See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/368911804

Customer Churn Prediction in the Banking Sector Using Machine Learning-

Based Classification Models

Article in Interdisciplinary Journal of Information · February 2023

DOI: 10.28945/5086

CITATIONS READS
18 2,427

3 authors, including:

Van-Ho Nguyen
University of Economics and Law (VNU-HCM)
14 PUBLICATIONS 83 CITATIONS

SEE PROFILE

All content following this page was uploaded by Van-Ho Nguyen on 02 May 2023.

The user has requested enhancement of the downloaded file.

 Volume 18, 2023

CUSTOMER CHURN PREDICTION IN THE BANKING

SECTOR USING MACHINE LEARNING-BASED
CLASSIFICATION MODELS
Hoang Tran University of Economics and Law, hoangtd21411c@st.uel.edu.vn
Ho Chi Minh City, Vietnam
Vietnam National University,
Ho Chi Minh City, Vietnam
Ngoc Le University of Economics and Law, ngocltl21413c@st.uel.edu.vn
Ho Chi Minh City, Vietnam
Vietnam National University,
Ho Chi Minh City, Vietnam
Van-Ho Nguyen* University of Economics and Law, honv@uel.edu.vn
Ho Chi Minh City, Vietnam
Vietnam National University,
Ho Chi Minh City, Vietnam
* Corresponding author

ABSTRACT
Aim/Purpose Previous research has generally concentrated on identifying the variables that
most significantly influence customer churn or has used customer segmentation
to identify a subset of potential consumers, excluding its effects on forecast ac-
curacy. Consequently, there are two primary research goals in this work. The ini-
tial goal was to examine the impact of customer segmentation on the accuracy
of customer churn prediction in the banking sector using machine learning
models. The second objective is to experiment, contrast, and assess which ma-
chine learning approaches are most effective in predicting customer churn.
Background This paper reviews the theoretical basis of customer churn, and customer seg-
mentation, and suggests using supervised machine-learning techniques for cus-
tomer attrition prediction.
Methodology In this study, we use different machine learning models such as k-means cluster-
ing to segment customers, k-nearest neighbors, logistic regression, decision tree,
random forest, and support vector machine to apply to the dataset to predict
customer churn.
Accepting Editor Dirk Frosch-Wilke│ Received: November 9, 2022│ Revised: January 2, January 28, February
19, February 20, 2023 │ Accepted: February 21, 2023.
Cite as: Tran, H., Le, N., & Nguyen, V.-H. (2023). Customer churn prediction in the banking sector using
machine learning-based classification models. Interdisciplinary Journal of Information, Knowledge, and Management, 18,
87-105. https://doi.org/10.28945/5086
(CC BY-NC 4.0) This article is licensed to you under a Creative Commons Attribution-NonCommercial 4.0 International
License. When you copy and redistribute this paper in full or in part, you need to provide proper attribution to it to ensure
that others can later locate this work (and to ensure that others do not accuse you of plagiarism). You may (and we encour-
age you to) adapt, remix, transform, and build upon the material for any non-commercial purposes. This license does not
permit you to use this material for commercial purposes.
Customer Churn Prediction

Contribution The results demonstrate that the dataset performs well with the random forest
model, with an accuracy of about 97%, and that, following customer segmenta-
tion, the mean accuracy of each model performed well, with logistic regression
having the lowest accuracy (87.27%) and random forest having the best
(97.25%).
Findings Customer segmentation does not have much impact on the precision of predic-
tions. It is dependent on the dataset and the models we choose.
Recommendations The practitioners can apply the proposed solutions to build a predictive system
for Practitioners or apply them in other fields such as education, tourism, marketing, and human
resources.
Recommendations The research paradigm is also applicable in other areas such as artificial intelli-
for Researchers gence, machine learning, and churn prediction.
Impact on Society Customer churn will cause the value flowing from customers to enterprises to
decrease. If customer churn continues to occur, the enterprise will gradually lose
its competitive advantage.
Future Research Build a real-time or near real-time application to provide close information to
make good decisions. Furthermore, handle the imbalanced data using new tech-
niques.
Keywords churn prediction, machine learning, banking industry, classification models,
SMOTE

INTRODUCTION
Customer churn is a business term known as customer agitation. In business, when customers are
dissatisfied with the service or product that you provide, attrition will occur, or in other words, they
will stop connecting or cooperating with your business. Nowadays, more and more businesses are
concerned with customer retention. Customer churn hurts businesses since it can result in large pre-
mium losses, decreased profit margins, and possibly lost referral business from loyal clients. Accord-
ing to Baran and Galka (2017), “The pioneering work of F. F. Reichheld and W. E. Sasser Jr. found a
strong relationship between customer retention and company profits. They found that just a 5 per-
cent increase in customer retention yielded improved profitability in net present value from 20 to 85
percent across a wide range of businesses.” Additionally, many types of research demonstrated that
keeping current clients costs about five times less than acquiring new ones (Dawes & Swailes, 1999).
According to Reichheld (1996), a client’s average net present value improves by 35% for software
companies and 95% for advertising agencies with a 5% increase in customer retention rates. To limit
this, businesses must anticipate specific customers at risk of leaving to adjust their strategies accord-
ingly, such as improving the quality of products and services or increasing their benefits. Therefore, it
is essential to create predictive models that could aid in client retention. Machine learning methods
can successfully anticipate customer turnover, as Dolatabadi and Keynia (2017) show. Data collection
and analysis allow one to identify customers who may be about to leave a business, thereby focusing
on customer retention rather than attracting new clients. In addition, machine learning models are
also applied in retail and economic sectors. For retailers, understanding the needs of customers is ex-
tremely necessary; machine learning models are applied to analyzing shopping behavior based on the
customer’s purchase history, thereby finding out the relationship between items. This is the key to
boosting purchase rates and optimizing profits for businesses (Reddy & Reddy, 2021).
A few studies have also integrated the use of customer segmentation and machine learning tech-
niques to improve the predictive model’s accuracy. Customer segmentation is the process of dividing
customers into groups based on common target customer characteristics so that companies can mar-

88
 Tran, Le, & Nguyen

ket to each group effectively and appropriately. This is a pretty important step, a factor that helps in-
crease the conversion rate for businesses. If businesses take this step well, it can help them divide
their advertising budgets better and save more. Additionally, customer segmentation will assist them
in better understanding their consumers, identifying the target customer category for the business to
focus on, and then considering the factors that affect that customer’s churn.
Previous research has generally concentrated on identifying the variables that most significantly influ-
ence customer churn or has simply used customer segmentation to identify a subset of potential con-
sumers, excluding its effects on forecast accuracy. Consequently, there are two primary research goals
in this work. The initial goal is to examine the impact of customer segmentation on the accuracy of
customer churn prediction in the banking sector using machine learning models. The second objec-
tive is to experiment, contrast, and assess which machine learning approaches are most effective in
predicting customer churn. Relationships with customers have a big impact on long-term growth,
thus, understanding their behavior allows banks to easily improve existing policies. Building a model
to predict credit card discontinuation has important significance for banks. In this study, we use dif-
ferent machine learning models such as k-means clustering (Kanungo et al., 2000), k-nearest neigh-
bors (Cunningham & Delany, 2021), logistic regression (Sperandei, 2014), decision tree (Quinlan,
1990), random forest (Biau & Scornet, 2016), and support vector machine (Jakkula, 2006) to predict
credit card churn.
The remainder of this paper is structured as follows. The history of the study and the related studies
are covered in the next section. Then the theoretical foundation is presented, followed by the data
and methodology. Next, the model results are shown with some discussion. Finally, the paper out-
lines the research’s limitations.

RELATED WORK
C H URN P REDICTION U SING M ACH INE L EARNING
Various machine learning models, including logistic regression (LR), decision tree (DT), k-nearest
neighbor (KNN), random forest (RF), were used in this study (Kaur & Kaur, 2020) to estimate the
likelihood that a client will leave. Performance measures including memory, accuracy, and others are
compared. Support vector machines (SVM), which might increase system performance, were not
taken into account in this study. Inspired by this, we added the SVM model to our work.
Guliyev and Tatoğlu (2021) employed SHapley Additive exPlanations (SHAP) values to help the ma-
chine learning model evaluation and interpretability for customer churn analysis. It focused on partic-
ularly explainable machine learning models. Utilizing actual banking data, the research aimed to esti-
mate the explainable machine learning model and assess a variety of machine learning models using
test data. The XG-boost model fared better than other machine learning techniques in categorizing
clients who churn, according to the data. According to Yaseen (2021), they used the feature selection
technique to determine the most important variables in predicting customer attrition. They used the
wrapper-based feature selection approach, where particle swarm optimization (PSO) was applied for
searches, and different classifiers, such as decision tree, naive bayes, k-nearest neighbor, and logistic
regression, were applied for evaluation to judge the enactment on optimally sampled and condensed
datasets. Last, but not least, simulations showed that their recommended strategy did well for fore-
casting churners and might therefore be helpful for the telecommunications sector’s constantly ex-
panding rivalry. Rahman and Kumar (2020) proposed an approach for predicting client turnover in a
bank using machine learning methods, a subfield of artificial intelligence by examining consumer be-
havior, the study encourages investigation into the possibility of churn. In their work, the classifiers
KNN, SVM, DT, and RF are employed. Additionally, several feature selection techniques have been
used to identify the features that are more pertinent and to assess system performance. On the
Kaggle churn modeling dataset, the experiment was run. To discover a suitable model with greater
precision and predictability, the results are compared. As a result, the accuracy of the random forest

89
Customer Churn Prediction

model following oversampling is superior to other models. The dataset used by Karvana et al. (2019),
which consisted of 57 attributes, was used to test 5 different categorization techniques. Using com-
parisons between several classes, experiments were conducted repeatedly. The most effective ap-
proach for forecasting customer attrition at an Indonesian private bank uses support vector machines
in comparison to data from class sample split 50:50. The company that plans to take strategic action
to stop customer churn might make use of the modeling results. When there is a lack of a lot of
data, this research uses k-fold cross-validation to assess machine learning models. This helps to avoid
the model being evaluated incorrectly.
Elyusufi and Ait Kbir (2022) analyzed several specific machine learning models that had been put
forth in the literature to address this issue and contrasted them with some recently developed models
that are based on ensemble learning techniques. As a result, they developed predictive churn strate-
gies that analyzed customer history data, determined who was active after a specific period, and then
developed models that pinpointed the points at which a client might stop using a certain firm service.
When completing the training step with traditional models like multi-layer perception neural net-
works, ensemble learning methods were also employed to locate pertinent features to reduce their
quantity. The proposed methodologies could obtain accuracy levels of up to 89% when other re-
search studies using the same dataset only manage to reach 86%. According to this study, the SVM
model has an accuracy rating of 86.18%, which is rather good. A high accuracy for a particular model
would mean that the model is capable of predicting the choice that a client can make (leave the
bank/remain with the bank).
With the use of the three intelligence models random forest, adaboost, and support vector machine,
Muneer et al. (2022) developed a method for predicting client attrition (SVM). When the synthetic
minority oversampling technique (SMOTE) was used to overcome the unbalanced dataset and the
combination of undersampling and oversampling, the method produces the best results. Excellent
results were obtained utilizing the approach on SMOTED data, with a 91.90% F1 score and an over-
all accuracy of 88.7% when employing RF. Additionally, the experimental findings demonstrated that
RF produced good outcomes for the entire feature-selected datasets. SMOTE was utilized in this
study to reduce sample size disparity. This improves the forecast results’ accuracy as well.

C H URN P REDICTION U SING M ACH INE L EARNING C OMBINES C USTOMER

S EGMENTATION
Sivasankar and Vijaya (2017) compared different unsupervised learning methodologies employing al-
gorithms like fuzzy c-means (FCM), probabilistic fuzzy c-means (PFCM), and k-means clustering (k-
means), which aggregated similar types of customers into clusters and forecast improved customer
segmentation. By using the Holdout approach, the clusters were sorted into training and testing
groups. The training was done using decision trees, and testing was done using the produced models.
The experimental findings of this study have demonstrated that the k-means method provides better
cluster quality than fuzzy grouping and that the classification accuracy is improved when the decision
tree is used in combination with the k-means algorithm. Therefore, in this work, we integrated the
decision tree and the k-means algorithm.
Using the clustering technique and the k-means clustering algorithm, Olaniyi et al. (2020) analyzed
consumer competency and sector continuity to anticipate customer behavior. The data were grouped
into three labels according to the inflow and outflow of transactions. A support vector machine was
used to classify the clustering findings, and an accuracy of 97% was reached. This study enabled
banking administrators to analyze client behavior, which might lead to appropriate methods for im-
proving consumer engagement and strengthening administrator conducts. Through customer seg-
mentation, Zhang et al. (2022) intended to create a churn prediction model for telecom clients. A tel-
ecom customer churn prediction model was created using data gathered from three significant Chi-
nese telecom providers, fisher discriminant equations, and logistic regression analysis. The results in-
dicate that the regression-based telecom customer churn model produced better outcomes and had a

90
 Tran, Le, & Nguyen

higher prediction accuracy (93.94%). By clustering and classifying the customers according to their
features, this study (Routray, 2021) suggested a market segmentation and customer behavior predic-
tion model. As a result, using the rapid-miner-9.07-toolkit, this study offered the random forest algo-
rithm for predicting the output of churn and the decision tree for predicting the choice of the mo-
bile plan of consumers. These studies examined client segments to discover prospective future con-
sumer groups and, based on that information, to create efficient customer retention tactics.
Xiahou and Harada (2022) suggested a loss prediction model that combines customer segmentation
using k-means with prediction using a support vector machine. The technique identified the primary
customer groups and divides the consumer base into three categories. To forecast customer attrition,
the support vector machine and logistic regression were evaluated. The findings demonstrated the
importance of k-means clustering segmentation by demonstrating a significant improvement in each
prediction index following customer segmentation. SVM predictions were more accurate than those
made using logistic regression. To assist businesses in effectively segmenting their customers, Zhuang
(2018) developed a customer value model that incorporated the importance of social networks.
Then, they forecast customer turnover before and after the subdivision using the machine learning
technique XG-boost. The study discovered that consumer segmentation increases prediction accu-
racy. The XG-boost algorithm was also superior to other algorithms in terms of benefits.
The empirical findings of the above studies demonstrated that consumer segmentation would greatly
enhance each predictor, so we experimented with the data from the banking industry to support this.

DATA AND METHODOLOGY

The research method will comprise phases from data collection to data preparation and pre-pro-
cessing so that it can be used in the model, including customer segmentation to verify the hypothesis
and handling imbalanced data using SMOTE. SMOTE is a common sampling technique to enhance
the sample size of the minority group in cases when there is an imbalance in the sample size. To in-
crease the sample size, we will select the k nearest neighbor samples for each sample belonging to the
minority group, and then perform linear combination to create the simulated sample. This pipeline in
Figure 1 demonstrates the step-by-step process we employ in this section to anticipate client attrition.

DATA P REPARATION
Data preparation is crucial for improved data analysis and management across all data-driven applica-
tions, not just data science (Hameed & Naumann, 2020). This section provides some background in-
formation on the dataset and outlines how it was utilized for the prediction process. The Kaggle
website (https://www.kaggle.com/datasets/anwarsan/credit-card-bank-churn) provides a freely ac-
cessible dataset for the prediction job. The final two columns of the 23 variables should be elimi-
nated because they are useless for categorizing. After removing the final two columns, the dataset
now has 21 variables, with a total of 10,127 entries. Table 1 contains and describes a list of the varia-
bles in the dataset.

Machine learning scaling is part of data pre-processing as this technique brings data points that are
far from each other closer to increase the algorithm effectiveness and speed up the Machine Learning
processing. Scaling data enables the model to learn and understand the problem. There are two main
data scaling methods commonly used in algorithms when scaling is indispensable techniques are nor-
malization and standardization. As a result, shown in Figure 2, multiple histograms have been gener-
ated to show the distribution of the required variables so that a suitable scaling technique may be se-
lected. The variables that were used to determine the distribution for data scaling among the quanti-
tative data have a wide range. Since the remaining quantitative data has a narrow range of variance as
shown in Figure 3, it will be maintained.

91
Customer Churn Prediction

Figure 1. Process of research and implementation

92
 Tran, Le, & Nguyen

Table 1. Data description

Feature Description
CLIENTNUM Client identification number
Attrition_Flag Account status, Existing = stay, Attrited = churn
Customer_Age Customer Age
Gender Customer Gender, M=Male, F=Female
Dependent_count Number of dependents
Education_Level Customer education level
Marital_Status Customer marital status
Income_Category Customer’s Annual Income Category
Card_Category Type of card
Months_on_book Relationship duration with the bank
Total_Relationship_Count Total number of goods owned by the consumer
Months_Inactive_12_mon Number of inactive months in the last 12 months
Contacts_Count_12_mon Number of Contacts in the last 12 months
Credit_Limit The credit limit on the credit card
Total_Revolving_Bal Total revolving balance on the credit card
Avg_Open_To_Buy Average of last 12 months
Total_Amt_Chng_Q4_Q1 Change in Transaction amount (Q4 over Q1)
Total_Trans_Amt Total Transaction Amount (Last 12 Months)
Total_Trans_Ct Total Transaction Count (Last 12 Months)
Total_Ct_Chng_Q4_Q1 Change in Transaction count (Q4 over Q1)
Avg_Utilization_Ratio The average credit card use ratio

93
Customer Churn Prediction

Figure 2. Data distribution

94
 Tran, Le, & Nguyen

Figure 3. Quantitative data chart

DATA P REPROCESSING
Data preprocessing is a critical stage in machine learning that improves the quality of the data to en-
courage the extraction of valuable insights from the data (Alexandropoulos et al., 2019). Preparing
(cleaning and arranging) raw data to make it acceptable for creating and training Machine Learning
models is known as data preprocessing in machine learning. Data preprocessing in machine learning
is, to put it simply, a data mining approach that converts raw data into a format that is legible and in-
telligible. First, the Unknown values discovered in the data preparation phase will be eliminated from
the dataset. In the same stage, Customer_Age and Months_on_Book were figured out to have a dis-
tribution similar to the normal distribution. Thus, the standardization method will be applied to Cus-
tomer_Age and Months_on_Book. With the rest of the quantitative data that needs to be trans-
formed, normalization will be used. For the categories with two values, such as Gender and Attri-
tion_Flag, label encoding will then be used. For category data with numerous values, such as Educa-
tion_Level, Marital_Status, Income_Category, and Card_Category, one hot encoding will be utilized.
The final steps include replacing the Divorced value with the Single value and the College value with
the Graduate value, as was specified during the data preparation process. A list of the transformation
methods we employ for each variable in the dataset is provided in Table 2.
Table 2. Transformation method
Feature Transformation method
CLIENTNUM Not used in prediction
Attrition_Flag Label encoding
Customer_Age Standardization
Gender Label encoding
Dependent_count Unchanged
Education_Level One-hot encoding
Marital_Status One-hot encoding
Income_Category One-hot encoding
Card_Category One-hot encoding

95
Customer Churn Prediction

Feature Transformation method

Months_on_book Standardization
Total_Relationship_Count Unchanged
Months_Inactive_12_mon Unchanged
Contacts_Count_12_mon Unchanged
Credit_Limit Normalization
Total_Revolving_Bal Normalization
Avg_Open_To_Buy Normalization
Total_Amt_Chng_Q4_Q1 Unchanged
Total_Trans_Amt Normalization
Total_Trans_Ct Normalization
Total_Ct_Chng_Q4_Q1 Unchanged
Avg_Utilization_Ratio Unchanged

K-MEANS AND C USTOMER S EGMENTATION

In numerous research, they use k-means clustering to separate customers into several segments
(Pradana & Ha, 2021). To increase accuracy, they frequently combine elbow techniques to determine
the ideal number of clusters for data division. As a result, the number of clusters is 6 as shown in
Figure 4:

Figure 4. Distortion score elbow for k-means clustering

96
 Tran, Le, & Nguyen

DATA B ALANCING
The results before and after using SMOTE are shown in Tables 3 and 4:

Table 3. The ratio of data before using SMOTE

Dataset Churn Non-churn Ratio
Original 1113 5968 1:5.4
Cluster I 249 840 1:3.4
Cluster II 354 1323 1:3.4
Cluster III 104 930 1:8.9
Cluster IV 182 934 1:5.1
Cluster V 126 820 1:6.5
Cluster VI 98 1121 1:11.4

Table 4. The ratio of data after using SMOTE

Dataset Churn Non-churn Ratio
Original 5968 5968 1:1
Cluster I 840 840 1:1
Cluster II 1323 1323 1:1
Cluster III 930 930 1:1
Cluster IV 934 934 1:1
Cluster V 820 820 1:1
Cluster VI 1121 1121 1:1

RESULT AND DISCUSSION

Data from each cluster and the sample dataset were balanced, then divided into a 70/30 ratio and
stratified k-fold cross-validation with k equal to 10 was used to evaluate the model before applying
the test set to the models for prediction. The result of each cluster is shown in Table 5.
Remember that accuracy is the proportion of correctly categorized data. The findings shown in Fig-
ure 5 demonstrate that the results for the sample and the cluster average using the hold-out approach
is 86.26% and 87.57% respectively, indicating that logistic regression has the lowest accuracy. In con-
trast, random forest produced the greatest results of the five models, with sample accuracy and clus-
ter mean accuracy of 97.4% and 97%, respectively, for the hold-out approach. Three out of the five
models had sample accuracy that is higher than the cluster average, which lends credence to the con-
clusion.
The disadvantage of the accuracy evaluation metric is that it does not explain how each class is classi-
fied or which class is classified most properly, it just provides the percentage of data that is correctly
classified. Therefore, some other metrics for evaluating the performance of the model that have been
used are precision, recall, and the F1 score.
The cluster average precision outcomes for each model are shown in Figure 6. With random forest
and support vector machine scores of 96.62% and 96.15% for the sample, and 96.56% and 96.63%
for the cluster average respectively, these two models demonstrate their efficacy in accurately identify-

97
Customer Churn Prediction

ing clients with churn inclinations when the percentage of actual churn to predicted churn is rela-
tively high. An accuracy of 83.08% for the sample and 84.86% for the cluster average is achieved by
the k-nearest neighbor, the least precise of the five models.
However, the k-nearest neighbors had a superior outcome with a recall of 99.15% for the sample and
98.81% for the cluster average in Figure 7, which shows the cluster average recall statistics. Addition-
ally, even if the outcome is only second with a recall of 97.94% and 97.45% corresponding to the
sample and cluster average, the random forest still demonstrates that it is a strong model for this data
set.
The cluster average F1 score findings, which are the harmonic mean of the model’s precision and re-
call, are displayed in Figure 8. The random forest performed the best on the dataset with a score of
97.43% and 97% matching the sample and cluster average, while the support vector machine came in
second with a score of 93.89% and 93.36%.
Table 5. Clusters results with hold out
No. Model Cluster Accuracy (%) Precision (%) Recall (%) F1 Score (%)
1 KNN I 89.68 83.78 98.41 90.51
II 89.8 84.08 98.51 90.72
III 92.83 88.89 98.63 93.51
IV 87.52 81.14 98.61 89.03
V 92.68 88.18 99.62 93.55
VI 89.6 83.08 99.1 90.38
2 LR I 88.69 88.24 89.29 88.76
II 90.81 90.72 90.49 90.6
III 84.59 85.17 85.17 85.17
IV 84.67 85.07 85.07 85.07
V 86.18 86.83 85.43 86.12
VI 90.49 90.06 91.62 90.83
3 DT I 92.66 93.68 91.86 92.76
II 93.45 92.54 94.42 93.47
III 91.76 89.57 93.61 91.54
IV 90.02 88.49 92.76 90.57
V 95.33 95.98 94.84 95.41
VI 93.76 92.25 96.37 94.26
4 RF I 95.24 94.47 95.98 95.22
II 97.48 97.18 97.68 97.43
III 98.57 98.56 98.56 98.56
IV 93.76 92.63 94.96 93.78
V 98.17 98.04 98.43 98.23
VI 98.81 98.46 99.07 98.77

98
 Tran, Le, & Nguyen

No. Model Cluster Accuracy (%) Precision (%) Recall (%) F1 Score (%)
5 SVM I 93.06 96.61 89.41 92.87
II 91.44 94.97 86.29 90.42
III 95.7 97.45 94.01 95.7
IV 91.44 96.33 85.82 90.77
V 93.9 97.02 90.84 93.83
VI 96.43 97.4 95.74 96.56

Figure 5. The sample and cluster average accuracy with hold out

99
Customer Churn Prediction

Figure 6. The sample and cluster average precision with hold out

100
 Tran, Le, & Nguyen

Figure 7. The sample and cluster average recall with hold out

101
Customer Churn Prediction

Figure 8. The sample and cluster average F1 score with hold out

The results from this research only go as far as testing, contrasting, and deciding which machine
learning technique is best for categorizing customer attrition and how the accuracy of churn predic-
tion is impacted by customer segmentation in the banking industry. Xiahou and Harada’s (2022) find-
ings demonstrate that each prediction index improved considerably following customer segmenta-
tion, demonstrating the necessity of k-means clustering segmentation. The research, however, only
examines the performance of LR and SVM and just on a certain dataset. As a result, it is still not
quite clear that segmentation enhances churn prediction. When comparing the outcomes of churn
forecasting before and after applying customer segmentation, Zhuang (2018) employed just one
model, XG-boost, and that study mainly concentrated on forecasting possible customer groups after
clustering. As a result, it cannot be conclusively shown that customer segmentation truly improves
the accuracy of churn prediction. Our study examined the effects of customer segmentation on
churn prediction using a variety of classification techniques. The results show that the proportion of
models with lower churn prediction results after using segmentation with evaluation metrics of accu-
racy, precision, recall, and F1-score are 3:5, 2:5, 4:5, and 3:5. Thus, it is clear that the churn prediction
result depends on the dataset, the model’s tuning, and the model’s intended usage.

102
 Tran, Le, & Nguyen

CONCLUSION
This paper reviewed the theoretical basis of customer churn, and customer segmentation, and sug-
gested the use of supervised machine-learning techniques for customer attrition prediction. The ex-
perimental findings indicate that the two best training methods are random forest and support vector
machines and discovered that customer segmentation has no impact on the ability to anticipate cus-
tomer attrition. This study serves as a benchmark for the implementation of customer churn predic-
tion in industries including banking, e-commerce, telecommunications, and insurance. However,
there are still limitations in this study that may be continued or studied in the future.
First, this study exclusively uses the Kaggle website to collect data for its analysis. By linking to a local
bank and assisting them in analyzing their customer attrition, research may also extend to gathering
the client’s transaction history or customer information.
Second, the study is restricted to determining whether or not a consumer quits the business. The
next study will focus on determining which feature is most responsible for customer attrition.
Finally, this research simply looks at how attrition is classified and how customer segmentation influ-
ences that outcome. The increased research approach will concentrate on developing an application
to help a financial company, such as a bank, keep its customers.

ACKNOWLEDGMENT
This research is funded by University of Economics and Law, Vietnam National University Ho Chi
Minh City.

REFERENCES
Alexandropoulos, S. A. N., Kotsiantis, S. B., & Vrahatis, M. N. (2019). Data preprocessing in predictive data
mining. The Knowledge Engineering Review, 34, 1–33. https://doi.org/10.1017/S026988891800036X
Baran, R. J., & Galka, R. J. (2017). Customer relationship management: The foundation of contemporary marketing strategy.
Routledge.
Biau, G., & Scornet, E. (2016). A random forest guided tour. TEST, 25(2), 197–227.
https://doi.org/10.1007/s11749-016-0481-7
Cunningham, P., & Delany, S. J. (2021). k-Nearest neighbour classifiers – A tutorial. ACM Computing Surveys,
54(6), 1–25. https://doi.org/10.1145/3459665
Dawes, J., & Swailes, S. (1999). Retention sans frontieres: Issues for financial service retailers. International Journal
of Bank Marketing, 17(1), 36–43. https://doi.org/10.1108/02652329910254037
Dolatabadi, S. H., & Keynia, F. (2017, July). Designing of customer and employee churn prediction model
based on data mining method and neural predictor. Proceedings of the 2nd International Conference on Computer
and Communication Systems, Krakow, Poland, 74–77. https://doi.org/10.1109/CCOMS.2017.8075270
Elyusufi, Y., & Ait Kbir, M. (2022). Churn prediction analysis by combining machine learning algorithms and
best features exploration. International Journal of Advanced Computer Science and Applications, 13(7), 615–622.
https://doi.org/10.14569/IJACSA.2022.0130773
Guliyev, H., & Tatoğlu, F. Y. (2021). Customer churn analysis in banking sector: Evidence from explainable ma-
chine learning models. Journal of Applied Microeconometrics, 1(2), 85–99.
https://doi.org/10.53753/jame.1.2.03
Hameed, M., & Naumann, F. (2020). Data preparation: A survey of commercial tools. ACM SIGMOD Record,
49(3), 18–29. https://doi.org/10.1145/3444831.3444835
Jakkula, V. (2006). Tutorial on support vector machine (SVM). School of EECS, Washington State University.
https://course.ccs.neu.edu/cs5100f11/resources/jakkula.pdf

103
Customer Churn Prediction

Kanungo, T., Mount, D. M., Netanyahu, N. S., Piatko, C., Silverman, R., & Wu, A. Y. (2000). The analysis of a
simple k-means clustering algorithm. Proceedings of the 16th Annual Symposium on Computational Geometry,
100–109. https://doi.org/10.1145/336154.336189
Karvana, K. G. M., Yazid, S., Syalim, A., & Mursanto, P. (2019, October). Customer churn analysis and predic-
tion using data mining models in banking industry. Proceedings of the International Workshop on Big Data and
Information Security, Bali, Indonesia, 33–38. https://doi.org/10.1109/IWBIS.2019.8935884
Kaur, I., & Kaur, J. (2020, November). Customer churn analysis and prediction in banking industry using ma-
chine learning. Proceedings of the Sixth International Conference on Parallel, Distributed and Grid Computing, Wak-
naghat, India, 434–437. https://doi.org/10.1109/PDGC50313.2020.9315761
Muneer, A., Ali, R. F., Alghamdi, A., Taib, S. M., Almaghthawi, A., & Abdullah Ghaleb, E. A. (2022). Predicting
customers churning in banking industry: A machine learning approach. Indonesian Journal of Electrical Engi-
neering and Computer Science, 26(1), 539–549. https://doi.org/10.11591/ijeecs.v26.i1.pp539-549
Olaniyi, A. S., Olaolu, A. M., Jimada-Ojuolape, B., & Kayode, S. Y. (2020). Customer churn prediction in bank-
ing industry using k-means and support vector machine algorithms. International Journal of Multidisciplinary
Sciences and Advanced Technology, 1(1), 48–54.
Pradana, M. G., & Ha, H. T. (2021). Maximizing strategy improvement in mall customer segmentation using k-
means clustering. Journal of Applied Data Sciences, 2(1), 19–25. https://doi.org/10.47738/jads.v2i1.18
Quinlan, J. R. (1990). Decision trees and decision-making. IEEE Transactions on Systems, Man, and Cybernetics,
20(2), 339–346. https://doi.org/10.1109/21.52545
Rahman, M., & Kumar, V. (2020, November). Machine learning based customer churn prediction in banking.
Proceedings of the 4th International Conference on Electronics, Communication and Aerospace Technology, Coimbatore,
India, 1196–1201. https://doi.org/10.1109/ICECA49313.2020.9297529
Reddy, V. N., & Reddy, P. S. S. (2021). Market basket analysis using machine learning algorithms. International
Research Journal of Engineering and Technology, 8(7), 2570–2572.
Reichheld, F. F. (1996). The loyalty effect: The hidden force behind growth, profits, and lasting value. Harvard Business
School Press.
Routray, S. K. (2021). Marketing strategy through machine learning techniques: A case study at telecom indus-
try. International Journal of Innovation Engineering and Science Research, 5(3), 21–30.
Sivasankar, E., & Vijaya, J. (2017). Customer segmentation by various clustering approaches and building an
effective hybrid learning system on churn prediction dataset. In H. Behera, & D. Mohapatra (Eds.), Compu-
tational intelligence in data mining (pp. 181–191). Springer. https://doi.org/10.1007/978-981-10-3874-7_18
Sperandei, S. (2014). Understanding logistic regression analysis. Biochemia Medica, 24(1), 12–18.
https://doi.org/10.11613/BM.2014.003
Xiahou, X., & Harada, Y. (2022). B2C e-commerce customer churn prediction based on k-means and SVM.
Journal of Theoretical and Applied Electronic Commerce Research, 17(2), 458–475.
https://doi.org/10.3390/jtaer17020024
Yaseen, A. (2021). Next-wave of e-commerce: Mobile customers churn prediction using machine learning. La-
hore Garrison University Research Journal of Computer Science and Information Technology, 5(2), 62–72.
https://doi.org/10.54692/lgurjcsit.2021.0502209
Zhang, T., Moro, S., & Ramos, R. F. (2022). A data-driven approach to improve customer churn prediction
based on telecom customer segmentation. Future Internet, 14(3), 94. https://doi.org/10.3390/fi14030094
Zhuang, Y. (2018). Research on e-commerce customer churn prediction based on improved value model and
XG-boost algorithm. Management Science and Engineering, 12(3), 51-56. https://doi.org/10.3968/10816

104
 Tran, Le, & Nguyen

AUTHORS
Hoang Tran is currently a sophomore student majoring in E-commerce
at the Faculty of Information Systems, University of Economics and
Law, Vietnam National University, Ho Chi Minh City, Vietnam. His cur-
rent research interests include business analytics, business intelligence,
data analytics, and machine learning. He can be contacted by email at ho-
angtd21411c@st.uel.edu.vn

Ngoc Le is currently a sophomore at the Faculty of Economic Mathe-

matics, University of Economics and Law, Vietnam National University,
Ho Chi Minh City, Vietnam. Her current research interests include busi-
ness analytics, data analytics, and machine learning. You can contact her
via email at ngocltl21413c@st.uel.edu.vn

Van-Ho Nguyen received a B.S. degree in Management Information

System (MIS) from the Faculty of Information Systems, University of
Economics and Law (VNU–HCM), Vietnam in 2015, and a Master’s de-
gree in MIS from the University of Economics Ho Chi Minh City, Vi-
etnam in 2020. He is currently a lecturer at the Faculty of Information
Systems, University of Economics and Law, Vietnam National University,
Ho Chi Minh City, Vietnam. His research was published in international
journals and conferences, such as the Journal of Information Processing Sys-
tems, Business Research Systems, and Journal of Hospitality and Tourism Technol-
ogy, 9th NAFOSTED Conference on Information and Computer Science in 2022,
and the 4th International Conference on Business (ICB) in 2021. He is currently
a reviewer for the Journal of Marketing Analytics and Business Research Systems.
His current research interests include business analytics, business intelli-
gence, data analytics, and machine learning. He can be contacted by email
at honv@uel.edu.vn or nguyenvanho.uel@gmail.com

105

View publication stats