A Proposed Churn Prediction Model

Essam Shaaban, Yehia Helmy, Ayman Khedr, Mona Nasr / International Journal of Engineering
Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com
Vol. 2, Issue 4, June-July 2012, pp.693-697
693 | P a g e
A Proposed Churn Prediction Model
Essam Shaaban*, Yehia Helmy, Ayman Khedr, Mona Nasr**
*(Department of Information Systems, MUST University, 6 October, Cairo, Egypt
** (Department of Information Systems, Helwan University, Cairo, Egypt
ABSTRACT
Churn prediction aims to detect customers
intended to leave a service provider. Retaining one
customer costs an organization from 5 to 10 times
than gaining a new one. Predictive models can
provide correct identification of possible churners in
the near future in order to provide a retention
solution. This paper presents a new prediction model
based on Data Mining (DM) techniques. The
proposed model is composed of six steps which are;
identify problem domain, data selection, investigate
data set, classification, clustering and knowledge
usage. A data set with 23 attributes and 5000
instances is used. 4000 instances used for training the
model and 1000 instances used as a testing set. The
predicted churners are clustered into 3 categories in
case of using in a retention strategy. The data mining
techniques used in this paper are Decision Tree,
Support Vector Machine and Neural Network
throughout an open source software name WEKA.
Keywords:-Churn prediction, classification,
clustering, data mining, prediction model
1. INTRODUCTION
Churn prediction process is a highly debated
research area for more than ten years. Researchers from
different disciplines have tried to analyze this problem
from their own perspectives to figure out a clear
understanding and to recommend an effective solution
for churners in many business areas. Abbasimehr et al.
[1] state that churn prediction is a useful tool to predict
customer at churn risk. Conventional churn prediction
techniques have the advantage of being simple and
robust with respect to defects in the input data, they
possess serious limitations to the interpretation of
reasons for churn. Therefore, measuring the
effectiveness of a prediction model depends also on how
well the results can be interpreted for inferring the
possible reasons of churn [2]. The purpose of prediction
is to anticipate the value that a random variable will
assume in the future or to estimate the likelihood of
future events [3]. Most DM techniques derive their
predictions from the value of a set of variables
associated with the entities in a database. DM models
may be employed to predict customer churn developed
in many disciplines such as demographic data and/or
behavioral data. There are many DM techniques that can
be used in classification and clustering customer data to
predict churners in the near future. These
techniques may use Decision Tree (DT), Support Vector
Machine (SVM) in addition to Neural Networks (NN),
Genetic Algorithms (GA) or Fuzzy Logic (FL) to predict
churners.
This paper is organized as follows. Section 2
describes the types of churners. Section 3 shows the
existing prediction models rather than the techniques of
developing a predictive model. Section 4 describes the
proposed churn prediction model besides the results of
an implemented case study. Finally; conclusion and
future work are presented.
2. TYPES OF CHURNERS
As figure 1 depicts; There are two main
categories of churners which are voluntary and
involuntary [4]. Involuntary churners are the easiest to
identify. These are the customers that Telco decides to
remove from subscribers list. Therefore’ this category
includes people that are churned for fraud, non-payment
and customers who don’t use the phone. Voluntary
churner is more difficult to determine; it occurs when a
customer makes a decision to terminate his/her service
with the provider. When people think about Telco churn
it is usually the voluntary kind that comes to mind.
Figure 1: churn taxonomy
Voluntary churn can be sub-divided into two
main categories, incidental churn and deliberate churn.
Incidental churn occurs, not because the customers
planned on it but because something happened in their
lives. For example: change in financial condition churn,
change in location churn, etc. Deliberate churn happens
for reasons of technology (customers wanting newer or
better technology), economics (price sensitivity), service
quality factors, social or psychological factors, and
convenience reasons. Deliberate churn is the problem
that most churn management solutions try to solve [4]
[5].

694 | P a g e
3. PREDICTIVE MODELS
Predictive modeling is mainly concerned with
predicting how the customer will behave in the future by
analyzing their past behavior. Predicting customers who
are likely to churn is one example of the predictive
modeling [6]. Predictive modeling is used in analyzing
Customer Relationship Management (CRM) data
and DM to produce customer-level models that describe
the likelihood that a customer will take a particular
action. The actions are usually sales, marketing and
customer retention related. There are many models that
can used to define distinguish between churners and non-
churners in an organization. These models can be
classified into traditional models or techniques (RA and
DT) and soft computing techniques (FL and NN) [2].
3.1 Traditional techniques
3.1.1 Decision trees
DT is most popular type of predictive model. It
has become an important knowledge structure, used for
the classification of future events [7]. DT usually
consists of two main steps, tree building and tree
pruning. The tree-building step consists of recursively
partitioning the training sets according to the values of
the attributes. The partitioning process continues until
all, or most of the records in each of the partitions
contain identical values. Some branches may be removed
because it could consist of noisy data. The pruning step
involves selecting and removing the branches containing
the largest estimated error rate. Tree pruning is known to
enhance the predictive accuracy of the decision tree,
while reducing the complexity [8].
3.1.2 Regression Analysis
RA is another popular technique used to deal
with predicting customer satisfaction it is based on
supervised learning models. Regression models deal
with a dataset consisting of past observations, for which
both the value of the explanatory attributes and the value
of the continuous numerical target variable are known
[3].
3.2 Soft computing techniques
3.2.1 Neural Networks
NN has been successfully used to estimate
intricate non-linear functions. A NN is an analogous data
processing structure that possesses the ability to learn.
The concept is loosely based on a biological brain and
has successfully been applied to many types of
problems, such as classification, control, and prediction
[9]. NN is different from DT and other classification
techniques because they can provide a prediction with its
likelihood. Various neural network approaches have
emerged over time, each with varying advantages and
disadvantages (Liao et al., 2004), however greater detail
into these variances is beyond the scope of this paper.
Research suggests that neural networks outperform
decision trees and regression models for churn prediction
[8].
3.2.2 Fuzzy Logic (FL)
FL is a conceptually easy to understand. The
mathematical concepts behind fuzzy reasoning are very
simple. Naturalness of the approach makes it preferable
to the other techniques. FL is flexible, tolerant of
imprecise data, and it can model nonlinear functions of
arbitrary complexity. It can be blended with
conventional control techniques. In many cases fuzzy
systems expends the concept of the conventional control
techniques and simplify their implementation. Regarding
the telecom industry; there is no work achieved related
to churn prediction using the fuzzy techniques [10].
There are a lot of studies have implemented in the area
of telecom churn prediction. Summary of latest churn
prediction studies is shown in table 1.
Table 1: churn prediction studies
Year Author Technique
2001 Datta et al. [11] DT
2002 Ping and Tang [12] DT induction
2003 Au et al. [13] GA
2006 Ahn et al. [14] partial defection
2007 Junxiang [15]
Survival Analysis
Modeling
2008 Piotr [16] rough-sets
2008
Seo, and Ranganathan
[17]
Two-level model
2009 Jahromi et al [18] NN and DT
2010 Gotovac [4] DT
2011 Lee et al. [19]
partial least squares
(PLS) model
2011 Yeshwanth et al [20] Hybrid Learning
2011
Fasanghari and
Keramati [21]
Local Linear Model
Tree (LOLIMOT)
algorithm.
4. THE PROPOSED MODEL
The proposed model is composed of six steps.
As shown in figure 1, these steps are: identify problem
domain, data selection, investigate data set,
classification, clustering and knowledge usage. As figure
2 depicts; the classification step produces two types of
customers (churners and non-churners) while the
clustering step produces 3 clusters which are used to be
evaluated according to the retention strategy in further
usage.

695 | P a g e
Figure 2: The Proposed Churn Prediction model
The proposed model can produce more than 3
clusters based on the types of acquired knowledge.
Knowledge usage receives the produced clusters for
assign a retaining solution for each type of churners.
Churners can be clustered according to many criteria
such as profitability or dissatisfactory of customers.
4.1 Churn Prediction with the Proposed Model
During this case study an open source DM tool
named WEKA [22] is used in addition to a data set of
5000 instances. The data set is obtained from an
anonymous mobile service provider. As figure 2 depicts;
the data set is divided into a training set and a testing set.
The training set is 4000 (80%) instances and the testing
set is 1000 (20%) instances. The training data contains
3200 (80%) instances are labeled non-churners bot the
others are 800 (20%) are labeled churners.
Figure 2: the distribution of the data set
Figure 3: Training set
Figure 4: Testing set
The attributes for this data set are 23 attributes
as shown in table 2. The class of evaluating the status of
each instance in the a data set is named churn. Churn
attribute is labeled churn=True if a customer left the
service provider but it is labeled churn=False if he/she is
still continuing with the service provider. churn=True if
a customer left the service provider but it is labeled
churn=False if he/she is still continuing with the service
provider.
Table 2: attributes of the data set.
Attribute Data
Type
Attribute Description
Age Number
Customer’s age categorized
into five groups
Gender Number
Customer’s gender (1=male,
0=female)
mar_st Number
Customer’s marital status
(1=yes, 0=no)
M_in_sm
_MOU
Number
mean of in minutes of use from
the same service provider
M_out_s
m_MOU
Number
mean of out minutes of use
from the same service provider
M_sm_M
OU
Number
mean of all minutes of use
from the same service provider
M_in_oth
_MOU
Number
mean of in minutes of use from
other service provider
M_out_ot
h_MOU
Number
from other service provider
M_oth_
MOU
Number
mean of all minutes of use
from other service provider
M_in_M
OU
Number
mean of in minutes of use
either from the same or other
service provider
M_out_
MOU
Number
either from the same or other
service provider
MOU Number
mean of all in and our minutes
of use either from the same or
other service provider
chng_M
OU
Nominal
Change in minutes of use from
one month to another during
the time of the experiment
(decreased, Normal or
increased)
M_sms Number
The mean number of messages
during the time of the
experiment
M_M_re
v
Number Mean monthly revenues
Ass_prod Number Associated product (yes or no)
Ass_ser Number Associated services (yes or no)
M_CC_c
alls
Number
Mean number of customer care
calls
M_drop_
calls
Nominal Mean number of dropped calls
Compaint
s
Nominal
Mean number of complaints
no_chng_
tif_plan
Number
Number of changes in tariff
plan

696 | P a g e
max_call
_distance
Number
The maximum time distance
between 2 calls
CHURN Nominal Churn=True or False
4.2 Interpretation of Results
The result of the classification step is represented in
table 3.
Table 3: Confusion Matrix of results
Actual
Class
Actual Prediction
Decision
Tree
Non-
churners
Churners
Non-
churners
376/500
(75.2%)
124/500
(14.8%)
Churners
97/500
(19.4%)
403/500
(80.6%)
Neural
Networks
Non-
churners
420/500
(84%)
80/500
(16%)
Churners
83/500
(16.6%)
417/500
(83.4%)
Support
Vector
Machine
Non-
churners
416/500
(83.2%)
84/500
(16.8)
Churners
79/500
(15.8%)
421/500
(84.2%)
The accuracy and error rate of the predicted
results are shown is table 4. The accuracy rate and error
rate are computed as shown in the following equations
Accuracy = (no. of correct / Total no. of predictions)
Error rate = (no. of wrong / Total no. of predictions)
Table 4: Accuracy and error rates comparison for
DT, NN, SVM techniques.
Technique Accuracy Error rate
Decision Tree
(DT)
77.9%
(779/1000)
22.1%
(221/1000)
Neural
Networks (NN)
83.7%
(837/1000)
16.3%
(163/1000)
Support Vector
Machine (SVM)
83.7%
(837/1000)
16.3%
(163/1000)
The best classification results can be extracted
from tables 3 and 4 which is found in SVM classification
technique that is the most relevant with NN classification
technique for the data set in hand. SVM classification
technique predicts 421 churners from 500. The predicted
churners are used to be clustered in the next step.
During the clustering step the 421 predicted churners are
used to be clustered using simple K Means algorithm. 22
attributes are used in clustering as shown in table 5. The
clustering process use 3 types of clusters which one can
name Cluster 0 (Low), Cluster 1 (Medium) and Cluster 2
(High) according to the suitable situation of clustering.
The 3 resulting clusters can be assigned for profitability,
priority for retaining, or dissatisfactory.
Table 5: Clustering Output
3 Clusters with 421 instances
Cluster 0
(88) 20.9
%
Cluster 1
(208)
49.4%
Cluster 2
(125) 29.7%
Age 12 43 25
Gender Female Male Female
mar_st Single Married Single
M_in_sm_MOU 47 29 23
M_out_sm_MOU 12 45 12
M_sm_MOU 698 43 52
M_in_oth_MOU 41 45 65
M_out_oth_MOU 14 25 28
M_oth_MOU 77 52 158
M_in_MOU 50 54 69
M_out_MOU 730 142 85
MOU 194 232 210
chng_MOU Decreased Decreased Decreased
M_sms 5 25 1
M_M_rev 21.4 112.14 29
Ass_prod NO NO NO
Ass_ser YES NO NO
M_CC_calls 2 6 5
M_drop_calls 0 0 0
Compaints 2 1 0
no_chng_tif_plan 0 2 3
max_call_distance 0 0 0
The output of the clustering step can be used for the
knowledge usage step in order to assign a retention
strategy for a specific cluster or customer. In case of
clustering for a specific purpose such as customer’s
profitability, dissatisfactory, or cross selling this requires
a specific attribute selection from table 5.
5. CONCLUSION
Many churn prediction models and techniques
have been presented to date. However, a simple model is
required to distinguish churners from non- churners then
clustering the resulted churners for providing retention
solutions. In this paper, a simple model based on DM
techniques was introduced to help a CRM department to
keep track its customers and their behavior against
churn. A data set of 5000 instances with 23 attributes is
used to train and test the model. Using 3 different
techniques which are DT, SVM, and NN for
classification and simple K Means techniques for
clustering results indicate that the best output for the data
set in hand is SVM technique. The next stage of the
authors’ research will involve performing a deeper
analysis into the customer data to try to establish new
churn prediction retention model that will use the
predicted and clustered data to assign a suitable retention
strategies for each churner type.
6. REFERENCES
[1] H. Abbasimehr, M. Setak, M. Tarokh. A Neuro-
Fuzzy Classifier for Customer Churn Prediction.

697 | P a g e
International Journal of Computer Applications,
vol. 19, no. 8, pp. 35-41, April, 2011
[2] V. Lazarov and M. Capota. Churn
Prediction. Business Analytics Course. TUM
Computer Science, December 2007.
http://home.in.tum.de/~lazarov/files/research/pap
ers/churn-prediction.pdf
[3] Carlo Vercellis, Business Intelligence: Data
Mining and Optimization for Decision Making,
John Wiley & Sons, Ltd. 2009 ISBN: 978-0-470-
51138-1
[4] S. Gotovac. “Modeling Data Mining
Applications for Prediction of Prepaid Churn in
Telecommunication Services,” vol. 51, no. 3, pp.
275-283, 2010
[5] H. Kim, and C. Yoon, “Determinants of
subscriber churn and customer loyalty in the
Korean mobile telephony market.”
Telecommunications Policy. Vol. 28 No.: PP.
751-765, 2004.
[6] M. Hassouna, Agent Based Modelling and
Simulation: An Examination of Customer
Retention in the UK Mobile Market. PhD thesis,
Brunel University, UK, 2012.
[7] K. Muata, and O. Bryson, Evaluation of
Decision Trees: AMulti Criteria Approach,
Computers and Operational Research, 31, 1933-
1945, 2004
[8] W. Au, C. Chan, and X. Yao, A Novel
Evolutionary Data Mining Algorithm with
Applications to Churn Prediction, IEEE
transactions on evolutionary computation, 7, 6,
532-545, 2003.
[9] R. Behara, W. Fisher, and J. Lemmink,
Modelling and Evaluating Service Quality
Measurement Using Neural Networks,
International journal of operations and
production management, 22, 10, 1162-1185,
2002.
[10] Ö. SELVİ, Traffic Accident Predictions Based
On Fuzzy Logic Approach For Safer Urban
Environments, Case Study: Izmir Metropolitan
Area, PhD thesis, 2009.
[11] P Datta, B. Masand, D. Mani, and B. Li,
Automated Cellular Modeling and Prediction on
a Large Scale, Issues on the application of data
mining, 14, 485- 502, 2001
[12] C. Wei and I. Chiu, “Turning
telecommunications call details to churn
prediction: a data mining approach,” Expert
Systems with Applications, Elsevier , vol. 23,
no. 2, pp. 103-112, Aug. 2002.
[13] W. Au, , C. Chan, and X. Yao, A Novel
Evolutionary Data Mining Algorithm with
Applications to Churn Prediction, IEEE
transactions on evolutionary computation, 7, 6,
532-545, 2003
[14] H. Ahn, , P. Hana, and S. Lee, Customer churn
analysis: Churn determinants and mediation
effects of partial defection in the Korean mobile
telecommunications service industry.
552-568, 2006)
[15] L. Junxiang. Predicting Customer Churn in the
Telecommunications Industry – An Application
of Survival Analysis Modeling Using SAS. Data
mining techniques. SUGI 27. Paper 114, 2007
[16] S. Piotr. Global Perspectives Mobile Operator
Customer Classification in Churn Analysis.
Technical University of Szczecin, Poland SAS
Global Forum, pp. 1-5, 2008.
[17] B. Seo, C. Ranganathan, and Y. Babad,. “Two-
level model of customer retention in the US
mobile telecommunications service market.”
182-196, 2008
[18] A. tamaddoni, M. Moeini, I. Akbari, A.
Akbarzadeh, "A dual-step multi-algorithm
approach for churn prediction in Pre-paid
telecommunications service providers", the 6th
International Conference on Innovation &
Management, SÃO PAULO, Brazil, 2009.
[19] H. Lee, Y. Lee, H. Cho, K. Im, and Y. S. Kim,
“Mining churning behaviors and developing
retention strategies based on a partial least
squares (PLS) model,” Decision Support
Systems, July 2011.
[20] V. Yeshwanth, V. Vemal Raj, and M.
Sharavanan, “Evolutionary Churn Prediction in
Mobile Networks Using Hybrid Learning,”
Proceedings of the Twenty-Fourth International
Florida Artificial Intelligence Research Society
Conference, pp. 471-476, 2011
[21] M. Fasanghari and A. Keramati. Customer
Churn Prediction Using Local Linear Model
Tree for Iranian Telecommunication Companies.
pp. 25-37, July 2011.
[22] http://www.cs.waikato.ac.nz/ml/weka/

A Proposed Churn Prediction Model

More Related Content

What's hot (19)

Similar to A Proposed Churn Prediction Model (20)

Recently uploaded (20)

A Proposed Churn Prediction Model