Indonesian Journal of Electrical Engineering and Informatics (IJEEI)

Vol. 10, No. 2, June 2022, pp. 431~440

ISSN: 2089-3272, DOI: 10.52549/ijeei.v10i2.2985  431

Customer Churn Prediction in Telecommunication Industry

Using Classification and Regression Trees and Artificial Neural
Network Algorithms
Sulaiman Olaniyi Abdulsalam1, Micheal Olaolu Arowolo2,3 Yakub Kayode Saheed4, Jesutofunmi
Onaope Afolayan5
1
Department of Computer Science, Kwara State University, Malete, Nigeria.
2,5
Department of Computer Science, Landmark University, Omuaran, Nigeria.
3
Landmark University SDG 9 (Industry, Innovation and Infrastructure).
4
School of Information Technology & Computing, American University of Nigeria.

Article Info ABSTRACT

Article history: Customer churn is a serious problem, which is a critical issue encountered by
large businesses and organizations. Due to the direct impact on the company's
Received Jan 25, 2021 revenues, particularly in sectors such as the telecommunications as well as the
Revised Mar 2, 2022 banking, companies are working to promote ways to identify the churn of
Accepted Mar 11, 2022 prospective consumers. Hence it is vital to investigate issues that influence
customer churn to yield appropriate measures to diminish churn. The major
objective of this work is to advance a model of churn prediction that helps
Keyword: telecom operatives to envisage clients that are most probable to be subjected
to churn. The experimental approach for this study uses the machine learning
Telecoms procedures on the telecom churn dataset, using an improved Relief-F feature
Relief-F selection algorithm to pick related features from the huge dataset. To quantify
ANN the model's performance, the result of classification uses CART and ANN, the
CART accuracy shows that ANN has a high predictive capacity of 93.88% compared
Churn to the 91.60% CART classifier.
Copyright © 2022 Institute of Advanced Engineering and Science.
All rights reserved.

Corresponding Author:
Micheal Olaolu Arowolo
Department of Computer Science,
Landmark University, Omuaran Nigeria.
Email: [email protected]

1. INTRODUCTION
Data-driven sectors have been able to carry out analysis of data and fetch out extensive knowledge
through technology advances. Methods of data mining has helped in achieving the prediction of certain future
customer behaviors [1]. Customer churn, is categorized as customer attrition, it is amongst the most critical
issues that reduces a company's profit. The procedures of business intelligence for locating customers who
wants to change from a company to its other competitor can be described as customer churn [2]. The telecoms
industry is a highly technological industry that has grown enormously in the past, as a consequence of the
emergence and commercial success of both mobile telecommunications, these two decades [3-4].
For many telecoms’ firms, customer churn or customer attrition is a major problem, it occurs when a
customer terminates his subscription and moves to another rival. There are several variables that impact the
decision of the client to turn to another rival. In general, these variables were related to the high cost, bad jobs,
fraud and privacy issues related to customer service[5]. Customer turnover causes significant loss of profit
when those thresholds are surpassed. Companies know that gaining fresh clients can be costly than retaining
old ones [6].

Journal homepage: http://section.iaesonline.com/index.php/IJEEI/index

432  ISSN: 2089-3272

In several sectors, such as telecom providers, credit cards, Internet service providers, e-commerce,
newspaper publishing firms, banking sectors, among others, Consumer Churn Prediction (CCP) has been raised
as a key concern in telecommunication firms [7].
In recent years, Consumer Churn Prediction has become an increasingly common research problem
and therefore, telecom suppliers have commonly used strategies to classify potential churn customers based on
their historical records, previous behaviors and offering some services to convince them to live[8]. Long-term
customers, on the other hand, are more lucrative for service providers because they are more focused on
purchasing additional goods and spreading the satisfaction of the customer within their radius, thereby
attracting more and more customers indirectly [9].
Businesses must have a thorough knowledge of why churn emerges to retain their clients. There are
certain factors to be discussed, such as organization discontent, certain businesses' competitive costs, customer
migration, the necessity for better services for clients that can encourage users switching to their present service
provider and moving to a different one [4]. Companies however, understand that winning new customers is a
great deal More costly than current ones being retained [10].
In general, churn prediction obtained data are imbalanced, instances in non-churner customer may
outstrip churners class instances. Typical classification techniques seem to achieve relevant accuracy results
for huge classes and miss smaller ones, this is regarded one of the most challenging and significant issues.
Different methods have been suggested to handle the issue of imbalanced churn prediction data. These
techniques comprises of sufficiently common evaluation metrices, use of cost-sensitive learning, modification
of training set distributions by method sampling, the use of approaches to reduce dimensionality, among others
[11][12].
Reducing dimensionality is of essence in data mining, it is motivated by developing feature
dimensionality in specified concerns and increasing interest in innovative yet costly computational
methodologies capable of modeling complex associations. Feature Selection is one of the methods of
preprocessing to classify the data sub-set from large-dimensional data. In particular, feature selection
techniques such as Relief-F, Genetic Algorithm, among others, are computationally efficient, but responsive
to complex association patterns, such as associations, so that prior to downstream modeling, informative
features are not mistakenly excluded [13]. Relief-F-based algorithms, a distinct category of filter-based feature
selection algorithms which have received attention by achieving an efficient balance among these goals
optimally adapting to different data features [14][15][16].
In attempt to fine-tune developed models, recent investigations for churn analysis have proposed and
suggested that methods such as SVM, ANN, CART, among others classification methods are the most
commonly utilized. Numerous optimization techniques and strategies have been explored and recommended
to have been identified to operate the best and make studies in several fields such as telecommunications
companies, banking, business and insurance, among others improved in productivity of these sectors [17][18].
In this study, the major contribution is adopting an enhanced Relief-F feature selection algorithm, is
created with an innovative learning method by using the subsets of the relevant churn prediction method data
based on CART and ANN classifiers using optimal predictors that increase the predictive output of variables.
Using diverse evaluation metrics, and related to traditional prediction methods as well as other relevant
processes presented in the literature, the presented methodology will be evaluated.

2. LITERATURE REVIEW
Classification of churners and non-churners are considered to be a predominant problem for telecom
providers, it is characterized as losing clients as they flee for contenders. To be able to pre-classify customer
churn offers the telecom business an appreciated insight into retaining its customer base. In recent years, large
ranges of churn classification methods have been explored. Most creative models uses advanced machine
learning classifiers and have found that the roots of customer churn are reviewed in relation to service quality,
customer satisfaction/dissatisfaction and economic value variables.
Prediction investigation utilizing Random Forest with discriminating features method of analysis for
prediction of churners in the telecommunications industry was proposed [3], grounded on developmental
search Random Forest, they predicted churners and non-churners in the telecommunication sector that use
discriminant feature investigation as innovation postponement of the traditional Random Forest to learn tilted
Developmental Detection tree. The suggested approach controls the benefit of two methods of discriminant
investigation to measure the project index used in PPtree construction. they used Support Vector Machines
with Linear Discriminant Analysis to obtain linear division of variables and developed specific classifiers that
are stronger and more flexible than traditional Random Forest in oblique PPtree development. The detection
techniques are proven to outperform in terms of Accuracy. The prediction model, PPForest based on LDA
delivers efficient evaluators.

IJEEI, Vol. 10, No. 2, June 2022: 431 – 440

IJEEI ISSN: 2089-3272  433

A comparative investigation of customer churn prediction by means of Negative Correlation Learning

have been suggested [4], by utilizing an ensemble based Multilayer Perceptrons. Training for predicting
consumer churn in a telecom market is gained using negative correlation learning. The test findings suggested
that the NCL-MLP-ensemble can improve overall classification efficiency (high churn rate) compared to the
non-NCL-MLP-ensemble and other traditional data mining strategies utilized in churn study.
Investigating synthetic and ensemble approaches in the telecommunications industry for users churn
prediction system was suggested [19], by presenting a detailed study on churn-based machine learning
prediction of the telecommunications industry for 8 years. The issues and problems in telecommunication churn
were were observed with eliminated predictions and the suggestions and solutions published. The study and
overview enable researchers or data experts in telecommunication fields to fetch optimum and suitable
techniques with design methods to improve novel models for future churn prediction.
A findings of consumers’ purchase decision-making using churn prediction support framework was
carried out [20], their investigation showed a wide range of characteristics used to generate consumer churn
model by many scholars. It demonstrates specific methodologies used in churn prediction up to date. Methods
of modeling, for instance; Neural Network, Logistic Regression, Decision Tree, Random Forest, Support
Vector Machine among other methods are employed for churn discovery. The discoveries indicate that through
predictive analytics, the customer churn forecast can achieve more precise results as compared with other
related prediction approaches. In Customer churn prediction and customer retention by predictive analytics,
there is a wide scope of research.
Churn analysis for telecommunications Sector utilizing Decision Tree was proposed [21]. Decision
Tree classification method used a large client dataset evaluated for churn. After the implementation of all
possible decision tree variants in SPSS, it was noted that Exhaustive CHAID method demonstrated to be further
consistent and reliable to envisage the likely customers to churn rather than impending ones.
Telecom consumer analysis churn prediction with machine learning in large amount of data was
proposed [22], they constructed a new approach to develop and choose features. The Area Under Curve sole
criterion is utilized to determine the model's efficiency, and the obtained AUC value is 93%. The use of
customer communal networks in the prediction model by mining features of social network study is another
key contribution. The use of system network study increased the model's output against the AUC benchmark
from 84 to 93%. Via Spark setting, the model was systematized and verified by employing a huge dataset
generated by transforming large raw information obtained from SyriaTel telecommunication corporation. The
database comprises certain customer data across a span of nine months which was used by SyriaTel to train,
test and assess the classification. Four algorithms were experimented with in the methodology: Decision Tree,
Random Forest, Gradient Boost Tree and Extreme Gradient Boost. The better outcomes, therefore, are obtained
by implementing the XGBOOST algorithm. In this churn-predictive framework, this algorithm was used for
classification.
An innovative attribute selection approach and framework telecommunications churn evaluation
correlation was proposed [23], the information used for the analysis contains information of actual customer
phone metadata collected from a large telecom industry in Turkey for the years 2013 and 2014. In addition to
an overall attribute selection procedure, for the development of five datasets, four various methodologies called
R-correlation coefficient-based feature selection, ¥-correlation coefficient-based feature selection, Relief-F,
and Gain Ratio were utilised. Four classifier algorithms were subsequently implemented, with Random Forest,
Decision Tree, Naive Bayes and AdaBoost. Evaluation criteria consisting Accuracy, Sensitivity, Specificity,
F-score, and run-time were used to evaluate the results obtained. The findings of the correlations indicate that
on consumer churn estimation, the projected feature selection algorithm outdoes the state-of-the-art
approaches.
Telecom market consumer churn prediction analysis using the CART algorithm was proposed [24],
The estimation of customer attrition in the telecommunications industry has been the most relevant subject for
research in recent years. Since it helps detect which client is likely to change or cancel their service
subscription. Review of data collected from telecommunications providers can help find the consumer churn
reasons and also use the information to attract customers. Thus, for telecommunication companies to maintain
their customers, predicting churn is extremely important. This study developed the call tree classification
model, evaluated the output indicators, and compared its performance with the model of logistic regression.

3. RESEARCH METHOD
The goal of the proposed study is to construct a classification model to indicate that the customer in
Telecom datasets is a likely churner or non-churner. By implementing the key retention policies that are likely
to retain and attract consumers who have the most propensity to churner and pursue them to stay, this procedure
would aid customer relationship management. The feedback for suggesting the customer churn prediction

Customer Churn Prediction in Telecommunication Industry…. (Sulaiman Olaniyi Abdulsalam et al)

434  ISSN: 2089-3272

model includes information for each mobile subscriber from past calls, along with all the person and business
information held by the provider of telecom services. Fully trained with the training dataset after the prediction
model, the test dataset and the model have to be able to predict churners. Figure 1 shows the technique for the
prediction of churners and the description of the steps proposed.

Figure 1. Customer churn prediction Approach using Relief-F with CART and ANN models

Machine learning is a method of understanding strategies from big data to find useful knowledge. To
obtain and analyze beneficial information from various huge datasets, it uses analytical tools, arithmetic,
artificial intelligence, and data science, it presents it for advanced, valuable knowledge and information.
Machine learning can solve problems relating to data learning theory of classification, regression, clustering,
and correlation depending on the intent of research. The pattern of data is In this method, descriptively and
intelligently presented.

3.1. Datasets
Telecom datasets produced by Telecom Industry operators collected from the Francisco gallery of
bigml.com are the realistic part of this analysis, it comprises of 20 attributes and 3333 instances. A dataset
pertaining to functionality and use of telephony account features and whether or not the customer has churned
[25]. The main characteristics of the dataset attributes comprises of; name, account length, zone code, global
plan, voicemail, number vmail messages, entire day minutes, entire day calls, entire day charge, entire eve
minutes, churn, among others [25].

3.2. Feature Selection based on enhanced Relief-F

Identifying attributes that are definitely applicable to the target variable is the most critical step in data
pre-processing. Not all features, nevertheless, are well-contributed to the classifier learner model. The feature
selection method became important to improve efficiency and make the customer churn prediction model easier
to interpret, minimize overfitting, remove variables that are redundant and do not provide any information or
contribution to the model's production because of the wide-scale datasets in telecom provider services. In
addition, it decreases the size of the prediction problem and allows classification algorithms to generate results
as quickly as possible [3].
Instance-based learning inspired the initial Relief algorithm. Relief computes a deputation statistic for
respective feature that can be utilized to approximate the eminence or significance of the feature to the target
definition as a distinct assessment filtering feature selection process. Such feature information is represented
as feature weights (feature weight 'A' = W[A]), or informally as 'scores' features that can vary from -1 (worst)
to +1 (best). Particularly, the unique Relief algorithm was restricted to problems with binary classification and
had no method for handling misplaced data. Approaches to extend relief to issues with multi-class or
continuous endpoints are required [13]. The Pseudocode for the traditional Relief-F Algorithm emphasized the
sequences for training selected instances with no substitutes with user-defined parameters [13]. Relief-F have
proven to be the best known variant and most utilized, it relies on number of neighbors, increasing weight
estimate reliability and noisy problems, it can handle missing data values, and handle multi-class endpoints
[26].

Algorithm1: Pseudocode for the enhanced Relief-F Algorithm

n = numbers of the trained instances
a = feature numbers (attributes)
m = parameter numbers random training instances out of n used to update W
c = constant

set all feature weights W[A] = 0.0

for i: = 1 to m do
arbitrarily pick a ‘target’ instance Ri
discover an adjoining hit ‘H’ and adjoining miss ‘M’ (instances)
for A: = 1 to a do

IJEEI, Vol. 10, No. 2, June 2022: 431 – 440

IJEEI ISSN: 2089-3272  435

W[A] = W[A] di ff (A, Ri, H)/m + di ff (A, Ri, M)/m

𝒇𝒇(𝑨,𝒓𝒊,𝒉𝒋)
W[A] = W[A]∑𝒌𝒋=𝟏 𝒅𝒊 + O (n2.a) (m+c32c.n. a)
𝒎.𝒄
end for
end for
return vector W of feature scores that estimate the feature values

The Relief computes the ratings of features based on modifications in feature and class values among
neighbor instances. If a set of neighbor instances has positive variations for a feature but the same class value,
then ReliefF reduces the score of that feature. Additionally, ReliefF improves the score of the function if
adjacent instances have positive variations for a feature and different class values. For a set of experimented
instances and their nearest neighbours, this is repeated to determine an average score for each characteristic
[16][27]. In this study an enhanced Relief-F for fetching the missed fits and best fits is suggested to fetch
relevant information from the churned telecom dataset. The results of the relief-f fetch a relevant subset of the
data and it is used as a reduced preprocessed data for classification.

3.3. Classification based on CART (Classification and Regression Tree)

For the continuous dependent variable and categorical predictor variable, the CART approach is more
appropriate. CART recursively divides the function space into non-overlapping areas. A classification tree is
generated to predict the value of a dependent categorical variable. Incorporates CART to determine the
goodness of fit more reliably, including checking with a reference data set and cross-validation. In distinct
sections of the tree, CART can use the same variables more than once. This ability can reveal complex
interdependencies of variables between sets. CART may be used to pick the input set of variables in
combination with other prediction techniques [21].
Pruning is performed after the CART algorithm is qualified. As the basis for pruning, the total error
rate is used. The smallest tree provides the most effective classification (trees with least number of layers). For
a target variable with constant and definite data, the CART algorithm is applicable. If continuous data is
represented by the goal variable, then the regression tree can be used. A classification tree may be used if the
target variable includes unconditional data [28]. An active threshold value is calculated as a state for each node
in the CART algorithm. At each node, a sole input variable function splits the data and constructs a binary DT.
To estimate the metrics, the Gini index is used. The presence of several groups in the data is demonstrated by
a high level of dispersed indicators. By comparison, the presence of a single group is suggested by a low level
of indicators.

3.4. Classification based on ANN

The Neural Networks Model is used to build features such as non-linear features. Due to its
comparable data processing system, the model retains the capacity to learn. After applying several concerns,
such as grouping, these methods provide good results. Due to its possible range estimation, the model is
dissimilar to the classification model and decision tree. Has the neural network multiple approaches of merits
and demerits. The investigator claims that the deep neural network is stronger than the churn prediction model
of decision tree and regression analysis [29].
Neural Networks is a methodology for datamining that has the capacity to learn from mistakes. The
brain stimulates Neural Networks. This happens in the sense that a few new things are learned by the brain,
which will then be transmitted through neurons. The neural network neuron may also learn from training data
with learning algorithms; this makes them denoted to as Artificial Neural Neurons [30].
It is possible to distinguish neural networks into single-layer perception and multilayer perception
(MLP) networks. The perception of multiple layers comprises of multiple layers of plain, two-state, sigmoid
transfer mechanism with processing element or neurons that communicate using weighted links. In reality, the
neural network involves one or more intermediaries’ secret layers of neurons in between the layers of input
and output.
These intermediate layers are known as hidden layers, and nodes embedded in these layers are known
as hidden nodes because they do not take inputs directly from outside [31].

3.5. Evaluation Criteria and Experimantal Setup

In this study, modifications of analytical data have been implemented in MATLAB data mining tool
[32][33]. In order to find a definitive decision, effective functional set-up and use of study variables and
effective performance metrics are important. To determine the efficiency of the churn prediction model, the
telecom industry represents various methods of performance measures [34].

Customer Churn Prediction in Telecommunication Industry…. (Sulaiman Olaniyi Abdulsalam et al)

436  ISSN: 2089-3272

Accuracy: Calculate the right predictions made over all sorts of predictions made by the prediction model. In
general, how frequently is the classifier model exact.
Accuracy = TP+TN/TP+TN+FP+FN

Precision: The number of confirmed samples that have been identified properly.
Precision = TP/TP+FP

Sensitivity: The amount of real positive instances that have been identified right.
Sensitivity = TP/TP+TN+FP+FN

Specificity: The number of real negative instances accepted appropriately.

Specificity = FP/FP+FN

F-Score: Precision is vital for evaluating the efficiency of datamining classifiers, but it definitely leaves out
details and will also be complicated for that purpose. The Recall is a part of the true optimistic predictions in
the dataset for overall positive observations. Calculate the proportion of the churn rate that is correctly labeled
as churn/non-churn. The low-recall prediction models indicate that a significant number of positive cases are
miss-classified.
F-Score = 2 X precision* sensitivity/ precision+sensitivity

4. RESULTS AND DISCUSSION

In order to create a friendly user experience, this study was carried out and built using Matlab
programming (MATLAB 2016A) with connected components of the MATLAB graphical user interface
framework. In order to develop an output result of the data mining task with data filtering, feature selection
using Relief-F, classification using ANN and CART, and performance evaluation, the built systems used
different component environments in Matlab. Figure 2 shows the user interface and the loaded telecom dataset
used in this syudy, 3333 samples with 21 attributes were loaded, Relief-F ranking was used as a feature
selection technique to select relevant information from the data and passed into ANN and CART classifier
separately.

Figure 2. Loaded Telecom Customer Churn Data

For the input data matrix and response vector, the relief-f computes ranks and weights of attributes
(predictors), the Releif-f filter selection method was able to identify the predicting variables according to their
respective weight score with respect to the class mark. The characteristics on the positive response variable
scale were chosen as the characteristics, totaling fourteen characteristics. Figure 3 shows the selected features
using Relief-F algorithm, 14 features were selected from the given data as a subset dataset.

IJEEI, Vol. 10, No. 2, June 2022: 431 – 440

IJEEI ISSN: 2089-3272  437

Figure 3. Selected Features Using Relief-F Algorithm

The selected data were passed to the training and testing set, the data was splitted into the training set
and the data set was tested. For both the ANN and the CART classification algorithm, the system used 75% of
the data for training. The loaded class mark indicates the split rate set at 0.25, which is an indicator of the data
for both algorithms being 25% kept out.

4.1. ANN Training Approach

The 14 inputs, 10 neurons in the hidden layer, 1 neuron in the output layer with an activation role with
1 output were used for artificial neural network architecture. The 14 inputs reflect the churn dataset input data
provided to the ANN with adjustable weight and bias (W,b), with 10 neurons the hidden layer was processed
while the output layer was processed with one neuron in order to predict a single churners or non-churners
outcome. 42.2313Sec the training computational time was used in processing the ANN for training the dataset,
it is measured in terms of the total seconds used for executing the training process.
Based on each classification algorithm, as well as the comparative assessment of the two algorithms,
the experimental results are mentioned. The evaluation parameter displays the KNN and SVM classifier
combination results. With the True Positive rate (TP), False Positive (FP), True Negative (TN) and False
Negative (FN), accuracy and error rate as shown in Table 1, the research (probing) assessment was achieved.
Using False Acceptance Rate (FAR), False Rejection Rate (FRR), Accuracy (Recognition Rate), and Error
Rate, the assessment parameters for classification rate were achieved.

Table 1. ANN Analysis per class

Analysis per class. True Positive True False False
Negative Positive Negative
Class 1 88 694 18 33
Class 2 694 88 33 18

The confusion matrix is used as a description of the prediction results of this study on a classification
issue. The number of correct and incorrect predictions is summarized and broken down by each class by
counting values. Class 1 is true, which is the consumer who is likely to churn, while class 2 is false, which is
the class of non-churners. Class 1 gives a total of 121 out of the test observation set, a total of 88 were correctly
classified and 33 were misclassified, while the class of non-churners described by mark 2 gives a total of 712
out of the test observation set, a total of 694 were correctly classified and 18 were misclassified. Table 2 shows
the Confusion matric used in ANN, with 88 =TP, 694=TN, FP= 18 and FN=33.

Table 2. ANN Confusion Matrix

ANN Confusion Matrix
1 2
1 88 33
2 18 694

4.2. CART Training Approach

The CART analysis per each class based on the churners and non-churners class is shown in the table
3.

Customer Churn Prediction in Telecommunication Industry…. (Sulaiman Olaniyi Abdulsalam et al)

438  ISSN: 2089-3272

Table 3. CART Analysis per class.

Analysis per class. True Positive True False False
Negative Positive Negative
Class 1 90 673 39 31
Class 2 673 90 31 39

To summarize the prediction outcomes on a classification problem, the Confusion matrix is used. The
number of correct and incorrect predictions is summarized and broken down by each class by counting values.
Class 1 is true, which is the consumer who is likely to churn, while class 2 is false, which is the class of non-
churners. Class 1 gives a total of 121 out of the test observation set, a total of 90 were correctly classified and
31 were misclassified, while the class of non-churners described by mark 2 gives a total of 712 out of the test
observation set, a total of 673 were correctly classified and 39 were misclassified. Table 4 shows the CART
confusion matrix wherr TP=90, TN=673, FP=39 and FN=31. The real computing time used in the processing
of the CART for the dataset training is taken and 7,811 seconds are used, which is calculated in terms of the
cumulative usage time of seconds for the training phase to be performed.

Table 4. CART Confusion Matrix

ANN Confusion Matrix
1 2
1 90 31
2 39 673

Table 5 shows a comparative result of the evaluation performance metrics for the classification of
telcom churn prediction using ANN and CART classifier. The comparative results for the Artificial Neural
Network and CART are shown in Table 5, which indicates that the ANN classification algorithm exceeded the
CART classification algorithm for the telecom churn dataset, as shown in the table, as it gives a higher
classification accuracy of 93.88% compared to 91.6% of the CART.

Table 5. Performance Metrics for ANN and CART

Performance Metrics (%) ANN (%) CART (%)
Accuracy 93.88 91.6
Sensitivity 72.7 74.38
Specificity 97.47 94.52
Precision 83.02 69.77
F-Score 77.53 72
Matthews Correlation 74.22 67.11
Coefficient

In this study a feature selection approach using Relief-F was used to select relevant features from a
huge churn telecom dataset, the relevant features were classified using ANN and CART, however the results
of the classification show that ANN outperformed CART approach and suggested that this approach is an
efficient one for this study compared with other existing works from literature, Table 6 compares the work
with existing works.

Table 6. Comparison with Existing Works

Authors and Years Work Done Results (%)
Khalid et al., 2021 [35] FE + Random 91
Forest
Saini et al, 2017 [36] CHAID + DT 91
Ahmad et al, 2019 [37] X-Boost 89

The comparative analysis using uncertainty matrix research was conducted between Relief-F-ANN
and Relief-F-CART. In order to verify the achievement, the assessment highlighted the accuracy
About R-F-ANN. Finally, device architecture that adopted MATLAB execution was then protected
by the RF-ANN prediction procedures mechanism. In order to provide a better overview of telecommunications
decision-making activities, the R-F-ANN prediction method was developed for data mining.

5. CONCLUSION
This research applied to the selection algorithm of a Relief-F function with ANN and CART classifiers
on telecom customer churn prediction results. The issue of customer churn prediction is simultaneously
important and difficult. In order to assist them in developing successful customer retention strategies,

IJEEI, Vol. 10, No. 2, June 2022: 431 – 440

IJEEI ISSN: 2089-3272  439

telecommunications companies invest more in creating accurate churn prediction model. An analysis of the
application of Relief-F with ANN and CART was tested in this study and trained to predict customer churn in
a telecommunications business. Experimental findings confirm that, compared to Relief-F-CART machine
learning models, Relief-F-ANN achieves better generalization efficiency in terms of churn rate prediction with
a highly reasonable precision rate.

REFERENCES
[1] U. Sivarajah, M. M. Kamal, Z. Irani, and V. Weerakkody, “Critical analysis of Big Data challenges and analytical
methods,” J. Bus. Res., vol. 70, pp. 263–286, Jan. 2017, doi: 10.1016/j.jbusres.2016.08.001.
[2] B. He, Y. Shi, Q. Wan, and X. Zhao, “Prediction of Customer Attrition of Commercial Banks based on SVM Model,”
Procedia Comput. Sci., vol. 31, pp. 423–430, 2014, doi: 10.1016/j.procs.2014.05.286.
[3] A. M. Naser alzubaidi and E. S. Al-Shamery, “Projection pursuit random forest using discriminant feature analysis
model for churners prediction in telecom industry,” Int. J. Electr. Comput. Eng., vol. 10, no. 2, p. 1406, Apr. 2020,
doi: 10.11591/ijece.v10i2.pp1406-1421.
[4] A. Rodan, A. Fayyoumi, H. Faris, J. Alsakran, and O. Al-Kadi, “Negative Correlation Learning for Customer Churn
Prediction: A Comparison Study,” Sci. World J., vol. 2015, pp. 1–7, 2015, doi: 10.1155/2015/473283.
[5] K. O. Kadiri and S. O. Lawal, “Comparative Analysis of Per Second Billing System of GLO, MTN, Etisalat, Airtel
and Visafone in Nigeria,” Curr. J. Appl. Sci. Technol., pp. 1–8, Apr. 2019, doi: 10.9734/cjast/2019/v34i230125.
[6] P. K. Banda and S. Tembo, “Factors Leading to Mobile Telecommunications Customer Churn in Zambia,” Int. J.
Eng. Res. Africa, vol. 31, pp. 143–154, Jul. 2017, doi: 10.4028/www.scientific.net/JERA.31.143.
[7] M. Singh, S. Singh, N. Seen, S. Kaushal, and H. Kumar, “Comparison of learning techniques for prediction of
customer churn in telecommunication,” in 2018 28th International Telecommunication Networks and Applications
Conference (ITNAC), Nov. 2018, pp. 1–5, doi: 10.1109/ATNAC.2018.8615326.
[8] K. Kim, C.-H. Jun, and J. Lee, “Improved churn prediction in telecommunication industry by analyzing a large
network,” Expert Syst. Appl., vol. 41, no. 15, pp. 6575–6584, Nov. 2014, doi: 10.1016/j.eswa.2014.05.014.
[9] A. Keramati and S. M. S. Ardabili, “Churn analysis for an Iranian mobile operator,” Telecomm. Policy, vol. 35, no.
4, pp. 344–356, May 2011, doi: 10.1016/j.telpol.2011.02.009.
[10] T. Hennig-Thurau and U. Hansen, Eds., Relationship Marketing. Berlin, Heidelberg: Springer Berlin Heidelberg,
2000.
[11] M. Galar, A. Fernandez, E. Barrenechea, H. Bustince, and F. Herrera, “A Review on Ensembles for the Class
Imbalance Problem: Bagging-, Boosting-, and Hybrid-Based Approaches,” IEEE Trans. Syst. Man, Cybern. Part C
(Applications Rev., vol. 42, no. 4, pp. 463–484, Jul. 2012, doi: 10.1109/TSMCC.2011.2161285.
[12] J. Brank et al., “Feature Selection,” in Encyclopedia of Machine Learning, Boston, MA: Springer US, 2011, pp. 402–
406.
[13] R. J. Urbanowicz, M. Meeker, W. La Cava, R. S. Olson, and J. H. Moore, “Relief-based feature selection: Introduction
and review,” J. Biomed. Inform., vol. 85, pp. 189–203, Sep. 2018, doi: 10.1016/j.jbi.2018.07.014.
[14] R. P. L. DURGABAI and R. B. Y, “Feature Selection using ReliefF Algorithm,” IJARCCE, pp. 8215–8218, Oct.
2014, doi: 10.17148/IJARCCE.2014.31031.
[15] D. M. D. Raj and R. Mohanasundaram, “An Efficient Filter-Based Feature Selection Model to Identify Significant
Features from High-Dimensional Microarray Data,” Arab. J. Sci. Eng., vol. 45, no. 4, pp. 2619–2630, Apr. 2020, doi:
10.1007/s13369-020-04380-2.
[16] R. Spencer, F. Thabtah, N. Abdelhamid, and M. Thompson, “Exploring feature selection and classification methods
for predicting heart disease,” Digit. Heal., vol. 6, p. 205520762091477, Jan. 2020, doi: 10.1177/2055207620914777.
[17] T. Vafeiadis, K. I. Diamantaras, G. Sarigiannidis, and K. C. Chatzisavvas, “A comparison of machine learning
techniques for customer churn prediction,” Simul. Model. Pract. Theory, vol. 55, pp. 1–9, Jun. 2015, doi:
10.1016/j.simpat.2015.03.003.
[18] Y. Qu, Y. Fang, and F. Yan, “Feature Selection Algorithm Based on Association Rules,” J. Phys. Conf. Ser., vol.
1168, p. 052012, Feb. 2019, doi: 10.1088/1742-6596/1168/5/052012.
[19] J. Pamina, T. Dhiliphan, Rajkumar, S. Kiruthika, T. Suganya, and F. Femila, “Exploring Hybrid and Ensemble
Models for Customer Churn Prediction in Telecom Sector,” Int. J. Recent Technol. Eng., vol. 8, no. 2, pp. 299–308,
Jul. 2019, doi: 10.35940/ijrte.A9170.078219.
[20] J. Britto and Gobinath, “A Detailed Review For Marketing Decision Making Support System In A Customer Churn
Prediction,” Int. J. Sci. Technol. Res., vol. 9, no. 4, pp. 3698–3702, 2020.
[21] Nisha Saini, Monika, and Dr. Kanwal Garg, “Churn Prediction in Telecommunication Industry using Decision Tree,”
Int. J. Eng. Res., vol. V6, no. 04, Apr. 2017, doi: 10.17577/IJERTV6IS040379.
[22] A. K. Ahmad, A. Jafar, and K. Aljoumaa, “Customer churn prediction in telecom using machine learning in big data
platform,” J. Big Data, vol. 6, no. 1, p. 28, Dec. 2019, doi: 10.1186/s40537-019-0191-6.
[23] F. Kayaalp, M. S. Basarslan, and K. Polat, “TSCBAS: A Novel Correlation Based Attribute Selection Method and
Application on Telecommunications Churn Analysis,” in 2018 International Conference on Artificial Intelligence
and Data Processing (IDAP), Sep. 2018, pp. 1–5, doi: 10.1109/IDAP.2018.8620935.
[24] S. Rai, N. Khandelwal, and R. Boghey, “Analysis of Customer Churn Prediction in Telecom Sector Using CART
Algorithm,” 2020, pp. 457–466.
[25] Francisco, “Churn in The Telecom Industry Dataset,” 2017.
https://bigml.com/user/cesareconti89/gallery/dataset/58cfbada49c4a13341003cba.

Customer Churn Prediction in Telecommunication Industry…. (Sulaiman Olaniyi Abdulsalam et al)

440  ISSN: 2089-3272

[26] T. T. Le et al., “Differential privacy-based evaporative cooling feature selection and classification with relief-F and
random forests,” Bioinformatics, vol. 33, no. 18, pp. 2906–2913, Sep. 2017, doi: 10.1093/bioinformatics/btx298.
[27] Z. M. Hira and D. F. Gillies, “A Review of Feature Selection and Feature Extraction Methods Applied on Microarray
Data,” Adv. Bioinformatics, vol. 2015, pp. 1–13, Jun. 2015, doi: 10.1155/2015/198363.
[28] C.-L. Lin and C.-L. Fan, “Evaluation of CART, CHAID, and QUEST algorithms: a case study of construction defects
in Taiwan,” J. Asian Archit. Build. Eng., vol. 18, no. 6, pp. 539–553, Nov. 2019, doi:
10.1080/13467581.2019.1696203.
[29] A. SIMION-CONSTANTINESCU, A. I. DAMIAN, N. TAPUS, L.-G. PICIU, A. PURDILA, and B.
DUMITRESCU, “Deep Neural Pipeline for Churn Prediction,” in 2018 17th RoEduNet Conference: Networking in
Education and Research (RoEduNet), Sep. 2018, pp. 1–7, doi: 10.1109/ROEDUNET.2018.8514153.
[30] S. A. Qureshi, A. S. Rehman, A. M. Qamar, A. Kamal, and A. Rehman, “Telecommunication subscribers’ churn
prediction model using machine learning,” in Eighth International Conference on Digital Information Management
(ICDIM 2013), Sep. 2013, pp. 131–136, doi: 10.1109/ICDIM.2013.6693977.
[31] X. Jiang et al., “Genome analysis of a major urban malaria vector mosquito, Anopheles stephensi,” Genome Biol.,
vol. 15, no. 9, p. 459, Sep. 2014, doi: 10.1186/s13059-014-0459-2.
[32] M. O. Arowolo, M. O. Adebiyi, A. A. Adebiyi, and O. J. Okesola, “Predicting RNA-seq data using genetic algorithm
and ensemble classification algorithms,” Indones. J. Electr. Eng. Comput. Sci., vol. 21, no. 2, p. 1073, Feb. 2021,
doi: 10.11591/ijeecs.v21.i2.pp1073-1081.
[33] M. O. Arowolo, M. O. Adebiyi, A. A. Ariyo, and O. J. Okesola, “A genetic algorithm approach for predicting
ribonucleic acid sequencing data classification using KNN and decision tree,” TELKOMNIKA (Telecommunication
Comput. Electron. Control., vol. 19, no. 1, p. 310, Feb. 2021, doi: 10.12928/telkomnika.v19i1.16381.
[34] M. O. Arowolo, M. Adebiyi, A. Adebiyi, and O. Okesola, “PCA Model For RNA-Seq Malaria Vector Data
Classification Using KNN And Decision Tree Algorithm,” in 2020 International Conference in Mathematics,
Computer Engineering and Computer Science (ICMCECS), Mar. 2020, pp. 1–8, doi:
10.1109/ICMCECS47690.2020.240881.
[35] L.F. Khalid, A.M. Abdulazeez, Y.H. Falah, D. Zeebaree, D.A. Zebari, " Customer Churn Prediction in
Telecommunications Industry Based on Data Mining" IEEE Symposium on Industrial Electronics and Applications,
2021.
[36] N. Saini, Monika, K. Garg, " Churn Prediction in TelecommunicationIndustry using Decision Tree", International
Journal of Engineering Research and Technology, Vol. 6, nno. 4, 2017. DOI: 10.17577/IJERTV6IS040379.
[37] A.K. Ahmad, A. Jafar, K. Aljoumaa, " Customer churn prediction in telecom using machine learning in big data
platform", Journal of Big Data, Vol. 6, no. 28, 2019. doi.org/10.1186/s40537-019-0191-6.

IJEEI, Vol. 10, No. 2, June 2022: 431 – 440