Customer Churn Prediction in Mobile Operator Using
Combined Model
Jelena Mamčenko and Jamil Gasimov
Information Technologies Department, Vilnius Gediminas Technical University, Saulėtekio al. 11, Vilnius, Lithuania
Keywords: Data Mining, Churn Detection, Neural Network, Cox Regression, Decision Tree, Combined Model.
Abstract: Data Mining technologies are developing very rapidly nowadays. One of the biggest fields of application of
data mining is prediction of churn in service provider companies. Customers who switch to another service
provider are called churned customers. In this study are described main techniques and processes of Data
Mining. Customer churn is defined, different types and causes of churn are discussed. Social aspects of
churn are brought to attention and specifically related to realities of Azerbaijan.
1 INTRODUCTION
Mobile market is very competitive and changing all
the time. Due to those changes companies have to
spend more resources to prevent customers from
switching service provider because it is getting much
more expensive to attract new customer rather than
to retain existing one.
Relevance of this work is defined by high
competition on mobile operator market after third
company entered with lower prices. It is also
planned to implement MNP (Mobile Number
Portability) service this year what will enable
subscribers to switch service provider and keep their
old number. Churn is a worldwide problem because
it is very difficult to win customers’ loyalty in
modern virtualized world. There are almost none
personal bonds between mass service providers and
their subscribers.
This work, like all other researches, also has its
limitations. One of the major limitations of this
research was data classification and data
confidentially in mobile operator that prevented
from having access to a part of customer‘s data such
as billing and credit data as well as call details
records. This helped to build social graph of people
interaction and apply social analysis techniques for
determining strong ties and how people’s decision to
churn is affected by their social group. For
companies, the cost of acquiring new customers is
increasing day by day. Therefore, a new era has
begun in marketing industry. Instead of organizing
campaigns to win new customers, companies are
searching different variety of programs to emphasis
on customer satisfaction, to increase customer-based
earnings and to have higher customer loyalty. The
only method to achieve those goals is preventing
customer churn before it happens. At this point,
customer churn modeling has created an important
competitive advantage and a new workspace. A
good modeling reveals which customer is close to
churn and which is loyal. With the development in
database systems and the variability of customer
behavior, an extraordinary increase in the size of the
data has occurred. This causes to extract previously
unknown information and relationships in huge
amount of data. This information requires applying
different techniques according to the structure of the
data sets to be analyzed. The results of the analysis
are used to plan a comprehensive promotional
campaigns and new strategies (Huang, 2012).
1.2 Churn and Its Prediction
Customer churn is term to denote the customers
which are willing to leave for competing companies.
It is estimated that to attract new customer is five
times more expensive than to keep existing one.
Customer churn is accepted as inevitable part of the
market (Geppert, 2003).
There are several concepts and methods to detect
customers who are about to switch to another
operator. A good churn prediction system should not
only detect at potential churners, but also provide a
sufficiently long term forecast. When potential
233
Mam
ˇ
cenko J. and Gasimov J..
Customer Churn Prediction in Mobile Operator Using Combined Model.
DOI: 10.5220/0004896002330240
In Proceedings of the 16th International Conference on Enterprise Information Systems (ICEIS-2014), pages 233-240
ISBN: 978-989-758-027-7
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
churners are identified, the marketing department
usually contacts them and, if the customers are
established as high churn risk, takes appropriate
actions to prevent loss of customers.
Modern churn-analysis tools are divided into two
categories: tools based on patterns of usage and ones
based on billing data. Usage-based tools watch for
customer usage patterns and try to predict when a
customer will leave. Diagnostic models based on
usage can be as simple as tracking price-plan
changes or as complicated as watching certain
customers' day-to-day usage. A usage-based model,
for example, might record that a customer on a $75
monthly service plan switched to a $49 plan. The
next month, he switched to a $29 plan. That pattern
signifies the customer is significantly altering his
usage and likely will churn soon (Geppert, 2003).
The next generation of churn-analysis tools is
more complex and uses more variables. It takes into
account thousands of factors, such as where a
customer called, when he called and his cell-site
movement patterns, to predict when he is likely to
churn. These churn-analysis models have not been
fully implemented by any mobile operator yet. Parts
of this model were implemented by some
companies, but the models they use are proprietary.
Churn analysis of next generation requires a heavy
commitment and careful planning, which many
operators do not want to do.
2 CUSTOMER CHURN
2.1 Churn Definition
If a customer stops the contract with one company
and becomes a customer of a competitor, this
customer is considered lost customer or churn
customer. Customer loss is very closely related with
customer loyalty. Today’s economic trend dictates
that price cuts are not the only way to build
customer loyalty. Accordingly, adding new value
added services to the products has become an
industry norm to have loyal customer. The main goal
of customer lost study is to figure out a customer
who will likely be lost and is to calculate cost of
obtaining those customers back again. During the
analysis, the most important point is the definition of
the churner customer. In some cases, to make a
definition is very difficult. A credit card customer,
for instance, can easily start using another bank’s
credit card without cancelling credit card of current
bank. In this specific case, a decrease in spending
can be taken into consideration to understand the
customer’s loss. Customer’s loss is a major problem
for companies which are likely to lose their
customers easily. Banks, insurance and
telecommunication companies can be given as
examples (Lazarov, 2007).
2.2 Types of Churn
Active/Deliberate – customer decides to quit his
contract and to switch to another provider.
Reasons for this may include: dissatisfaction
with the quality of service, too high prices, no
rewards for customer loyalty, bad support, no
information about reasons and predicted
resolution time for service problems, privacy
concerns.
Rotational/Incidental – the customer quits
contract without the aim of switching to a
competitor. It usually happens because of
changes in the circumstances that make it
impossible to use the service, e.g. financial
problems, when customer can’t pay; or change of
the geographical location which is not covered
by company.
Passive/Non-voluntary – the company
discontinues the contract itself. Reason can be
fraud, debt or long period of inactivity (Tuğba,
2010).
There are two categories of rotational churn:
when subscriber stops paying after contract ends or
while it is still active. Jonathan Burez calls them
commercial and financial churn respectively (Burez,
2008).
Voluntary churn (active, rotational) is hard to
predict. And while incidental churn only explains a
small fraction of overall churn it is very useful to
predict and react taking appropriate action to prevent
deliberate churn. To prevent voluntary churn
operator has to identify churner with high
probability and to find reasons why he wanted to
switch mobile operator.
Furthermore, churning can be divided also into
three other groups:
Total – the agreement is officially cancelled;
Hidden – the contract is not cancelled, but the
customer is not actively using the service since a
long period of time;
Partial – the agreement is not cancelled, but the
customer is not using the services to a full extent
and is using only parts of it, and is instead using
constantly a service of a competitor.
Depending on the company, the contract type
and the business model that is being applied hidden
or partial churning can lead to considerable money
ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems
234
loss and also needs to be identified and action should
be taken in order not to lose completely the customer
(Khalatyan, 2010).
2.3 Problems and Threats
Preventing customer churn is critical for the survival
of mobile service providers because it is estimated
that the cost of acquiring a new customer is about
$300 or more if the advertising, marketing, and
technical support are all taken into consideration. On
the other hand, the cost of retaining a current
customer is usually as low as the cost of a single
customer retention call or a single mail solicitation
(Berson et al., 2002). High acquisition cost makes it
imperative for mobile service providers to devise
ways to predict the churn behavior and execute
appropriate proactive actions before customers leave
the company (Berson et al., 2002).
In addition to lost revenue, customer churn
means increased activation and deactivation costs. In
the global wireless industry, these amount to $10
billion per year, according to an August 2001 study
by International Data Corporation (Geppert, 2003).
Geppert (2003) indicated that a high churn rate
also puts pressure on companies to win new
customers. The cost of acquiring each new customer
ranges from $350 to $475 and providers need to
retain these new customers for more than four years
to break even (Geppert, 2003).
Replacing old customers with new ones carries
other burdens. In addition to marketing and
advertising, companies incur costs associated with
provisioning new customers, as well as increased
risks associated with billing issues and other revenue
assurance matters. Customer churn also generates
soft costs: loss of brand value when dissatisfied
customers tell others about their experiences, lost
opportunities for cross-selling of complementary
products and services, and a potential domino effect
with respect to the carrier‘s remaining customer
base. Further, the deactivation and disconnection of
customers brings inherent risk of revenue and
margin deterioration, particularly when multiple
service providers are involved. Finally, the potential
impacts on profitability that come from inactive,
underutilized, and otherwise unprofitable network
facilities must be considered
2.4 Techniques for Churn Prediction
Marcin Owczarczuk (2010) in his article “Churn
Models for Prepaid Customers in the Cellular
Telecommunication Industry Using Large Data
Marts” described methods to predict churn for
prepaid customers where big data marts are available
for analyses. He used datasets with 1381 variables
for each of about 80000 customers.
“In this article, we evaluated usefulness of
regression and decision trees approach to the
problem of modeling churn in the prepaid sector of
the cellular telecommunication company. Linear
models are more stable than decision trees that get
old quickly and their performance weakens in time,
especially in top deciles of the score. Nevertheless,
we showed that prepaid churn can be effectively
predicted using large data mart” (Owczarczuk,
2010).
Situation described in Marcin Owczarczuk’s
work is somehow similar to the one in this work. In
this work is also used big datamart of prepaid
customers containing 637 fields. But an attempt to
reduce amount of variables is made to make
understanding of model easier and to reduce time
needed to build a model. It was also experimentally
proven that small number of variables is enough to
churn with high accuracy (Verbeke et al., 2012).
“Customer Churn Analysis in
Telecommunication Sector” by Umman Tuğba
Şimşek Gürsoy. used similar techniques like
Decision Tree and Logistic Regression Analysis but
focuses mostly on determining the reasons why
customers decide to churn. He compared and
analyzed different parameters and variables for
churning and non-churning customers and got
interesting results. He discovered that incoming calls
have big influence on customers’ decision as well as
discount offers which they get (Tuğba et al., 2010).
In their article “Turning Telecommunications
Call Details to Churn Prediction: a Data Mining
Approach” Chih-Ping Weia and I-Tang Chiub were
using call detail data to determine customer
behavior: “.. we propose, design, and experimentally
evaluate a churn-prediction technique that predicts
churning from subscriber contractual information
and call pattern changes extracted from call details.
This proposed technique is capable of identifying
potential churners at the contract level for a specific
prediction time-period. In addition, the proposed
technique incorporates the multi-classifier class-
combiner approach to address the challenge of a
highly skewed class distribution between churners
and non-churners“ (Weia et al., 2002).
V. Yeshwanth et al. and Ying Huang, Tahar
Kechadi
paper presents predictive modeling of
customer behavior based on the application of
hybrid learning approaches for churn prediction in
the mobile network: “Our proposed framework deals
CustomerChurnPredictioninMobileOperatorUsingCombinedModel
235
with a better and more accurate churner prediction
technique compared to the existing ones as it
incorporates hybrid learning method which is a
combination of tree induction system and genetic
programming to derive the rules for classification
based on the customer behavior” (Yeshwanth et al.,
2011).
Next work has similar objectives: “To obtain
more accurate predictive results, we present a novel
hybrid model-based learning system, which
integrates the supervised and unsupervised
techniques for predicting customer behavior”
(Huang et al., 2013).
These articles helped to make decision to use
neural network, cox regression and decision tree
techniques in conjunction to build model that
predicts churn customers with high probability like
neural networks and which decisions could be
explained like decision tree model.
2.5 Social Ties and Their Influence
Churn is not only statistical phenomenon; it also
should be discussed from sociological point of view.
Customers make decision to churn based not only on
their personal preferences or some objective reasons
such as price and quality of the service but also
based on their social surrounding, influence from
family members and friends.
First in the list, the oldest and most cited article
is “Social Ties and their Relevance to Churn in
Mobile Telecom Networks” by Koustuv Dasgupta et
al. (2008). In this article authors used detailed call
record data of mobile operator for one month. Data
contains detailed information about voice calls,
SMS, value-added calls of users. They built graph
for all connections between subscribers based on
calls made between them. To reduce graph’s size
and eliminate biased data they excluded one-
direction only connections and short numbers. They
possessed only this CDR files and no additional
information about customers like demographics,
when he started using service, how much spent
during last months. Such practical limitations made
the problem very challenging, but authors succeed to
demonstrate how reasonable prediction accuracy can
still be achieved using only link information.
In another work “An Efficient Method of
Building the Telecom Social Network for Churn
Prediction” by Pushpa and G Shobha, authors made
accent on finding groups of customers within social
graph. Contrary to the previous article they paid a lot
of attention on the types of relationships between
nodes. They wroite about two types of social
networks: Homogeneous and Heterogeous.
Homogeneous social networks are those where is
only one kind of relationship between the customer
for example the relationship may be friendship
between the two customers are linked heterogeneous
social networks represent several kinds of
relationship between customers, and can be called as
Multi-relational social networks. Example of
different relationship types may be: friendship,
acquaintance, professional, family. Based on the
duration of voice calls, call frequency etc. for each
of these relationship types it is possible to define
unique behavioral pattern. Authors concluded that
the accuracy of the churning model can be increased
by considering the multiple relationship between the
customers while construction of the telecom social
network to extract the hidden communities of the
churners and non-churners (Pushpa et al., 2012).
Next work is called “Predicting customer churn
in mobile networks through analysis of social
groups” by Yossi Richter et al. (2010). In this work
author implemented opposite approach to on in
previous article by concentrating on social groups
first and eliminating weak ties between groups. By
doing so he got several completely separated groups.
Richter calls it the group-first social networks
approach because first he calculates churn prediction
for the group rather than individual customer as is
done in most researches. He used decision tree
algorithm for scoring each group’s churn based on
defined KPI’s of the group. After that author
calculated churn prediction for individual
subscribers by first computing their relative churn
score (Tuğba et al., 2010).
Xiaohang Zhang (2012) “Predicting Customer
Churn through Interpersonal Influence” used
methods for social network analysis described in
aforementioned article by Koustuv Dasgupta (2008)
but also combined it with personal characteristic of
the customer. He built three models based only on
network attributes, only traditional attributes and
combination of both. He applied three popular data
classification techniques including logistic
regression (LR), decision tree (DT) and neural
network (NN) methods. Then author compared the
prediction results of traditional attributes-based
models, network attributes-based models and
combined attributes models and found that
incorporating network attributes into predicting
models can greatly improve prediction accuracy. In
addition he proposed a novel prediction model based
on the propagation process that accounts for
interpersonal influence and customers’ personalized
characters. The empirical results show that the
ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems
236
proposed propagation model outperforms traditional
classification models (Zhang, 2012).
And the last article in this review of social aspect
of churn is called “Estimating the effect of word of
mouth on churn and cross-buying in the mobile
phone market with Markov logic networks” by
Torsten Dierkes et al. (2011). His main goal was to
optimize network analysis process by introducing
Markov logic networks (MLNs) as this method have
recently been suggested as a significant step forward
in this field. The method draws on Markov Random
Fields and ILP (inductive logic programming) and is
able to handle larger data sets compared to earlier
ILP implementations such as FOIL. They collected
customers’ call data and built social graph just like
researchers from previous articles that was
described. But this time authors inserted this data in
a relational database creating multiple relations
between nodes. They used Alchemy – open source
software for learning Markov logic networks from
data. Markov logic networks (MLNs) are a
collection of formulas from first-order logic, to each
of which a weight is assigned. In other words, it
describes a probabilistic logic. Ideas from estimating
Markov networks are then applied to learn the
weights of the formulas. The vertices of the MLN
graph are atomic formulas, and the edges are the
logical connectives used to construct the logical
formula. A Markov network is a model for the joint
distribution of the properties of underlying objects
and relations among them. It was established that
MLNs have higher predictive accuracy (+8%) and
sensitivity (+19.7%) than the benchmark logistic
regression (Dierkes et al., 2011).
2.6 Concept of Customer Retention
In order to reduce amount of people who stop using
service of the company different customer retention
techniques are being used. Marketing department
should use information provided by data analysis
team and offer to customers who are predicted to
churn new services to keep them. Because
companies have limited human resources to call or
somehow interact with customers who are suspected
to churn some bonuses or discounts could be offered
to larger group of customers without significant
effort. But even for these campaigns there should be
reasonable amount of subscribers to whom
campaign is offered. Usually top 10% of predicted
customers who have highest probabilities of churn
and most value for the company are contacted
personally and offered some discounts. Next 2-3
deciles are offered some free minutes or time-limited
discounts.
Usually each customer is considered individually
during churn prediction. The goal is to predict each
customer’s likelihood of churning in the near future,
where usually a forecast horizon of a month to three
months is considered. To this end, dozens to
hundreds of complex Key Performance Indicators
(KPIs) are generated per customer; these KPIs span
the customer’s personal characteristics as well as
trends in their call activities over a long period. The
information then serves as input to a statistical
regression model (usually a logistic regression
variant) that outputs a churn score. In other words,
this approach focuses on identifying patterns that are
uncommon to a given customer, and are correlated
with churn (Kim et al., 2012).
Other system try to solve churn prediction
problem by monitoring customers’ calls to the
mobile carrier’s call center, such systems apply
speech and emotion analysis to the calls, and
together with additional information (number and
length of calls by the customer, number of transfers,
hold period, etc.) try to quantify the customer’s
dissatisfaction level and hence the associated churn
risk. The system can then react by prioritizing
pending ‘churners’, even suggesting retention
packages. This approach has a major disadvantage:
although it may accurately pinpoint the potential
churners, the forecast horizon it provides is very
short as the system identifies customers that have
already expressed dissatisfaction with the service. At
this stage, retention prospects are lower while cost is
significantly higher. Even when combining the long
term and ad-hoc churn prediction systems, one
drawback is fairly obvious: we clearly rely on the
assumption that a churning customer either changes
calling patterns or contacts the carrier’s call center to
express dissatisfaction prior to switching carriers.
While this may be true in some cases, there are
certainly many scenarios in which these assumptions
are violated. For example, this may occur when
customers come to believe that they have found a
better deal with a competitor and churn immediately.
Another, less obvious, drawback of traditional
solutions is that they focus exclusively on the
individual customer without taking into account any
social influence. Clearly, there are many social
aspects to churn, as witnessed in other consumer
areas, where a dominant example is when a churning
customer influences other customers to churn as
well. Thus, developing churn prediction systems that
take social aspects into account poses an emerging
theoretical challenge with potentially great practical
implications.
CustomerChurnPredictioninMobileOperatorUsingCombinedModel
237
The nature of the churn prediction problem
dictates a specific non-standard performance
measure. Recall that once the prediction system
produces its churn scores, the retention department
makes contact with the subscribers that are most
likely to churn, in an attempt to preserve each
customer that is established to be a churn risk.
Naturally, only a small fraction of the subscriber
pool can be contacted at any given time, and the
subscribers with the highest churn scores are
assigned top priority. Therefore, a churn prediction
system should be measured by its ability to identify
churners within its top predictions. Formally,
performance is measured using the lift metric. For
any given fraction 0 < T < 1, lift is defined as the
ratio between the number of churners among the
fraction of T subscribers that are ranked highest by
the proposed system, and the expected number of
churners in a random sample from the general
subscribers pool of equal size. For example, a lift of
3 at a fraction T = 0:01 means that if we contact the
1% of subscribers ranked highest by the proposed
system, we expect to see three times more people
who planned to churn in this population than in a
0:01-fraction random sample of the population (Kim
et al., 2012).
3 BUILDING CHURN MODELS
In this part models based on the data that were
prepared are being built. For modeling it was
planned to use several available modeling nodes.
For modeling and evaluating of models data set
was divided into training and testing partitions.
Separating data into training and testing sets is an
important part of evaluating data mining models. By
using similar data for training and testing it is
possible to minimize the effects of data
discrepancies and better understand the
characteristics of the model.
After running C5.0 model it produced output in a
form of decision tree. Predictor importance chart is
also shown in the model output window. As can be
observed from the picture below the most important
predictor is chosen total credit amount on the
subscriber’s account during week 1. Among other
most important fields are Count of outgoing
destination calls on week 1, duration of incoming
weekend calls for week 1 and sum of consecutive
two-day periods without outgoing calls.
Strange thing here is that all subscribers with
total credit of less than 112.532 were classified as
non-churners. Even though prediction is remarkably
precise and results hold for other data sets model
should be revised often to eliminate over fitting
problem.
Precision of prediction is shown on Figure 1.
Cox regression model was used as a specific
implementation of survival analysis.
Normally one model is used to predict a target. It
is possible to try several types of models, but in the
end there will be one model left. Most of data-
mining projects also develop one model, but data
mining implies usage of several approaches to
analysis. Neural Network and Cox regression
models’ outcomes are combined since they had
lower accuracy than C5.0 model.
At the picture below on the left side are results of
full data set analysis and on the right – for reduced
fields. It can be seen that results for reduced fields
are slightly worse than for full set but not
significantly.
4 TESTS AND RESULTS
4.1 Evaluation of Models
There is 91.44% accuracy against 87.13% for
Figure 1: Neural network model results.
ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems
238
reduced fields for testing data partition (Figure 2).
Analysis revealed that the best results are achieved
using C5.0 decision tree model. Principal
components analysis reduction technique was used
to reduce number of fields and compared how the
same model performed on full set of fields
against reduced data set. Results for neural network
and Cox regression models are significantly less
even though they were built over all available fields
except those that were removed during data
preparation phase. On the Figure 3 results for both
models are reflected. On the left side is shown result
for neural network which is 79.04% accurate on
testing partition. And on the right side Cox
regression output is presented with 77% accuracy.
One way to increase accuracy of the model is
already described method of combining models. Pay
attention to the ‘Comparing agreement with Churn
Status’ table and if to be more precisely, to the
correct percentage in the testing column. Their
combined accuracy increased to 83.84% compared
to 79.04% and 77% they had respectively.It is also
possible to evaluate model with Evaluation node
which shows how model performed on graph.
5 CONCLUSIONS AND
RECOMMENDATIONS
Churn is directly related to customer loyalty and
defines a process of switching service provider.
Churn can be of several categories which are defined
based on the reasons why it happened. These types
are active or deliberate churn, rotational or
incidental churn and passive or non-voluntary.
Active churn is initiated by subscriber which wants
to change service provider, rotational happens
without intention to switch but for different reasons.
It is not easy to distinguish those two. And passive
churn covers cases when provider disconnects
customer for inactivity. It is the most dangerous type
because is difficult to discover. Various reasons
which lead to customer churn are also discussed.
Establishing reasons of churn usually happen based
on questionnaires and surveys.
Social network and interpersonal relationship
specific to Azerbaijan were discussed. As a result
few points related to Azerbaijani society were
brought to attention.
Achieved results showed that C5.0 is still the most
precise model while neural network and cox
regression perform worse. After combining last two
models result was improved but was still worse
compared to C5.0 model. Main downside of C5.0
model is that it can be over fitted to the expected
result. That’s why this model should be reviewed
and calibrated for new data sets.
During preparation of this work several
recommendations to mobile operator came up which
are worth mentioning. From social network analysis
Figure 2: C5.0 for all fields vs reduced factors results.
Figure 3: Neural network (left), Cox regression (right) results.
CustomerChurnPredictioninMobileOperatorUsingCombinedModel
239
part where mentality of Azerbaijani society was
discussed recommendations were to consider giving
higher value to married, working men because they
have more influence on the family and can be cause
of his family members’ churn if he decides to
change mobile operator himself; other important
factor is prestige and willingness to show it which
can be used to create positive impression around the
brand and particular product; parents can have
strong influence on their children even if they are
not underage anymore because most of young
people live with parents till marriage and respect
their opinion very much. Other approach of using
social information could be creation of social
network graph of the customers using call data
records.
REFERENCES
Ballings, M., Poel, D., V., 2012. Customer Event History
for Churn Prediction: How Long Is Long Enough?,
Expert Systems with Applications: An International
Journal archyve, Vol. 39 Issue: 13517-13522.
Berson, A., Smith S and Thearling K, New York, 2002,
Building Data Mining Applications for CRM, ISBN
978-0071344449.
Burez, J., Poel, D., 2008. Separating financial from
commercial customer churn: A modeling step towards
resolving the conflict between the sales and credit
department, Expert Systems With Applications, 35(1-
2): 497-514
Dasgupta, K., Singh, R., Viswanathan, B., et al, 2008.
Social Ties and their Relevance to Churn in Mobile
Telecom Networks, EDBT: 668-677
Dierkes, T., Bichler, M., Krishnan, M., 2011. Estimating
the effect of word of mouth on churn and cross-buying
in the mobile phone market with Markov logic
networks, Decision Support Systems archyve, Vol. 51
Issue 3: 361-371.
Geppert, KPMG & Associates, 2003. Customer Churn
Management: Retaining High-Margin Customers with
Customer Relationship Management Techniques,
Journal of International Business Disciplines, Volume
4, Number 2: 18-19.
Huang, B., Kechadi, M., T., Buckley, B., 2012. Customer
Churn Prediction in Telecommunications, Expert
Systems with Applications: An International Journal
archyve, Vol. 39 Issue 1: 1414-1425.
Huang, Y., Kechadi, T., 2013. An Effective Hybrid
Learning System for Telecommunication Churn
Prediction, Expert Systems with Applications.
Yeshwanth, V., Raj, V., Saravanan, M., 2011.
Evolutionary Churn Prediction in Mobile Networks
Using Hybrid Learning, FLAIRS Conference, AAAI
Press.
Khalatyan, A., 2010. Churn Management in
Telecommunications, iSCHANNEL 05: 21-26.
Kim, Y.,S., Moon, S., 2012. Measuring the Success of
Retention Management Models Built on Churn
Probability, Retention Probability, And Expected
Yearly Revenues, Expert Systems with Applications,
Vol. 39, Issue 14: 11718–11727.
Lazarov, V., Capota, M., 2007. Churn Prediction, TUM
computer science.
Owczarczuk, M., 2010. Churn Models for Prepaid
Customers in the Cellular Telecommunication
Industry Using Large Data Marts, Expert Systems with
Applications: An International Journal archyve, Vol.
37 Issue 6: 4710-4712.
Pushpa, Shobha, G., 2012. An Efficient Method of
Building the Telecom Social Network for Churn
Prediction, International Journal of Data Mining &
Knowledge Management Process, Vol. 2, Number 3.
Richter, Y., Yom-Tovy, E., Slonim, N., 2010. Predicting
customer Churn in Mobile Networks through Analysis
of Social Groups, SDM: 732-741.
Tuğba, U., Gürsoy, Ş, 2010. Customer Churn Analysis in
Telecommunication Sector, Journal of The School of
Business Administration
, Vol.39, Number 1: 35-49.
Verbeke, W., Dejaeger, K., Martens, D., Hur, J., Baesens,
B., 2012. New Insights into Churn Prediction in The
Telecommunication Sector: A Profit Driven Data
Mining Approach, European Journal of Operational
Research, Vol. 218, issue 1: 211-229.
Weia, C.-P., Chiub, I.-T., 2002. Turning
Telecommunications Call Details To Churn
Prediction: A Data Mining Approach, Expert Syst.
Appl. 23(2): 103-112.
Zhang, X., Zhu, J., Xu, S., Wan, Y., 2012. Predicting
Customer Churn Through Interpersonal Influence,
Knowl.-Based Syst. 28: 97-104.
ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems
240
