International Journal of Innovative Technology and Exploring Engineering (IJITEE)

ISSN: 2278-3075 (Online), Volume-9 Issue-6, April 2020

Predicting Churn Customer in Telecom using

Peergrading Regression Learning Technique
M.HemaLatha, S.Mahalakshmi

Abstract: Customer churn is an important issue which is In this telecommunication, churn is a major issue where
faced by telecom industries daily. It is an essential concern for Churn is the activity of the telecommunication industry is
enterprises. Most of the telecom companies suffer a lot for the customers leaving the current company and moving to
voluntary churn. Here the churn rate is a significant impact for another telecom company. The technical progress and the
the industries on the lifetime value, and it affects the length of increasing number of operators raised the level of
the service and also the revenue of the company. Because of the competition. Many companies are finding the reasons of
direct effect of the company revenue, especially in the telecom losing customers by measuring customer loyalty to regain
field, the companies are requesting to implement the development the lost customers. To keep up with the competition and to
of predicting the potential customer to churn. Here the churn is acquire as many customers, most operators invest a huge
an essential factor. Hence, analyzing the churn factor of amount of revenue to expand their business in the beginning
increasing customer churn is as necessary as it is needed to [2]. In the telecmmunication industry each company
reduce the churn. Recently telecom service facing the churn provides the customers with huge incentives to attract them
problem, for this analysis this research article focused on to switch to their services, it is one of the reasons that
prediction of customer churn. In this research work, the main
customer churn is a big problem in the industry nowadays.
Churn management is very important for reducing churns as
contribution is to develop a prediction model of churn, that
acquiring a new customer is more expensive than retaining
supports the telecom operators to predict the customer, who is
the existing ones [3]. To prevent this, the company should
going to be churn. Here the model used is the Machine learning
know the reasons for which the customer decides to move
technique with the big data analysis. Our methodology used here
on to another telecom company. Prediction is the only
is Peer Grading Regression (PGR). To provide a prediction of
process to analyze the problem of churn activity. Customers
churn customer in telecom industries. To validate the churn prediction in telecom is generally a challenging task
performance of this proposed model, the Area under Curve for the large feature space and imbalanced nature of the
(AUC) is used, and it is one of the standardized approaches. This dataset. Churn prediction is used to recognize customers
model will be tested under some benchmark dataset; working on who are most probable to churn. Churn prediction and
these vast datasets creates a transformation of raw data by analysis can help a company to develop a sustainable
telecom companies. The dataset values are provided for testing strategy for customer retention programs [4].
and training in the ensemble classifier. For experimentation, Churn prediction and management have become of great
four algorithms are considered here: Peer Grading Regression concern to the mobile operators. Mobile operators wish to
(PGR), Random Forest, Decision Tree, and k-NN classifier. retain their subscribers and satisfy their needs. Hence, they
Moreover, the best outcomes are attained with the use of the need to predict the possible churners and then utilize the
boosting algorithm. Here this algorithm is applied for the limited resources to retain those customers. The basic layer
classification of the churn prediction model. for predicting future customer churn is data from the past.
Keywords: Machine Learning, Telecommunication, Customer We look at data from customers that already have churned
Churn, Area Under Curve and Prediction (response) and their characteristics / behavior (predictors)
before the churn happened. By fitting a statistical model that
I. INTRODUCTION relates the predictors to the response, we will try to predict
the response for existing customers. This method belongs to
Today numerous telecom companies are prompt all over the the supervised learning category, just in case you needed
world. The telecommunications sector has become one of one more buzzing expression [5]. The minority class
the main industries in developed countries. comprises of fewer instances in the whole dataset that leads
Telecommunication market is facing a severe loss of to biased training of classifiers due to dominating presence
revenue due to increasing competition among them and loss of majority class instances. The telecom companies also
of potential customers [1]. Companies are working hard to acquire lots of information about customers including
survive in this competitive market depending on multiple billing, payments, call records, demographics, etc. which
strategies. turns large number of features to account for in the dataset.
Consequently, the predictors suffer from the curse of
dimensionality and imbalanced distribution of the telecom
Revised Manuscript Received on April 30, 2020.
dataset for predicting churners. Thus, a churn prediction
* Correspondence Author method is highly desirable that can affectively mitigate the
Dr. M. Hema Latha*, Professor, Ramakrishna College of Arts & imbalanced nature and high dimensionality issues present in
Science Coimbatore the telecom dataset [6].
S.Mahalakshmi, Research Scholar, Bharathiar University,Coimbatore.
Churn occurs mainly due to customer dissatisfaction.
© The Authors. Published by Blue Eyes Intelligence Engineering and Identifying customer dissatisfaction requires several
Sciences Publication (BEIESP). This is an open access article under the parameters.
CC-BY-NC-ND license (http://creativecommons.org/licenses/by-nc-
nd/4.0/)

Retrieval Number: F3861049620/2020©BEIESP Published By:

DOI: 10.35940/ijitee.F3861.049620 Blue Eyes Intelligence Engineering
Journal Website: www.ijitee.org 1025 & Sciences Publication
 Predicting Churn Customer in Telecom using Peergrading Regression Learning Technique

A customer usually does not churn due to a single designed to improve accuracy by reducing error, not all of
dissatisfaction scenario [7]. them take into account the class balance, and that may give
There usually exist several dissatisfaction cases before a bad results. In general, classes are considered to be balanced
customer completely ceases to do transactions with an in order to be given the same importance in training.
organization. Several properties associated with the
customer and their mode of operations with the organization II. RELATED WORK
are recorded by the organizations. This represents the
Many approaches were applied to predict churn in telecom
customer’s behavior data. Analyzing this data would present
companies. Most of these approaches have used machine
a clear view of the customer’s current status [8]. Hence this
learning that were discussed in this section.
can be used as the base data for churn prediction. The major
difficulty arising from this mode of operation is that the data
Adnan Amin et al 2015 [12] study the problem of Customer
under discussion tends to be very huge. The hugeness can be
churn in telecom industry and understand that predicting the
attributed to the behavioral nature of the data, depicting all
customer behavior and to preserve those customers that will
the product lines dealt with by the organization. Further, due
churn or conceivable may churn. Author’s another attempt is to
to the requirement of structural representation of the data, all
make use of rough set a rule based decision making technique
the instances are bound to contain all the properties
for mining the rules for predicting the customer. For this
corresponding to a generic customer in the organization [9].
approach author used four different algorithms (Exhaustive,
This leads to data sparseness, since customers will be
Genetic, Covering, and LEM2). From this 4 algorithm genetic
associated with only a few properties and not all the
algorithm produces most suitable performance. Dataset was
properties pertaining to the organization. The hugeness of
used here is publicly available dataset. This approach fully
data and sparsity acts as the major difficulties in the process
predicts the customers who will churn or possibly may churn.
of churn prediction.
Hanif, E., 2019 [13] presents a predictive analytics approach to
Many research confirmed that machine learning technology
improve customer churn in the telecom industry for “cross-
is highly efficient to predict this situation. This technique is
selling” or “market basket analysis”. Here 4 algorithms were
applied through learning from previous data. Machine
used K-Nearest Neighbor, Decision Tree, Naïve Bayes and
learning is subset of artificial intelligence which enables
Random Forest and predict customer churn in Rapid Miner.
computers to learn (i.e. improvise) from data without any
Datasets was used here is publically available datasets. From
intervention. Machine learning is often coined with the
their implementation Decision Tree and Random Forest are the
terms pattern recognition and computational learning theory.
two algorithm which predict most accurate. Overall, the key
Machine learning involves constructing algorithms that can
drivers of churn are identified in this study and useful
learn from data available and can be used to make
associations between products are established. This information
predictions on data. Here a model is built from the given
can be used by companies to create personalized offers and
input data which is then used to make predictions on new
campaigns for customers who are at risk of churning.
data. As the data being collected is drastically increasing
Abdelrahim Kasem Ahmad 2019 [14] is to develop a churn
each day, this calls for the need of machine learning. In
prediction model which assists telecom operators to predict
machine learning, the learning process is extracting
customers who are most likely subject to churn. The model
knowledge from the given data thus there is no need for a
developed in this work uses machine learning techniques on
human to specify any kind of knowledge [10]. Because of
big data platform. For experimenting their performance Area
the hierarchy of concepts, complex concepts can be learned
Under Curve (AUC) standard measure is adopted and obtained
by building them from simpler ones thus reducing the
93.3%. then another contribution is extracting Social Networ
complexity. Prediction of churn customer is analyzed and
Analysis. This model used four algorithms: Decision Tree,
implemented using machine learning approach. Large
Random Forest, Gradient Boosted Machine Tree “GBM”
companies interact with their customers to provide a variety
and Extreme Gradient Boosting “XGBOOST”. However,
of services to them [11]. Customer service is one of the key
the best results were obtained by applying XGBOOST
differentiators for companies. The ability to predict if a
algorithm. This algorithm was used for classification in this
customer will leave in order to intervene at the right time
churn predictive model.
can be essential for pre-empting problems and providing
Vafeiadis, T., et al 2015 [15], present a comparative study
high level
on the most popular machine learning methods for the
of customer service. The problem becomes more complex as
problem of customer churning prediction in telecom. They
customer behavior data is sequential and can be very
performed a series of Monte Carlo simulations to determine
diverse. Churn is an unavoidable process in any industry.
the most efficient parameter combinations. They analyze the
However, though difficult, it is possible to identify the
classifier algorithm by using SVM-POLY in AdaBoost with
causes of churn using Machine learning approaches.
accuracy of almost 97% and F-measure over 84%.
Having a good churn prediction model becomes extremely
Amin, A., Khan, C., Ali, I. and Anwar, S., 2014 [16], to
useful in order to minimize the churn rate because tailored
formalize customer churn prediction where rough set theory
promotions can be offered to specific customers that are not
is used as one-class classifier and multi-class classifier to
satisfied. Thus, there is no formal notification from the
investigate the trade-off in the selection of an effective
customer of ending a contract term. Our goal is to predict
classification model for customer churn prediction.
the customer churn. The sooner these changing patterns are
Experiments were performed to explore the performance of
detected the more opportunities and time the company will
four different rule generation algorithms (i.e. exhaustive,
have to retain the customer. For this accuracy Area Under
genetic, covering and LEM2).
Curve (AUC) is used in classification analysis in order to
determine which of the used models predicts the classes
best. Although machine learning algorithms are usually

Retrieval Number: F3861049620/2020©BEIESP Published By:

DOI: 10.35940/ijitee.F3861.049620 Blue Eyes Intelligence Engineering
Journal Website: www.ijitee.org 1026 & Sciences Publication
 International Journal of Innovative Technology and Exploring Engineering (IJITEE)
ISSN: 2278-3075 (Online), Volume-9 Issue-6, April 2020

It is observed that rough set as one-class classifier and participation. Author analyze suggested two Genetic
multi-class classifier based on genetic algorithm yields more Algorithm(GA) based neural network (NN) models to
suitable performance as compared to the other three rule predict the customer churn, improved balance random forest
generation algorithms. (IBFR) model which is a combination of balanced random
Bi, W., Cai, M., Liu, M. and Li, G., 2016 [17], find he problem forests and weighted random forests in order to overcome
of dealing with big data in the industry, existing churn the data distribution problems, generalized additive models
prediction models cannot work very well and decision makers and many other researchers work, hence many approaches
are always faced with imprecise operations management. So of churn prediction models has low accuracy and prediction.
proposed clustering algorithm called semantic-driven Later a good prediction model is required in order to avoid
subtractive clustering method (SDSCM). Author results the customer churn problem. This can be achieved by
indicate that SDSCM has stronger clustering semantic strength combining SVM with boosting algorithms for higher
than subtractive clustering method (SCM) and fuzzy c-means accuracy and performance which can be considered as a
(FCM). Then, a parallel SDSCM algorithm is implemented future work for Churn prediction.
through a Hadoop MapReduce framework. From their
approach, the proposed algorithm enjoys a fast running speed Gavril et al. 2016 [22] presented an advanced methodology of
when compared with the other methods. data mining to predict churn for prepaid customers using
Lu, N., Lin, H., Lu, J. and Zhang, G., 2012 [18] in their dataset for call details of 3333 customers with 21 features, and
research work conducts a real-world study on customer churn a dependent churn parameter with two values: Yes/No. Some
prediction and proposes the use of boosting to enhance a features include information about the number of incoming and
customer churn prediction model. Most of the research work
outgoing messages and voicemail for each customer. The
that uses boosting as a method to boost the accuracy of a given
author applied principal component analysis algorithm “PCA”
basis learner. Lu, N., Lin, H., Lu, J. and Zhang, G., tries to
to reduce data dimensions. Three machine learning algorithms
separate customers into two clusters based on the weight
assigned by the boosting algorithm. From this a higher risk were used: Neural Networks, Support Vector Machine, and
customer cluster has been identified. Logistic regression is used Bayes Networks to predict churn factor. The author used AUC
as a basis learner in this research, and a churn prediction model to measure the performance of the algorithms. The AUC values
is built on each cluster, respectively. The result is compared were 99.10%, 99.55% and 99.70% for Bayes Networks, Neural
with a single logistic regression model. Experimental networks and support vector machine, respectively. The dataset
evaluation reveals that boosting also provides a good separation used in this study is small and no missing values existed.
of churn data; thus, boosting is suggested for churn prediction Huang F, Zhu M, Yuan K, Deng EO. 2015 [23] studied the
analysis. problem of customer churn in the big data platform. The
Prashanth, R., Deepak, K. and Meher, A.K., 2017 [19], find goal of the researchers was to prove that big data greatly
the Churn prediction is an important factor to consider for enhance the process of predicting the churn depending on
Customer Relationship Management. This paper statistical
the volume, variety, and velocity of the data. Dealing with
and data mining technique were used for churn prediction.
data from the Operation Support department and Business
Prashanth, R.,et al use linear (logistic regression) and non-
linear techniques of Random Forest and Deep Learning Support department at China’s largest telecommunications
architectures including Deep Neural Network, Deep Belief company needed a big data platform to engineer the
Networks and Recurrent Neural Networks for prediction. fractures. Random Forest algorithm was used and evaluated
This is the first time that a comparative study of using AUC.
conventional machine learning methods with deep learning Burez, J. and Van den Poel, D., 2009 [24] studied the
techniques have been carried out for churn prediction. It is problem of unbalance datasets in churn prediction models
observed that non-linear models performed the best. Such and compared performance of Random Sampling, Advanced
predictive models have the potential to be used in the Under-Sampling, Gradient Boosting Model, and Weighted
telecom industry for making better decisions and customer Random Forests. They used (AUC, Lift) metrics to evaluate
management. the model. the result showed that under sampling technique
Huang, B., Kechadi, M.T. and Buckley, B., 2012 outperformed the other tested techniques.
[20] presents a new set of features for land-line customer Awang, M.K., Rahman, M.N.A. and Ismail, M.R., 2012 [25]
churn prediction, including 2 six-month Henley A regression based churn prediction model was presented.
segmentation, precise 4-month call details, line information, This method identifies churn by using multiple regressions
bill and payment information, account information, analysis. This technique utilizes the customer’s feature data
demographic profiles, service orders, complain information, for analysis and proposes to provide good performance. This
etc. Then the seven prediction techniques (Logistic technique differs from the other proposed techniques by the
Regressions, Linear Classifications, Naive Bayes, Decision fact that most of the techniques are only able to identify the
Trees, Multilayer Perceptron Neural Networks, Support customers who will instantaneously churn.
Vector Machines and the Evolutionary Data Mining Zhu, B., Xiao, J. and He, C., 2014 [26]. Class imbalance
Algorithm) are applied in customer churn as predictors, plays a major role in affecting the reliability of a classifier.
based on the new features. Finally, the comparative The major issue existing due to class imbalance is that the
experiments were carried out to evaluate the new feature set minority class is not well represented and hence the
and the seven modelling techniques for customer churn classifier is undertrained on the minority classes. The
prediction. technique proposed by Zhu et al. proposes to eliminate this
Saran Kumar, A. and Chandrakala, D., 2016 [21] reviews issue by using transfer learning techniques.
the most popular machine learning algorithms used by
researchers for churn predicting, not only in banking sector
but also other sectors which highly depends on customer

Retrieval Number: F3861049620/2020©BEIESP Published By:

DOI: 10.35940/ijitee.F3861.049620 Blue Eyes Intelligence Engineering
Journal Website: www.ijitee.org 1027 & Sciences Publication
 Predicting Churn Customer in Telecom using Peergrading Regression Learning Technique

The approach presented operates by training the classifier C.Feature Selection

using customer related behavioral data obtained from related Once noisy data is removed, the next step is to select the
domains. This approach has its major focus on the banking important features from the sample taken for processing. Here,
industry and the results are proposed to exhibit enhanced we suggest that firstly the features should be selected on the
performance. basis of domain knowledge [29, 30]. For instance, there are
some common churn indicators such as spend rate decline,
III. PROPOSED METHODOLOGY outbound call decrease, increase on days with zero balance,
decrease in recharge frequency, balance burn out rate, increase
Customer churn is the action of the customer who is like to
in calls to one competitor, decrease in voice call frequency and
leave the company and it is one of the mounting issues of increase in inactive days. Such indicators are very helpful in
today’s rapidly growing and competitive telecommunication filtering out the variables which are not relevant to the
industry. To minimize the customer churn, prediction prediction task then to correlate the data. In general, these are
activity to be an important part of the telecommunication the main indicators which help analysts and domain experts in
industry’s vital decision making and strategic planning identifying the potential churners before they actually decide to
process. Customers’ churn is a considerable concern in churn. With the help of these indicators, analysts can easily
service sectors with high competitive services. On the other remove the non-informative features. Then select the
hand, predicting the customers who are likely to leave the information rich variables to process further. Then feature
company will represent potentially large additional revenue correlation will be done, this can be extract by using 2 variables
source if it is done in the early phase. The main objective of A and B to the correlation parameter ρ .
this research is to produce a predictive model with better period of time a specific customer is likely to churn or
results that assess customer churn rate of telecommunication receive some probability estimate of churn per customer
companies. The model developed in this work uses machine [34]. √ ( ) ( )

learning techniques on customer churn data. Proposed

approach of peer grading regression method where used to
predict the customer churn. Peer grading regression is the
process of customer churn reviewing work which is D.Imbalance in output class labels
scientifically and empirically investigation and testing of An important observation from our data-sets is the
prediction model which encompassing relationships between imbalance in data. On an average about 5-10% customers
different constructs such as important churn related churn year-on-year basis depending on the segment we are
variables, switching reasons, service looking at [31, 32]. This imbalance in distribution consisting
usage, costs and behavior. This is based on the method of returning/non-returning customers is a good recipe for
X ⊆C and B is an equivalence relation in information learning algorithms to classify a large number of customers
system IS= (C, B). X belongs to customer; C indicates the
customer of the specific telecom industry. Here B specifies under returning and still attain high overall accuracy. We
the balancing customer, which shows the will be churn or employ cross-validation technique based on confusion
not churn. matrix AUC curve to provide a best predictor model. Main
A. Dataset advantage of this model is proportion responders. It gives
The data was processed to convert it from its raw status into the rate of successful classification. It is the metric that is
features to be used in machine learning algorithms. This used to measure how well the model can distinguish two
process took the longest time due to the huge numbers of classes.
columns. Firms in telecommunication sector have detailed call
records. These firms can segment their customers by using call
records for developing price and promotion strategies from that
complaint call also registered, by analyzing the customer
complaints and their customer care executive’s answers rating
we can easily identify the customer who will be going to churn. Number of occurrence C represents the total number of
Table 1 and Table 2 shows some of the data from 2018- 2019, occurrences of samples belonging to a given class
which is used for processing this implementation from cell 2 E.Classification using Peer Grading Regression
cell telecom industries the link where referred in [ 27]. In this step, identified churners are categorized based on
B.Data Filter and Noise Removal complaints data provided by the operator. Complaints
In the next step, we remove the noisy data from the sample dataset contain the information like complaint type, number
taken in our last step. Noise hinders in most of the data of complaints and resolution time that is whether complain
analysis, so removing noise is most important goal of data resolved within agreed time or not. Complaint type is
cleaning. Noisy data can lead to inappropriate and erroneous marked by operator either critical or normal, based on nature
results, so removing such data before training a classifier is of complaint registered by customer [33].
very important step [28]. Duplicate or missing data may also In general, identified churners with critical complaints have
cause incorrect or even misleading results. In telecom dataset, been categorized into ‘High’ whereas the customer with
generally, there exist a lot of garbage, missing, inappropriate ‘Normal’ complaint types are categorized as ‘Medium’ and
and incomplete values like spaces, ‘Null’ or any other special rest are marked as ‘Low’. There are three factors involved in
characters that need to be removed or replaced with appropriate deciding the categorization of a churner. First is complaint
values. For instance, most of the CDR data columns have type, followed by the number of complaints registered by
“Null” values and those could be replacing with “0” for the same customer and finally the time taken to resolve the
efficient processing. In order to remove inconsistencies, outliers problem.
are identified and removed from sample at this stage.

Retrieval Number: F3861049620/2020©BEIESP Published By:

DOI: 10.35940/ijitee.F3861.049620 Blue Eyes Intelligence Engineering
Journal Website: www.ijitee.org 1028 & Sciences Publication
 International Journal of Innovative Technology and Exploring Engineering (IJITEE)
ISSN: 2278-3075 (Online), Volume-9 Issue-6, April 2020

In case complaint is not resolved within agreed time and Msg < AvgMsg add PC to MsgList Add Submission
customer has registered the same complaint more than once s1 check till s30) // process continues daily up to 30
with type ‘Critical’, then churner severity is marked as ‘High’. days
Similarly, churn severity is marked as ‘Medium’ against which IF PC drop in more than 8 List then
there is only a single complaint is registered under complaint and add to separate List Grade G which belongs
type ‘Normal’ and resolved within agreed time. to Churns
Churn customer classification is applied by a peer grading ELSE IF PC Complaint Severity is Medium
regression task, Regression analysis is a statistical technique to getMaxUsage(Voice_Call, Data, Sms)
estimate the relationship between a target variable and other and add to separate List
data values that influence the target variable, expressed in ELSE
continuous values. If that’s too hard – the result of regression is No Campaign
always some number, while classification always suggests a END_IF
category. In addition, regression analysis allows for estimating END_IF
how many different variables in data influence a target variable. END_IF
With regression, businesses can forecast in what END_IF
Peer grading is the process of reviewing each variable of END FOR
data. Reviewing of data will helps to analyze the task fully. Return List[Voice], List[Data] and List[Msg]
Here customer care call is to be analyzed then if the
customer provides bad review for that call then we apply the In order to validate and demonstrate the accuracy of our
peer grading algorithm to get the updates after that customer proposed model, implemented and that were shown in next
call, here we collect all the records like their call activity, section with the telecom dataset. Proposed work flow
data activity and message activity for verifying that the diagram is shown in Fig 1.
customer will be churn or not to churn.
We assume that the true scores of all submissions S. Each
grader g has an inherent bias(g) and a certain reliability
customer(C).

True difference in means is not equal to 0

Fix analysis of customer after their bad review calls and
provide the submission s1,s2,... finding their activity for
minimum 1 month. To this exercise have been handed in by
peer grading mechanism. Consider a set of graders g1,g2,.... By
score(s,g) we denote the score given to submission s by
grader g. here the grades tells that the activity to our
predictor churn. Here the average prediction is calculated by
following equation.
Average Prediction = Average of Labeled data - Average
predicted data
Following algorithm shows the peer grading algorithm Fig 1: Flow diagram of PGR approach
INPUT:
Average Non-Churn Voice Data IV. EXPERIMENTAL RESULT
Average Non-Churn Message Data
We focused on evaluating and analyzing the performance of
Average Non-Churn Data Data
the machine learning methods and algorithms for predicting
Complaint Cruelty (low, High, Medium)
churn in telecommunications companies. We have
Possible_Churners PC
experimented a number of algorithms such as Peer Grading
VoiceCus, DataCus, and SmsCus of all PCs
Regression (PGR), Random Forest, Decision Tree and k-
OUTPUT:
NN classifier to build the predictive model of customer
Churner Data_Classification (‘Voice’, ‘Sms’, and
Churn after developing Noise Removal, Feature Selection,
‘Data’)
Imbalance in output class labels and Classification methods.
FOR each NC Here, simulation is carried out in MATLAB environment as
// Finding the priority of customer complaints it is a user friendly tool. This tool establishes proper
Check Complaint_Seriousness (High, Medium, Low) validation of classification in terms of accuracy, sensitivity,
// Analyzing the complaints and if the complaints is specificity, F-measure and so on. This software runs on Intel
in high then further process is continuing core i3 processor with certain configuration. To evaluate the
IF Complaint_Seriousness is High then performance of tested classifiers, we use the churn dataset
Check their activity(peer) from the UCI Machine Learning Repository. These
// Check condition of Voice, SMS and Data techniques analyze the data and identify reasons behind
IF (Voice_call < AvgVoice_Call add PC to VoiceList customer churning.
AND
Data < AvgData add PC to DataList AND

Retrieval Number: F3861049620/2020©BEIESP Published By:

DOI: 10.35940/ijitee.F3861.049620 Blue Eyes Intelligence Engineering
Journal Website: www.ijitee.org 1029 & Sciences Publication
 Predicting Churn Customer in Telecom using Peergrading Regression Learning Technique

CRM can employ these techniques to maximize their profit. convention information like Customer ID, Monthly
Furthermore, it may be used to design retention strategies to Minutes, Roaming Calls, Blocked Calls, Customer Care
reduce the ratio of customers that are going to churn. Calls, Received Calls, Outbound Calls, Inbound Calls, Call
Forwarding, Months in Service, Service Area. Some of the
A.Dataset data were attached here which is shown in table 1.

Here, data were collected from online available Telecom

industries of cell 2 cell dataset for validation. Customer data
contains all data related to customer’s services and
Table I: Dataset parameters of telecom customer
Call Months

Customer Monthly Roaming Blocked Customer Received Outbound Inbound Forwarding In

ID Minutes Calls Calls Care Calls Calls Calls Calls Calls Service Service Area
3000006 483 0 1 1.7 55.3 46.3 6.3 0 56 MILMIL414
3000018 570 0 0.7 8.7 106.3 14.7 0.7 0 57 NNYSYR315
3000034 1039 0 3 11.3 213.9 103.3 0.3 0 55 MILMIL414
3000070 153 0 0.7 0 33.1 8 4.7 0 59 SLCSLC801
3000074 1213 1.3 6 4.3 490.1 50.7 8.3 0 52 OKCTUL918
3000086 1424 0 0 0.3 939.3 7.7 39 0 54 SEAPOR503
3000098 300 0 0.7 1 77.5 6.3 5 0 52 SLCSLC801
3000110 972 0 0 0 244.6 20 8 0 58 PITHOM412
3000246 406 0 1.7 0 47.7 6.3 4.3 0 50 SANMCA210
3000254 2961 62.3 9.3 0 871.5 210.7 96.3 2.3 52 NSHNSH615
3000258 244 0 2.7 0 48.8 13 1 0 54 KCYKCK913
3000274 380 0 0.7 0 76.2 21.7 4.7 0 56 DENDEN303
3000334 1088 0 1 3.7 335.1 11 6 0 54 OKCLRK501
3000338 1348 1.9 7.3 0 436.5 59.3 23 0 56 SANAUS512

3000342 1804 0 11 0 742.4 133 31.7 0 50 NSHNSH615

3000366 365 0 7 3.3 99.9 9.3 2 0 50 KCYWIC316
3000370 1306 0 1 1.7 809.1 25.3 0 0 52 KCYKCM816
3000438 1656 0 9.7 9.3 197.9 42.7 0.3 0 52 DALSHR903
3000450 25 0 0 0 1.6 0 0 0 50 NSHNSH615

There are many things brands may do wrong, from complicated onboarding when customers aren’t given easy-to-understand
information about product usage and its capabilities to poor communication, e.g. the lack of feedback or delayed answers to
queries. Another situation: Longtime clients may feel unappreciated because they don’t get as many bonuses as the new ones. And
also longtime customer may feel for dissatisfied customer care support call. For this Churn Analysis is applied to this research why
customersswitch service provider based on some data like customer call data. In a Churn analysis application, the first thing is to
access to the customer data. Table 1 shows the customer data to process or identify the churn. Then, factors are classified to decide
which factor or factors affect customer churn decision. In general, it’s the overall customer experience that defines brand perception
and influences how customers recognize value for money of products or services they use.

TableII: Churn identification dataset through customer care calls

Total Customer Director Un Not Un

Customer Recurring Care Assisted answered respond satisfied Unique Active
ID Charge Calls Calls Calls calls customer Subs Subs Service Area Credit Rating
3000002 22 6.3 0.25 0.7 0.7 0 2 1 SEAPOR503 5-satisfied
3000010 17 2.7 0 0.3 0 0 1 1 PITHOM412 5-satisfied
3000014 38 0 0 0 0 0 1 1 MILMIL414 3-Good
3000022 75 76 1.24 52 7.7 4.3 2 2 PITHOM412 4-Medium
3000026 17 0 0 0 0 0 2 2 OKCTUL918 1-Highest

Retrieval Number: F3861049620/2020©BEIESP Published By:

DOI: 10.35940/ijitee.F3861.049620 Blue Eyes Intelligence Engineering
Journal Website: www.ijitee.org 1030 & Sciences Publication
 International Journal of Innovative Technology and Exploring Engineering (IJITEE)
ISSN: 2278-3075 (Online), Volume-9 Issue-6, April 2020

3000030 52 13 0.25 9 1.7 0.7 1 1 OKCTUL918 3-Good

3000038 30 2.3 0.25 0 1 0 2 2 OKCTUL918 5-satisfied
3000042 66 4 2.48 0 0.3 4 2 2 OKCOKC405 0- Dissatisfied
3000046 35 1 0 0 0 0 3 3 SANMCA210 5-satisfied
3000050 75 0 0 0 0 0 1 1 PITHOM412 3-Good
3000054 25 0.3 0 0 0 0 2 2 SANMCA210 5-satisfied
3000058 85 43.7 2.23 9 0 0.3 5 1 SLCSLC801 5-satisfied
3000062 37 7.7 0.25 3.3 1.7 1 2 2 OKCOKC405 3-Good
3000066 60 17.3 0 5 0 3.7 1 1 SLCSLC801 3-Good
3000078 70 9 0 1.7 0.3 0.3 1 1 MILMIL414 3-Good
3000082 100 114.3 0 7.3 18 4 2 1 LOULOU502 0- Dissatisfied
3000102 30 2.3 0 0.3 0 4 3 3 SLCSLC801 0- Dissatisfied
3000118 30 0.7 0 0 0 0 1 1 SANMCA210 0- Dissatisfied
3000122 17 0 0 0 0 0 2 1 KCYKCK913 1-Highest
3000126 30 5.3 0 2 1.7 0 1 1 SANMCA210 4-Medium
3000130 35 3.3 0.25 0 0.7 0 1 1 OKCOKC405 1-Highest
3000134 17 7.3 0 1.3 3 0 1 1 KCYNEW316 1-Highest
3000138 30 1.3 0 0.3 0 0 2 1 SLCSLC801 1-Highest
3000142 75 53.3 0.5 11 5 2.7 1 1 KCYKCM816 1-Highest
3000146 30 3 0 2 2 0 2 2 KCYKCM816 1-Highest
3000158 30 6.3 0 6.7 0.3 0 2 1 DENDEN303 1-Highest
Table 2 shows the dataset of customer care call data like their applicability of these in a real world environment where large
satisfactory level, dissatisfactory level and the level of their amount of data is present. which don’t provide a true
opinion from customer care support, from this it can analyze representation of real world telecom data consisting of noise
that what problem is made and that solution is satisfied or not, and large number of missing values. The presence of noise and
from their opinion if the person provides dissatisfactory level missing values in the variables degrades the performance of the
then we have to watch their future response about the calls, prediction models. Here 50% of the ‘control list’ customers
data. If the response gets low means it is possible for churn and have churned, retention would be 50% effective. From C5
in that we can predict the customer churn easily. Model incorrectly predicted 19% of churners, who were not
B.Discussion actually churners, but it marked them as churners. For False
This section is going to discuss the comparison of the most cases approach, the model was able to correctly predict 71% of
prominent classification methods on churn prediction. There the non-churners correctly. As one of the important steps to
are a lot of forms of the boosting algorithms, but the most ensure the model generalizes well, the performance of the
popular is AdaBoost, where the weak classifiers are decision predictive churn model has to be evaluated. That is, the
trees. Linear Regression (LR), classification based on Support prediction rates of a predictor are needed to be considered. In
Vector Machine (SVM), Association Rules, advanced rule this work, the prediction rates refer to true churn rate (TP) and
induction, Decision Tree (DT), ensemble of hybrid methods, false churn rate (FP). The objective of the application is to get
churn prediction by feature selection techniques, Bayesian high TP with low FP.
network classifiers and improved balanced random forests. The To assess the classification results we count the number of
primary objective of these previously reported prediction True Positive (TP), True Negative (TN), False Positive (FP)
models is to utilize large amount of telecom data to identify and False Negative (FN). The FN value actually belongs to
potential churners. However, this suffer from a number of Positive P (e.g. TP + FN = P) but wrongly classified as
limitations which put strong barriers towards the direct Negative N (e.g. TN + FP = N).

TableIII: Gender Computation

Retrieval Number: F3861049620/2020©BEIESP Published By:

DOI: 10.35940/ijitee.F3861.049620 Blue Eyes Intelligence Engineering
Journal Website: www.ijitee.org 1031 & Sciences Publication
 Predicting Churn Customer in Telecom using Peergrading Regression Learning Technique

Table IV: Churn Age Factor Computation

Age No Churn presented Churn presented

15- 25 45 5

26- 35 35 4

36-45 18 3

46-55 22 1

Above 55 13 2

Table V: Prediction based on Customer complaint

Month Repeated Customer Rare Customer

Complaint call Complaint call

Jan- Mar 2 .2

Apr- Jun 1.1 .5

Jul- Sep 2.3 .4

Oct- Dec 2.4 .5

Table VI: Churner categorization based on usage pattern

Category Data SMS Voice

High 12 83 74

Medium 566 712 1,349

Low 812 834 2,209

Table VII: Prediction comparison

Prediction Peer Grading Existing work

Calculus Regression

89 72
Churn prediction = Number of cases satisfied – whole
Customer call count
Accuracy: Overall accuracy of the classifier can calculate by the given formula
Accuracy = TP + TN / P + N
TableVIII: Recall computation

Method Recall

Peer Grading Regression(PGR) 88.642

Random Forest 74.444
Decision Tree 80.678
k-NN classifier 82.909

Retrieval Number: F3861049620/2020©BEIESP Published By:

DOI: 10.35940/ijitee.F3861.049620 Blue Eyes Intelligence Engineering
Journal Website: www.ijitee.org 1032 & Sciences Publication
 International Journal of Innovative Technology and Exploring Engineering (IJITEE)
ISSN: 2278-3075 (Online), Volume-9 Issue-6, April 2020

Sensitivity(Recall): It measures the fraction of churn customers who are correctly identified as true churn
Recall = TP / P
Table IX: Precision computation

Method Precision

Peer Grading Regression(PGR) 89.022

Random Forest 74.444

Decision Tree 68.678

k-NN classifier 79.909

Precision: It is characterized the number of correctly predicted churns over the total number of churns predicted by proposed
approach. It can formally express as;
Precision = TP / TP + FP
Precision or recall alone cannot describe the efficiency of a classifier since good performance in one of those indices does
not necessarily imply good performance on the other. For this reason, F-measure, a popular combination.
Table X: F-measure computation

Method F-measure

Peer Grading Regression (PGR) 86.022

Random Forest 74.444

Decision Tree 78.678

k-NN classifier 69.909

F- Measure = precision * recall / precision + recall
Table XI: Computation of AUC Curve

TableXII: Churner categorization based on complaints

Table XIII: AUC performance measure for churn predictive model

Months Jan, Mar, May, Jul, Sep, Nov,

Feb Apr Jun Aug Oct Dec

Retrieval Number: F3861049620/2020©BEIESP Published By:

DOI: 10.35940/ijitee.F3861.049620 Blue Eyes Intelligence Engineering
Journal Website: www.ijitee.org 1033 & Sciences Publication
 Predicting Churn Customer in Telecom using Peergrading Regression Learning Technique

PGR 5 5.5 4.2 3.6 6 8

Random 2 2.3 3.4 3.9 1.9 4

Forest

Decision 1 4 3 .9 2 2
Tree

k-NN 4 3 5 4 2 2
classifier
A classified data sample is taken from original CDR dataset. Proposed approach exactly find the consistent ratio of churner and
non-churners from the dataset. There are total 20000 datasets were selected in which 9.1% are churners. Below the chart shows the
graphical representations of our work.
F-Measure: A composite measure of precision and recall to
compute the test’s accuracy. It can be interpreted as a
weighted average of precision and recall.

Fig 1: Gender Computation

Fig 3: Prediction based on Customer complaint

Fig 2: Churn Age Factor Computation Fig 4: Churner categorization based on usage pattern

Retrieval Number: F3861049620/2020©BEIESP Published By:

DOI: 10.35940/ijitee.F3861.049620 Blue Eyes Intelligence Engineering
Journal Website: www.ijitee.org 1034 & Sciences Publication
 International Journal of Innovative Technology and Exploring Engineering (IJITEE)
ISSN: 2278-3075 (Online), Volume-9 Issue-6, April 2020

Fig 7: Precision computation

Fig 5: Prediction comparison

Fig 6: Recall computation Fig 8: F-measure computation

Fig 9: AUC Computation

Retrieval Number: F3861049620/2020©BEIESP Published By:

DOI: 10.35940/ijitee.F3861.049620 Blue Eyes Intelligence Engineering
Journal Website: www.ijitee.org 1035 & Sciences Publication
 Predicting Churn Customer in Telecom using Peergrading Regression Learning Technique

Fig 10: Churner categorization based on complaints

Fig 11: AUC performance measure for churn predictive model

V. CONCLUSION
REFERENCES
1. Idris, A. and Khan, A., 2017. “Churn prediction system for telecom
Churn is an unavoidable procedure in the telecom industry. In using filter–wrapper and ensemble classification”. IEEE, 60(3),
any case, however, troublesome, it is conceivable to recognize pp.410-430.
the reasons for churn usages by a few methodologies. 2. Ullah, I., Raza, B., Malik, A.K., Imran, M., Islam, S.U. and Kim,
S.W., 2019. “A churn prediction model using random forest: analysis
Prediction plays a vital role in identifying the churn customer. of machine learning techniques for churn prediction and factor
In this research work, the prediction model has been proposed identification in telecom secto”r. IEEE Access, 7, pp.60134-60149.
for the telecommunication sector for the customer churn using 3. Azeem, M. and Usman, M., 2018. A fuzzy based churn prediction and
retention model for prepaid customers in telecom industry.
peer grading regression method. This model provides that International Journal of Computational Intelligence Systems, 11(1),
reviewing the customer process in daily activity and analyze the pp.66-78.
4. Anjum, A., Zeb, A., Afridi, I.U., Shah, P.M., Anjum,A., Raza, B.
maximum possible churn activity and predict the customer and Anwar, Z., 2017. “Optimizing Coverage of Churn Prediction in
churn process accurately. Telecommunication Industry”. International Journal of Advanced
By applying our proposed method, it predicts Computer Science and Applications, 8(5), pp.179-188.
5. https://towardsdatascience.com/hands-on-predict-customer-churn-
the churn customer exactly by analyzing the activity of the 5c2a42806266
customer. As many approach said that holding the existing 6. Adwan, O., Faris, H., Jaradat, K., Harfoushi, O. and Ghatasheh, N.,
customer is more precious than getting new customer. So to 2014. Predicting customer churn in telecom industry using multilayer
preceptron neural networks: Modeling and analysis. Life Science
provide a enhancing approach our proposed model exploits the Journal, 11(3), pp.75-81.
classifiers to accurately predict the churners from an enormous 7. Seo, D., Ranganathan, C. and Babad, Y., 2008. Two-level model of
customer retention in the US mobile telecommunications
arrangement of customer records dependent on the machine
servicemarket. Telecommunications policy, 32(3-4), pp.182-196.
learning approach. Our result of experimental performed on the 8. Hung, S.Y., Yen, D.C. and Wang, H.Y., 2006. Applying data mining
CDRs dataset of a cell two cell telecom organization proved to telecom churn management. Expert Systems with Applications,
31(3), pp.515-524.
that the proposed model achieves best result by applying AUC 9. Coussement, K., Benoit, D.F. and Van den Poel, D., 2015. Preventing
score of true positive range of 89% accuracy. customers from running away! Exploring generalized additive models
for customer churn prediction. In The Sustainable Global Marketplace
(pp. 238-238). Springer, Cham.

Retrieval Number: F3861049620/2020©BEIESP Published By:

DOI: 10.35940/ijitee.F3861.049620 Blue Eyes Intelligence Engineering
Journal Website: www.ijitee.org 1036 & Sciences Publication
 International Journal of Innovative Technology and Exploring Engineering (IJITEE)
ISSN: 2278-3075 (Online), Volume-9 Issue-6, April 2020

10. Qureshii SA, Rehman AS, Qamar AM, Kamal A,Rehman A. Industry. International Journal of Database Theory and Application,
2013 “Telecommunication subscribers’ churn prediction model using 9(8), pp.211-232.
machine learning”. In:Eighth international conference on digital 35. T. Verbraken, W. Verbeke, B. Baesens, “A novel profit maximizing
information management.. p. 131–6. metric for measuring classification performance of customer churn
11. Tiwari, A., Hadden, J. and Turner, C., 2010, March. A new neural prediction models”,IEEE Transaction on Knowlee and Data
network based customer profiling methodology for churn prediction. Engineering 25 (2013) 961–973
In International Conference on Computational Science and 36. Koen W. De Bock, Dirk Van den Poel, “An empirical evaluation of
ItsApplications (pp. 358-369). Springer, Berlin, rotation-based ensemble classifiers for customer churn prediction”,
12. Heidelberg. Expert Systems withApplications 38 (2011) 12293–12301
13. Adnan Amin, Shehzad, S., Khan, C., Ali, I. and Anwar,S., 2015.
“Churn prediction in telecommunication industry using rough set
approach”. In New trends in computational collective intelligence (pp.
AUTHORS PROFILE
83-95). Springer, Cham. Dr. M. Hemalatha completed M.Sc.,
14. Hanif, E., 2019. “Applications of data mining techniques for churn M.C.A., M. Phil., Ph.D (Ph. D, Mother Terasa
prediction and cross-selling in the telecommunications industry” women's University, Kodaikanal). She is
(Doctoral dissertation, Dublin Business School).
currently working as a professor in Ramakrishna
15. Ahmad, A.K., Jafar, A. and Aljoumaa, K., 2019.“Customer churn
prediction in telecom using machine learning in big data platform”. College of Arts & Science,Coimbatore .Having
Journal of Big Data, 6(1), p.28. Springer. fifteen years of experience in teaching and
16. Vafeiadis, T., Diamantaras, K.I., Sarigiannidis, G. andChatzisavvas, published more than two hundred papers in
K.C., 2015. “A comparison of International Journals and also presented more than hundred papers in
machinelearningtechniquesforcustomerchurnprediction”. Simulation various national and international conferences. She received best
Modelling Practice and Theory, 55, pp.1-9. researcher award in the year 2012 from Karpagam University. Her research
17. Amin, A., Khan, C., Ali, I. and Anwar, S., 2014,November. areas include Data Mining, Image Processing, Computer Networks, Cloud
“Customer churn prediction in telecommunication industry: With and Computing, Software Engineering, Bioinformatics and Neural Network.
withoutcounter-example”. In Mexican international conference on
She is a reviewer in several National and InternationalJournals.She
artificial intelligence (pp. 206-218). Springer, Cham.
received Women Excellance Award for the year 2020.
18. Bi, W., Cai, M., Liu, M. and Li, G., 2016. “A big data clustering
algorithm for mitigating the risk of customerchurn”. IEEE
Transactions on IndustrialInformatics, 12(3), S.Mahalakshmi Completed MCA.,Mphil.,SET and
pp.1270-1281. currently pursuing PhD in Bharathiar
19. Lu, N., Lin, H., Lu, J. and Zhang, G., 2012. “A customer churn University,Coimbatore.Published and presented
prediction model in telecom industry using boosting”. IEEE several papers in various national and International
Transactions on Industrial Informatics, 1(2), pp.1659-1665. conferences.Her field of interest include Data
20. Prashanth, R., Deepak, K. and Meher, A.K., 2017, July.“High Mining ,Machine Learning,Software Engineering
accuracy predictive modelling for customer churn prediction in etc.Having seven years of experience in the field of
telecom industry”. In International Conference on Machine Learning teaching as a Asst.Professor.
and Data Mining in Pattern Recognition (pp. 391-402). Springer,
Cham.
21. Huang, B., Kechadi, M.T. and Buckley, B.,
2012.“Customerchurnpredictionin telecommunications”. Expert
Systems with Applications, 39(1), pp.1414-1425.
22. Saran Kumar, A. and Chandrakala, D., 2016. “A Survey on
Customer Churn Prediction using Machine Learning Techniques”.
International Journal of Computer Applications, 975, p.8887.
23. Brandusoiu I, Toderean G, Ha B. Methods for churn prediction in the
prepaid mobile telecommunications industry. In: International
conference on communications. 2016. p. 97–100.
24. Huang F, Zhu M, Yuan K, Deng EO. 2015 “Telco churn prediction
with big data. In: ACM SIGMOD international conference on
management of data”. p.607–18.
25. Burez, J. and Van den Poel, D., 2009. Handling class imbalance in
customer churn prediction. Expert Systems with Applications, 36(3),
pp.4626-4636.
26. Awang, M.K., Rahman, M.N.A. and Ismail, M.R., 2012. Data mining
for churn prediction: multiple regressions approach. In Computer
Applications for Database, Education, and Ubiquitous Computing
(pp. 318-324). Springer, Berlin, Heidelberg.
27. Zhu, B., Xiao, J. and He, C., 2014. A balanced transferlearning model
for customer churn prediction. In Proceedings of the Eighth
International Conference on Management Science and Engineering
28. Management (pp. 97-104). Springer, Berlin,
29. Heidelberg.https://www.kaggle.com/vpfahad/telecom-churn-data-sets
30. Qureshi, S.A., Rehman, A.S., Qamar, A.M., Kamal, A.,Rehman, A.:
Telecommunication subscribers’ churnprediction model using
machine learning. In: Eighth International Conference on Digital
Information Management (ICDIM 2013), pp. 131–136. IEEE(2013)
31. Kandel, Ibrahem Hamdy Abdelhamid. "A comparative study of tree-
based models for churn prediction: a case study in the
telecommunication sector." PhD diss., 2019.
32. Le, M., Nauck, D., Gabrys, B. and Martin, T., 2013, December.
KNNs and sequence alignment for churn prediction. In International
Conference on Innovative Techniques and Applications of Artificial
Intelligence (pp. 279-285). Springer, Cham.
33. Zhu, B., Xiao, J. and He, C., 2014. A balanced transfer learning
model for customer churn prediction.In Proceedings of the Eighth
International Conference on Management Science and
EngineeringManagement (pp. 97-104). Springer, Berlin, Heidelberg.
34. Jayaswal, P., Prasad, B.R., Tomar, D. and Agarwal, S., 2016. An
Ensemble Approach for Efficient Churn Prediction in Telecom

Retrieval Number: F3861049620/2020©BEIESP Published By:

DOI: 10.35940/ijitee.F3861.049620 Blue Eyes Intelligence Engineering
Journal Website: www.ijitee.org 1037 & Sciences Publication