Faculty of Information and Communication Technology (Kampar Campus), UTAR
x
Figure 5.18 RFM distribution 55
Figure 5.19 RFM Age vs Spending 56
Figure 5.20 RFM Income vs Spending 56
Figure 5.21 RFM Education vs Spending 57
Figure 5.22 RFM Education vs Spending 57
Figure 5.23 RFM Numwebvisitmonth vs Spending 58
Figure 5.24 RFM Familysize vs Spending 58
Figure 5.25 RFM Days_enrolled vs Spending 59

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
xi
 LIST OF TABLES
Table Number Title Page

Table 1.1 Types of Customer Segmentation 3

Table 2.1 Comparative study for performance metric by authors 14
Table 3.1 Formulas for various performance metrics 24
Table 4.1 Spec Used 38

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
xii
 LIST OF SYMBOLS
k Number of cluster / components
ε epsilon

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
xiii
 LIST OF ABBREVIATIONS
PCA Principal Component Analysis
GMM Gaussian Mixture Model
DBSCAN Density-Based Spatial Clustering of Applications with Noise
RFM Recency, Frequency and Monetary
MICE Multiple Imputations by Chained Equations
AIC Akaike Information Criterion
BIC Bayesian information criterion
SOM Self-Organizing Map
t-SNE t-Distributed Stochastic Neighbor Embedding
CRM Customer Relationship Management
GUI Graphical User Interface
EM Expectation-Maximization

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
xiv
1 Introduction
1.1 Background Information
In today's competitive market, businesses are constantly looking for ways to gain a
competitive advantage. One such way is by understanding their customers better and tailoring
their marketing strategies to meet their needs. Customer segmentation is a powerful tool that
can help businesses achieve this goal.

Customer segmentation involves dividing customers into groups based on common

characteristics such as demographics, buying behavior, and preferences. These groups can
then be targeted with tailored marketing messages and offers, leading to improved customer
engagement and loyalty.

Cluster analysis is a popular technique for customer segmentation, and it involves grouping
customers based on their similarity in terms of certain attributes. There are several clustering
algorithms that can be used for this purpose, including k-means, DBSCAN, and GMM. Each
algorithm has its own strengths and weaknesses, and the choice of algorithm depends on the
nature of the data and the research question.

RFM analysis is another popular technique for customer segmentation, and it involves
analyzing customers based on their purchasing behavior. The analysis involves three metrics:
recency, frequency, and monetary value, which are used to segment customers into different
groups based on their purchasing patterns.

In this report, we apply several clustering algorithms and RFM analysis to segment customers
of an e-commerce company based on their transactional and demographic data. The aim of
the study is to identify distinct customer groups with unique characteristics, and to provide
insights into their behavior. The results of this study could potentially be used by the e-
commerce company to improve their marketing strategies and customer engagement [7].

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
1
1.1.1 Customer Segmentation
Customer segmentation is a vital aspect of customer relationship management (CRM) and
business intelligence. By dividing customers into groups based on shared characteristics such
as behavior, age, gender, education, location, and socioeconomic status, businesses can gain
insights into customer behavior, tailor marketing messages to specific segments, and improve
customer engagement and retention. Marketing personas, which personify a customer
segment, are often created from customer segmentation data and must be tightly matched to
customer categories for effective results. However, while customer segmentation can be a
powerful tool, it also has its limitations, such as the challenge of maintaining segmentation as
customer behavior changes over time. By understanding and implementing different types of
customer segmentation techniques, businesses can potentially improve their marketing
campaigns, target specific customer groups, and increase sales.

Customer segmentation is a vital component of business intelligence that enables

organizations to create, maintain, and grow long-term customer relationships. Traditional
marketer-designed segmentation models have certain limitations and may not fully uncover
insights and data patterns. The following stage is to put in place efficient campaign
management. The steps utilized for targeted consumer segmentation are depicted in Figure
1.1 [9].

Figure 1.1 Targeted market segmentation phrases

While traditional demographic characteristics are no longer sufficient for market

segmentation, non-demographic characteristics such as values, interests, and preferences are
critical to a successful marketing plan. For example, all the subcategories within the
demographic segment are comparable to one another, providing businesses with the data they
want to accurately target a certain client base with specific marketing strategies. All the
customer segmentation dimensions are summarized in Table 1.1 [7].

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
2
Segmentation Variables
Demographic Gender
Age
Occupation
Education
Generation
Family Size
Family life Cycle
Ethnicity
Religion
Nationality
Social class
Behavioral Occasions
Rate of usage
Brand loyalty

Status of user

Benefits desired
Readiness to buy
Psychographic Opinions

Attitudes
Activities
Values
Interest

Geographic Climate

Religion
Area Size
Population density
Table 1.1 Types of Customer Segmentation
Therefore, a mix of attributes from different dimensions of segmentation is essential for
developing a robust customer segmentation model. Overall, the ultimate purposes of
Bachelor of Computer Science (Honours)
Faculty of Information and Communication Technology (Kampar Campus), UTAR
3
customer segmentation are to maximize customer value to the company, optimize marketing
strategies, and improve customer experience.

1.1.2 Clustering
Clustering is a form of unsupervised learning strategy in the field of machine learning.
Unsupervised learning approach aims to discover intrinsic data patterns and actionable
insights from data sets that do not have a labelled output variable. It’s a type of exploratory
data analysis which involves grouping data sets into a specified number of clusters with
comparable features among the data points inside each cluster as shown in Figure 1.2 [9].

Figure 1.2 Unlabeled examples grouped into three cluster.

Clusters are made up of data points that are grouped together in such a way that the distance
between them is kept to a minimum. In other words, clusters are areas with a high density of
similar data points [9].

Clustering, also known as cluster analysis, is a vital data analysis technique and a field
within data mining. It involves grouping the dataset into clusters, where data points with
similar attributes are placed in the same group or cluster. The aim is to identify connections
between data points based on their raw data qualities. However, the main challenge is to
determine the appropriate number of clusters that are relevant and informative for analysis.
This is a systematic and iterative process that involves analyzing large volumes of raw data
for patterns and commonalities. The disorganized data is sifted for meaningful insights, and
then data points are assigned to clusters. To achieve good results in a market domain, a
particular clustering technique coupled with detailed examination of cluster properties based
on the clustering results may be ideal.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
4
 These are several clustering algorithms that differ from one another in terms of the
method they take to group items according to their attributes. Figure 1.3 [11] provides the
overview of clustering types following three different classes of categorization in this project.

Figure 1.3 An Overview of clustering taxonomy

In summary, clustering algorithms identify internally homogeneous and externally diverse
groupings. Customers have distinct behaviors, requirements, needs, and traits, and the main
aim of clustering is to identify different customer types and divide the customer base into
clusters with similar profiles, so that targeted marketing can be done more efficiently.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
5
1.2 Problem Statement

Effective customer segmentation is crucial for businesses seeking to maximize profits and
improve customer experiences. In this study, some alternatives were explored when it comes
to feature engineering, identifying significant variables, clustering algorithms implementation
and comparison and cluster interpretation. These steps are vital in the customer segmentation
pipeline. Addressing these challenges is essential for businesses to obtain accurate and
meaningful customer segmentation results, which can inform targeted marketing strategies,
improve customer engagement, and ultimately drive revenue growth [13].

1.2.1 Selection of sample size, features, and variable

Effective customer segmentation is a critical task for businesses, but selecting the appropriate
variables, determining the number of variables to consider, and determining the appropriate
sample size can be challenging. Improper selection of features, including irrelevant or
redundant features, can lead to inaccurate or meaningless cluster solutions, while an
inadequate sample size can lead to biased or incomplete results. Thus, there is a need to
identify best practices and guidelines for selecting variables, determining the number of
variables to consider, and determining the appropriate sample size to ensure accurate and
meaningful customer segmentation [15].

1.2.2 Data cleaning and preprocessing

Data cleaning and preprocessing are critical steps in preparing data for analysis. However,
many organizations face challenges in these areas, including dealing with missing values,
inconsistent formatting, duplicate entries, and outliers. Additionally, selecting the appropriate
preprocessing techniques can be difficult, as different methods may have varying effects on
the quality and accuracy of the resulting analysis. As a result, there is a need for effective
data cleaning and preprocessing strategies that can ensure data integrity and optimize data for
analysis [13].

1.2.3 Use appropriate clustering algorithms

Clustering is an essential technique for customer segmentation and market analysis, but
selecting the appropriate clustering algorithm is a complex task. Marketers and data analysts
must choose from various clustering algorithms, each with its own strengths and weaknesses,
and evaluate their suitability for the specific data set and business problem at hand. The

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
6
selection of inappropriate algorithms can lead to inaccurate or incomplete cluster solutions,
rendering the marketing efforts ineffective. Therefore, identifying the most appropriate
clustering algorithm that can provide accurate and meaningful insights for the marketing
decision-making process is critical for successful customer segmentation and targeted
marketing campaigns [15].

1.2.4 Calculate optimal number of clusters.

Determining the optimal number of clusters is a crucial step in clustering analysis. However,
it is a challenging task as it requires selecting the right method for calculating the number of
clusters and interpreting the results. The choice of the wrong method can lead to incorrect
conclusions, which can have significant consequences. Therefore, there is a need for a
systematic and reliable method for calculating the optimal number of clusters that takes into
account the characteristics of the data set and the clustering algorithm used. The problem is to
develop a method that can accurately determine the optimal number of clusters for a given
data set and clustering algorithm, considering the sample size, the number of variables, and
other relevant factors [15].

1.2.5 Validation test for clustering

Effective validation of clustering results is essential to ensure that the clusters obtained from
the data are meaningful and useful. However, determining the optimal validation test for a
given clustering problem is challenging, as there are various validation measures and
methods available, and each has its own strengths and weaknesses. Furthermore, the
effectiveness of the chosen validation test may depend on the type of clustering algorithm
used, the characteristics of the data, and the desired outcome of the clustering analysis.
Therefore, identifying the appropriate validation test to use in each clustering problem is
crucial to ensure the reliability and validity of the results.

1.2.6 Cluster interpretation and description

Despite the usefulness of clustering analysis in identifying distinct groups of customers or

objects, effectively interpreting and describing the resulting clusters remains a significant

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
7
challenge. This is due to the complex and often subjective nature of interpreting cluster
solutions, which involves making sense of the underlying patterns and relationships between
the variables and the clusters. Moreover, different clustering algorithms and validation
measures may produce different cluster solutions, making it difficult to choose the best one.
As a result, there is a need for more robust and standardized methods for cluster interpretation
and description that can help researchers and practitioners gain better insights into the
characteristics and behaviors of the clusters and make more informed decisions based on the
clustering results [15].

1.3 Motivation
Customer segmentation is a crucial aspect of any business that aims to better understand its
customers and tailor its products or services to meet their needs. However, selecting the
appropriate clustering algorithm for customer segmentation is a challenging task that requires
a thorough understanding of the strengths and weaknesses of different techniques. This report
aims to provide a comprehensive analysis of four popular clustering algorithms, namely K-
Means, DBSCAN, GMM, and RFM+k-means, in the context of customer segmentation. By
comparing the performance of these algorithms on a real-world dataset, this report will
provide valuable insights into their effectiveness in identifying meaningful customer
segments. Such insights can help businesses make more informed decisions about how to
allocate their resources and develop targeted marketing strategies that can improve customer
satisfaction and loyalty.

1.4 Objective
a) To compare the effectiveness of k-means, DBSCAN, GMM, and RFM clustering
algorithms in segmenting customers based on their purchasing behavior.
b) To determine the optimal number of clusters for each algorithm using internal
validation measures such as silhouette, AIC, BIC, Calinski-Harabaz, and Davies-
Boudin.
c) To interpret and describe the resulting clusters in terms of customer characteristics,
preferences, and behaviors.
d) To provide recommendations for marketers and business owners on how to use
clustering analysis to improve customer targeting and retention strategies.
Bachelor of Computer Science (Honours)
Faculty of Information and Communication Technology (Kampar Campus), UTAR
8
1.5 Project Scope and Direction

The scope of this project is to explore and compare four popular clustering algorithms for
customer segmentation: K-means, DBSCAN, GMM, and RFM. The project will involve
conducting a comprehensive literature review on each algorithm and its applications in
customer segmentation. The project will also involve implementing each algorithm on a
dataset of customer transactions and comparing the results obtained.
The direction of the project will be as follows:

a) Conduct a literature review of K-means, DBSCAN, GMM, and RFM algorithms and
their applications in customer segmentation.
b) Pre-process the customer transaction data, including data cleaning, missing value
imputation, and data normalization.
c) Implement each algorithm on the pre-processed dataset and obtain the corresponding
clusters.
d) Evaluate and compare the results obtained from each algorithm in terms of cluster
quality metrics such as Silhouette score, Dunn index, and Davies-Bouldin index.
e) Perform a detailed analysis of the cluster characteristics and interpret the results
obtained from each algorithm.
f) Discuss the strengths and weaknesses of each algorithm and provide
recommendations for their appropriate use in customer segmentation.
g) Summarize the findings of the project and provide future research directions.
h) Develop a user-friendly GUI using MiniBatchKmeans
The project aims to provide insights into the strengths and weaknesses of each algorithm and
their suitability for customer segmentation. The project will be useful for marketers and
business analysts who are interested in using clustering algorithms for customer
segmentation.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
9
1.6 Contributions

This report aims to contribute to the field of customer segmentation by providing a

comprehensive comparison of five popular clustering algorithms: K-means, DBSCAN,
GMM, RFM, and MiniBatch K-means but the MiniBatch K-means only be used in GUI
purpose. By incorporating MiniBatch K-means, a variant of K-means algorithm that is
specifically designed for large datasets, we expand the scope of our analysis and evaluate its
performance in identifying distinct customer segments.

Through the application of these clustering algorithms to a real-world dataset, we will

analyze their effectiveness in uncovering meaningful customer segments and extracting
valuable insights for businesses. By examining the strengths and limitations of each
algorithm, this report will enable marketing professionals, business owners, and data analysts
to make informed decisions regarding the selection of the most suitable algorithm for their
specific needs.

Furthermore, this report will provide practical guidance on selecting relevant variables,
determining optimal cluster numbers, validating cluster solutions, and interpreting cluster
descriptions. By bridging the gap between theoretical concepts and practical applications, this
report offers a valuable resource to those interested in leveraging customer segmentation and
clustering algorithms to improve customer targeting, enhance marketing campaigns, and
optimize overall business performance.

Overall, this study provides a comprehensive analysis and comparison of five clustering
algorithms, including MiniBatch K-means, offering valuable insights and practical guidance
for the field of customer segmentation.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
10
1.7 Report Organization

This report is organized into six chapters, each addressing different aspects of the project.
Chapter 1, the Introduction, provides a comprehensive overview of the project. It covers the
background information, problem statement, project motivation, scope, objectives, project
contribution, highlights of project achievements, and the overall organization of the report.
This chapter sets the stage for the subsequent chapters, establishing the context and purpose
of the study.

Chapter 2 focuses on the Literature Review, where an extensive analysis of existing customer
algorithm techniques in the market is conducted. The aim is to evaluate and compare the
strengths and weaknesses of these techniques. This chapter provides valuable insights into the
current state of the field and serves as a foundation for the subsequent chapters.

In Chapter 3, the Methodology, the overall model design and methods employed in the
project are discussed. This chapter offers a detailed explanation of the chosen approach,
providing readers with a clear understanding of the methodologies and techniques utilized.

Chapter 4, the Experiment/Simulation chapter, centers around the hardware used and the
development of the graphical user interface (GUI) for the project. It outlines the technical
aspects of the implementation, highlighting the tools, equipment, and software utilized in the
experiments or simulations conducted.

Chapter 5 delves into Model Evaluation and Discussion. This chapter focuses on the analysis
of clusters and the examination of the clustering output. It involves a thorough evaluation and
discussion of the results obtained, allowing for a comprehensive assessment of the
performance and effectiveness of the clustering algorithms.

Finally, Chapter 6, the Conclusion and Recommendations chapter, provides a summary of the
project's findings. It also offers recommendations for future work and identifies potential
areas for further exploration and improvement in the field of customer segmentation using
clustering algorithms.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
11
2 Literature Review
2.1 Previous works on customer segmentation

Customer segmentation has been a vital area of research in marketing and business analytics,
as it helps to identify different groups of customers with similar characteristics and behaviors
and create targeted marketing strategies to meet their specific needs. In this chapter, we
review previous works on customer segmentation techniques in the literature.

One of the most used customer segmentation techniques is the RFM (Recency, Frequency,
Monetary value) method, which is based on the principle that customers who have made
recent purchases, frequent purchases, and high-value purchases are more valuable to the
business. [1] used the RFM method to segment customers of an online retailer and found that
it was effective in identifying high-value customers and improving the business's overall
profitability.

In addition to the RFM method, clustering algorithms have also been widely used for
customer segmentation. [2] compared the performance of different clustering algorithms,
including K-means, SOM, and DBSCAN, on an e-commerce dataset. They found that the K-
means algorithm outperformed the others in terms of precision, recall, and F1-score.

[3] used K-means, GMM, and DBSCAN clustering algorithms to segment customers based
on their online shopping behavior. They found that the K-means algorithm was the most
accurate in identifying customer groups with distinct behavior patterns.

[4] conducted a comparative study of different customer segmentation methods based on

online shopping behavior. They evaluated the performance of K-means, GMM, and
hierarchical clustering algorithms using homogeneity, completeness, and V-measure metrics
and found that the hierarchical clustering algorithm outperformed the other two methods.

[5] used to cluster techniques, including K-means, DBSCAN, and Spectral clustering, to
detect credit card fraud. They found that the Davies-Bouldin index was a reliable metric for
evaluating the performance of the clustering algorithms.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
12
Overall, previous studies have shown that both RFM and clustering techniques can be
effective in customer segmentation, and the choice of a particular method depends on the
nature of the data and the business problem at hand. In the following section, we discuss the
different clustering algorithms used in our study for customer segmentation.

2.1.1 Clustering techniques for customer segmentation

Customer segmentation is a widely used marketing strategy that involves dividing customers
into groups based on their characteristics and behaviors. Clustering techniques are a popular
tool for customer segmentation as they allow for the identification of distinct groups of
customers with similar attributes. There are several clustering techniques that can be used for
customer segmentation, including k-means, DBSCAN, GMM, and RFM clustering.

K-means clustering is a popular unsupervised learning algorithm that involves dividing a

dataset into k clusters. The algorithm works by assigning each observation to the cluster with
the closest centroid, then re-calculating the centroids based on the mean of all the
observations in the cluster. The process is repeated until the centroids converge. K-means
clustering has been widely used in customer segmentation due to its simplicity and speed.

DBSCAN is another clustering algorithm that is often used for customer segmentation. This
algorithm groups together points that are close to each other and separates points that are
farther away. DBSCAN can handle datasets with irregular shapes and is not sensitive to
outliers. It has been shown to be effective in identifying dense regions in customer data,
which can be useful in identifying customer preferences and behaviors.

GMM is a probabilistic clustering algorithm that assumes that the data points are generated
from a mixture of Gaussian distributions. GMM can identify complex patterns in customer
data and can handle datasets with overlapping clusters. It has been shown to be effective in
identifying hidden customer segments, which may not be apparent from traditional
demographic data [1].

RFM clustering is a customer segmentation technique that is based on three key metrics: how
recently a customer has made a purchase, how frequently a customer makes purchases, and

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
13
how much money a customer spends. RFM clustering is often used in e-commerce and retail
industries to identify high-value customers and to tailor marketing strategies to individual
customers based on their purchasing behavior.

Overall, these clustering techniques have been widely used in customer segmentation and
have been shown to be effective in identifying distinct customer segments. Each technique
has its own strengths and weaknesses, and the choice of technique will depend on the specific
characteristics of the dataset and the research questions at hand.

A comprehensive comparison of different clustering algorithms developed under various

methodologies, considering various characteristics of clustering. We added annotations for
each algorithm, which offer a good picture of the benefits and drawbacks of each
approach[9].
Author Clustering Dataset Performance metric
techniques
Lee et al. (2015) [1] K-means, Online retail Silhouette
Hierarchical transactions Coefficient
Kumar et al. (2016) K-means, SOM, E-commerce Precision, Recall,
[2] DBSCAN transactions F1-score
Xiang and Gong K-means, GMM, E-commerce user Accuracy
(2018) [3] DBSCAN behavior
Chen et al. (2020) K-means, GMM, Online shopping Homogeneity,
[4] Hierarchical behavior Completeness, V-
measure
Jadhav and K-means, DBSCAN, Credit card Davies-Bouldin
Sonawane (2021) Spectral transactions Index
[5]
Yeh, Y.-L., & RFM Online retailer Recency, Frequency,
Huang, C.-C. (2018) customer purchase Monetary Value,
[6] Latency
Table 2.1 Comparative study for performance metric by authors

2.2 Limitation of previous study

For a marketer, the first and most fundamental constraint of cluster analysis is that you must
have access to relevant customer data. For instance, if we work for a service company, we
may have a reasonable client database that we can use to perform cluster analysis and
discover market categories. Larger companies are more likely to have access to relevant
marketing research survey data. However, most businesses, particularly smaller and newer
Bachelor of Computer Science (Honours)
Faculty of Information and Communication Technology (Kampar Campus), UTAR
14
ones, will lack access to relevant data and hence will be unable to use cluster analysis.
Cluster analysis is also limited by the fact that it is merely a statistical technique that
presupposes no prior knowledge of the market or how consumers might react. In other words,
it's simply clustering data around a set of core points, which may or may not make sense after
the analysis is completed. The technique's talent is not in doing the analysis, but in
understanding and applying the results to identify appropriate market segments and then
developing a successful marketing strategy that targets one or more of these segments.
Besides, some cluster analysis methods produce somewhat different results each time the
statistical analysis is done. This can arise because there is no one-size-fits-all method to data
analysis. Some type of random or arbitrary approach to "guessing" the possible locations
(means) of the various data centers (that is, market segments) is chosen at the start, especially
when there are many variables to consider. As a result, the results can vary based on the
original starting point (seed) of the data for each of the segments. Even if you use the same
statistical program to run the same data set, this can happen since the underlying statistical
technique can be to utilize a random starting point. [7]

Despite the promising results of the previous studies, there are several limitations that need to
be addressed. First, most of the studies used a limited number of clustering algorithms, which
might not be suitable for all types of datasets. Second, some studies did not include domain
knowledge in the clustering process, which might result in less effective segmentation. Third,
some studies used a limited number of performance metrics, which might not fully capture
the effectiveness of the segmentation. Finally, some studies focused on a specific industry or
dataset, which might limit the generalizability of the findings to other contexts.

These limitations highlight the need for a more comprehensive and systematic approach to
customer segmentation using clustering algorithms. In this study, we aim to address these
limitations by comparing multiple clustering algorithms, incorporating domain knowledge,
using a variety of performance metrics [8].

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
15
3 System Methodology
3.1 Data Pre-processing
This study outlines the methodology used to conduct the study on customer segmentation
using clustering techniques. In this section, we will describe the data collection process, data
preprocessing steps, clustering algorithms used, and evaluation metrics. The methodology
section is crucial as it provides a detailed explanation of how the study was conducted, and it
enables the reader to understand the study's reliability and validity.

3.1.1 Data Import

The dataset used in this study is the "Customer Personality Analysis" dataset obtained from
Kaggle ([Link] The
dataset contains information about comprehensive analysis of a company's preferred
customers which enables businesses to gain a better understanding of their customers and
customize their products to meet the specific needs, behaviors, and concerns of different
customer segments and includes demographic information, product information, and
responses to marketing campaigns. By analyzing the different customer segments, businesses
can identify which segment is more likely to purchase a particular product, which allows
them to market the product effectively to that segment instead of spending money on
marketing to all customers in the database.

The following code was used to import and load the dataset into a pandas dataframe:
#Imports
import numpy as np
import pandas as pd

#Set the working directory to where the data file is located

#sep='\t' was a delimiter to act as tab function
df = pd.read_csv('marketing_campaign.csv', sep = '\t')

The dataset has 29 columns and 2,240 rows. The columns and their respective data types are
listed in figure 3.1:

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
16
 Figure 3.1 Feature name with Datatype

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
17
3.1.2 Data Cleaning
When working with a dataset, it is important to ensure that the data is clean and ready for
analysis. This process is known as data cleaning or data preprocessing. In this step, I handle
missing or duplicate data, check for outliers, and ensure that the data is in the correct format
for analysis.

Figure 3.2 Schematic diagram of Data cleaning

In our study, data cleaning performed on the 'marketing_campaign.csv' dataset that was
imported using pandas. First, missing data was checked using the isnull() function in pandas.
There were total of 24 missing value in feature ‘Income’.

To handle the missing data, we decided to impute the missing values using the MICE
imputation method. This method replaces missing data with predicted values based on other
variables in the dataset. mice() function from the impyute library used to impute the missing
values. The IterativeImputer (MICE) performns multiple regressions over random sample
ofthe data, then takes the average ofthe multiple regression values and uses that value to
impute the missing value. MICE selected to use due to it was one type of univariate
imputation algorithm, which imputes values in the i-th feature dimension using only non-
missing values in that feature dimension (e.g. [Link]). By contrast,
multivariate imputation algorithms use the entire set of available feature dimensions to
estimate the missing values (e.g. [Link]). The MICE code was shown in
Appendix.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
18
Next, checked for duplicates in the dataset using the duplicated() function in pandas. There
was no duplicated features and data in this dataset.

After handling missing value and duplicated features, the next step is performing the
following cleaning steps:
a) Restructuring Datatypes: the date format converted of the 'Dt_Customer' column to a
datetime format using the pd.to_datetime() function. the age of each customer
calculated by subtracting their birth year from the current year.
b) Combining features: Created a new feature called 'Total_AcceptedCampaign' which is
the sum of all accepted campaigns and responses.
c) Calculating important metrics: Calculate the days enrolled by subtracting the last
enrollment date from the 'Dt_Customer' column. Calculate the total spending of each
customer by adding up the spending on different products. Additionally, combined the
'Marital_Status', 'Kidhome', and 'Teenhome' columns to create a new 'Familysize'
feature.
d) Removing features not used: Remove columns that were not relevant to our analysis,
such as 'Year_Birth', 'MntWines', 'MntFruits', 'MntMeatProducts', 'MntFishProducts',
'MntSweetProducts', 'MntGoldProds','Kidhome', and 'Teenhome'. Columns related to
the accepted campaigns, complaints, and purchase types were dropped.
e) These cleaning steps were performed using pandas functions such as drop(), sum(),
replace(), astype(), and fillna(). The cleaned dataset was saved as a new CSV file for
further analysis.

Finally, The data was ensured in the correct format for analysis. The 'date' column was
conveted to a datetime format using the to_datetime() function in pandas, and the categorical
columns was encoded using label encoding from the [Link] library.

Overall, the data cleaning process helped to ensure that the dataset was ready for analysis and
that any errors or inconsistencies in the data were handled appropriately.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
19
3.1.3 Feature Engineering
The following feature engineering techniques were used to generate new features for the
dataset:

 Zero variance features like Z_CostContact and Z_Revenue were dropped

 Deriving the minimum number of household members from attributes Marital_Status,
Kidhome, and Teenhome
 Computing the total number of accepted offers by summing up the Response and five
of the AcceptedCmp features
 Generating a correlation heatmap to visualize the relationships between features.
 Dropping highly correlated features

After the feature extraction was handled above, outliers in the numerical columns were
checked using boxplots and histograms. Outliers were found in the 'Age' and 'Income'
columns, which were removed using the IQR (Interquartile Range) method as shown in
Figure 3.3. The removal of outliers was justified as extreme values outside the range of
typical values for these features were observed, which can occur due to measurement errors,
data entry errors, or other anomalies in the data. Since outliers can have a significant impact
on the model's performance, it was decided to remove them using the commonly used IQR
method. By removing these outliers, the model's accuracy and reliability were improved.

Figure 3.3 Detect Outliers

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
20
In Figure 3.4, some observations were made during the analysis of the data. For instance, it
was noted that some customers had an income greater than 150000 and a few even had an
income greater than 600000. Additionally, two shaded dots appeared above 100 in the age
feature, which was deemed impossible since it is highly unlikely for a human to live beyond
100 years old. However, it was observed that the other features of the data were reasonable.
As a result, the decision was made to drop the data with impossible features to ensure the
accuracy of the analysis. By dropping the data with extreme values, the risk of skewing the
results was reduced, leading to more reliable and accurate findings.

Figure 3.4 Drop Outliers

Figure 3.4 shows that the total number of data after removing the outliers are 2212 and thus
the final number of data to segment after cleaning and engineering was 2212 rows.

The data cleaning and feature engineering steps resulted in a cleaned and optimized dataset
that can be used for customer segmentation analysis. By removing outliers and dropping low
variance features, the dataset became more accurate and reliable for analysis. The new
features created were also useful for uncovering insights and trends in the data.

Overall, the data cleaning and feature engineering steps are critical in preparing the data for
analysis and ensuring that the results obtained are trustworthy and valuable.

3.1.4 Data Scaling and Encoding

To prepare the data for modeling, we performed data scaling and encoding. First, we
identified the columns with numerical data types, including both integers and floats, and
applied standard scaling using the StandardScaler function from the [Link]

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
21
library. This rescaled the numerical data to have a mean of 0 and a standard deviation of 1,
which can improve the performance of some machine learning models.

Next, we encoded the categorical columns using LabelEncoder from the

[Link] library. This transformed categorical data into numerical format,
allowing us to include these features in our machine learning models.

By scaling and encoding the data, it was the last step before performing any machine learning
model and it was ready to implement any clustering algorithms.

3.1.5 Dimensionality Reduction

To aid in the visualization of the pre-processed dataset, Principal Component Analysis (PCA)
was applied to reduce the dimensionality of the feature space. PCA is a commonly used
technique for dimensionality reduction that transforms the original features into a new set of
orthogonal features called principal components. These new features are ranked based on the
amount of variance they explain in the original dataset, with the first principal component
explaining the most variance and subsequent components explaining progressively less.

It is important to note that in this project, PCA was only used for visualization purposes and
not for feature selection. By reducing the dimensionality of the feature space, the data could
be more easily visualized and explored. However, the original features were retained for
modelling and analysis purposes. By utilizing PCA for visualization, the data was able to be
presented in a more clear and concise manner.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
22
 Figure 3.5 PCA with 3 components

In this analysis, PCA was performed on the preprocessed dataset to reduce the feature space
while retaining as much information as possible. The number of principal components was
chosen based on the amount of variance explained by each component and the cumulative
variance explained. In Figure 3.5, the dimensionality of the data was reduced from the
original 29 features to just 3 principal components using PCA. The resulting scatter plot
shows that the data points are now more tightly clustered and appear to form two distinct
groups. This suggests that the two clusters identified in the k-means clustering algorithm may
indeed be representative of underlying subgroups in the data. Overall, this analysis using
PCA provides a useful visualization of the data and supports the findings of the k-means
clustering analysis.

After running PCA, the dataset was reduced to n components, and these components were
used as input features for the subsequent modeling steps. The resulting PCA components
were then examined to identify the most important variables in each component and their
contributions to the overall variance in the dataset.

Overall, the use of PCA for dimensionality reduction resulted in a more manageable feature
space, while still retaining much of the original information present in the dataset.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
23
3.2 Performance Metric (clarify internal and why not external)
Index Name Formula Variables
Akaike Information AIC = 2k – 2ln(L) k: number of parameters; L:
Criterion (AIC) maximum likelihood
function of the model
Bayesian Information BIC = k*ln(n) – 2ln(L) k: number of parameters; L:
Criterion (BIC) maximum likelihood
function of the model; n:
sample size
Calinski-Harabaz Index CH = (B / (k - 1)) / (W / (n - B: between-cluster scatter;
k)) W: within-cluster scatter; k:
number of clusters; n: total
number of data points
Davies-Bouldin Index DB = (1/k) * sum(max(Rij + k: number of clusters; Rij:
Rik) / dij) average distance between
cluster i and j; dij: distance
between the centroids of
clusters i and j
Silhouette Score Silhouette Score = (b - a) / a: mean distance between a
max(a, b) sample and all other points
in the same cluster; b: mean
distance between a sample
and all other points in the
nearest cluster.
Table 3.1 Formulas for various performance metrics

The internal validation metrics for clustering, such as the ones mentioned in the previous
question (AIC, BIC, CH Index, DB Index, and Silhouette Score), are used to evaluate the
quality of the clustering structure without reference to external information or a ground truth.
These metrics assess how well the data points within each cluster are like each other and how
dissimilar they are to the points in other clusters.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
24
In contrast, external validation metrics evaluate the quality of clustering by comparing it to a
known ground truth or external information, such as predefined class labels. These metrics
include measures such as precision, recall, and F1-score.

While external validation metrics provide a more objective evaluation of clustering quality,
they are often not feasible in practice because the true class labels or external information
may not be available. Additionally, even when external information is available, it may not
necessarily be the best representation of the true structure of the data, which makes internal
validation metrics more generalizable.

Therefore, internal validation metrics are often preferred in practice for evaluating clustering
algorithms since they can provide insight into the clustering structure and its quality, even
when external information is not available. Hence, this study only used for the internal
validation metrics due to the dataset was no ground truth.

These indices are commonly used for evaluating the quality of clustering results. The AIC
and BIC indices are used for model selection, where a lower value indicates a better model
fit. The Calinski-Harabasz and Davies-Bouldin indices are used for evaluating the cluster
quality, where a higher value indicates better clustering performance.

The Silhouette Score is another commonly used metric for evaluating the quality of clustering
results. The score measures how similar an object is to its own cluster compared to other
clusters. The score ranges from -1 to 1, where a score closer to 1 indicates better clustering
performance.

3.3 Selection of optimal number of clusters, k

In order to determine the optimal number of clusters, a range of k values were tested and
evaluated using several performance metrics, including the Davies-Bouldin Index, Silhouette
Score, Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), and
Calinski-Harabasz Score. The Davies-Bouldin Index measures the average similarity between
each cluster and its most similar cluster, while the Silhouette Score measures the average
distance between each point in a cluster and points in other clusters. The AIC and BIC
evaluate the trade-off between the goodness of fit and the complexity of the model, with
Bachelor of Computer Science (Honours)
Faculty of Information and Communication Technology (Kampar Campus), UTAR
25
lower values indicating a better model fit. Finally, the Calinski-Harabasz Score measures the
ratio of between-cluster variance to within-cluster variance, with higher values indicating
better clustering. The optimal number of clusters was chosen based on the highest scores for
each of these performance metrics.

The optimal number of clusters was determined by evaluating several performance metrics
including the Davies-Bouldin index, silhouette score, AIC, BIC, and Calinski-Harabaz score
for a range of values of k from 2 to 10. The results for each metric were plotted and analyzed.
Based on the analysis, it was found that the optimal number of clusters, k was 2. This was
determined by the metric with the highest score for k=2. Other metrics were also considered,
and they were consistent with this result. The selection of the optimal number of clusters was
based on a careful evaluation of multiple performance metrics to arrive at a robust and
reliable solution.

3.4 Clustering Algorithms

K-means clustering is a popular unsupervised learning algorithm used to group data points
into a k. It works by iteratively assigning each data point to the nearest cluster centroid and
then updating the centroids based on the newly assigned data points. The algorithm stops
when the centroids no longer move, or a maximum number of iterations is reached.

GMM is a probabilistic model that assumes that the data is generated from a mixture of
Gaussian distributions. It is a soft clustering algorithm, which means that instead of assigning
each data point to a single cluster, it assigns probabilities to each data point belonging to each
of the clusters. The algorithm uses an EM algorithm to iteratively estimate the parameters of
the Gaussian distributions and the probabilities of the data points belonging to each cluster.

DBSCAN is a density-based clustering algorithm that groups together data points that are
closely packed together (high density) and separates out data points that are far apart (low
density). It works by defining a radius around each data point and grouping together data
points that fall within that radius. The algorithm can handle noise and outliers by classifying
them as not belonging to any cluster.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
26
RFM is a customer segmentation technique used in marketing to identify and group
customers based on their purchasing behaviour. It involves analysing three factors: recency
(how recently a customer made a purchase), frequency (how often a customer makes
purchases), and monetary (how much money a customer spends). The data is then clustered
into groups based on these factors to identify different customer segments with different
characteristics and behaviours. While not a traditional clustering algorithm, RFM is often
used in marketing to group customers based on their behaviour.

3.4.1 K-means Clustering

The partitioning of a dataset into a predetermined number of clusters is achieved by using the
popular unsupervised learning algorithm known as K-means clustering. The aim of the
algorithm is to minimize the sum of squared distances between data points and their assigned
cluster centers, also known as centroids. The basic steps of the k-means algorithm include the
choice of the number of clusters, random initialization of centroids, assigning data points to
the nearest centroid, recalculating centroids as the mean of all data points assigned to each
cluster, and repeating until convergence.

To implement the algorithm, the data was first standardized using z-score normalization to
ensure that all variables had the same scale. The k-means clustering was performed using the
scikit-learn library with different values of k and random initializations. To determine the
optimal number of clusters, several metrics such as the Davies-Bouldin index, silhouette
score, AIC, BIC, and Calinski-Harabaz score were evaluated. These metrics allowed us to
assess the quality of the resulting clusters and determine the optimal number of clusters.

Based on the analysis, k was determined to be 2 as it had the lowest BIC and AIC scores. The
quality of the clusters was also assessed using the silhouette score and the Davies-Bouldin
index. The results indicated that the identified clusters were well-separated and distinct.
To mitigate the sensitivity of the k-means algorithm to initial centroid placement, the
algorithm was run multiple times with different initializations, and the solution with the
lowest sum of squared distances was chosen. In addition, other clustering algorithms such as
GMM clustering and DBSCAN were explored to compare their performance with k-means
clustering.
Bachelor of Computer Science (Honours)
Faculty of Information and Communication Technology (Kampar Campus), UTAR
27
K-means clustering is a useful tool for exploratory data analysis and pattern recognition, and
it can be applied to a wide range of applications such as market segmentation, image
processing, and anomaly detection. However, it is important to carefully choose the number
of clusters and evaluate the quality of the resulting clusters to avoid suboptimal results. The
pipeline was as below:
1. Standardize the data using z-score normalization
2. Choose the number of clusters
3. Initialize centroids randomly
4. Assign data points to the nearest centroid
5. Recalculate centroids as the mean of all data points assigned to each cluster
6. Repeat until convergence
7. Evaluate the quality of the resulting clusters using metrics such as Davies-Bouldin
index, silhouette score, AIC, BIC, and Calinski-Harabaz score
8. Choose the solution with the lowest sum of squared distances
9. Compare the performance of k-means clustering with other clustering algorithms such
as GMM and DBSCAN

The following figure was the output of k-means in this study:

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
28
 Figure 3.6 k-means output

3.4.2 GMM Clustering

GMMis a probabilistic model that represents the distribution of data as a mixture of multiple
Gaussian distributions. In GMM, each cluster is modeled by a Gaussian distribution with a
mean and covariance matrix. The probability density function of GMM is defined as the sum
of the individual Gaussian distributions weighted by their corresponding probabilities.

GMM clustering involves finding the optimal parameters for the Gaussian distributions that
best fit the data. This is typically done using the Expectation-Maximization (EM) algorithm,
which is an iterative algorithm that alternates between computing the expected probability of
each data point belonging to each cluster (the "E" step) and updating the parameters of the
Gaussian distributions based on these probabilities (the "M" step). The EM algorithm
continues until convergence, which is usually defined as a small change in the likelihood
function or the parameters.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
29
One advantage of GMM over k-means is that it allows for clusters of different shapes and
sizes, whereas k-means assumes that the clusters are spherical and of equal size. Additionally,
GMM can handle overlapping clusters and can assign each data point a probability of
belonging to each cluster, rather than a hard assignment as in k-means.

To perform GMM clustering, the number of clusters (k) and the initialization method for the
parameters was first to find and determine. Then, fit the GMM model to the data using the
EM algorithm, and obtain the cluster assignments and the parameters for each Gaussian
distribution. Then use these cluster assignments to label new data points and analyze the
characteristics of each cluster.

Like k-means clustering, various performance metrics can be used to evaluate the quality of
the GMM clustering results. These include the Davies-Bouldin index, silhouette score, AIC,
BIC, and Calinski-Harabaz score. Compare the results of GMM clustering to those of k-
means clustering and other clustering algorithms to determine the most appropriate method
for the dataset.

Overall, GMM clustering is a powerful tool for identifying clusters in data with complex
distributions and can provide additional insights beyond what is possible with simpler
clustering algorithms such as k-means.

In this project, the data was first standardized using z-score normalization to ensure that all
variables had the same scale. Then, the GMM model was fitted to the normalized data using
the expectation-maximization algorithm with a full covariance matrix and random state of 42.

To determine the optimal number of clusters, the Bayesian Information Criterion (BIC) and
the Akaike Information Criterion (AIC) were evaluated for a range of values of k, the number
of clusters. The BIC and AIC measures indicate the trade-off between the goodness of fit and
the complexity of the model, with lower values indicating a better fit with fewer parameters.

Based on the analysis, k was determined to be 2 as it had the lowest BIC and AIC scores. The
quality of the clusters was also assessed using the silhouette score and the Davies-Bouldin
index. The results indicated that the identified clusters were well-separated and distinct.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
30
Finally, the clusters were visualized using a scatter plot of the standardized data points
colored by their assigned cluster label. This allowed insights to be gained into the
characteristics of each cluster and their relationship to each other. The pipeline was as below:
1. Standardize the data using z-score normalization
2. Fit the GMM model to the standardized data using the expectation-maximization
algorithm with a full covariance matrix and random state of 42
3. Evaluate the BIC and AIC measures for a range of values of k, the number of clusters
4. Determine the optimal number of clusters based on the lowest BIC and AIC scores
5. Assess the quality of the clusters using the silhouette score and the Davies-Bouldin
index
6. Visualize the clusters using a scatter plot of the standardized data points colored by
their assigned cluster label

The following figure are the output for GMM in this study:

Figure 3.7 GMM output

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
31
3.4.3 Dbscan Clustering
The DBSCAN (Density-Based Spatial Clustering of Applications with Noise) method was
utilized to identify clusters in the dataset. DBSCAN is a density-based clustering algorithm
that groups together data points that are close to each other in terms of a specified distance
metric. The hyperparameters of the DBSCAN algorithm, namely the minimum number of
samples (min_samples) and the maximum distance between two points (eps), were tuned to
obtain the optimal values.

In this study, DBSCAN algorithm modified by using the value of the min_samples feature
column multiplied by 2 as the value for the min_samples hyperparameter. This was done to
increase the minimum number of data points required to form a cluster and ensure that the
clusters were well-defined.

To determine the optimal value for the eps hyperparameter, GridSearchCV, a cross-validation
method was utilized that systematically searches for the best combination of
hyperparameters. A range of values for eps was tested and evaluated the results using the
silhouette score and the Davies-Bouldin index. The silhouette score measures the similarity
of a data point to its own cluster compared to other clusters, while the Davies-Bouldin index
measures the average similarity between each cluster and its most similar cluster. Lower
values of the Davies-Bouldin index indicate better clustering performance.

After hyperparameter tuning, the optimal value for eps found was 0.1. The value for the
modified min_samples hyperparameter also optimized to ensure that each cluster had a
minimum number of data points. The optimal value for the modified min_samples
hyperparameter was 18.

Finally, the clusters visualized using a scatter plot of the data points colored by their assigned
cluster label. These outlines gain insights into the characteristics of each cluster and their
relationship to each other.

However, multiple performance metrics like Calinski-Harabaz score, Davies-Bouldin index

and Silhouette score also implemented to find the best optimal eps in range of 0.1 to 0.5 and
best min_samples in range of 2 to 6 to improve the accuracy of the clustering performance. In

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
32
the result, the best eps was 0.3 and 4 as its Calinski-Harabaz and SIhouette was biggest and
the Davies-Bouldin was smallest. The pipeline was as below:
1. Set the hyperparameters min_samples and eps to default values
2. Use the modified min_samples method to increase the minimum number of data
points required to form a cluster
3. Search for the optimal value of eps in a range of 0.1 to 0.5 and min_samples in a
range of 2 to 6
4. Evaluate the results using the Calinski-Harabaz score, Davies-Bouldin index, and
Silhouette score
5. Set the optimal values of min_samples and eps based on the highest Calinski-Harabaz
and Silhouette scores and lowest Davies-Bouldin index
6. Use the DBSCAN algorithm to cluster the dataset with the optimal hyperparameters
7. Visualize the clusters using a scatter plot of the data points colored by their assigned
cluster label

The following figure was the output for DBSCAN in this study:

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
33
 Figure 3.8 DBSCAN output

3.4.4 RFM+k-means Clustering

The relevant data for RFM analysis was extracted from the database, which included the
transaction date, customer ID, and purchase amount. The Recency, Frequency, and Monetary
values for each customer were then calculated based on the most recent transaction date, the
total number of transactions, and the total amount spent, respectively.

The k-means algorithm, a popular unsupervised learning technique, was used to cluster the
customers based on their RFM scores. The RFM values were first standardized using z-score
normalization to ensure that all variables had the same scale. The optimal number of clusters
to use for k-means clustering was determined using the elbow method. Different values of k
were tested, and the resulting within-cluster sum of squares (WCSS) was evaluated to
identify the k value that minimized WCSS while still preserving meaningful clusters.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
34
After determining the optimal number of clusters, k-means clustering was applied to the
standardized RFM data. The RFM mapping and distribution of each cluster were visualized
to gain insights into the behavior of the different customer segments. The average RFM
values for each cluster were also analyzed to identify characteristics of each group and tailor
marketing strategies accordingly.

Other clustering algorithms such as GMM and DBSCAN were explored to compare their
performance with k-means clustering. However, k-means clustering was found to provide the
best balance of performance and interpretability for the RFM analysis.

In conclusion, RFM clustering with k-means is a powerful technique for segmenting

customers based on their purchase behavior, which can provide valuable insights into
customer behavior and preferences. By understanding the characteristics of different
customer segments, marketing strategies can be tailored to maximize customer engagement
and profitability.

In this study, the recency feature already consists of and the Frequency and Monetary that
suitable in the dataset was the total purchases and total spending. Thus, this RFM algorithms
was implemented separately from other algorithms above as it needed to use different data
pre-processing like RFM mapping and distribution.

Figure 3.9 RFM Mapping

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
35
Based on Figure 3.9, the x-axis represents the Recency score, the y-axis represents the
Frequency score, and the z-axis represents the Monetary score. The color of each data point
represents the Monetary score as well, with higher scores indicated by warmer colors and
lower scores indicated by cooler colors.

Its pipeline was as follows:

1. Calculate the Recency, Frequency, and Monetary values for each customer based on
the most recent transaction date (Recency), the total number of transactions (Total
Purchases), and the total amount spent (Spending), respectively
2. Standardize the RFM values using z-score normalization
3. Determine the optimal number of clusters using the elbow method and the within-
cluster sum of squares (WCSS)
4. Apply k-means clustering to the standardized RFM data with the optimal number of
clusters
5. Visualize the RFM mapping and distribution of each cluster to gain insights into
customer behaviour
6. Analyse the average RFM values for each cluster to identify characteristics of each
group and tailor marketing strategies accordingly
7. Compare the performance of k-means clustering with other clustering algorithms such
as GMM and DBSCAN
8. Use the RFM clusters to develop targeted marketing strategies and improve customer
engagement and profitability.

The following figure was the output for RFM in this study:

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
36
 Figure 3.10 RFM output

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
37
4 Experiment/Simulation
4.1 Hardware
Description Specifications
Model Vivobook_ASUSLaptop X530FN_S530FN
Processor Intel(R) Core(TM) i7-8565U CPU @ 1.80GHz 1.99 GHz
Operating System Windows 10 Home Single Language 64-bit
Graphic Intel® UHD Graphics 620
Memory 4096MB RAM
Storage 931.51GB
Table 4.1 Spec Used

4.2 Software

4.2.1 Jupyter Notebook (Python3)

Python has the ability to connect to database systems. It can also read and alter files, as well
as manage large amounts of data and do sophisticated calculations.

Jupyter Notebook was chosen to used in this project since it was a popular integrated
development environment (IDE) for Python that offers several benefits, some of which are:
 Interactive Computing: Jupyter Notebook allows you to write and execute code
snippets in an interactive way. This means that you can run code cell by cell, making
it easy to test and debug code.

 Data Visualization: Jupyter Notebook comes with built-in visualization tools that
enable you to create and display charts, graphs, and other visualizations of your data
directly in the notebook.
 Markdown Support: Jupyter Notebook supports Markdown, which allows you to add
formatted text, links, and images to your notebook. This makes it easy to add
documentation, comments, and explanations to your code.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
38
4.2.2 Anaconda Prompt
Anaconda is a popular data science platform that provides a collection of powerful tools for
working with Python. One of the key features of Anaconda is the Anaconda Prompt, which is
a command-line interface for running Python code and managing Anaconda environments.

The Anaconda Prompt provides an easy way to launch the graphical user interface (GUI) of
Python applications. To run a Python GUI application using the Anaconda Prompt, it can
simply navigate to the directory containing the application's code and run the command to
launch the GUI. For example, the python script named “[Link]” that has a
GUI can be navigated via this prompt with the command “python [Link]”. to
run the GUI

Overall, the Anaconda Prompt is a convenient tool for running Python code and managing
Anaconda environments, and it can help streamline the process of launching Python GUI
applications.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
39
4.3 Kmeans extension
An experiment was conducted to compare the performance of k-means clustering algorithms
in a GUI environment for user interaction. The study focused on the traditional k-means
algorithm, as well as two extensions: incremental k-means and mini-batch k-means.

The traditional k-means algorithm was implemented using the KMeans class from scikit-
learn, with 4 clusters and the default initialization parameters. The incremental k-means
algorithm was implemented using the IncrementalKMeans class, with a batch size of 100 and
a maximum of 10 iterations. The mini-batch k-means algorithm was implemented using the
MiniBatchKMeans class, with a batch size of 100 and 1000 maximum iterations.

Incremental k-means updates the clustering model with new data points one at a time, and
adapts the centroids of the existing clusters accordingly. This allows the model to learn from
new data without having to reprocess the entire dataset, which can be time-consuming for
large datasets. However, incremental k-means may be less accurate than traditional k-means
due to the smaller batch size.

Mini-batch k-means, on the other hand, uses random subsets (or "mini-batches") of the data
to update the clustering model, instead of processing the entire dataset at once. This can be
much faster than incremental k-means, as well as traditional k-means, because the algorithm
only needs to process a small batch of data at a time. Mini-batch k-means can also be more
accurate than incremental k-means, because it uses a larger batch size to update the centroids.

The benefit of mini-batch k-means over incremental k-means is primarily in terms of

computational efficiency. By using larger batches of data, mini-batch k-means can process
more data in a shorter amount of time, making it well-suited for large datasets or real-time
data streams. Additionally, mini-batch k-means can be more accurate than incremental k-
means due to the larger batch size. However, mini-batch k-means may require more
hyperparameter tuning than incremental k-means to obtain optimal performance. Overall,
MiniBatchKmeans was selected in the GUI development. The full source code for the GUI
will included in the appendix chapter.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
40
4.4 GUI
In order to facilitate user interaction with the RFM clustering model, a graphical user
interface (GUI) was developed using the PyQt5 library in Python. The GUI was designed to
allow users to input the recency, frequency, and monetary values for a customer, and predict
which cluster the customer belongs to using mini-batch k-means clustering.

The GUI consists of three input fields for the recency, frequency, and monetary values,
respectively, as well as a "Predict" button and a "Plot" button. When the user inputs the
values and clicks the "Predict" button, the GUI uses the MiniBatchKMeans class from the
scikit-learn library to update the clustering model with the new data point and predict its
cluster label. The resulting label is displayed to the user in a text label on result section.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
41
Additionally, when the user clicks the "Plot" button, the GUI will prompt a plot window of
3D plot of the clustering results, with each cluster represented by a different color. The plot
also displays the new data point input by the user, marked by a red "x". The plot also can
zoom in/out and can navigate with the mouse to see more clearer the red ‘x’.

Furthermore, the function of CRUD also be implemented like user can add new customers
data, delete new/old customers data and predict the cluster results. The read function was
design as a table shown in Figure , users can straight away select the customers in the table to
predicts clusters label or delete the customer data.

Overall, the GUI provides a user-friendly interface for interacting with the RFM clustering
model, allowing users to input new data points and explore the clustering results in an
intuitive way.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
42
5 Model Evaluation and Discussion
As this analysis was exploratory in nature, traditional evaluation metrics such as accuracy
were not applicable, as there was no ground truth for comparison. Therefore, evaluation
metrics such as elbow method, calinski-harabaz and silhouette score were used to assess the
performance of the segmentation algorithms.

In this study, different approaches were separated from determining the number of clusters
and more ease to do the project. Three different files will included in this study which were
“Customer segmentation using k-means, GMM, and DBSCAN clustering”, “Customer
segmentation using RFM clustering”, and “Visualization after data-preprocessing”. There
were various clustering algorithms used in this project as mentioned in chapter 3. The reason
why RFM algorithms used in different files are it was a behavior segmentation based
compare with others algorithms.

5.1 Determine optimal number of clusters

5.1.1 k-means, GMM, DBSCAN

Figure 5.1 Silhouette score and Calinski-Harabaz

Figure 5.1 shows the result is elbow on k=2, the calinski-harabaz was 5076.938 and the
silhouette score was 0.600.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
43
 Figure 5.2 Comparison of all metric prediction
Figure 5.2 shows the comparison results of clusters for the dataset as it shows k = 2 is the
best optimal clusters.

5.1.2 RFM+kmeans
There were two metrics evaluation used which were elbow and silhouette score:

Figure 5.3 Elbow Method for RFM

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
44
 Figure 5.4 Silhouette Score for RFM

The optimal cluster based on the elbow method is 4, while according to silhouette score is 2.
Since silhouette score method considers intra and inter clusters, it might produce more
separated clusters. However, two clusters might be too general for customer segmentation
which require more specific cluster to build personalized offers. Thus, k=4 will be used to
perform clustering [1].

5.2 Clustering Evaluation

5.2.1 k-means and GMM

Figure 5.5 Clustering Metrics Summary for k-means and GMM

To identify the optimal number of clusters in our customer segmentation analysis, several
evaluation metrics were used to assess the performance of different clustering algorithms.
Specifically, looked for the number of clusters that best satisfied the following criteria: a
Bachelor of Computer Science (Honours)
Faculty of Information and Communication Technology (Kampar Campus), UTAR
45
lower value of AIC or BIC, a higher value of the Calinski-Harabasz score, a Silhouette score
approaching 1, and a lower Davies-Bouldin Index score. A lower value of AIC or BIC
indicates a better fit for the model, while a higher value of the Calinski-Harabasz score
indicates better clustering performance. The Silhouette score, which ranges from -1 to 1, can
be used to evaluate the quality of clustering, with a score of 1 denoting the best clustering and
a score approaching 1 being desirable. The Davies-Bouldin Index, which ranges from 0 to 1,
can be used to evaluate the compactness and separation of clusters, with a lower score
indicating better clustering performance. By using these evaluation metrics, the optimal
number of clusters were able to identify for the customer segmentation analysis, gain insights
into the different customer segments and their characteristics.

5.2.2 DBSCAN

Figure 5.6 GridSearch for eps

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
46
 Figure 5.7 Clustering Metrics Summary for DBSCAN

The choice of clustering metrics differs between DBSCAN and k-means/GMM due to the
unique hyperparameters of DBSCAN, namely ε and min_samples, which are distinct from
the traditional number of clusters or components. In Figure 5.6, the clustering results were
analyzed using GridSearchCV() based on the Silhouette Score, resulting in an optimal ε value
of 0.1. Furthermore, Figure 5.7 highlights the best combination of ε and min_samples as 0.3
and 4, respectively. Consequently, these values were selected to achieve more accurate
clustering results with DBSCAN compared to the previous settings.

5.2.3 3D Plot Performance

In 3D plots, clusters can be represented by distinct shapes or colors, making it easier to

visually identify and explore the separation between clusters. This can help in understanding
the distinct characteristics of different clusters and how they are distributed in the data space.
And the following figures was show all the three algorithms concerns the clusters of two with
different colors for more appealing.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
47
 Figure 5.8 Comparison between kmeans, GMM and DBSCAN
Bachelor of Computer Science (Honours)
Faculty of Information and Communication Technology (Kampar Campus), UTAR
48
 Figure 5.9 3D plot for RFM

5.3 Cluster Analysis (k-means. GMM and DBSCAN)

Figure 5.10 Tabulate total number of clusters

The result shows that there were total number of 999 customers in cluster 1 and total number
of 1213 customers in cluster 2 and hence the number of clusters 2 is more than cluster 1.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
49
 Figure 5.11 Age vs spending

Figure 5.12 Education vs Spending

Bachelor of Computer Science (Honours)
Faculty of Information and Communication Technology (Kampar Campus), UTAR
50
 Figure 5.14 Marital Status vs Spending

Figure 5.13 NumwebvisitsMonth vs Spending

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
51
 Figure 5.15 Familysize vs Spending

The summary insight in figure 5.11 5o figure 5.15 were:

Cluster 1: Highly Active and Moderately Active Customers
Customers in this cluster have a professional or postgraduate (PG) background. They
generally have higher incomes compared to the Least Active Customers. Their ages range
from 25 to 75, with the majority falling between 35 to 60 years old. The number of children
varies, with Highly Active Customers having, on average, more children than Moderately
Active Customers. In terms of spending, Moderately Active Customers tend to spend more
than Highly Active Customers. Both segments have been engaged with the company for a
longer period compared to the Least Active Customers. Additionally, they exhibit a tendency
to accept at least one campaign.

Cluster 2: Least Active Customers

Customers in this cluster typically have an undergraduate (UG) educational background.
Their income is very low or almost negligible. They tend to be younger, with ages ranging
from 15 to 30. They have very few or no children. Their spending on the company's products
or services is minimal. Furthermore, their level of engagement with the company over an
Bachelor of Computer Science (Honours)
Faculty of Information and Communication Technology (Kampar Campus), UTAR
52
extended period is limited. These customers demonstrate low activity in accepting any
campaign.

5.3.1 Cluster Analysis (RFM)

Figure 5.16 Tabulate number of customers in each cluster

The figure 5.16 concern that number of customers in cluster 1 is the highest while the number
of customers in cluster 2 is the lowest.

Figure 5.17 Summary boxplot

Based on figure 5.17, The clusters separate the values of RFM into two groups, low and high,
as observed from the boxplot and centroid analysis.

The Recency values are categorized as follows: Low (22-23 days) and High (73 days). The
Frequency values are divided into Low (7 purchases) and High (19 purchases). Monetary

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
53
values show a separation between Low (146-151 USD) and High (1181-1184 USD) ranges.
These RFM differentiations form a combination of four customer segments.

 Cluster 0 represents the group of Low-Spending Active Customers with low recency,
low frequency, and low monetary values. This segment comprises 619 customers,
which accounts for 28% of the total customer base.

 Cluster 1 consists of Churned Low-Spending Customers with high recency, low

frequency, and low monetary values. These customers display low engagement and
minimal spending. Unfortunately, this segment comprises a high number of
customers, specifically 631, representing 28% of the total customer base.

 Cluster 2 identifies the Best Active Customers with low recency, high frequency, and
high monetary values. These customers are highly valuable to the company, as they
make frequent high-value purchases. However, their number is relatively small,
comprising only 466 customers or 21% of the total customer base.

 Cluster 3 represents the Churned Best Customers, characterized by high recency, high
frequency, and high monetary values. Although these customers have contributed
significantly to the company's revenue, their last purchase occurred a long time ago,
potentially indicating churn. This cluster consists of 524 customers, equivalent to 23%
of the total customer base.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
54
 Figure 5.18 RFM distribution

Based on Figure 5.18, it is evident that Cluster 2 stands out as the best customer segment
when compared to the other clusters. This is primarily due to their lower recency, higher
frequency, and higher monetary values. Specifically, customers in this cluster tend to visit the
company for shorter periods but exhibit higher visit frequency and spending patterns. These
customers may visit the company multiple times within a week and spend significantly more
on their purchases.

On the other hand, customers in Cluster 3 demonstrate higher values in terms of recency,
frequency, and monetary factors, indicating that they are highly valuable customers. They
make substantial purchases and spend more, but their visit intervals are longer, suggesting a
longer gap between their visits to the company.

In contrast, customers in Cluster 0 exhibit considerably lower values across all three
dimensions—recency, frequency, and monetary. This cluster represents the worst-performing
group within the dataset, as these customers display lower levels of engagement, visit the
company less frequently, and make smaller purchases compared to the other clusters.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
55
 Figure 5.19 RFM Age vs Spending

Figure 5.20 RFM Income vs Spending

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
56
 Figure 5.21 RFM Education vs Spending

Figure 5.22 RFM Education vs Spending

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
57
 Figure 5.23 RFM Numwebvisitmonth vs Spending

Figure 5.24 RFM Familysize vs Spending

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
58
 Figure 5.25 RFM Days_enrolled vs Spending

The output of Figure 5.19 to figure 5.35 was concluded at below:

1. Low-Spending Active Customers:
 These customers have an income range of 5000 to 40000 and typically spend
between 0 to 500 units. They fall within the age range of 25 to 50 and come
from various educational backgrounds. They can be both married and
unmarried individuals, with many of them being parents, some with one child.
These customers have been with the company for at least 300 days and tend to
show rare acceptance of promotions. Additionally, they rarely complete
purchases using discounts.
2. Churned Low-Spending Customers:
 Customers in this segment have a higher income range of 65000 to 85000 but
still exhibit low spending patterns, ranging from 550 to 2000 units. Their age
ranges from 30 to 60, and the majority have completed graduation. Most of
them are married individuals without children. They have been customers for
at least 250 days, with a moderate promotion acceptance ratio of 0.5.
Completing purchases using discounts is rare among these customers.
3. Best Active Customers:
 The Best Active Customers segment comprises customers with an income
range of 50000 to 80000, demonstrating higher spending levels ranging from

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
59
 250 to 1800 units. Their age ranges from 35 to 60, and most have completed
graduation. They are predominantly married individuals who are parents, with
the majority having one child. These customers have been with the company
for at least 400 days. Although they show a poor promotion acceptance ratio,
they are highly interested in completing purchases using discounts.
4. Churned Best Customers:
 Customers in this segment have an income range of 40000 to 60000 and
exhibit lower spending patterns, ranging from 0 to 500 units. Their age ranges
from 40 to 65, and most have completed graduation. They can be both married
and unmarried individuals, and like the previous segment, they are also
parents, with the majority having two children. These customers have been
with the company for at least 150 days but show rare acceptance of
promotions. However, they display a high level of interest in completing
purchases using discounts.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
60
6 Conclusion and Recommendations
6.1 Summary
This report explored the application of various clustering techniques, including K-means,
GMM, DBSCAN, and RFM analysis, for customer segmentation. The objective was to
identify distinct customer groups based on their purchasing behavior and preferences using
cluster analysis. Through the analysis and evaluation of these techniques, valuable insights
into the segmentation process were gained. The dataset used lacked accuracy and ground
truth, making it inherently challenging.

Initially, after preprocessing the data, K-means clustering was employed, which is a widely
used method for partitioning data into groups. It determined clusters based on the proximity
of customers in the feature space. However, K-means assumes spherical-shaped clusters with
equal variances, which may not always hold true in real-world scenarios.

Subsequently, GMM was investigated as a probabilistic clustering algorithm. GMM

considers the underlying distribution of data points and assigns probabilities to each point
belonging to a specific cluster. This approach allows for more complex cluster shapes and
accommodates overlapping clusters. GMM effectively captured hidden patterns within the
dataset.

Next, DBSCAN, a density-based clustering algorithm, was examined. DBSCAN identifies

dense regions of points separated by sparser areas. It does not assume any specific cluster
shape and handles outliers effectively. DBSCAN revealed clusters of varying densities and
exposed non-linear structures in the data.

Lastly, RFM analysis was employed, a valuable technique for customer segmentation based
on recency, frequency, and monetary value. RFM provided insights into customer
transactional behavior, allowing for the classification of customers into distinct segments
with unique characteristics and marketing implications.

Throughout the analysis, each clustering technique demonstrated strengths and limitations. K-
means offered a straightforward and interpretable approach but struggled with complex and

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
61
non-linear structures. GMM provided more flexibility but required careful consideration of
the number of components. DBSCAN excelled in identifying clusters with varying densities
but could be sensitive to parameter selection, specifically k, epsilon, and min_samples. RFM
analysis complemented the clustering techniques by incorporating transactional data and
offering insights into customer value. In RFM analysis, the selection of the optimal number
of clusters, k, was a critical consideration. A comparison of different performance metrics
was lacking in previous studies, and the elbow method was commonly used.

In addition to the clustering techniques, a user-friendly GUI was developed based on RFM
analysis and the Mini-Batch K-means algorithm. This GUI provides businesses with an
intuitive platform to perform customer segmentation on their own datasets. By incorporating
RFM analysis, the GUI enables users to analyze customer transactional behavior and identify
valuable customer segments. The integration of the Mini-Batch K-means algorithm ensures
efficient and scalable clustering, making it suitable for large datasets. The GUI's
visualizations and interactive features enhance the interpretation and exploration of the
segmentation results, empowering businesses to make data-driven decisions.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
62
6.2 Further Study
1. Dimensionality Reduction
 One technique for reducing dimensionality is t-SNE. This nonlinear method
focuses on preserving local structures and clusters within the data, making it
valuable for visualizing high-dimensional data. By employing t-SNE, it
becomes easier to identify hidden patterns or relationships that may not be
immediately apparent in the original feature space. Reducing dimensionality
with t-SNE provides insights into customer grouping and separability, thereby
assisting subsequent clustering processes.
2. Clustering Algorithms
 SOM or Kohonen maps, are unsupervised neural networks that utilize
competitive learning to organize and visualize high-dimensional data. SOM
constructs a grid of neurons, where each neuron represents a cluster or
prototype. Throughout training, the neurons adapt their weights to accurately
represent various regions within the input data space. SOM excels at
preserving the topological structure of the data, revealing relationships and
clusters that may not be initially evident. It is commonly employed for
visualizing and exploring high-dimensional data.
 Autoencoders are neural network models employed for reconstructing input
data from a compressed or bottleneck representation. In the context of
clustering, autoencoders serve as a dimensionality reduction technique, with
the bottleneck layer representing a lower-dimensional representation of the
input data. The encoder part of the autoencoder maps the input data to the
bottleneck layer, while the decoder part reconstructs the original data. By
training an autoencoder on customer data, the bottleneck layer captures
essential features for clustering. Subsequent clustering algorithms can then be
applied to the reduced-dimensional representation. Autoencoders excel at
capturing non-linear relationships and identifying complex clusters within the
data.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
63
 REFERENCES
Article:
Journal article
[1] Lee, Y., Kwon, O., & Lee, Y. (2015). Customer segmentation using purchase data for an
online retailer. Expert Systems with Applications, 42(1), 332-341. DOI:
10.1016/[Link].2014.08.020

[2] Kumar, A., Jain, A., Jain, S., & Jain, S. (2016). Comparative study of clustering
algorithms for customer segmentation in e-commerce. In 2016 International Conference on
Computing, Analytics and Security Trends (CAST) (pp. 300-305). IEEE. DOI:
10.1109/CAST.2016.79

[3] Xiang, Y., & Gong, Y. (2018). Online Shopping Behavior Analysis Based on K-means,
GMM and DBSCAN Clustering Algorithm. In 2018 International Conference on
Computational Science and Computational Intelligence (CSCI) (pp. 1122-1126). IEEE. DOI:
10.1109/CSCI46756.2018.00209

[4] Chen, X., Zuo, X., Wu, Z., & Liu, X. (2020). A Comparative Study of Customer
Segmentation Methods Based on Online Shopping Behavior. In 2020 IEEE International
Conference on Industrial Engineering and Engineering Management (IEEM) (pp. 791-795).
IEEE. DOI: 10.1109/IEEM47687.2020.9378743

[5] Jadhav, M., & Sonawane, K. (2021). Credit card fraud detection using clustering
techniques. In 2021 11th International Conference on Cloud Computing, Data Science &
Engineering (Confluence) (pp. 1-6). IEEE. DOI:
10.1109/CONFLUENCE51715.2021.9461624

[6] Yeh, Y.-L., & Huang, C.-C. (2018). Customer segmentation of bicycle-sharing users
based on their usage behavior. Sustainability, 10(5), 1579. DOI: 10.3390/su10051579

[7] Allenby, G., Fennell, G., Bemmaor, A., Bhargava, V., Christen, F., Dawley, J., Dickson,
P., Edwards, Y., Garratt, M., Ginter, J., Sawyer, A., Staelin R. & Yang, S. (2002) Market
Bachelor of Computer Science (Honours)
Faculty of Information and Communication Technology (Kampar Campus), UTAR
64
segmentation research: beyond within and across group differences. Marketing Letters, 13, 3,
pp. 233–243.

[8] Hyunseok Hwang, Taesoo Jung, Euiho Suh

An LTV model and customer segmentation based on customer value: A case study on the
wireless telecommunication industry
Expert Systems with Applications, 26 (2004), pp. 181-188

[9] Dolničar, S. & Leisch, F. (2004) Segmenting markets by bagged clustering. Australasian
Marketing Journal, 12, 2, pp. 51–65.

[10] Dimitriadou, E., Dolničar, S. & Weingessel, A. (2002) An examination of the number of
indexes for determining the number of clusters in binary data sets. Psychometrika, 67, 2, pp.
137–160.

[11] Decker, R., Wagner, R. & Scholz, S.W. (2005) Growing clustering algorithms in market
segmentation: defining target groups and related marketing communication. In H.-H. Bock,
W. Gaul & M. Vicki (eds) Data Analysis, Classification and the Forward Search. Berlin:
Springer, pp. 23–30.

Online Article:
[12] Qualtrics, "Customer Segmentation: Definition & Methods", Qualtrics AU, 2020.
[Online].
Available:
[Link]
segmentation/?rid=ip&prevsite=en&newsite=au&geo=MY&geomatch=au.

[13] D. Gong, "Clustering Algorithm for Customer Segmentation", Medium, 2021. [Online].
Available:
[Link]
e2d79e28cbc3.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
65
Book:
[14] Aldenderfer, M.S. & Blashfield, R.K. (1984) Cluster Analysis. Beverly Hills: Sage
Publications.

[15] A. Hiziroglu, “Soft computing applications in customer segmentation: State-of-artreview

and critique,” Expert Syst Appl.

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
66
 Appendix

The following code was used to impute missing value:

#Mice Imputer
df_mice = [Link](deep=True)
mice_imputer = IterativeImputer()
df['Income'] = mice_imputer.fit_transform(df[['Income']])

#Check the present of missing values

present = 0
for col in [Link]:
if df[col].isnull().sum() > 0:
print(f"Column '{col}' has {df[col].isnull().sum()} missing values")
present = 1
if not present:
print("No missing values are present in the dataset")

Here is an example of the code used for some of the cleaning steps:
# change date format
df['Dt_Customer'] = pd.to_datetime(df['Dt_Customer'], format='%d-%m-%Y')

#Gather all the AcceptedCmps and response in one features

df['Total_AcceptedCampaign'] =
df[['AcceptedCmp1','AcceptedCmp2','AcceptedCmp3','AcceptedCmp4','AcceptedCmp5','Res
ponse']].sum(axis=1)

# convert categorical features to factor

categorical_features = ["Education", "Marital_Status", "AcceptedCmp1",
"AcceptedCmp2", "AcceptedCmp3", "AcceptedCmp4",
"AcceptedCmp5", "Complain", "Response"]
for feature in categorical_features:
df[feature] = df[feature].astype('string')

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
67
# Calculate age
df['Age'] = [Link]().year - df['Year_Birth']

# Calculate last enrollment date

last_enrollment = df['Dt_Customer'].max()
# Calculate days enrolled
df['Days_Enrolled'] = (last_enrollment - df['Dt_Customer']).[Link]

# total spendings
df['Spend'] = df[['MntWines', 'MntFruits', 'MntMeatProducts', 'MntFishProducts',
'MntSweetProducts', 'MntGoldProds']].sum(axis=1)

df['Familysize'] = df['Marital_Status'].replace({"Married":"Relationship",
"Together":"Relationship",
"Widow":"Single", "Divorced":"Single",
"Single":"Single","Alone":"Single", "Absurd":"Single",
"YOLO":"Single"}).replace({'Single': 1, 'Relationship' : 2}).fillna(0).astype(int) +
df['Kidhome'] + df['Teenhome']

#Total purchases
df['Total_Purchases'] =
df[['NumDealsPurchases','NumWebPurchases','NumCatalogPurchases','NumStorePurchases']
].sum(axis=1)

# remove features not used

df = [Link]([ 'Year_Birth', 'MntWines', 'MntFruits', 'MntMeatProducts', 'MntFishProducts',
'MntSweetProducts', 'MntGoldProds','Kidhome', 'Teenhome'], axis=1)

[Link](columns=['Dt_Customer','AcceptedCmp1', 'AcceptedCmp2', 'AcceptedCmp3',

'AcceptedCmp4', 'AcceptedCmp5', 'Complain',
'Response','NumDealsPurchases','NumWebPurchases','NumCatalogPurchases','NumStorePur
chases'], inplace=True)
# Identify numerical columns and apply standard scaling
Bachelor of Computer Science (Honours)
Faculty of Information and Communication Technology (Kampar Campus), UTAR
68
int_list = []
for col in [Link]:
if df[col].dtypes == int or df[col].dtypes == float:
int_list.append(col)

autoscaler = StandardScaler()
df[int_list] = autoscaler.fit_transform(df[int_list])

# Identify categorical columns and apply label encoding

obj_col = []
for col in [Link]:
if df[col].dtypes == object:
obj_col.append(col)

label_encoder = LabelEncoder()
for col in obj_col:
df[col] = label_encoder.fit_transform(df[col])

Performance Metric:
# k-means clustering
kmeans = KMeans(n_clusters=2, random_state=42)
kmeans_preds = kmeans.fit_predict(df_norm)
plot_km = df
plot_km['KMeans Cluster'] = kmeans_preds
plot_km=plot_km.sort_values(by='Cluster')
plot_km['KMeans Cluster'] = plot_km['Cluster'].astype(str).apply(lambda x: 'Outliers' if x ==
'-1' else x)

# # Reverse normalization
# plot_km['Income'] = plot_km['Income'] * 100
# plot_km['Spend'] = plot_km['Spend'] * 1000

# Plot of clusters
temp = [Link]()
[Link].plot_bgcolor = '#F9F9F9'
[Link].paper_bgcolor = '#F9F9F9'
[Link] = '#4D4D4D'
[Link] = '#4D4D4D'
fig = [Link](plot_km, x="Income", y="Spend", color="KMeans Cluster",
color_discrete_sequence=[Link].T10[2:])
fig.update_traces(marker=dict(size=11, opacity=0.85, line=dict(width=1, color='#F7F7F7')))
Bachelor of Computer Science (Honours)
Faculty of Information and Communication Technology (Kampar Campus), UTAR
69
fig.update_layout(template=temp, title="KMeans Cluster Profiles,<br>Customer Spending
vs. Income",
width=700, legend_title='Cluster',
paper_bgcolor='rgb(229, 236, 246)',
title_font_size=22,
xaxis=dict(title='Spending', showline=True, zeroline=False, range=[0, None]),
yaxis=dict(title='Income, $', ticksuffix='k', showline=True, range=[0, None]))

[Link]()

K-means Imputation:
# k-means clustering
kmeans = KMeans(n_clusters=2, random_state=42)
kmeans_preds = kmeans.fit_predict(df_norm)
plot_km = df
plot_km['KMeans Cluster'] = kmeans_preds
plot_km=plot_km.sort_values(by='Cluster')
plot_km['KMeans Cluster'] = plot_km['Cluster'].astype(str).apply(lambda x: 'Outliers' if x ==
'-1' else x)

# # Reverse normalization
# plot_km['Income'] = plot_km['Income'] * 100
# plot_km['Spend'] = plot_km['Spend'] * 1000

# Plot of clusters
temp = [Link]()
[Link].plot_bgcolor = '#F9F9F9'
[Link].paper_bgcolor = '#F9F9F9'
[Link] = '#4D4D4D'
[Link] = '#4D4D4D'
fig = [Link](plot_km, x="Income", y="Spend", color="KMeans Cluster",
color_discrete_sequence=[Link].T10[2:])
fig.update_traces(marker=dict(size=11, opacity=0.85, line=dict(width=1, color='#F7F7F7')))
fig.update_layout(template=temp, title="KMeans Cluster Profiles,<br>Customer Spending
vs. Income",

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
70
 width=700, legend_title='Cluster',
paper_bgcolor='rgb(229, 236, 246)',
title_font_size=22,
xaxis=dict(title='Spending', showline=True, zeroline=False, range=[0, None]),
yaxis=dict(title='Income, $', ticksuffix='k', showline=True, range=[0, None]))

[Link]()

GMM Imputation:
# Gaussian Mixture Model clustering
gmm = GaussianMixture(n_components=2, covariance_type='full', random_state=42)
[Link](df_norm)
gmm_preds = [Link](df_norm)
plot_gmm = df
plot_gmm['GMM Cluster'] = gmm_preds
plot_gmm = plot_gmm.sort_values(by='Cluster')
plot_gmm['GMM Cluster'] = plot_gmm['GMM Cluster'].astype(str).apply(lambda x:
'Outliers' if x == '-1' else x)

# Plot of clusters
temp = [Link]()
[Link].plot_bgcolor = '#F9F9F9'
[Link].paper_bgcolor = '#F9F9F9'
[Link] = '#4D4D4D'
[Link] = '#4D4D4D'
fig = [Link](plot_gmm, x="Income", y="Spend", color="Cluster",
color_discrete_sequence=[Link].T10[2:])
fig.update_traces(marker=dict(size=11, opacity=0.85, line=dict(width=1, color='#F7F7F7')))
fig.update_layout(template=temp, title="Gaussian Mixture Model Cluster
Profiles,<br>Customer Spending vs. Income",
width=700, legend_title='Cluster',
paper_bgcolor='rgb(229, 236, 246)',
title_font_size=22,
xaxis=dict(title='Spending', showline=True, zeroline=False),
yaxis=dict(title='Income, $', ticksuffix='k', showline=True))
[Link]()

DBSCAN Imputation:
# DB Scan clustering
db=DBSCAN(eps=0.3, min_samples=4, metric='euclidean')
db_preds=db.fit_predict(df_norm)
plot_db=df
plot_db['DB Cluster'] = db_preds
Bachelor of Computer Science (Honours)
Faculty of Information and Communication Technology (Kampar Campus), UTAR
71
plot_db=plot_db.sort_values(by='Cluster')
plot_db['DB Cluster'] = plot_db['Cluster'].astype(str).apply(lambda x: 'Outliers' if x == '-1'
else x)

# Plot of clusters
temp = [Link]()
[Link].plot_bgcolor = '#F9F9F9'
[Link].paper_bgcolor = '#F9F9F9'
[Link] = '#4D4D4D'
[Link] = '#4D4D4D'
fig = [Link](plot_db, x="Income", y="Spend", color="Cluster",
color_discrete_sequence=[Link].T10[2:])
fig.update_traces(marker=dict(size=11, opacity=0.85, line=dict(width=1, color='#F7F7F7')))
fig.update_layout(template=temp, title="DBSCAN Cluster Profiles,<br>Customer Spending
vs. Income",
width=700, legend_title = 'Cluster',
paper_bgcolor='rgb(229, 236, 246)',
title_font_size=22,
xaxis=dict(title='Spending',showline=True, zeroline=False),
yaxis=dict(title='Income, $',ticksuffix='k',showline=True))
[Link]()

Comparison algorithms in 3D plot:

# # Define color palette
colors = [Link].T10[2:]

# Initializing figure with 3 3D subplots

fig = make_subplots(rows=3, cols=1,
vertical_spacing=0.1,
specs=[[{'type': 'scatter3d'}],
[{'type': 'scatter3d'}],
[{'type': 'scatter3d'}]],
subplot_titles=("K-Means Clustering with 2 clusters",
"Gaussian Mixture Model Clustering<br>with 2 clusters",
"DBSCAN<br>with 2 clusters")
)

# Adding clusters to scatterplots

# Clean up the 'KMeans Cluster' column
plot_km['KMeans Cluster'] = plot_km['KMeans Cluster'].[Link]('Cluster ', '').astype(int)
plot_km=plot_km.sort_values(by='KMeans Cluster')
for i in range(0,3):
marker_symbol = 'diamond'
if i == 1:
marker_symbol = 'circle'
elif i == 0:
Bachelor of Computer Science (Honours)
Faculty of Information and Communication Technology (Kampar Campus), UTAR
72
 marker_symbol = 'cross'
fig.add_trace(go.Scatter3d(x = plot_km[plot_km['KMeans Cluster'] == i]['Spend'],
y = plot_km[plot_km['KMeans Cluster'] == i]['Age'],
z = plot_km[plot_km['KMeans Cluster'] == i]['Income'],
mode = 'markers+lines', marker=dict(
size=7,
color=colors[i+1],
line_width = 1,
line_color='#F7F7F7',
symbol=marker_symbol,
opacity=0.9),
name = str('Cluster '+str(i)), legendgroup = 1),
row=1, col=1)

plot_gmm['GMM Cluster'] = plot_gmm['GMM Cluster'].astype(int)

plot_gmm = plot_gmm.sort_values(by='GMM Cluster')
for i in range(0, 2):
marker_symbol = 'circle'
if i == 1:
marker_symbol = 'diamond'
fig.add_trace(go.Scatter3d(x=plot_gmm[plot_gmm['GMM Cluster'] == i]['Spend'],
y=plot_gmm[plot_gmm['GMM Cluster'] == i]['Age'],
z=plot_gmm[plot_gmm['GMM Cluster'] == i]['Income'],
mode='markers+lines', marker=dict(
size=7,
color=colors[i],
line_width=1,
line_color='#F7F7F7',
symbol=marker_symbol,
opacity=0.7),
name=str('GMM Cluster ' + str(i+1)), legendgroup=2),
row=2, col=1)

for i, j in enumerate(plot_db['DB Cluster'].unique()):

marker_symbol = 'circle'
if i == 1:
marker_symbol = 'diamond'
fig.add_trace(go.Scatter3d(x=plot_db[plot_db['DB Cluster'] == j]['Spend'],
y=plot_db[plot_db['DB Cluster'] == j]['Age'],
z=plot_db[plot_db['DB Cluster'] == j]['Income'],
mode='markers+lines', marker=dict(
size=7,
color=colors[i+1],
line_width=1,
line_color='#F7F7F7',
symbol=marker_symbol,
opacity=0.8),
name=str('DB Cluster ' + str(j)), legendgroup=3),
row=3, col=1)

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
73
fig.update_traces(hovertemplate='Customer Spending Score: %{x}<br>Income:
$%{z}<br>Age: %{y}')
fig.update_layout(title="Customer Segments based on Income, Spending, and Age",
paper_bgcolor='rgb(229, 236, 246)',
title_font_size=22,
template=temp, height=1800, legend_tracegroupgap = 500,
scene=dict(aspectmode='cube',
xaxis = dict(title='Spending Score',
backgroundcolor="#F3F3F3",
gridcolor="white",
showbackground=True,
zerolinecolor="white",),
yaxis = dict(title='Age, in years',
backgroundcolor="#E4E4E4",
gridcolor="white",
showbackground=True,
zerolinecolor="white"),
zaxis = dict(title='Income, $',
ticksuffix='k',
backgroundcolor="#F6F6F6",
gridcolor="white",
showbackground=True,
zerolinecolor="white")),
scene2=dict(aspectmode='cube',
xaxis = dict(title='Spending Score',
backgroundcolor="#F3F3F3",
gridcolor="white",
showbackground=True,
zerolinecolor="white",),
yaxis = dict(title='Age, in years',
backgroundcolor="#E4E4E4",
gridcolor="white",
showbackground=True,
zerolinecolor="white"),
zaxis = dict(title='Income, $',
ticksuffix='k',
backgroundcolor="#F6F6F6",
gridcolor="white",
showbackground=True,
zerolinecolor="white")),
scene3=dict(aspectmode='cube',
xaxis = dict(title='Spending Score',
backgroundcolor="#F3F3F3",
gridcolor="white",
showbackground=True,
zerolinecolor="white",),
yaxis = dict(title='Age, in years',
backgroundcolor="#E4E4E4",

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
74
 gridcolor="white",
showbackground=True,
zerolinecolor="white"),
zaxis = dict(title='Income, $',
ticksuffix='k',
backgroundcolor="#F6F6F6",
gridcolor="white",
showbackground=True,
zerolinecolor="white"))
)
[Link]()

RFM Imputation:
df['Frequency'] =
df['NumWebPurchases']+df['NumCatalogPurchases']+df['NumStorePurchases']

df['Monetary'] =
df['MntWines']+df['MntFruits']+df['MntMeatProducts']+df['MntFishProducts']+df['MntSweet
Products']+df['MntGoldProds']

#Data distribution
from [Link] import download_plotlyjs, init_notebook_mode, plot, iplot
init_notebook_mode(connected=True)

#plotting rfm distribution before scaling

fig = [Link]([Link](df_rfm), x='variable', y='value',
title='<b>RFM Data Distribution Before Scaling</b>',
color='variable', color_discrete_sequence=colors_rfm,
boxmode='overlay', points='all')

fig.update_layout(showlegend=False,paper_bgcolor='rgb(229, 236, 246)',

title_font_size=22)

[Link]()

# RFM Mapping
# plotting rfm
fig = px.scatter_3d(df_rfm_scaled, x='Recency', y='Frequency', z='Monetary',
title='<b>RFM Mapping</b>',
opacity=0.5, color='Monetary', color_continuous_scale='electric')

fig.update_traces(marker=dict(size=5))

fig.update_layout(paper_bgcolor='rgb(229, 236, 246)', title_font_size=22)

[Link]()

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
75
Determine k:
#plotting elbow method
fig = [Link](x=K, y=inertias,
title='<b>Optimal Number of Clusters by Elbow Method</b>',
color_discrete_sequence=['rgb(5, 60, 94)'])

fig.update_traces(line_width=4)

fig.update_layout(paper_bgcolor='rgb(229, 236, 246)',title_font_size=22,

xaxis_title='k (No. of Clusters)', yaxis_title='Inertia')
fig.add_vline(x=4, line_width=3, line_dash='dash', line_color='rgb(183, 9, 76)')

fig.add_annotation(x=4.1, y=1800, text='<i>optimal k=4</i>', font_size=16,

showarrow=True, ax=60, ay=-30, arrowhead=2, arrowsize=1,
arrowwidth=2)

fig.add_shape(type='circle', xref='x', yref='y',

x0=3.9, y0=1650, x1=4.1,y1=1914,
line_color='rgb(183, 9, 76)')

[Link]()

#plotting silhouette score

fig = [Link](x=K, y=silhouette,
title='<b>Optimal Number of Clusters by Silhouette Score</b>',
color_discrete_sequence=['rgb(5, 60, 94)'])

fig.update_traces(line_width=4)

fig.update_layout(paper_bgcolor='rgb(229, 236, 246)',title_font_size=22,

xaxis_title='k (No. of Clusters)',
yaxis_title='Silhouette Score')

fig.add_vline(x=2, line_width=3, line_dash='dash', line_color='rgb(183, 9, 76)')

fig.add_annotation(x=2.1, y=0.44, text='<i>optimal k=2</i>', font_size=16,

showarrow=True, ax=60, ay=-30, arrowhead=2, arrowsize=1,
arrowwidth=2)

fig.add_shape(type='circle', xref='x', yref='y',

x0=1.9, y0=0.432, x1=2.1,y1=0.442,
line_color='rgb(183, 9, 76)')

[Link]()

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
76
RFM plot:
df["Clusters"] = pd.to_numeric(df["Clusters"])

fig = px.scatter_3d(df, x='Recency', y='Monetary', z='Frequency',

color='Clusters', category_orders=dict(Clusters=[0,1,2,3]),
title='<b>Cluster Results</b>', opacity=0.5,
color_continuous_scale=colors_cluster,
labels={'Clusters': 'Cluster'})

fig.update_traces(marker=dict(size=6, opacity=0.6))

fig.update_layout(showlegend=True, paper_bgcolor='rgb(229, 236, 246)',

title_font_size=22, legend_title='Cluster')

[Link]()

#plotting income & monetary by clusters

#one customer with income of 666666 is excluded because it's obscuring the pattern
fig = [Link](df[df['Income']<500000], x='Monetary', y='Income',
color='Clusters',
color_continuous_scale='electric',
title='<b>Clusters by Income & Monetary</b>',
opacity=0.5)

fig.update_traces(marker=dict(size=11, opacity=0.85, line=dict(width=1, color='#F7F7F7')),

text=df['Clusters'])

fig.update_layout(paper_bgcolor='rgb(229, 236, 246)',title_font_size=22,

legend_title='Clusters', xaxis_title='Monetary', yaxis_title='Income')

[Link]()

Cluster Analysis (kmeans, GMM, DBSCAN):

#Visualize all the feature based on spending
df["Marital_Status"] = df["Marital_Status"].astype("category")
df["Education"] = df["Education"].astype("category")

Personal = ["Age", "Education", "Marital_Status", "NumWebVisitsMonth", "Familysize"]

for i in Personal:
fig = [Link](data_frame=df, x=i, y="Spend", color="Cluster", marginal_y="violin",
marginal_x="box",
color_discrete_sequence=[Link])
fig.update_layout(title=[Link]() + " vs Spending",
xaxis_title=[Link](),
yaxis_title="Total Expense",
paper_bgcolor='rgb(229, 236, 246)',
plot_bgcolor='rgb(229, 236, 246)',
Bachelor of Computer Science (Honours)
Faculty of Information and Communication Technology (Kampar Campus), UTAR
77
 width=800,
height=600,
template="simple_white")
[Link]()

Cluster Analysis (RFM):

#Visualize all the feature based on spending
df["Marital_Status"] = df["Marital_Status"].astype("category")
df["Education"] = df["Education"].astype("category")

Personal = ["Age", "Income", "Education", "Marital_Status", "NumWebVisitsMonth",

"Familysize", 'Days_Enrolled']

for i in Personal:
fig = [Link](data_frame=df, x=i, y="Spend", color="Clusters", marginal_y="violin",
marginal_x="box",
color_discrete_sequence=colors_cluster,
category_orders=dict(Clusters=[0,1,2,3]))
fig.update_layout(title=[Link]() + " vs Spending",
xaxis_title=[Link](),
paper_bgcolor='rgb(229, 236, 246)',
plot_bgcolor='rgb(229, 236, 246)',
template="simple_white")
[Link]()

GUI Source code:

import sys
import numpy as np
import pandas as pd
import matplotlib
import [Link]
import [Link] as plt
from [Link] import Figure
from [Link].qt_compat import QtWidgets as QC
from [Link].backend_qt5agg import FigureCanvasQTAgg
from [Link].backend_qt5agg import NavigationToolbar2QT as
NavigationToolbar
from mpl_toolkits.mplot3d import Axes3D
from [Link] import MiniBatchKMeans
from [Link] import StandardScaler
from [Link] import make_pipeline
from [Link] import QWidget, QLabel, QLineEdit, QPushButton, QGroupBox,
QFormLayout, QVBoxLayout, QGridLayout, QHBoxLayout, QSizePolicy, QTableWidget,
QAbstractItemView, QTableWidgetItem, QDialog, QApplication, QMainWindow,
QMessageBox
from [Link] import QIcon, QPalette, QColor
Bachelor of Computer Science (Honours)
Faculty of Information and Communication Technology (Kampar Campus), UTAR
78
from [Link] import QPropertyAnimation, QRect, Qt
from [Link] import QPainter, QBrush, QColor, QPen

data = pd.read_csv('marketing_campaign.csv',delimiter='\t')
numeric_cols = ['Year_Birth', 'Income', 'Kidhome', 'Teenhome', 'Recency', 'MntWines',
'MntFruits', 'MntMeatProducts', 'MntFishProducts', 'MntSweetProducts', 'MntGoldProds',
'NumDealsPurchases', 'NumWebPurchases', 'NumCatalogPurchases', 'NumStorePurchases',
'NumWebVisitsMonth']
data_numeric = data[numeric_cols]
X = StandardScaler().fit_transform(data_numeric)

class CustomerSegmentation(QWidget):
def __init__(self):
super().__init__()
[Link] = pd.read_csv('marketing_campaign.csv',delimiter='\t')

[Link]('Customer Segmentation')
[Link](100, 100, 800, 600)

header = QLabel('RFM')
[Link]('font-size: 36px; font-weight: bold; color: #fff;')
subheader = QLabel('Customer Segmentation')
[Link]('font-size: 18px; color: #666; margin-bottom: 20px;')

layout = QGridLayout()
[Link](20, 20, 20, 20)
[Link](10)

[Link](header, 0, 0, 1, 2, [Link])
[Link](subheader, 1, 0, 1, 2, [Link])

recency_label = QLabel("Recency:")
recency_icon = QIcon('[Link]')
self.recency_input = QLineEdit()
self.recency_input.setPlaceholderText("Days Since Last Purchase")
self.recency_input.setStyleSheet('font-size: 14px; color: #666;')
recency_label.setPixmap(recency_icon.pixmap(24, 24))
[Link](recency_label, 2, 0)
[Link](self.recency_input, 2, 1)

frequency_label = QLabel("Frequency:")
frequency_icon = QIcon('[Link]')
self.frequency_input = QLineEdit()
self.frequency_input.setPlaceholderText("Number of Purchases")
self.frequency_input.setStyleSheet('font-size: 14px; color: #666;')
frequency_label.setPixmap(frequency_icon.pixmap(24, 24))
[Link](frequency_label, 3, 0)
[Link](self.frequency_input, 3, 1)

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
79
 monetary_label = QLabel("Monetary:")
monetary_icon = QIcon('[Link]')
self.monetary_input = QLineEdit()
self.monetary_input.setPlaceholderText("Total Amount Spent")
self.monetary_input.setStyleSheet('font-size: 14px; color: #666;')
monetary_label.setPixmap(monetary_icon.pixmap(24, 24))
[Link](monetary_label, 4, 0)
[Link](self.monetary_input, 4, 1)

button_layout = QHBoxLayout()

self.add_button = QPushButton('Add')
self.add_button.setIcon(QIcon('[Link]'))
self.add_button.setToolTip('Click to add a new customer')
self.add_button.setStyleSheet('font-size: 14px; color: #fff; background-color: #3b9cff;
border: none; padding: 10px 20px;')
self.add_button.setFixedSize(120, 50)
self.add_button.[Link](self.add_customer)
button_layout.addWidget(self.add_button)

self.del_button = QPushButton('Del')
self.del_button.setIcon(QIcon('[Link]'))
self.del_button.setToolTip('Click to delete the selected customer')
self.del_button.setStyleSheet('font-size: 14px; color: #fff; background-color: #3b9cff;
border: none; padding: 10px 20px;')
self.del_button.setFixedSize(120, 50)
self.del_button.[Link](self.del_customer)
button_layout.addWidget(self.del_button)

self.clear_button = QPushButton('Clear')
self.clear_button.setIcon(QIcon('[Link]'))
self.clear_button.setToolTip('Click to clear the input fields')
self.clear_button.setStyleSheet('font-size: 14px; color: #fff; background-color: #3b9cff;
border: none; padding: 10px 20px;')
self.clear_button.setFixedSize(120, 50)
self.clear_button.[Link](self.clear_inputs)
button_layout.addWidget(self.clear_button)

self.predict_button = QPushButton('Predict')
self.predict_button.setIcon(QIcon('[Link]'))
self.predict_button.setToolTip('Click to predict the customer segment')
self.predict_button.setStyleSheet('font-size: 14px; color: #fff; background-color:
#3b9cff; border: none; padding: 10px 20px;')
self.predict_button.setFixedSize(120, 50)
self.predict_button.[Link]([Link])
button_layout.addWidget(self.predict_button)

self.plot_button = QPushButton('3D Plot')

self.plot_button.setIcon(QIcon('[Link]'))

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
80
 self.plot_button.setToolTip('Click to plot the customer segments in 3D')
self.plot_button.setStyleSheet('font-size: 14px; color: #fff; background-color: #3b9cff;
border: none; padding: 10px 20px;')
self.plot_button.setFixedSize(120, 50)
self.plot_button.[Link](self.show_plot)
button_layout.addWidget(self.plot_button)

[Link](button_layout, 5, 0, 1, 2, [Link])

table_button_layout = QHBoxLayout()

table_button_layout.addLayout(button_layout)

table_button_layout = QHBoxLayout()
[Link] = QTableWidget()
columns_to_drop = ['Year_Birth', 'Education', 'Marital_Status', 'Income', 'Kidhome',
'Teenhome', 'Dt_Customer', 'MntWines',
'MntFruits','MntMeatProducts','MntFishProducts','MntSweetProducts','MntGoldProds',
'NumWebPurchases', 'NumCatalogPurchases', 'NumStorePurchases', 'NumDealsPurchases',
'NumWebVisitsMonth', 'AcceptedCmp1', 'AcceptedCmp2', 'AcceptedCmp3',
'AcceptedCmp4', 'AcceptedCmp5', 'Complain', 'Response', 'Z_CostContact', 'Z_Revenue']

[Link](columns=columns_to_drop, inplace=True)

[Link](len([Link]))
[Link](len([Link]))

# Set the header labels of the table to match the column names of your data
[Link](list([Link]))

[Link]([Link])
[Link]([Link])
[Link]([Link])
[Link](self.select_customer)
[Link]('font-size: 14px;')
[Link](600, 400)

table_button_layout.addWidget([Link])

for i in range(len([Link])):
row_items = []
for col_name in ['ID', 'Recency', 'Frequency', 'Monetary']:
col_index = [Link].get_loc(col_name)
row_items.append(int(float([Link][i, col_index])))

[Link](i)
for j, item in enumerate(row_items):
[Link](i, j, QTableWidgetItem(str(item)))

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
81
 [Link]([Link], 6, 0, 1, 2, [Link])
table_button_layout.addWidget([Link])

result_layout = QHBoxLayout()
self.result_label = QLabel("Result:")
self.result_label.setStyleSheet('font-size: 14px; color: #fff;')
result_layout.addWidget(self.result_label)
[Link](result_layout, 8, 0, 1, 2)
[Link] = [Link]([[0, 0, 0]])
[Link](layout)
recency_label.setBuddy(self.recency_input)
frequency_label.setBuddy(self.frequency_input)
monetary_label.setBuddy(self.monetary_input)

[Link]('Fusion')
palette = QPalette()
[Link]([Link], QColor(53, 53, 53))
[Link]([Link], [Link])
[Link]([Link], QColor(25, 25, 25))
[Link]([Link], QColor(53, 53, 53))
[Link]([Link], [Link])
[Link]([Link], [Link])
[Link]([Link], [Link])
[Link]([Link], QColor(53, 53, 53))
[Link]([Link], [Link])
[Link]([Link], [Link])
[Link]([Link], QColor(42, 130, 218))
[Link]([Link], QColor(42, 130, 218))
[Link]([Link], [Link])
[Link](palette)

def add_customer(self):
try:
recency = int(self.recency_input.text())
frequency = int(self.frequency_input.text())
monetary = int(self.monetary_input.text())
except ValueError:
[Link](self, 'Invalid Input', 'Please enter valid numeric values for
recency, frequency, and monetary.')
return

id = int([Link]['ID'].max()) + 1

new_customer = [Link]({
'ID': [id],
'Recency': [recency],
'Frequency': [frequency],
'Monetary': [monetary]
})

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
82
 row_num = 0
for i in range(len([Link])):
if [Link][i, 'ID'] > id:
break
row_num += 1

[Link](row_num)
for i, value in enumerate([id, recency, frequency, monetary]):
[Link](row_num, i, QTableWidgetItem(str(value)))
[Link] = [Link]([[Link][:row_num], new_customer,
[Link][row_num:]], ignore_index=True)

[Link](self, 'Success', 'New customer added to the dataset.')

self.recency_input.clear()
self.frequency_input.clear()
self.monetary_input.clear()

def del_customer(self):
# Get the selected row index
row_index = [Link]()
if row_index >= 0:
[Link](row_index)
else:
[Link](self, 'Error', 'No customer selected.')

def select_customer(self, row, col):

recency = [Link](row, 1).text()
frequency = [Link](row, 2).text()
monetary = [Link](row, 3).text()
self.recency_input.setText(str(recency))
self.frequency_input.setText(str(frequency))
self.monetary_input.setText(str(monetary))

def predict(self):
data['Frequency'] = data['NumWebPurchases'] + data['NumCatalogPurchases'] +
data['NumStorePurchases']
data['Monetary'] = data['MntWines'] + data['MntFruits'] + data['MntMeatProducts'] +
data['MntFishProducts'] + data['MntSweetProducts'] + data['MntGoldProds']

recency_str = self.recency_input.text().strip()
frequency_str = self.frequency_input.text().strip()
monetary_str = self.monetary_input.text().strip()

if not recency_str or not frequency_str or not monetary_str:

self.result_label.setText("Please provide values for all inputs")
return

try:

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
83
 recency = int(float(recency_str))
frequency = int(float(frequency_str))
monetary = int(float(monetary_str))
except ValueError:
self.result_label.setText("Please provide numeric values for inputs")
return

X = data[['Recency', 'Frequency', 'Monetary']]

[Link] = MiniBatchKMeans(n_clusters=4, batch_size=100, random_state=42)
[Link].partial_fit(X)
[Link] = [Link]([[recency, frequency, monetary]])
cluster_label = [Link]([Link](1, -1))[0]

customer_segments = {
0: "Low-Spending Active Customers",
1: "Mid-Spending Active Customers",
2: "High-Spending Active Customers",
3: "Churned or Inactive Customers"
}
segment = customer_segments.get(cluster_label)
self.result_label.setText("Cluster {}: {}".format(cluster_label, segment))

def plot(self):
X = data[['Recency', 'Frequency', 'Monetary']]
kmeans = MiniBatchKMeans(n_clusters=4, batch_size=100, random_state=42)
kmeans.partial_fit(X)
fig = Figure()
ax = fig.add_subplot(111, projection='3d')
colors = ['#3C1642', '#086375', '#1DD3B0', '#AFFC41']
legend_elements = [
[Link].Line2D([0], [0], marker='o', color='w', label='Low-Spending Active',
markerfacecolor=colors[0], markersize=10),
[Link].Line2D([0], [0], marker='o', color='w', label='Mid-Spending Active',
markerfacecolor=colors[1], markersize=10),
[Link].Line2D([0], [0], marker='o', color='w', label='High-Spending Active',
markerfacecolor=colors[2], markersize=10),
[Link].Line2D([0], [0], marker='o', color='w', label='Churned or Inactive',
markerfacecolor=colors[3], markersize=10),
]
[Link](handles=legend_elements, loc='upper right')

for i in range(4):
cluster_points = X[kmeans.labels_ == i]
[Link](cluster_points['Recency'], cluster_points['Frequency'],
cluster_points['Monetary'],
c=colors[i], label=f'Cluster {i}', s=50, alpha=0.5)

[Link]([Link][:, 0], [Link][:, 1], [Link][:, 2], c='red', marker='x',

s=100, linewidths=2, alpha=0.8)

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
84
 ax.set_xlabel('Recency')
ax.set_ylabel('Frequency')
ax.set_zlabel('Monetary')
canvas = FigureCanvasQTAgg(fig)
toolbar = NavigationToolbar(canvas, self, coordinates=True)
plot_window = QMainWindow()
plot_window.setWindowTitle('Customer Segmentation Plot')
plot_window_layout = QVBoxLayout()
plot_window_layout.addWidget(canvas)
plot_window_layout.addWidget(toolbar)
central_widget = QWidget()
central_widget.setLayout(plot_window_layout)
plot_window.setCentralWidget(central_widget)
plot_window.setGeometry(100, 100, 800, 600)
self.plot_window = plot_window

def show_plot(self):
[Link]()
self.plot_window.show()

def clear_inputs(self):
self.recency_input.clear()
self.frequency_input.clear()
self.monetary_input.clear()

if __name__ == '__main__':
app = QApplication([Link])
segmentation = CustomerSegmentation()
[Link]()
[Link](app.exec_())

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
85
 FINAL YEAR PROJECT WEEKLY REPORT
(Project II)

Trimester, Year: Y3S3 Study week no.: 1

Student Name & ID: Toh Wei Xuan 19ACB03568
Supervisor: Dr Lim Jia QI
Project Title: Customer Segmentation on Clustering Algorithms

1. WORK DONE
Check for the previous work.

2. WORK TO BE DONE
Copy and paste chapters 1 and 2 to new fyp2 template.

3. PROBLEMS ENCOUNTERED
Previous source code got some bugs and error

4. SELF EVALUATION OF THE PROGRESS

Not bad Progress done until the visualization

_________________________ _________________________
Supervisor’s signature Student’s signature

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
86
Trimester, Year: Y3S3 Study week no.: 2
Student Name & ID: Toh Wei Xuan 19ACB03568
Supervisor: Dr Lim Jia QI
Project Title: Customer Segmentation on Clustering Algorithms

1. WORK DONE
Learnt how to Determine the k value.

2. WORK TO BE DONE
Encounter all the clustering algorithms.

3. PROBLEMS ENCOUNTERED
Various methods to find the optimal k.

4. SELF EVALUATION OF THE PROGRESS

Still fine in start Clustering implementation

_________________________ _________________________
Supervisor’s signature Student’s signature

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
87
Trimester, Year: Y3S3 Study week no.: 3
Student Name & ID: Toh Wei Xuan 19ACB03568
Supervisor: Dr Lim Jia QI
Project Title: Customer Segmentation on Clustering Algorithms

1. WORK DONE
Optimal k = 4
Implement clustering algorithms k-means.

2. WORK TO BE DONE
Implement GMM and DBSCAN

3. PROBLEMS ENCOUNTERED
Different k value encountered 2 and 4 when used different metrices.

4. SELF EVALUATION OF THE PROGRESS

Still fine

_________________________ _________________________
Supervisor’s signature Student’s signature

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
88
Trimester, Year: Y3S3 Study week no.: 4
Student Name & ID: Toh Wei Xuan 19ACB03568
Supervisor: Dr Lim Jia QI
Project Title: Customer Segmentation on Clustering Algorithms

1. WORK DONE
Implemented k-means, GMM and DBSCAN

2. WORK TO BE DONE
Performs summary comparison of performance metrices to determine k

3. PROBLEMS ENCOUNTERED
Epsilon of DBSCAN is difficult to determine.

4. SELF EVALUATION OF THE PROGRESS

Still fine

_________________________ _________________________
Supervisor’s signature Student’s signature

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
89
Trimester, Year: Y3S3 Study week no.: 5
Student Name & ID: Toh Wei Xuan 19ACB03568
Supervisor: Dr Lim Jia QI
Project Title: Customer Segmentation on Clustering Algorithms

1. WORK DONE
Pivot table of summary metrices was implemented, optimal k was 2 and perform the
three clustering algorithms except DBSCAN.

2. WORK TO BE DONE
Use GridSearchCV() to determine the epsilon ok DBSCAN

3. PROBLEMS ENCOUNTERED
Hyperparameter of DBSCAN difficult to determine.

4. SELF EVALUATION OF THE PROGRESS

Mad on the progress since is stucked

_________________________ _________________________
Supervisor’s signature Student’s signature

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
90
Trimester, Year: Y3S3 Study week no.: 6
Student Name & ID: Toh Wei Xuan 19ACB03568
Supervisor: Dr Lim Jia QI
Project Title: Customer Segmentation on Clustering Algorithms

1. WORK DONE
Three algorithms, k-means, GMM and DBSCAN successfully implemented and plot the
result in 2D and 3D form. The comparison between the clustering algorithms.

2. WORK TO BE DONE
learn new approach/ algorithm, RFM.

3. PROBLEMS ENCOUNTERED
Difficult to categorize the group of customers if two cluster only

4. SELF EVALUATION OF THE PROGRESS

Still fine

_________________________ _________________________
Supervisor’s signature Student’s signature

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
91
Trimester, Year: Y3S3 Study week no.: 7
Student Name & ID: Toh Wei Xuan 19ACB03568
Supervisor: Dr Lim Jia QI
Project Title: Customer Segmentation on Clustering Algorithms

1. WORK DONE
Distribution of RFM, RFM modelling and plotting

2. WORK TO BE DONE
Cluster Analysis to categorize the clusters groups.

3. PROBLEMS ENCOUNTERED
Got two k values, 2 and 4 and confirm to use k=4 due to its was behavioral based
clustering and easier to identify.

4. SELF EVALUATION OF THE PROGRESS

Still ok

_________________________ _________________________
Supervisor’s signature Student’s signature

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
92
Trimester, Year: Y3S3 Study week no.: 8
Student Name & ID: Toh Wei Xuan 19ACB03568
Supervisor: Dr Lim Jia QI
Project Title: Customer Segmentation on Clustering Algorithms

1. WORK DONE
Cluster Analysis and RFM modelling and categorizing.

2. WORK TO BE DONE
GUI development using Incremental k-means.

3. PROBLEMS ENCOUNTERED
Naming of the categorized customer groups.

4. SELF EVALUATION OF THE PROGRESS

Still fine.

_________________________ _________________________
Supervisor’s signature Student’s signature

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
93
Trimester, Year: Y3S3 Study week no.: 9
Student Name & ID: Toh Wei Xuan 19ACB03568
Supervisor: Dr Lim Jia QI
Project Title: Customer Segmentation on Clustering Algorithms

1. WORK DONE
GUI using MiniBatch K-means was developed and got the function of input integer value
recency, frequency and monetary.

2. WORK TO BE DONE
Add CRUD function on the GUI.

3. PROBLEMS ENCOUNTERED
Incremental k-means module cannot implement due to the scikit learn version problem of
my pc and I replaced it to use MiniBatch k-means.

4. SELF EVALUATION OF THE PROGRESS

Still fine

_________________________ _________________________
Supervisor’s signature Student’s signature

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
94
Trimester, Year: Y3S3 Study week no.: 10
Student Name & ID: Toh Wei Xuan 19ACB03568
Supervisor: Dr Lim Jia QI
Project Title: Customer Segmentation on Clustering Algorithms

1. WORK DONE
User friendly GUI developed.

2. WORK TO BE DONE
Report writing from chapter 3.

3. PROBLEMS ENCOUNTERED
Report organization not sure

4. SELF EVALUATION OF THE PROGRESS

Glad due to almost done

_________________________ _________________________
Supervisor’s signature Student’s signature

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
95
 POSTER

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
96
 PLAGIARISM CHECK RESULT

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
97
 Universiti Tunku Abdul Rahman
Form Title : Supervisor’s Comments on Originality Report Generated by Turnitin
for Submission of Final Year Project Report (for Undergraduate Programmes)
Form Number: FM-IAD-005 Rev No.: 0 Effective Date: 01/10/2013 Page No.: 1of 1

FACULTY OF INFORMATION AND COMMUNICATION

TECHNOLOGY

Full Name(s) of TOH WEI XUAN

Candidate(s)
ID Number(s) 19ACB03568
Programme / Course CS

Title of Final Year Project CUSTOMER SEGMENTATION ON CLUSTERING ALGORITHMS

Similarity Supervisor’s Comments

(Compulsory if parameters of originality exceeds
the limits approved by UTAR)

Overall similarity index: 16 % Similarity

by source
Internet Sources: 8_%
Publications: 6 %
Student Papers: 12 %

Number of individual sources listed of

more than 3% similarity: 0
Parameters of originality required and limits approved by UTAR are as Follows:
(i) Overall similarity index is 20% and below, and
(ii) Matching of individual sources listed must be less than 3% each, and
(iii) Matching texts in continuous block must not exceed 8 words
Note: Parameters (i) – (ii) shall exclude quotes, bibliography and text matches which are less than 8 words.

Note Supervisor/Candidate(s) is/are required to provide softcopy of full set of the originality report
to Faculty/Institute

Based on the above results, I hereby declare that I am satisfied with the originality of the Final
Year Project Report submitted by my student(s) as named above.

______________________________ ______________________________
Signature of Supervisor Signature of Co-Supervisor

Name: __________Lim Jia Qi_______ Name:

Date: _________28/04/2023_________ Date:

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
98
 UNIVERSITI TUNKU ABDUL RAHMAN
FACULTY OF INFORMATION & COMMUNICATION TECHNOLOGY
(KAMPAR CAMPUS)
CHECKLIST FOR FYP2 THESIS SUBMISSION
Student Id 19ACB03568
Student Name TOH WEI XUAN
Supervisor Name DR. LIM JIA QI

TICK (√) DOCUMENT ITEMS

Your report must include all the items below. Put a tick on the left column after you have
checked your report with respect to the corresponding item.
Front Plastic Cover (for hardcopy)
√ Title Page
√ Signed Report Status Declaration Form
√ Signed FYP Thesis Submission Form
√ Signed form of the Declaration of Originality
√ Acknowledgement
√ Abstract
√ Table of Contents
√ List of Figures (if applicable)
√ List of Tables (if applicable)
√ List of Symbols (if applicable)
√ List of Abbreviations (if applicable)
√ Chapters / Content
√ Bibliography (or References)
√ All references in bibliography are cited in the thesis, especially in the chapter
of literature review
√ Appendices (if applicable)
√ Weekly Log
√ Poster
√ Signed Turnitin Report (Plagiarism Check Result - Form Number: FM-IAD-005)
√ I agree 5 marks will be deducted due to incorrect format, declare wrongly the
ticked of these items, and/or any dispute happening for these items in this
report.

*Include this form (checklist) in the thesis (Bind together as the last page)

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
99
I, the author, have checked and confirmed all the items listed in the table are included in my
report.

______________________
(TOH WEI XUAN)
Date: 25/4/2023

Bachelor of Computer Science (Honours)

Faculty of Information and Communication Technology (Kampar Campus), UTAR
100