The Architecture of a Churn Prediction

System Based on Stream Mining

Borja Balle a , Bernardino Casas a , Alex Catarineu a , Ricard Gavaldà a ,
David Manzano-Macho b
a
Universitat Politècnica de Catalunya - BarcelonaTech. e-mail:
{balle,bcasas,catarineu,gavalda}@lsi.upc.edu
b
Ericsson Spain. e-mail: [email protected]

Abstract. Churning is the movement of customers from a company to

another. For any company, being able to predict with some time which
of their customers will churn is essential to take actions in order to re-
tain them, and for this reason most sectors invest substantial effort in
techniques for (semi)automatically predicting churning, and data mining
and machine learning are among the techniques successfully used to this
effect. In this paper we describe a prototype for churn prediction using
stream mining methods, which offer the additional promise of detect-
ing new patterns of churn in real-time streams of high-speed data, and
adapting quickly to a changing reality. The prototype is implemented on
top of the MOA (Massive Online Analysis) framework for stream min-
ing. The application implicit in the prototype is the telecommunication
operator (mobile phone) sector.
A shorter version of this paper, omitting Section 5, was presented at
CCIA’13 (http://mon.uvic.cat/ccia2013/en/).
Keywords. Data stream mining, churn prediction, Hoeffding trees,
machine learning, MOA

1. Introduction

Customer churning is the movement of customers from a company to another.

For any company, being able to predict which of their customers will churn before
they actually leave is essential to take actions in order to retain them. For this
reason most sectors invest substantial effort in techniques for (semi)automatically
predicting churning as accurately and as early as possible.
Traditional data mining / machine learning techniques are often applied for
building predictive models that assign each customer a probability that s/he
churns in one or more prespecified periods of time. However, these techniques
have a number of limitations. First, they tend to be batch-based: all the data to
be used for the modeling must be available upfront for the model construction;
many of them cannot work incrementally, i.e., incorporate into the model infor-
mation arrived after it has been built; the model construction is computationally
costly, and often needs human intervention for e.g. parameter optimization, data
selection, or model evaluation; and, most importantly, methods assume that the
source of data is stationary, that is, does not change in nature over time.
The latter assumption is patently false in many scenarios: customers change
their behavior over time in reaction to market conditions, product or price
changes, or sociological phenomena. Sometimes change is gradual, but abrupt,
overnight change can also occur. The problem is particularly acute in sectors such
as the telecommunications market (say, companies to which customers subscribe
for mobile phone service), which are affected by high customer churn rate.
In this setting, models built sometimes in the past can quickly or gradually
get out of sync with the current customer patterns, resulting in suboptimal churn
prediction rate and therefore revenue loss. An obvious solution is to rebuild the
models periodically, but this typically either requires human analyst time (which
is slow and expensive) or automatizing decisions such as when and on which
amount data to retrain new models, which may be hard to do in abstract.
Data stream mining techniques have emerged in the last decade or so to
transport the benefits of data mining to such scenarios: they intend to provide e.g.
model induction techniques that work on streams of data that may potentially
never end, that can process thousands or millions of instances per second both
for training and for prediction, and that can maintain these models updated as
the statistical or logical nature of the data change over time.
In this paper we describe the architecture and a proof-of-concept implementa-
tion of a system for churn prediction based on stream mining, with mobile phone
subscriber churning as the background scenario. The use of stream mining meth-
ods offer the promise of detecting new patterns of churn in real-time streams of
high-speed data, and adapting quickly to changing realities.
The core stream mining platform used is MOA (Massive Online Analysis) [2],
and in particular we use its implementation of a variant of decision trees (Ho-
effding trees) that can be built and maintained efficiently and incrementally from
streams. Unfortunately, as usual in this field for e.g. privacy reasons, we could not
use real data for our experiments. We developed a synthetic data generator based
on descriptions of real data available in the literature, which may be of interest
in itself.
The main conclusion from our proof-of-concept is that the technology exists
to perform real-time, high-throughput prediction on streams of items containing
information similar to that contained in churn prediction literature. On further
work we will concentrate on several issues concerning scalability to international
scale data volumes.
The paper is structured as follows: Section 2 discusses in detail the problem
of churning, particularly in the telecommunications market and related previous
work. Section 3 describes the requirements for a churn prediction system in a
high-speed, highly volatile scenario such as mobile phone customer behavior anal-
ysis. Section 4 describes the architecture of the proposed solution and its main
elements, as well as some details of the implementation. Section 5 presents an
example of the workflow of events processed by the system. Section 6 describes
the generator of synthetic data used to test the system. Section 7 describes qual-
itatively the results of the prototype. Finally, Section 8 summarizes the lessons
learned during the implementation and testing, the main challenges to be solved
for a large-scale, fully scalable solution, and other future work.

2. Preliminaries

2.1. The Problem of Churning in the telecommunication market

Like in every business sector, operators are always looking to get a better under-
standing about their customers’ preferences and their satisfaction to offer them
better products and services. Despite of the maturity of this market and due to
subscriber growth down and revenues flat, churn management has become into
one of the most pressing problems faced by operators. There are many reasons
that may affect a subscriber on deciding to churn. Some of them may be:
• In contrast to post-paid customers, prepaid customers are not bound by
a contract. The central problem concerning prepaid customers is that the
actual churn date in most cases is difficult to assess.
• Customer loyalty is directly related to the customer service and service
experience. Lack of connection capabilities or quality in places where the
customer requires service can cause customers to abandon their current
service provider in favor of others with broader reach or a more robust
network. Besides, a slow response to customer complaints or billing errors
are sure paths to a customer relations disaster.
• Finally other factors such as the pricing, lack of features (customers may
switch carriers for features not provided by their current providers) or cov-
erage, new technology introduced by competitors (for example, high-speed
data) or the fact that new competitors enter the market, are also reason
that has their influence on the churn rate.
Churn prediction modeling techniques attempt to understand the precise cus-
tomer behaviors and attributes which signal the risk of customer churn. The ac-
curacy of the technique used is obviously critical to the success of any proactive
retention efforts. In order to have successful loyalty programs operators need to
analyze their customers based on several parameters such as spending and usage
of each service and product as well as their behavior and traffic patterns. All these
decisions should be made upon the analysis of the data related to customer. Time
is critical, which means that online approaches are gaining more momentum to
support operational decisions in almost real time. By means of online analysis, an
operator can launch and execute quickly some actions with aim of retaining some
customers, and measure the impact of the campaign in real time which allows
adapting them instantly as required.

2.2. Previous Work

There is substantial published literature on churn prediction techniques, although

probably most of the experiences have remained unpublished in the company.
We mention only a few representative ones of those using data mining, machine
learning, or, more in general, that build models from information gathered in
the past; see [8] for a good survey. References [1,11,5,10] use methods such as
decision trees, support vector machines, genetic algorithms, and multilayer neural
networks for churn prediction in various contexts, from finantial-service providers
to landline phones and mobile phones; [6] carries out an intensive and state-of-the
art comparison of this and other methods. [9] discusses the problem of evaluating
predictors in this context. [12] discusses the use of social network information
(and customer social circle) to help in churn prediction.
As mentioned in the introduction, all these works and those that we are aware
of use the batch paradigm: a model is built offline by the data analyst and then
used online to make predictions. Some of the works do mention the possibility of
retraining the model periodically. By contrast, we aim at techniques that allow
building the models online too and keeping them accurate as the customers change
their behavior.
We omit the many other papers discuss the features actually used in practice
and the “feature engineering” process, as well as the business implications of
churn detection and possible customer retention strategies once they have been
identified as possible churners.
For background on stream mining, see for example the books [3,4].

3. Requirements

The telecommunication market is highly and unpredictably dynamic. The ability

for immediately detecting such a trend on even a small segment of subscribers
may be essential to retaining many more leaving in the near future. However,
the reasons why subscribers churn may change suddenly. For many (and hard
to define a priori) reasons, subscribers that were not considering leaving their
company may suddenly consider another company very compelling and decide to
leave overnight. A scenario like this demands the application of online analytics,
to detect and react in timely manner to the changes in the expected behavior of
operators customers. Other requirements are:
• Accuracy. High recall (all churners are flagged as such) and relatively high
precision (not many non-churners are flagged) are both important.
• Performance. Time is critical. A right action taken out of its proper window
time is useless. Detect and react to any relevant change detected through
the analysis of the incoming data flows is a competitive advantage.
• Flexibility. Data sources may change and any churning system has to have
the ability to be adapted into the new scenario.
• Scalability. Churn prediction technology needs to handle difficult contexts,
in which there are big data flows related to customers activity, with real-
time requirements and prone to changes.
• Ability to segment customers and incorporate analyst’s existing knowl-
edge. Customer profiling demonstrates the direct business benefit gained
from analytics. Traditional parameters include customer type, spending,
subscription type and preferred services.
 Figure 1. Architecture and Design of a Platform for Adaptive, Real-time Churn Prediction

4. Architecture

4.1. General Description

In this paper, we describe the architecture proposed for our prototype. It is de-
picted in Figure 1, and its main components are described next.

4.2. Integrator

The system processes as input a number of streams with diverse information. For
example, a stream of call records, a stream with billing actions by the company
and bill payments by the subscribers, contents from social networking services
such as Twitter, etc. The Integrator module integrates all these streams, gener-
ating a logically unique stream of events. We distinguish several type of events,
including at least a subscriber joining the company, calls and SMSes, complaints,
bills emitted and bills paid, tweets by a subscriber, and “churn” events (e.g., a
user explicitly has left the company, or for a prepaid user, it has been declared as
a churner according to the company’s criterion).

4.3. Data Manager

The Data Manager module manages the customer information database and the
Pending Predictions queue.
The customer information database contains basic information about our sub-
scribers (age, address, type of contract, etc.) as well as highly dynamic infor-
mation (e.g. last numbers called, most frequently numbers called). The Pending
Predictions queue contains all predictions that are awaiting for confirmation or
refusal in the future (that is, whether the subscriber to which the prediction refers
to has churned or not within a specified time frame).
These two structures, the customer information database and the Pending
prediction queue, can easily become the two main bottlenecks in the system if
not carefully implemented, both for time and for memory usage. If they do not
fit in RAM, a write-optimized disk-resident database will be required.

4.4. Record Generator

The Record Generator receives the stream of Events generated by the integrator
and uses it for two purposes. First, it updates the customer information database
according to each Event. Second, it generates one or more Records from each
Event using information from the customer information database. Thus, it creates
a stream of Records passed downstream.
A Record is a vector of features, the first of which is a subscriber identifier,
and the rest contains all the information about that subscriber state that is con-
sidered relevant for prediction. Many of them cannot be directly derived from the
Event, but are aggregations of information about the customer precomputed in
the database. One of the features (say, the last one) indicates if the subscriber
has churned so far: it will be true when the Event originating the record was a
churning one.
The prototype currently use this set of features for prediction, which figure
among the most widely reported as useful in the literature:
• Age, sex, income range
• Contract type (mobile or landline, pre-paid or post-paid)
• Average call duration during last month
• Number of calls last week and last month
• Increase of decrease in number of calls in the last 2 weeks and the last 2
months
• % of calls by/to this subscriber where the caller/recipient belongs to an-
other company
• # of complaints in the last 2 months, and % of these that were resolved
satisfactorily
• Average bill value
• Increase or decrease of value last 2 bills
We made the following optimization for efficiency. Every event gives rise to at
least one record, with the exception that every day only one call by or from a
subscriber generates a record and a prediction. That is, if a subscriber gets or
receives 20 calls in a day, all of them will be used to update his/her statistics
in the customer database, but only the first one will generate a record and a
prediction. This introduces a delay of (at most) one day in flagging this customer
as churner, but reduces a lot of overhead.

4.5. Record Processor

The Record Processor is the heart of the system: it builds, maintains, and applies
the predictive models. It therefore contains the data mining or machine learning
algorithms that make prediction possible.
 When a record not indicating churn arrives, it passes the record through the
current model and makes a churn prediction for it. The record with its prediction
is stored the Pending Predictions queue, waiting for future confirmation. When
a record indicating churn arrives, records for that subscriber are searched in the
Pending Predictions queue and, if found, passed to the model trainer as positive
instances of churning. Expired records in the Pending Predictions queue (corre-
sponding to subscribers that did not churn within a specified time) are passed to
the model trainer as negative instances of churn. All records (describing current
states of subscribers) are passed a clustering submodule to build subscriber pro-
files. We have used both 1) clustering methods available in MOA and 2) the split
induced by the Hoeffding tree branches to define customer segments; they may
give alternate segmentations of potential use for analysis.
Additionally, a background process periodically scans customers for which
no Event has occurred and injects a special record indicating “no activity”, so
that 1) a prediction for the customer is generated from time to time (in case e.g.
inactivity may indicate churning propension) and 2) the system is also trained to
predict well on periods of customer inactivity.

4.6. Reporting and Interface to other Systems

The Record Processor module thus produces predictions, statistics and profiles of
the predicted churners. The predicted churners id’s and their current profiles are
passed to the user interface or other parts of the customer management system
so that adequate actions can be assessed and taken.
The subscriber profiles provide information for human analysts to build un-
derstandable portraits of churners and causes for their churning. Moreover it al-
lows to focus the retention action efforts, such as calling with a promotion, to sub-
sets of the subscribers with propensity to churning, even before they are flagged
as churners by the system.

4.7. Implementation

We implemented a prototype of this architecture using the Java language. The

prediction core of the system uses MOA (Massive Online Analysis), a platform
related to the popular WEKA machine learning package, which includes a col-
lection of machine learning algorithms (classification, regression, and clustering)
and tools for algorithm evaluation, visualization, synthetic data generation, and
stream management.
We used in particular MOA’s Hoeffding Adaptive Tree classifier, which is a
variation of the CVFDT method proposed in a seminal paper on stream mining by
Hulten, Spencer, and Domingos [7]. A Hoeffding tree is an incremental, anytime
decision tree induction algorithm that is capable of learning from massive data
streams. The Adaptive version in [7] and implemented in MOA is able to adapt to
changes of the data over time, updating and revising the structure and contents
of the tree to keep it accurate.
5. Example Workflow

The following illustrates by a few examples the process workflow.

Suppose that Alice, currently a company customer, calls Bob, who is currently
subscribed to another company. The (external) call management system passes
the identifiers (e.g. numbers) of Alice and Bob together with the information
deemed relevant (say, start time, duration, approximate location of Alice and
Bob if available) to the Integrator module, which generates a Call Event and
passes it to the Record generator. The Record generator queries its customer
database, determines that Alice is a customer but Bob is not, so it 1) updates
Alice’s information in the database recording that she participated in this call
and 2) generates a record with Alice as identifier, as described next. The record
generated has Alice’s identifier, and consists of a vector of “features” describing
as accurately as possible the state of Alice at this point in time as recorded in the
database. The record is passed to the Record processor. As it encodes a call, the
Record processor will input it to the current prediction model, who will output a
prediction about Alice’s probability of churning in some period of reference such
as a month; this is called the “prediction expiry date”. The record, together with
the prediction, is placed in the PendingPrediction queue, as it has to be verified
in the future, and also passed to the Loyalty management module. The latter
decides according to current rules if the customer merits some retention action,
and which one.
The process is similar for most other types for events generated by Alice
(SMS, bill payment, complaint, tweet, etc., but not those indicating customer
churn): One record associated to Alice is generated, for which a prediction is
generated. The pair (record,prediction) is stored in the PendingPrediction queue
and used by the Loyalty Management module to possibly generate actions aimed
at retaining Alice.
Concurrently with processing these incoming records the Record Processor
monitors the PendingQueue for expired predictions. These records are labeled
“noChurn”, meaning that a customer did not churn within a prescribed period,
and passed to the model updater, which will use to confirm, update, or revise the
current model (in the usually called “training” model). For example, a pending
prediction created on April 16th indicating that Alice may churn with probability
60% within one month, will be removed from the queue and processed as below
by May 16th, because it is known at that time that Alice has indeed not churned
in one month.
When customer Alice is determined to have churned (either by explicit action
on her part or by e.g. lack of any activity in prepaid contracts), the system should
receive a Churn event. The Record generator will generate a record with Alice’s
identifier and content “Churn” and the churning date and pass it to the Record
processor. The Record processor retrieves any existing records associated to Alice
in the PendingPrediction queue (which should be non-expired, as by the previous
paragraph), and labels them with the “Churn” prediction, meaning that the right
prediction at the time the record was produced should have been “Churn”. This
record is also passed to the model updater.
When either a prediction about Alice expires or a record indicating that Alice
has Churned, the Record Processor notifies the Loyalty management module so
that this information can be taken into account (by the module’s internal rules)
when actions have to be proposed to customers with a profile similar to Alice’s.
As a hypothetical example of what “model update” means, suppose that at
the time in which Alice’s call arrives, one of the rules used by the system to predict
churn is “customers who are female, aged 20-35, with two or more complaints
about service in the last month, paying an average monthly bill > 200, and who
have used roaming in the last 6 months have a probability of churning in the next
month of 80%”. The system additionally keeps track that it has recently seen 1000
such customers, of which 800 have indeed churned within one month (hence the
80% probability). If Alice fits into this profile, the system will predict that she will
churn in the next month within 80%. After one month, if she has indeed churned,
the system will update its statistics to 801/1001, so the rule will be updated to
“. . . with 80.1% probability”, and otherwise to 800/1001. More interestingly, the
system may realize that for customers in this profile living in urban areas, the
probability is in fact higher (say, 90%), and lower those living in rural areas is
in fact lower (say, 70%). Then the rule above would be split into two rules, the
first of which would be: “customers who are female, aged 20-35, with two or more
complaints about service in the last month, paying an average monthly bill >
200, who have used roaming in the last 6 months, and live in urban areas have a
probability of churning in the next month of 90%”. And the second would apply
to customers living in rural areas and have a probability of 70%. This description,
intended for comprehension, does not necessarily reflect any of the known stream
mining techniques that could be used in an implementation of the invention.

6. The Synthetic Data Generator

In order to test the proposed system, we developed a simulator able to generate

realistic synthetic data similar to what an operator gets from their customers.
Basically, the simulation is a probabilistic dynamical system containing a pre-
determined number of customers. The system tries to reproduce the major types
of interactions between customers and providers and their dynamics. For example,
customers can place calls to other customers, either having the same provider or
a different one. Customers are periodically billed for their calls, and big bills may
make them angry. Angry customers sometimes complain to their provider and
ask for cheaper rates, or maybe churn without further notice.
In particular, each user is modeled in the system using five parameters. Three
of them are socio-economic factors which are known to providers: gender, age, and
income. The other two are indices of how communicative (C) and impulsive (I) the
user is; these are hidden parameters unknown to the provider. Furthermore, these
indices evolve over time depending on the “mood” of the user, which is influenced
by several factors described below. At start users are generated by sampling these
parameters from a fixed probability distribution. In particular, gender, age and
income follow a distribution trained real data. The other indices are sampled from
a hand-crafted distribution, and in particular are not independent of the socio-
economic factors. As a result, the simulation is populated with users that will
behave differently from one another, and whose “profile” is loosely correlated with
the features known a priori by the provider. This scheme is depicted in Figure 2.
 Gender Age Income

C I

⋯ Moodt ⋯

Callt Complaint

Figure 2. Factors affecting mood and behavior

The actions of each user are governed by a dynamic markovian model whose
current state determines the user’s “mood”, which is in one of four states:
{happy,neutral,angry,churn}. The dynamics of this model are as follows:

1. Time between state changes is larger for smaller values of I

2. The more time spent in “angry” state, the higher probability of churning
3. A high bill (w.r.t. the subscriber income) or an unresolved complaint makes
you more angry (moves you one state towards “angry” or, if you are already
in “angry”, makes your churning probability higher), with a probability
that depends on I
4. A complaint resolved ok makes you go back towards “happy”, with a prob-
ability that depends on I

This internal mood state, which is unknown to the provider, affects the behavior
of the user in multiple ways.

1. A user only complains if ”Angry” (and the longer the time in ”Angry”,
the more s/he complains.
2. A user only churns if ”Angry” (and this becomes more likely the longer
time s/he’s been ”Angry”)
3. The longer time in ”Angry”, the less s/he calls.
4. The longer time in ”Happy”, the more s/he calls.
5. When s/he goes back to ”neutral”, the rate of calls per day goes back
slowly towards the default value for the user.
6. Both duration and number of calls depend on the hidden parameter C.

This scheme is depicted in Figure 3.

 pn,n
t

Neutral
pn,h
t pn,a
t

ph,n
t
pa.n
t

ph,h
t
Happy Angry pa,a
t

pa,c
t

Churn

Figure 3. Factors affecting mood and behavior

7. Results and Scalability

We obtain good levels of recall and precision, roughly to the point that the ran-
domness that we placed in the random generator allows. That is, we can predict
which users will churn with an accuracy that is close to the probability with which
(randomly) decide to churn or not given their internal states. In particular, if we
happen to make the subscribers absolutely deterministic, we get results close to
100%. Of course, the absolute “goodness” of these figures does not mean much,
other than how difficult or easy to predict we made our synthetic data. The point
is that the system is able to correctly remember and put to use the information
in the event stream for one particular purpose, that of churn prediction.
We also checked that Hoeffding trees are extremely good at adapting to
changes. Via the prototype GUI we can vary during the execution parameters
such as prices of our company and the competition, frequency of complaint calls
and % of those resolved satisfactorily, average number of calls per subscriber, etc.
which affect our subscribers’ churn rate. We verified that after a change, predic-
tion accuracy falls because the predictor gets out of sync, but after a few thousand
calls, the new behaviors are captured by the tree and accuracy rises to almost
optimal levels again.
On a commodity PC, the system processes about 10,000 records per second.
Average memory consumption is about 40Mbytes for each 1,000 subscribers with
more than realistic levels of average activities (40 calls day, 2% daily churn rate,
etc.). Thus, there is ample room for upscaling using higher-end machines.
For deployment by large operators, with possibly many million subscribers, it
is clear that scaling out by distributed processing would be necessary. Addition-
ally, the customer base would be geographically distributed over the planet, so
communication latencies among datacenters and traffic splitting and routing must
be taken into account. Finally, the emergence of new technologies and services,
as well as company culture, will undoubtedly put additional constraints on the
processing. From a data mining point of view, techniques for distributed model
building will have to be incorporated. In fact, building several models at geo-
graphically distinct location may be advantageous to capture different customer
patterns at different zones. Since the models themselves are compact, they could
possibly be exchanged among machines and sites and be used cooperatively (e.g.,
with ensemble methods) for better accuracy.

8. Conclusions and Extensions

We have hopefully shown that stream mining technology may help customer churn
prediction on high-volume streams originating from customer activity. The main
difference with exiting, batch-oriented, data mining approaches to the problem
is the ability of these technologies for reacting and adapting fast to changes in
customer behavior, without human intervention, which may have a direct impact
on revenue and image for companies. Although the system is a prototype far from
being deployable, we have shown that even on a single low-end machine we can
deal with quite high data speeds and gracefully handle all the churn prediction
process, including user segmentation and connecting with the customer relation
management subsystem.
Further work and additional research includes testing the system with real
subscriber data collected from live networks and combine additional data sources
from outside operators boundaries.

Acknowledgements

This work was supported by a collaboration agreement between Ericsson Research

and U. Politècnica de Catalunya. Research at UPC was also partially supported
by the BASMATI MICINN project (TIN2011-27479-C04-03). We thank Albert
Bifet and Germán Blanco, for their help at various stages.

References

[1] C. Archaux, H. Laanaya, A. Martin, and A. Khenchaf. An svm based churn detector in
prepaid mobile telephony. In Intl. Conf. on Information and Communication Technolo-
gies: from Theory to Applications (ICTTA), Damascus, Syria, 19-23 April 2004.
[2] Albert Bifet, Geoff Holmes, Richard Kirkby, and Bernhard Pfahringer. MOA: Massive
online analysis. J. Mach. Learn. Res., 11:1601–1604, August 2010.
[3] J. Gama. Knowledge Discovery from Data Streams. Data Mining and Knowledge Discov-
ery. Chapman & Hall/CRC, 2010.
[4] J. Gama and M.M. Gaber. Learning from Data Streams: Processing Techniques in Sensor
Networks. New generation computing. Springer, 2007.
[5] B. Q. Huang, T. M. Kechadi, B. Buckley, G. Kiernan, E. Keogh, and T. Rashid. A
new feature set with new window techniques for customer churn prediction in land-line
telecommunications. Expert Syst. Appl., 37(5):3657–3665, May 2010.
 [6] Bing Quan Huang, Mohand Tahar Kechadi, and Brian Buckley. Customer churn prediction
in telecommunications. Expert Syst. Appl., 39(1):1414–1425, 2012.
[7] Geoff Hulten, Laurie Spencer, and Pedro Domingos. Mining time-changing data streams.
In Proc. 2001 ACM SIGKDD Intl. Conf. on Knowledge Discovery and Data Mining,
pages 97–106, 2001.
[8] Sahand KhakAbi, Mohammad R. Gholamian, and Morteza Namvar. Data mining applica-
tions in customer churn management. In Proceedings of the 2010 International Conference
on Intelligent Systems, Modelling and Simulation, ISMS ’10, pages 220–225, Washington,
DC, USA, 2010. IEEE Computer Society.
[9] Scott A. Neslin, Sunil Gupta, Wagner Kamakura, Junxiang Lu, and Charlotte H. Mason.
Defection Detection: Measuring and Understanding the Predictive Accuracy of Customer
Churn Models. Journal of Marketing Research, 43(2):204–211, May 2006.
[10] P.C. Pendharkar. Genetic algorithm based neural network approaches for predicting churn
in cellular wireless network services. Expert Syst. Appl., 36(3):6714–6720, April 2009.
[11] Anita Prinzie and Dirk Van den Poel. Incorporating sequential information into traditional
classification models by using an element/position-sensitive SAM. Decis. Support Syst.,
42(2):508–526, November 2006.
[12] Yossi Richter, Elad Yom-Tov, and Noam Slonim. Predicting customer churn in mobile
networks through analysis of social groups. In SIAM Intl. Conf. on Data Mining (SDM),
pages 732–741. SIAM, 2010.