A Survey on Customer Churn Prediction using

Machine Learning and data mining Techniques

in E-commerce
2021 IEEE Asia-Pacific Conference on Computer Science and Data Engineering (CSDE) | 978-1-6654-9552-3/21/$31.00 ©2021 IEEE | DOI: 10.1109/CSDE53843.2021.9718460

Priya Gopal Dr. Nazri Bin MohdNawi

Under Graduate department Faculty of Computer science and
line 3: College of Banking and Information Technology
Financial Studies line 3: Universityi Tun Hussein Onn
Muscat,Oman Johor, Malaysia
[email protected] [email protected]
\ identifying customers fond of churning, thus increasing profit
of the enterprise, by performing customer churn reduction
Abstract—In recent era, major companies from e-commerce through necessary steps [2]. The increased rates of churn in
sector should focus on customer identification through churns
e-commerce compared to traditional methods are listed to be
in accordance to their business strategy as the market
saturation increases. Several e-commerce professionals and
its main disadvantage. Numerous researchers who performed
industrialists highlighted /enlisted/specified that identifying broad and in-depth research emphasizing on loss forecasts of
churn customers whose subscription period is about to end or e-commerce customers achieved a better result [3]. Artificial
those likely to migrate services from pre-existing company to Intelligence and Statistical Analysis are key prediction tools
another is known/represented as customer attrition. In order to for the categorization of e-commerce customers.
retain loyal customers, earlier prediction of client behavior
plays a vital part in real-time marketing. In this survey work, Cluster analysis, Bayesian network, time series, decision
the difficulties in the prediction of customer attrition in the trees, linear regression comprise statistical analysis. The
motor insurance sector are represented along with various data
customer loss in e-commerce represents high-dimensional
mining techniques comprising deep learning and machine
learning advancements. It also emphasizes on the churns within and nonlinear characteristics due to economics, minds of
the customer management cycle with surroundings. An customers, culture, buying behavior and other factors which
overview of the survey performed orderly provides construction should be reckoned in linear variation, therefore e-commerce
of churn prediction model, various methods of prediction customer churn characteristic is not fully reflected in
utilized and their application in the business sector. statistical analysis model [5]. The appropriate prediction rate
is attained by artificial intelligence technology [4] compared
Keywords—Customer churn; Customer Churn Prediction; to statistical analysis because of its nonlinear processing and
Data Mining Methods; Predictive models and Performance self-learning ability. The characteristics of data of actual
metrics. (key words) customer loss are high nonlinearity, highly uneven sample,
high dimensionality with noise. E-commerce prediction
I. INTRODUCTION accuracy of customer churn is enhanced by using multiple
In various economic sectors, the pace of globalization and predictive models which utilize the prediction advantages of
deregulation influences the competitive framework. The each model. This level of accuracy is difficult to attain by a
onset of new technologies and competitors created a sharp single predictive model if the difficulty of prediction is
hike in grounds of competition. A good competitor must increased [6].
withhold strong bonding with the customer for retaining their Data mining (DM) method and machine learning (ML)
investment in the shifting markets in order to compete with technique uses customer churn prediction in the traditional
other service providers. The enterprises mostly enhance their method which does not utilize the bid data advantage fully.
resources to hold the customers by providing higher returns In this work, a complete e-commerce churn prediction
on investment than new customer findings. The fundamental solution survey is presented which includes deep learning
elements in a competitive environment primarily includes (DL) techniques, comprising of ML and DM based analysis
concentrating on the launch of actions to retain customers and of customer churning to constitute customer churn predictive
prediction of buyer’s willingness to abandon, also known as model which attain real-time warning. In this work, the DM
churn. For business analytics, the concept of data mining techniques, are categorized separately after churn prediction
(DM) is utilized, where the succession stages are determined is examined. Thus, providing an insight about research scope
on the basis of its scripted methodology. [1]. in e-commerce sector for churn analysis, paved way for
prospective studies in this field. The paper is organized as
follows: Section 2 provides / represents an entire literature
A worldwide network of trade was constituted by enterprises survey and review of DM techniques, recently used for
via internet due to its enormous growth, rendering an performing churn analysis and based-applications, Section 3
outstanding choice to customers. The transitions in customer emphasizes on the result, while Section 4 adds conclusion to
behavior and the business was brought-in by e-commerce. the survey, and scope for future endeavors in e-commerce
Hotspot research in e-commerce mainly focused on sector.

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

Authorized licensed use limited to: Unitec Library. Downloaded on September 17,2022 at 05:57:52 UTC from IEEE Xplore. Restrictions apply.
 section. Generally, they are categorized into two as follows.
i) Hybrid Technique (Integrates soft computing along with
machine learning) ii) Machine learning-based methods. Fig.
Previous studies, proposed widespread methods for churn 1 represents classification of various techniques utilized for
prediction and analysis. Recent techniques extensively used churn prediction.
for analysis and prediction studies are discussed in this

(KLMM) was developed by Zhao [12]. The integral

connection between classification result and data attributes
II. REVIEW ON CHURN PREDICTION USING
was exposed by dimension of scalability of the data source.
MACHINE LEARNING
For clustering amongst credit card holders, a modified K-
Different data mining techniques are utilized for carrying out
means algorithm was designed [13]. The clustered data was
research on churn prediction system of credit cards.
further classified into training cluster and testing cluster by
Framework for precise churn prediction of e-commerce
hold-out method. Later, diverse algorithms such as decision
customers mainly comprises of four steps namely( you can
tree, k-nearest neighbor, vector machine and naïve bayes
include names)[7]. Initially, the collection of behavioral data
were used for classification studies. The prediction model
from customers and transfer to data warehouse was
displayed for telecom churn in our work precisely
performed by an Extract Transform Load (ETL). Thereby,
emphasizes on integrating uplifting based strategies with
data samples denoted as train objects precisely, were
ensemble stacking. Cost heuristic is often used for evaluation
processed from constituent warehouse. Different prediction
of performance than conventional methods.
algorithms are enabled for further training of received
samples. At last, customer forecasting was performed Rough Set Theory (RST) based smart rule-based decision-
selectively using an, extended version of prediction making technique was developed [15] to obtain essential
algorithm. An extended support vector machine (ESVM) decision rules corresponding to non-churn and churn
with parameters suitable for non-churner, nonlinear customer customers. The proposed approach was efficient in
churn outcome was used. The proposed model for clustering predictions involving current and future churn customer. It
of customers mainly emphasized on customer relationships also facilitated extraction of churned customer from non-
with four dimensions such as Relation length (L), recent churn customers. Fuzzy based prediction model for churn
transaction time (R), buying frequency (F), and monetary (M) was proposed later. [16] This model was primarily formalized
known as LRFM model[8]. The LRFM clustering model used by the real time date from a south Asian telecom platform. In
in the real-time textile business for clustering customers order to identify key features for accurate churn predictions
denoted a noticeable change with strong marketing strategy in telecom industry, an RST based methodology was
from the extracted result. A partial churn detection model developed. [17] Random subspace, boosting and bagging
was for predictions in retail grocery sector [9]. The proposed were the ensemble- classification methods for which the
model mainly predicted the resemblance between non- identified feature functioned as an input.
churner and churner sequence of products' first purchase.
Markov for discrimination was used to model the first Sivasankar& Vijaya [18] proposed a system for customer
purchase sequence. Random forest and logistic regression churn prediction, which is a hybrid technique integrating
were the two divisions utilized in empirical studies. artificial neural network with probabilistic possibilistic fuzzy
C-means clustering (PPFCM-ANN). The probabilistic
Telecom sector predominantly used ensemble and effective possibilistic fuzzy C-means clustering algorithm is used to
feature extraction techniques for building a smart prediction collect the data in a cluster using a clustering module. The
system of churn [10]. The prediction problem of churn for system comprises of two components namely) ANN-based
telecom sector is built using RotBoost, Random Forest, component for churn prediction ii) PPFCM based component
Rotation Forest and DECORATE ensembles integrated with for clustering.
minimum redundancy and maximum relevance (mRMR), F-
score method and Fisher’s ration. Efficient management of Another novel hybrid technique used for classification was
telecom datasets in turn provided/enabled high dimensional Logit leaf model (LLM) integrating logistic regression with
characteristic through the proposed system. This system decision tree [19]. The DT strength was balanced by
integratedmRMR features (CP-MRB) with RotBoost as reduction in weakness. To fit the logistic regression in every
depicted in the obtained simulation result. subset, decision tree subdivided the data into many unvarying
subsets in LLM. To discover the churn customer in the
Decision Tree (DT), K-Nearest Neighbors (KNN), Support telecom sector and the reason behind the customer churning
Vector Machine (SVM) and Artificial Neural Network a prediction model for churn which combines clustering
(ANN) are the four main categories of techniques utilized in technique with classification technique is proposed by
the dataset of Iranian telecommunication companies. Among Ullah[20]. Correlation attribute ranking filter and gained
the four techniques, ANN outperformed others. The hybrid information are utilized to carry out performance selection.
methodology system, which comprises these four techniques, Random Forest (RF) classification algorithm was used to
was proposed through early studies [11]. A detailed classify the data of churn customers with 88.63% correctly
examination of diversifying subspace partition rules paved classified instances.
way to build a corresponding potential structure. A new K-
local maximum margin feature extraction algorithm

Authorized licensed use limited to: Unitec Library. Downloaded on September 17,2022 at 05:57:52 UTC from IEEE Xplore. Restrictions apply.
By observing the training set hidden pattern, each customers To perform churn prediction and analysis for greater datasets,
individuality is discovered and the obtained knowledge is another technique was proposed [29].This method was based
applied to the test set to improve the performance of the on metaheuristic which used firefly algorithm as a classifier.
prediction in enhanced prediction system developed by The comparison block was the intensive component of firefly
Jafari-Marandi [21]. The error-driven learning and self- algorithm. This algorithm was used to identify the firefly
organized approaches (ChP-SOEDNN) advantage are used with brighter intensity of light compared to others .
by artificial neural network for designing the prediction
system of churn. V. ISSUES FROM EXISTING WORK
For proper functioning of ML–based models, detailed
III. REVIEW ON SOFT COMPUTING WITH emphasis is required on the key features of existing models.
MACHINE LEARNING-BASED CHURN PREDICTION The appropriate feature definition is required for each model
Chr-PmRF approach which consists of RF, mRMR and PSO which depends upon the problems. Traditional models failed
techniques to provide better result than the other existing to identify the behavioral pattern due to the unstructured
classification, feature reduction and sampling techniques.[22] format and an enormous volume of data. To overcome the
Based on AUC, the Random forest and KNN were used for failure, data mining and its application in customer churn are
primary examination and optimal selection from each subset. comes under research focus, lot of researches are carried out
Each optimal subset for which Fisher’s ratio, PCA, mRMR in this field to find optimal solutions. In the analysis of
techniques are applied individually, were further analyzed customer churn, family of deep learning algorithms also
using KNN and RF classifier. The data set is subsample by applied in the research of covering algorithms and its
an under sampling method of PSO. The power of prediction application which is evident from the literature survey. Deep
of various algorithms in Bayesian Network were investigated learning being an emerging field in churn prediction analysis,
by Verbraken [23] which is ranges from Naive Bayes paves way to greater research under machine learning.
classifier to General Bayesian Network classifiers. Markov TABLE 1. INFERENCE OF EXISTING MACHINE
blanket concept-based feature selection method is tested in LEARNING MODEL BASED ON CLASSIFICATION
this work which corresponded to Bayesian Networks. To use
the customer data in both related source and target transfer S. Methods Merits Demerits
learning theory is proposed in feature-selection-based No
dynamic transfer ensemble (FSDTE) model which is
developed by Xiao [24]. 1 extended Proficient of Constructed models for
support vector handling large scale, nonlinearity
The particle fitness calculation (PFC) and particle machine massive data and high-
classification optimization (PCO) are executed iterative (ESVM) dimensionality
manner in telecommunication BP network particle accuracy besides
classification optimization-based customer churn prediction generalization
(PBCCP) algorithm which is proposed by Yu [25]. To capability might not be
develop high-performance churn prediction system with assured.
superior ability in churn identification, AdaBoost
classification capability is combined with genetic 2 logistic can be used Doesn't perform well
programming (GP) searching capability. The learning-based regression and to solve both when feature space is
AdaBoost is used to evaluate the features selected frequently random forests classification too large.
from the expression of various GP which are noted and as well
as regression Restriction in
examined. Churn prediction system (ChP-GPAB) which is
problems investigation of people
developed by the integration of GP-AdaBoost with Particle
buying list in physical
swarm optimization based under sampling method is used to
discovers the hidden reason behind the behavior of customers stores
churning and provides better churners learning. Yu [25] focus
on a discussion of number of research work in the area of
accurate churn measurement in customer churn and various 3 RotBoost in Feature space Limitations is restricted
approaches to retention. 2Manivannan proposed Grey Wolf amalgamation is diminished to to application
Optimization approach (GWO)[27] algorithm whose with mRMR through domains in which
improved accuracy is 89.26% by matching the actual churn features (CP- mRMR datasets does not
and converge in minimal time compared to other ACO and MRB) efficiently, possesses high
PSO approaches which take more time of convergence for leading to dimensionality as well
appreciable churn prediction. CUPGO retains the valuable better as imbalance
customers with 34.81% customer retention which can be learning distribution.
achieved by processing the dataset gathered over two capabilities.
consistent years. The particle swarm optimization (PSO) with
three variants such as feature selection integrated PSO, 4 Rough K- Good Outcomes are not
simulated annealing integrated PSO and PSO integrated both means+ SVM accuracy satisfactory
are used for proposed a technique for the prediction of churn besides lesser
in telecom sector by Vijaya &Sivasankar [18]. The feature misclassificat
selection is used as a pre-processing mechanism. ion error in

Authorized licensed use limited to: Unitec Library. Downloaded on September 17,2022 at 05:57:52 UTC from IEEE Xplore. Restrictions apply.
 contradiction TABLE 2. INFERENCE OF EXISTING MACHINE
to single LEARNING MODEL BASED ON CLUSTRING
classifier
model S. Methods Merits Demerits
No
5 rough set theory Predicted An open research
(RST) churn at risk problem may exist due 1 LRFM clustering Able to Very less frequency
which to the sort of obtain irrespective of good
perhaps classification technique important customer performance
might churn. used to approach churn characteristic contribution.
prediction attributes for
customer
6 Fuzzy Classifier Fuzzy Domain knowledge relationship
Classifiers is Augmentation may management.
greatly prone to erroneous
utilized for results. 2 K- local Efficiency is Necessitates greater
Maximum maximum enhanced in computational resource
churners margin feature addition to overhead besides
captured by extraction churn extended computing
Fuzzy algorithm prediction time
Classifiers (KLMM) accuracy in
due to high telecom
TP rate.

7 Logit leaf More Restricted TABLE 3. INFERENCE OF EXISTING MACHINE

model (LLM) actionable consideration in LEARNING MODEL BASED ON ENSEMBLE / HYBIRD
model for customer churn MODEL.
which prediction
segments are S.No Methods Merits Demerits
produced
besides main 1 ensemble high Higher false negative
drivers are stacking performance rates.
identified on levels in
segment terms of cost.
level.
2 Hybrid Maximum Not very efficient with
8 ChP-SOEDNN Preeminent Imbalancing of ratio of probabilistic accuracy is large number of
balance for churn and non-churn possibilistic attained observations.
false positive customers is an another fuzzy C-means when
and false challenge. clustering associated to
negative (PPFCM) any single
misclassificat merged with model.
ion costs for artificial neural
each network
customers (PPFCM-ANN)
cluster.
3 Random Forest Guidelines Omission of number of
9 CUPGO minimal time Requires Feature (RF) on customer information-rich
to converge Scaling retention for features during
besides telecom modeling development
enriched companies
accuracy is decision-
supported makers.

10 PSO with PSO in More conservative in 4 GP-AdaBoost Incomplete Short learning

simulated addition its recognizing classes. results a churn knowledge efficiency, slow
annealing variants prediction for working convergence rate, local
accomplish system (ChP- optimum, etc.
well on GPAB)
massive
imbalanced 5 hybridized Very low Accuracy level as well
data. firefly time latency as churn prediction
algorithm which might be enhanced
Outclasses

Authorized licensed use limited to: Unitec Library. Downloaded on September 17,2022 at 05:57:52 UTC from IEEE Xplore. Restrictions apply.
 normal through integrating Deep learning is the enhanced research of machine learning,
firefly game theory. which uses a hierarchical learning process for obtaining high-
algorithm. level features of data and utilized to handle the enormous
amount of unstructured data. By eliminating the low-level
features, the proposed novel Stacked Auto-Encoder (SAE)
6 hybrid It can gain a Better precision may
incorporated in multi-layer feature selection obtained only
methodology considerably cause low Recall
high-level feature from the data of churn.
higher measure and vice versa
accuracy

VII. EXPERIMENTAL RESULTS AND DISCUSSION

The churn prediction model is design considered in this paper
utilize the dataset obtained from the web link
http://www.ics.uci.edu~mlearn/MLRepository.html. Vmail,
VI. CHURN PREDICTION WITH DEEP LEARNING day.charge, area, vmail.msgs, day.calls,day.mins are the
METHODS customer details which is stored in the telecom database as
customer information. F-score, accuracy, precision is the
The major problem of telecom operators in customer parameters taken in account as bar-plot to compare the
relationship management system is prediction of churn performance of PBCCP [25], rough K-means+ SVM [13] and
customer. For churn prediction, [30] proposed Convolutional PPFCM-ANN [18] the existing machine learning algorithms
Neural Network (CNN) based deep learning approach which with the deep learning model [30].
provides better accuracy. For the prediction of churn in
telecom customer, an efficient method of feature extraction is Precision: Equation gives the positive samples proportion
needed. Fisher’s ratio analysis and stacked auto-encoder which is classified correctly as expected:
which is the nonlinear feature extraction method, integrated 𝑇𝑃
to develop a hybrid feature extraction method proposed by 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =
𝐹𝑃+𝑇𝑃
[13].
Recall: The proportion of positive sample assigned properly
De Caigny[32] proposed a method for customer churn to the positive samples total number is known as classifier's
prediction (CCP) models which integrated textual data. It recall given in the equation as follows:
extended the convolutional neural networks (CNNs)
𝑇𝑃
proposed in previous literature by benchmarking it against 𝑅𝑒𝑐𝑎𝑙𝑙 =
𝑇𝑃+𝐹𝑁
with many practices for CCP textual data analysis by utilizing
European financial service provider's real-time data to F-measure: The harmonic mean of recall and precision is
formalize the framework which details the incorporation of given by F-measure which is also termed as F1-Score.
textual data in the predictive model.
2 ∗ (𝑅𝑒𝑐𝑎𝑙𝑙 ∗ 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛)
Compared with models like decision tree modelling, 𝐹 − 𝑚𝑒𝑎𝑠𝑢𝑟𝑒 =
(𝑅𝑒𝑐𝑎𝑙𝑙 + 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛)
autoencoders and deep conventional neural networks is
outperformed in understanding and prediction of Accuracy: The proportion of samples classified correctly in
telecommunication industry churns. To improve its the total number of samples is given by accuracy which is
efficiency further the input images is augmented and utilized to measure the classifier performance .
developed with a greater number of features because for each 𝑇𝑃+𝑇𝑁
customer only 12 temporal features are taken into account by 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁
the existing models. Stacked convolutional encoder is
utilized by deep convolutional neural network for pre- Where no-churn samples classified correctly denoted as True
training the weights to improve the AUC. [33]. Positive (TP), the no-churn sample classified as churn sample
incorrectly is denoted by False positive (FP), True negative
Deep learning (DL) technique-based analysis of customer (TN) denoted the churn sample classified correctly, the
churning is utilized by Cao [34] to constitute customer churn incorrect classification of churn as no churn is denoted by
predictive model which attain real-time warning. The False Negative (FN)
features are extracted from the data by stacked encoder
network and the customers are classified by logistic Precision Result comparison
regression (LR). Greedy layer-wise unsupervised learning
algorithm is utilized for the deep learning model training of
pre-train stacked autoencoder network. Back propagation
(BP) is utilized by stacked autoencoder network which is
formed by arranging every layer after pretraining to reverse
tuning parameters. After that logistic regression layer is
trained.
Currently, the researchers are not focused in deep learning
integrated with stack autoencoder model for churn prediction.

Authorized licensed use limited to: Unitec Library. Downloaded on September 17,2022 at 05:57:52 UTC from IEEE Xplore. Restrictions apply.
 Accuracy comparison

Fig.1. Precision performance comparison

The precision comparison between the methods CNN, Number of data
PPFCM-ANN, PBCCP and rough K-means+ SVM is given
Fig.3. Result of Accuracy
in Fig. 1. From the comparison, it is concluded the CNN has
a high precision rate of 99% compared to the existing The accuracy comparison of the methods CNN, PPFCM-
methods used to achieve better precision. The other existing ANN, PBCCP and rough K-means+ SVM is given in Fig. 3.
techniques PPFCM-ANN, PBCCP and rough K-means+ The accuracy rate of CNN, PPFCM-ANN, PBCCP and rough
SVM gives considerable precision rates of 98%, 99% and K-means+ SVM 98.5%, 97%, 98% and 96% respectively.
95% respectively which is lesser compared to CNN. From the result, it is concluded that the accuracy of the
proposed CNN is 98.5% which is higher than other methods.
F-measure Result Comparison The accuracy rate is directly proportional to the amount of
data. The proposed work prediction model uses only one
dataset. From the accuracy rate, it is known that the CNN
The result of the comparison of F-measure of the methods selects high accuracy cluster center during cluster center
CNN, PPFCM-ANN, PBCCP are 98%, 94%, 97% and 89%. selection. In telecom industry, among the methods, CNN has
From the result, it is concluded that the F-measure of CNN is the potential to obtained good decision for churn
higher than the other existing methods which provide the best management.
result in churn prediction compared to other methods.

VIII. CONCLUSION AND FUTURE WORK

In e-commerce, the detection of customers willing to cancel
subscription and orders termed as customer churn prediction
is vital to eliminate the loss confront by the companies in
terms of time and amount spend to search a new customer and
due to cancellation of order and subscription. The companies
and industries confront the loss of its brand value, image and
loss of revenue due to customer churn. New customer finding
process is very difficult. To mitigate these losses, the
company needs to constitute a model for customer churn
prediction in advance. This paperwork mainly focuses on
Number of data examining the existing models of customer churn prediction
implemented in many industries than introduced new model.
It provides details understanding and projects various
techniques and attributes of churn prediction model which is
Fig.2. Recall performance comparison utilized for customer retention. The prediction model of
customer churn is used to eliminate huge loss. The models
The result comparison of recall value of the methods CNN, are trained and prediction is performed based on the real-time
PPFCM-ANN, PBCCP and rough K-means+ SVM is given data obtained from various industries like
in Fig. 2. The recall rate of CNN, PPFCM-ANN, PBCCP and telecommunication, etc.. which is detailed in the study. The
rough K-means+ SVM is 99%, 97%, 98% and 95% churn models utilized the deep learning techniques gives
respectively. From the result, it is validated that the better performance according to this study. Standard
anticipated ALEBPEC CNN technique recall rate is higher evaluation metrics are used to validate the churn model which
than the other method. The detection result will be better for is used for data analytics. The current methods required
higher recall rate. The churn prediction result of CNN is further development when compared to the methods based on
better than other methods. machine learning discussed in section 2 which is utilized in

Authorized licensed use limited to: Unitec Library. Downloaded on September 17,2022 at 05:57:52 UTC from IEEE Xplore. Restrictions apply.
machine learning-based (MBD) field for providing better test for customer churn in e-commerce. Expert Systems with
result for real data. The difficulties faced by the methods Applications, 38(3), 1425-1430.
using machine learning in MBD analysis is given as follows: [8] Li, D. C., Dai, W. L., & Tseng, W. T. (2011). A two-
stage clustering method to analyze customer
1) Large-Scale. For attaining, better precision and efficiency characteristics to build discriminative customer
in methods using machine learning for MBD analysis, an management: A case of textile manufacturing
enormous amount of applications and real-time data is business. Expert Systems with Applications, 38(6), 7186-
required. 7191.
2) Generalization Problem. The scalability or the [9] Miguéis, V. L., Van den Poel, D., Camanho, A. S., & e
generalization ability plays a vital role in performance Cunha, J. F. (2012). Predicting partial customer churn
evaluation of trained deep learning or machine learning using Markov for discrimination for modeling first
model. It is appropriate for diverse data subspace. The entire purchase sequences. Advances in Data Analysis and
data gaining in impractical in the massive scale MBD, Classification, 6(4), 337-353.
although it belongs to a specific field. [10] Idris, A., Khan, A., & Lee, Y. S. (2013). Intelligent churn
prediction in telecom: employing mRMR feature
3) Multimodal Learning. It is a great challenge to acquire selection and RotBoost based ensemble
useful pattern and hidden knowledge from the multimodal classification. Applied intelligence, 39(3), 659-672.
learning which uses diverse input data and multimodal with [11] Keramati, A., Jafari-Marandi, R., Aliannejadi, M.,
deep and machine learning. Ahmadian, I., Mozaffari, M., & Abbasi, U. (2014).
Improved churn prediction in telecommunication
Scalability attains more attention which increases the inferred industry using data mining techniques. Applied Soft
classes number, in recognition and classification problem Computing, 24, 994-1012.
besides with generalization problem. To upfront the [12] Zhao, L., Gao, Q., Dong, X., Dong, A., & Dong, X.
requirement of the churn prediction analysis, it is necessary (2017). K-local maximum margin feature extraction
to enhance the scalability of the methods with high feasibility algorithm for churn prediction in telecom. Cluster
and accuracy. In the future work, for better churn prediction Computing, 20(2), 1401-1409.
deep learning approaches should be investigated. For trend [13] Rajamohamed, R., &Manokaran, J. (2018). Improved
analysis and prediction, soft computing and autoencoder are credit card churn prediction based on rough clustering
utilized in artificial intelligence techniques which is used to and supervised learning techniques. Cluster
discover the behavior pattern change of churn customer in Computing, 21(1), 65-77.
future work. [14] Ahmed, A. A., &Maheswari, D. (2019). An enhanced
ensemble classifier for telecom churn prediction using
cost based uplift modelling. International Journal of
REFERENCES Information Technology, 11(2), 381-391.
[15] Amin, A., Anwar, S., Adnan, A., Nawaz, M., Alawfi, K.,
[1] Berger, P., &Kompan, M. (2019). User modeling for
Hussain, A., & Huang, K. (2017). Customer churn
churn prediction in E-commerce. IEEE Intelligent
prediction in the telecommunication sector using a rough
Systems, 34(2), 44-52.
set approach. Neurocomputing, 237, 242-254.
[2] Ren, A. H., & Zhao, W. W. (2013, December).
[16] Azeem, M., Usman, M., & Fong, A. C. M. (2017). A
Electronic Commerce Based on Self-Organizing Data
churn prediction model for prepaid customers in telecom
Mining Customer Churn Prediction Model. In 2013
using fuzzy classifiers. Telecommunication
International Conference on Advances in Social Science,
Systems, 66(4), 603-614.
Humanities, and Management (ASSHM-13). Atlantis
[17] Vijaya, J., &Sivasankar, E. (2018). Computing efficient
Press.
features using rough set theory combined with ensemble
[3] Akter, S., &Wamba, S. F. (2016). Big data analytics in
classification techniques to improve the customer churn
E-commerce: a systematic review and agenda for future
prediction in telecommunication
research. Electronic Markets, 26(2), 173-194.
sector. Computing, 100(8), 839-860.
[4] Vanneschi, L., Horn, D. M., Castelli, M., &Popovič, A.
[18] Sivasankar, E., & Vijaya, J. (2019). Hybrid PPFCM-
(2018). An artificial intelligence system for predicting
ANN model: an efficient system for customer churn
customer default in e-commerce. Expert Systems with
prediction through probabilistic possibilistic fuzzy
Applications, 104, 1-21.
clustering and artificial neural network. Neural
[5] Yu, X., Guo, S., Guo, J., & Huang, X. (2011). An
Computing and Applications, 31(11), 7181-7200.
extended support vector machine forecasting framework
[19] De Caigny, A., Coussement, K., & De Bock, K. W.
for customer churn in e-commerce. Expert Systems with
(2018). A new hybrid classification algorithm for
Applications, 38(3), 1425-1430.
customer churn prediction based on logistic regression
[6] Wu, H. L., Zhang, W. W., & Zhang, Y. Y. (2010,
and decision trees. European Journal of Operational
August). An empirical study of customer churn in e-
Research, 269(2), 760-772.
commerce based on data mining. In 2010 International
[20] Ullah, I., Raza, B., Malik, A. K., Imran, M., Islam, S. U.,
Conference on Management and Service Science (pp. 1-
& Kim, S. W. (2019). A churn prediction model using
4). IEEE.
random forest: analysis of machine learning techniques
[7] Yu, X., Guo, S., Guo, J., & Huang, X. (2011). An
for churn prediction and factor identification in telecom
extended support vector machine forecasting framework
sector. IEEE Access, 7, 60134-60149.

Authorized licensed use limited to: Unitec Library. Downloaded on September 17,2022 at 05:57:52 UTC from IEEE Xplore. Restrictions apply.
[21] Jafari-Marandi, R., Denton, J., Idris, A., Smith, B. K., optimization based feature selection model with
&Keramati, A. (2020). Optimum profit-driven churn simulated annealing. Cluster Computing, 22(5), 10757-
decision making: innovative artificial neural networks in 10768.
telecom industry. Neural Computing and Applications, [29] Ahmed, A. A., &Maheswari, D. (2017). Churn
1-34. prediction on huge telecom data using hybrid firefly
[22] Idris, A., Rizwan, M., & Khan, A. (2012). Churn based classification. Egyptian Informatics
prediction in telecom using Random Forest and PSO Journal, 18(3), 215-220.
based data balancing in combination with various feature [30] Mishra, A., & Reddy, U. S. (2017, December). A novel
selection strategies. Computers & Electrical approach for churn prediction using deep learning.
Engineering, 38(6), 1808-1819. In 2017 IEEE International Conference on
[23] Verbraken, T., Verbeke, W., &Baesens, B. (2014). Profit Computational Intelligence and Computing Research
optimizing customer churn prediction with Bayesian (ICCIC) (pp. 1-4). IEEE.
network classifiers. Intelligent Data Analysis, 18(1), 3- [31] Li, R., Wang, P., & Chen, Z. (2016). A feature extraction
24. method based on stacked auto-encoder for telecom churn
[24] Xiao, J., Xiao, Y., Huang, A., Liu, D., & Wang, S. prediction. In Theory, methodology, tools and
(2015). Feature-selection-based dynamic transfer applications for modeling and simulation of complex
ensemble model for customer churn systems (pp. 568-576). Springer, Singapore.
prediction. Knowledge and information systems, 43(1), [32] De Caigny, A., Coussement, K., De Bock, K. W.,
29-51. &Lessmann, S. (2019). Incorporating textual
[25] Yu, R., An, X., Jin, B., Shi, J., Move, O. A., & Liu, Y. information in customer churn prediction models based
(2018). Particle classification optimization-based BP on a convolutional neural network. International
network for telecommunication customer churn Journal of Forecasting.
prediction. Neural Computing and Applications, 29(3), [33] Wangperawong, A., Brun, C., Laudy, O.,
707-720. &Pavasuthipaisit, R. (2016). Churn analysis using deep
[26] Idris, A., Iftikhar, A., &ur Rehman, Z. (2019). Intelligent convolutional neural networks and autoencoders. arXiv
churn prediction for telecom using GP-AdaBoost preprint arXiv:1604.05377.
learning and PSO undersampling. Cluster [34] Cao, S., Liu, W., Chen, Y., & Zhu, X. (2019, October).
Computing, 22(3), 7241-7255. Deep Learning Based Customer Churn Analysis. In 2019
[27] Manivannan, R., Saminathan, R., & Saravanan, S. 11th International Conference on Wireless
(2019). An improved analytical approach for customer Communications and Signal Processing (WCSP) (pp. 1-
churn prediction using Grey Wolf Optimization 6). IEEE.
approach based on stochastic customer profiling over a
retail shopping analysis: CUPGO. Evolutionary
.
Intelligence, 1-10.
[28] Vijaya, J., &Sivasankar, E. (2019). An efficient system
for customer churn prediction through particle swarm

Authorized licensed use limited to: Unitec Library. Downloaded on September 17,2022 at 05:57:52 UTC from IEEE Xplore. Restrictions apply.