Modeling Data Mining Applications for Prediction of Prepaid Churn

Document technical information

Format pdf
Size 1.4 MB
First found May 22, 2018

Document content analysis

Category Also themed
Language
English
Type
not defined
Concepts
no text concepts found

Persons

Carl Perkins
Carl Perkins

wikipedia, lookup

Organizations

Places

Transcript

ISSN 0005-1144
ATKAFF 51(3), 275–283(2010)
Goran Kraljević, Sven Gotovac
Modeling Data Mining Applications for Prediction of Prepaid
Churn in Telecommunication Services
UDK 654.034:004.42
IFAC 5.8.0
Professional paper
This paper defines an advanced methodology for modeling applications based on Data Mining methods that
represents a logical framework for development of Data Mining applications. Methodology suggested here for Data
Mining modeling process has been applied and tested through Data Mining applications for predicting Prepaid users
churn in the telecom industry. The main emphasis of this paper is defining of a successful model for prediction of
potential Prepaid churners, in which the most important part is to identify the very set of input variables that are
high enough to make the prediction model precise and reliable. Several models have been created and compared
on the basis of different Data Mining methods and algorithms (neural networks, decision trees, logistic regression).
For the modeling examples we used WEKA analysis tool.
Key words: Data Mining applications, Prepaid churn model, Neural networks, Decision trees, Logistic regression
Modeliranje Data Mining aplikacija za detekciju churna Prepaid korisnika telekomunikacijskih usluga.
U radu je definirana unapriježena metodologija za proces modeliranja aplikacija zasnovanih na metodama dubinske
analiza podataka (Data Mining), koja predstavlja logički okvir (framework) za razvoj različitih Data Mining aplikacija. Predložena metodologija je primjenjena na primjeru Data Mining aplikacije za detekciju churna Prepaid
korisnika telekomunikacijskih usluga. Glavno težište ovog rada je na definiranju uspješnog modela za predvižanje
potencijalnih Prepaid churn-era. Najvažniji dio izgradnje modela je definiranje onog skupa ulaznih varijabli koje
su u toj mjeri snažne da model predvižanja bude točan i pouzdan. Kroz navedeni primjer je kreirano i usporeženo
više modela zasnovanih na različitim Data Mining metodama i algoritmima (neuronske mreže, stabla odlučivanja,
logistička regresija). Za primjere modeliranja korišten je analitički alat WEKA.
Ključne riječi: aplikacije dubinske analize podataka, prepaid churn model, neuronske mreže, stabla odlučivanja,
logistička regresija
1 INTRODUCTION
th
In the late 20 century, scientists were focused on theoretical bases for Data Mining, and improvements and upgrading of the Data Mining algorithms and methods [2].
In the 21st century, the focus is moved toward scientific research on Data Mining applications, in other words
application of Data Mining methods in real environment,
which resulted in real necessity for these applications on
the market [1],[3],[16],[17].
Modeling and planning of development process of Data
Mining applications has been recognized as a new challenging field for research [5],[6],[7],[8],[9].
Currently, only a small number of users (less than
20%) apply one of the defined methodologies (CRISPDM, SEMMA etc.) for development of predictive Data
Mining applications [4],[11].
There is obviously available space and need for further
research that would result in new methodologies that could
AUTOMATIKA 51(2010) 3, 275–283
meet all requirements today’s business sets for applicative
solutions based on Data Mining methods and algorithms.
This paper will define an advanced methodology for
modeling of applications based on Data Mining analysis that represents a logical framework for development
of Data Mining applications. A focus is on a more precise definition of initial stages (defining Business and Data
Mining goals), as well as final stages of modeling processes of Data Mining applications. These stages are recognized as especially important and critical due to the transition between a Business and Data Mining domain (initial stage) and vice versa (final stage), that happen in these
steps.
Suggested methodology for modeling process of Data
Mining applications has been applied and tested on a Data
Mining application for prediction of Prepaid users churn in
telecommunications.
In contrast to Postpaid segment, Prepaid segment
275
Modeling Data Mining Applications for Prediction of Prepaid Churn in Telecommunication Services
doesn’t imply any contractual obligation between users
and a telecom operator, so the very definition of Prepaid
churn is not simple [13],[14],[15]. In addition, the data
available on Prepaid users are much more inadequate as
compared to Postpaid users, especially when it comes to
customer data domain. This fact adds to complexity of
modeling process of predictive Data Mining applications
related to Prepaid churn.
2 METHODOLOGIES AND MODELS FOR DEVELOPMENT OF DATA MINING APPLICATIONS
It is often heard in business world that there is a will to
use Data Mining applications since they result in visible
benefits, but process of knowledge discovering based on
Data Mining still seems to be not so easy to understand.
Precisely defined steps within a methodological framework and actions to be undertaken at each stage contribute
to demystification of the process of Data Mining application development. Also, big project teams cannot function
properly without such a clearly defined process of application development and modeling, since it considerably facilitates project design, time scheduling and project supervision.
According to the Gartner base, two most commercially
complete Data Mining tools, in terms of their visions and
potentials, are SPSS and SAS. SPSS uses CRISP-DM
methodology, while SAS uses its own SEMMA methodology (although it might be wrong to call SEMMA a methodology), and for this reason the two methodologies (CRISPDM and SEMMA) are best known and most commonly
used for development of Data Mining applications [4].
G. Kraljević, S. Gotovac
Data understanding, Data preparation, Modeling, Evaluation, Deployment) intended as a cyclical process (see
Fig. 1.) [12].
The CRISP-DM Special Interest Group was created
with the goal of supporting the model. At the moment this
groups has more than a hundred members participating actively in work on the version 2.0 [19].
SAS has defined SEMMA (Sample, Explore, Modify,
Model, Assess) model that was incorporated into the commercial Data Mining platform – SAS Enterprise Miner
(Fig. 2) [4].
Fig. 2. Steps in SEMMA process
Latest research prove that less than 20% of consumers
use one of the defined methodologies for development
of predictive Data Mining applications, among which
large majority relates to the most represented CRISP-DM
methodology. Half of the users apply “their own” methodology, and others either doesn’t use any specific one or rely
on consulting knowledge of the outside company and their
methodological frameworks (Fig. 3). [11].
Fig. 1. CRISP-DM process
Fig. 3. Based on 167 respondents who have implemented
predictive analytics.
CRISP-DM (CRoss-Industry Standard Process for Data
Mining) consists of six phases (Business understanding,
Cios and Kurgan adjusted CRISP-DM model to serve
the needs of research in an academic community, and they
276
AUTOMATIKA 51(2010) 3, 275–283
Modeling Data Mining Applications for Prediction of Prepaid Churn in Telecommunication Services
did it by adding several feedback mechanisms that did not
exist in the previous models, and putting emphasis on the
fact that the knowledge revealed within one domain can be
applied in other domains (areas) as well [5],[6].
Reinartz’s model combines the tasks of data selection,
cleaning and transformation as a single data preparation
task [8].
Berry and Linoff have defined a methodology in 11
steps, emphasizing a need for defining a methodology that
would help avoid situations in which things learned are not
true, or they are true but useless at the same time [1].
Recently there have been attempts to improve existing
analytics methodologies and allow them to become more
effective and reliable in providing useful insights in business contexts [9],[10].
However, it is important to stress that the potential value
of analytics has not been fully realised or utilised in business settings as yet.
2.1 Defining a Framework for the Process of Data
Mining Applications Development
There is obviously space and need for further research
that would result in newly defined methodologies that
would meet all requirements today’s business sets for applicative solutions based on Data Mining methods and algorithms.
We will define an advanced methodology for modeling
process of Data Mining applications which represents a
logical framework for development of Data Mining applications.
G. Kraljević, S. Gotovac
1)
2)
3)
Business goals definition
Data mining goals definition
Data preparation
a) Data selection
b) Data cleaning
c) Data transformation
4) Data modeling
5) Analysis of results
6) Deployment
7) Monitoring
Data selection, Data cleaning and Data transformation
can be named together as Data preparation.
We will present advantages of this methodology as compared to the existing one (described previously).
The first systems for knowledge discovering were intended primarily for experienced users familiar with Data
Mining methods and algorithms. Commercial success of
these systems was minimal.
With new tools with graphically intuitive interface, Data
Mining has become available to a large number of users,
but at the same time grew the number of meaningless data
analysis by “experts” without sufficient knowledge on the
data they analyzed, or on the Data Mining methods and algorithms available (neural networks, decision trees, logistic regression, fuzzy expert systems, clustering, Bayesian
networks, etc.).
Terms such as „data fishing“ or „data dredging“ were
empty terms for (sometimes desperate) attempts to find statistically indicative data, and they are definitely not part of
Data Mining analysis [5].
To avoid such “analyses” with no purpose, a change was
introduced relating to the initial development stages.
Firstly, each project has to have from the very beginning, clearly defined business goals. The second key step is
proper mapping of business goals into Data Mining goals.
Failure to properly translate the business problem into a
Data Mining problem leads to one of the dangers we are
trying to avoid - learning things that are true, but not useful.
Business goals and Data Mining goals have to be measurable, reachable, realistic and well-timed. It is important
to avoid words such as improve, optimize, clarify, help etc.
These words are vague and a person using them is obviously not capable of measuring their results.
Fig. 4. Framework for development of DM applications
This methodology defines the following stages of development (Fig. 4):
AUTOMATIKA 51(2010) 3, 275–283
Also, a disadvantage of the existing methodologies is
visible in the domain of integration of ready-made Data
Mining models into business and analytic information systems of companies (BI, CRM etc.). Even the very defining
of Data Mining should include thinking about presentation
of the analysis results to the target users, to ensure that the
model results become an integral part of a business process
277
Modeling Data Mining Applications for Prediction of Prepaid Churn in Telecommunication Services
of a company, comprehensible to the users even without
Data Mining knowledge.
A crucial step that is often forgotten is control and maintenance of the model after implementation.
It has been emphasized already that goals set at the beginning are to be measurable. Only clearly defined and
measurable things can be unambiguously controlled and
their efficiency tested (plan/goal vs. implementation).
Control can be on a daily, weekly, monthly or other basis, depending on the type of the analysis performed.
By observing duration of each step within development
methodology and its role within a whole development process, it is visible that most of the time (60-70%) is spent on
preparing data for analysis.
About 10% of time is consumed on setting business and
Data Mining goals, around 15% on creation of a Data Mining model and 10% on implementation, monitoring and
maintenance of the model (Fig. 5).
G. Kraljević, S. Gotovac
since there are no new users. There are only users of rival companies that are exposed to numerous, carefully designed marketing campaigns in attempts to win them over.
At the same time, a continuous work on customer retention and churn prevention becomes a necessity, because
the competition has similar acquisition issues. Retention
of the existing users is important since it is 5 up to 7 times
cheaper to retain a consumer than to acquire a new one
[16],[17].
Two basic categories of churners are voluntary and involuntary churners (Fig. 7.) [13].
CHURN
VOLUNTARY
Deliberate
I NVOLUNTARY
Incidental
Fig. 7. Churn taxonomy
Fig. 5. Estimate of time (without DW)
If a company owns a Data Warehouse (DW) and time for
data preparation, that will reduce time consumption considerably (30-40%).
Fig. 6. Estimate of time (with DW)
By applying described methodologies for the process of
modeling of applications based on Data Mining analysis,
we’ll perform an analysis for prediction of Prepaid users
churn in telecommunication industry (Section 4.).
3 CHURN TYPES IN TELECOMMUNICATION
INDUSTRY
In the most of the European countries, penetration of
mobile network users has gone beyond 100% (e.g. in Croatia 130%). Acquisition of new users is made more difficult,
278
Involuntary churners are the customers that telecommunication company decides to remove from the subscribers
list. This category includes people that are churned for
fraud (customers who cheat), non-payment (customers
with credit problem), and under-utilization (customers who
don’t use the phone).
Voluntary churn occurs when the customer initiates termination of the service contract. When people think about
Telco churn it is usually the voluntary kind that comes to
mind. Under the category of voluntary churn we recognize
two major types of voluntary churn: 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.
3.1 Postpaid and Prepaid Churn
When considering Postpaid churn, the deactivation date,
i.e. the date that a customer is disconnected from the network, is equal to the churn date [14]. After all, this is the
actual date a customer stops using the operator’s services.
AUTOMATIKA 51(2010) 3, 275–283
Modeling Data Mining Applications for Prediction of Prepaid Churn in Telecommunication Services
In contrast to Postpaid segment, Prepaid segment does
not imply contractual obligations between users and a telecom operator, so the very definition of Prepaid churn is not
that simple.
In the case of Prepaid churn however, the deactivation
date does not necessarily have to match the churn date.
In general, it takes a long period before a Prepaid customer is actually disconnected from the network. In many
cases customers are churned long before they are disconnected from the network. This is exactly the reason why
the deactivation date is not a suitable indicator for churn.
Thus, we are interested in a churn definition which indicates when a customer has permanently stopped using his
Prepaid SIM-card [18].
4 PREDICTING OF PREPAID CHURN
TELECOMMUNICATION INDUSTRY
IN
Successful detection of potential churners enables companies to define activities for their retention. Data Mining
enables us to predict behavior of mobile networks users
[13],[14],[15].
This example will show a Data Mining application for
prediction of Prepaid churn in the telecommunication company HT Mostar (Bosnia & Herzegovina) by using previously defined methodology (see section 2.1).
Data that are going to be analyzed relate to a six-month
period (from 1st July 2009 to 1st January 2010).
4.1 Business Goals Definition
As we mentioned earlier, a business goal is to be measurable, reachable, realistic and well-timed.
Business goal: To decrease churn rate of Prepaid users
by 20% in a three-month time.
What is missing in the definition above to be comprehensible to everybody without ambiguity is a clear definition of what churn is in a Prepaid segment.
A Prepaid churn user in this example is defined as a user
who has neither had any calls or incoming calls during
the last three months, nor used any other services (SMS,
MMS, mobile internet. . . ), nor any additional payment
within these three months.
4.2 Data Mining Goals Definition
One of the greatest risks in the project was definitely
mapping of business goals into Data Mining goals. Experts
from business and Data Mining domain have to cooperate
in defining business and Data Mining project goals.
Failure to properly translate the business problem into
a Data Mining problem leads to one of the dangers we are
trying to avoid - learning things that are true, but not useful.
Data mining goal: To create a model that will predict
potential Prepaid churners with 90% accuracy.
AUTOMATIKA 51(2010) 3, 275–283
G. Kraljević, S. Gotovac
4.3 Data Preparation (Data Selection, Cleaning and
Transformation)
The data set used throughout this work is from a
telecommunication company in Bosnia and Herzegovina
(HT Mostar).
HT Mostar company has a DW/BI system implemented
in its infrastructure since 2007. Data Warehouse (DW) is
organized as a “star-schema” data model (or snowflakeschema” data model when it was necessary), and it contains all available data of Postpaid/Prepaid users and services they used.
ETL procedure enabled data extraction from source systems, their cleaning, transformation and aggregation, as
well as data import into DW, so data preparation is considerably made shorter and simplified for the Data Mining
analysis.
Also, the data warehouse contained aggregated data (on
a monthly basis) on services by Prepaid users, which additionally facilitates necessary data preparation.
From DW we extracted data in the period between 1st
July 2009 and 1st January 2010.
There are three groups of variables plus target variable
that have been combined to create the dataset:
– customer data
– traffic (usage) data (outgoing/incoming)
– recharge data
Customer data:
–
–
RATE_PLAN (TARIFF MODEL)
CUSTOMER_MONTH_DURATION
Outgoing traffic (usage) data:
–
–
–
–
–
–
–
–
–
–
OUTGOING_SMS_NUMBER
∆ OUTGOING_SMS_NUMBER
OUTGOING_CALLS_NUMBER
∆ OUTGOING_CALLS_NUMBER
OUTGOING_CALLS_MINUTES
∆ OUTGOING_CALLS_MINUTES
SERVICE_CALLS_NUMBER
∆ SERVICE_CALLS_NUMBER
COMPETITION_SERVICE_CALLS_NUMBER
∆ COMPETITION_SERVICE_CALLS_NUMBER
Incoming traffic (usage) data:
–
–
–
–
–
–
INCOMING_SMS_NUMBER
∆ INCOMING_SMS_NUMBER
INCOMING_CALLS_NUMBER
∆ INCOMING_CALLS_NUMBER
INCOMING_CALLS_MINUTES
∆ INCOMING_CALLS_MINUTES
Recharge data:
279
Modeling Data Mining Applications for Prediction of Prepaid Churn in Telecommunication Services
–
–
–
–
NUMBER_OF_RECHARGES
∆ NUMBER_OF_RECHARGES
TOTAL_RECHARGED_AMOUNT
∆ TOTAL_RECHARGED_AMOUNT
DM
Algorithm
Training Set
Delta (∆) variables (VAR) relate to the change of a variable between the two periods (p1 and p2) according to the
following formula:
∆VAR. =
G. Kraljević, S. Gotovac
VAR.(p2) − VAR.(p1)
VAR.(p1)
(1)
Attr1
Attr2
...
...
Attr3 Target
...
0
...
...
...
1
Learn
Model
Model
Validation Set
Attr1
Attr2
...
...
Attr3 Target
...
Apply
Model
?
Fig. 8. Data modeling
where:
p2 = period 2 (October, November and December
2009.)
p1 = period 1 (July, August and September 2009.). E.g.
for the variable ∆ OUTGOING_CALLS_MINUTES (OCM):
∆OCM =
OCM(p2) − OCM(p1)
OCM(p1)
(2)
where:
OCM = Outgoing_Calls_Minutes
p2 = period 2 (October, November and December
2009.)
p1 = period 1 (July, August and September 2009.)
Naturally, the variables indicating changes between the
two periods (p1 i p2) are valid only if a condition is met.
CUSTOMER_MONTH_DURATION >= 6
th
th
th
Other variables relate to the period 2 (10 , 11 and 12
month in 2009). That means that analysis took into consideration last three months, i.e. variables relating to the
change (∆), and the change within last three months as
compared to the previous three month period.
In the data preparing process, most of the time was spent
on calculation of variables relating to the change (∆), since
these variables are not present in DW.
Fig. 9. Training and validation set for modeling
The most important thing to remember about model
building is that it is an iterative process. No single method
works best in all cases.
For training and model validation, we used a sample of
3,000 users, out of which 2/3 is training set (2,000 users)
and 1/3 the validation set (1,000 users) (Fig. 9).
Users who cannot be churners for any reason were
excluded from the sample (test users, employees of HT
Mostar).
The training set had 259 churners, and 1,741 nonchurners (Fig. 10). The validation set had 123 churners
and 877 non-churners (Fig. 11).
Target data
The target data for customer is represented by a 0
(FALSE) for non-churner or a 1 (TRUE) for churner.
4.4 Data Modeling
Churn predictive modeling refers to the task of building
a model for the target variable (churner or non-churner) as
a function of the explanatory variables [2].
Fig. 10. Training set for modeling
Modeling is simply the act of building a model in one
situation where you know the answer and then applying it
to another situation that you don’t (Fig. 8).
The predictive power of different Data Mining models
(Neural Network, Decision Tree and Logistic Regression)
were analyzed and compared.
280
AUTOMATIKA 51(2010) 3, 275–283
Modeling Data Mining Applications for Prediction of Prepaid Churn in Telecommunication Services
G. Kraljević, S. Gotovac
Table 4. Confusion matrix (decision tree model)
Fig. 11. Validation set for modeling
We decided to use Open source Data Mining software
for analyzing and our choice was Weka. Weka was developed at the University of Waikato in New Zealand, and
the name stands for Waikato Environment for Knowledge
Analysis. The system is written in Java and distributed under the terms of the GNU General Public Licence [2],[20].
It runs on almost any platform and has been tested under
Linux, Windows, and Macintosh operating systems.
Evolution of the performance of a prediction model is
based on the counts of test records correctly and incorrectly
predicted by the model. These counts are tabulated in a
table known as a confusion matrix (Table 1).
Table 1. Confusion matrix for a 2-class problem
4.5 Analysis of Results
Although a confusion matrix provides the information
needed to determine how well a prediction model performs, summarizing this information with a single number would make it more convenient to compare the performance of different models. This can be done using a
performance metric such as accuracy and error rate [2].
Accuracy =
Number of correct predictions
Total number of predictions
Error rate =
Number of wrong predictions
Total number of predictions
Table 5. Comparison of different DM models
Confusion matrix is shown for each model. Based on
the entries in the confusion matrix we can calculate the
total number of correct and incorrect predictions.
Table 2. Confusion matrix (neural network model)
Decision tree method proved to be the most successful
for concrete application in modeling Prepaid users churn
prediction (Table 5).
Decision trees are a Data Mining method aimed at classification of attributes with regard to the set target variable
(in this case CHURN).
Table 3. Confusion matrix (logistic regression model)
AUTOMATIKA 51(2010) 3, 275–283
A record enters the tree at the root node. The root node
applies a test to determine which child node the record will
encounter next. There are different algorithms for choosing the initial test, but the goal is always the same: To
choose the test that best discriminates among the target
classes. This process is repeated until the record arrives at
a leaf node. All the records that end up at a given leaf of
the tree are classified the same way. There is a unique path
from the root to each leaf. That path is an expression of the
rule used to classify the records (churn = TRUE or churn =
FALSE) [1].
281
Modeling Data Mining Applications for Prediction of Prepaid Churn in Telecommunication Services
The main advantage of this method is that it presents
its results in the form of easily readable rules that can be of
great value, with a possibility of churn prediction for an individual user. Besides, this method can indicate dominant
variable.
It is important, though, to stress that variable dominance
can be observed in both directions:
–
–
Domination toward churn (user churn)
Domination against churn (user retention)
4.6 Deployment
When defining goals we had a very important and sensitive issue of merging business and Data Mining environments (mapping business goals into Data Mining goals).
Now we have a similar goal to merge these two environments, only in an opposite direction. It is necessary to
transform results from Data Mining environment into business environment and present them in a more comprehensible way.
In this example we created an OLAP cube (by using
Microsoft SQL Server Analysis Services 2007) with all results gained from the analysis. Users had direct approach
to the cube through Excel 2007 environment.
Based on the data available and identified potential
churners, Marketing and Sales decided to undertake retention activities. E.g. an offer to switch to a more convenient
tariff model, an account bonus, more favorable price for
mobile phone purchase, etc...
Especially important is to undertake activities on retention of profitable Prepaid users (e.g. those with consumption of more than 15 EUR).
4.7 Monitoring
Data mining model building is an iterative process that
can not be fully automated due to the constant change in
business processes and environment.
Model control and maintenance in this example is going
to be performed on a monthly basis, when its accuracy will
be checked with the new set of data. From the DW data is
always extracted for a last six-month period.
According to the needs, this model can be adjusted by
e.g. adding new available variables that influence accuracy
of the model prediction. It is possible to apply a selected
algorithm ( a decision tree), since with a new variable and
new set of data some better predictive characteristics might
be revealed (neural networks, logistic regression).
Of course that after the first results it is possible to make
decisions on adjusting set business and Data Mining goals.
282
G. Kraljević, S. Gotovac
5 FUTURE RESEARCH
Future research will deal with building of the whole system for prevention of Prepaid churn.
Data Mining model defined and described here for detection of potential Prepaid churners is just one part of that
system.
Detection model results are to be matched with defined
segments of users, and for each segment it is necessary to
define appropriate action. This work is with Prepaid users
much more complex than in the case of Postpaid users who
are mostly bound by their contractual obligations (e.g. for
1 or 2 years), so the most critical points are expire dates
of the contracts. Prepaid doesn’t imply contractual obligations which means that users are able to stop using services
at any point, without previous notification.
In the Prepaid world it is common for users to have more
than one card from different telecom companies. It is then
great challenge to detect such users and keep them by additional cross-sell / up-sell offers within one network and
prevent churn in that way.
6 CONCLUSION
This paper emphasizes a necessity for development of
Data Mining applications in line with a clearly defined
methodological framework.
An understandable development methodology is one of
the key parameters for successful modeling of applications
for Prepaid users churn prediction in telecommunications,
which was proved in the section 4. of this paper.
The main emphasis of this paper was defining of a successful model for prediction of potential Prepaid churners,
in which the most important part was to identify the very
set of input variables that were high enough to make the
prediction model precise and reliable.
Definition of Prepaid churn and the very modeling of an
application is more complex than the same task for Postpaid users. More complexity is brought into modeling by a
lower data amount available for Prepaid users, so the variables are often set by using traffic data (calls, SMS, MMS,
mobile internet...) and data on additional payments.
A successful model for prediction and prevention of
Prepaid churn in telecommunication companies can influence very positively an overall profit of companies, due to
the fact that far less money needs to be invested into development of a predictive Data Mining model and marketing
preventive action to retain users, as compared to the possible loss cause by these users churn.
REFERENCES
[1] Michael J. A. Berry, Gordon S. Linoff: „Data Mining Techniques: For Marketing, Sales and Customer Relationship
Management“, Wiley Publishing, Indidana, 2004.
AUTOMATIKA 51(2010) 3, 275–283
Modeling Data Mining Applications for Prediction of Prepaid Churn in Telecommunication Services
[2] Ian H. Witten, Eibe Frank: „Data Mining: Practical Machine Learning Tools and Techniques“, Morgan Kaufmann
Publishers, San Francisco, 2005.
[3] Paolo Giudici, Silvia Figini: „Applied Data Mining for
Business and Industry“, John Wiley & Sons, 2009.
[4] David L. Olson, Dursun Delen: „Advanced Data Mining
Techniques“, Springer, 2008.
[5] L. A. Kurgan, P. Musilek: “A survey of Knowledge Discovery and Data Mining process models”, The Knowledge
Engineering Review, Vol.21, pp.1-24, 2006.
[6] K. Cios, L.A. Kurgan: „Trends in Data Mining and
knowledge discovery“– in N.Pal, L.Jain (eds): „Advanced
Techniques in Knowledge Discovery and Data Mining“,
Springer, pp. 1–26., 2005.
[7] T. Li, D. Ruan: „An extended process model of knowledge
discovery in databases“, Journal of Enterprise Information
Management, Volume 20, Number 2, pp. 169-177(9), 2007.
[8] T. Reinartz: „Stages of the discovery process“ – In
W.Klosgen, J.Zytkow (eds): „Handbook of DataMining
and Knowledge Discovery“, Oxford University Press, pp.
185–192., 2002.
[9] P. González-Aranda, E. Menasalvas, S. Millán, C.Ruiz, J.
Segovia: “Towards a Methodology for Data Mining Project
Development: The Importance of Abstraction“, Volume
118/2008, Springer, pp. 165–178., 2008.
[10] M. Van Rooyen: „An evaluation of the utility of two Data
Mining project methodologies“– In S.J.Simoff, G.Williams
(eds): „Proceedings of the 3rd Australasian Data Mining
Conference“, Cairns, Australia, pp. 85-94., 2004.
[11] Wayne W. Eckerson: „Predictive Analytics – Extending the
Value of Your Data Warehousing Investment“, TDWI (The
Data Warehousing Institute), 2007.
[12] C. Shearer: „The CRISP-DM Model: The new blueprint
for Data Mining“, Journal of Data Warehousing, Volume 5,
Number 4, Fall 2000, pp.13-22, 2000.
[13] R. Mattison: “The Telco Churn Management Handbook”,
XiT Press, Oakwood Hills, Illinois, 2005.
[14] G. Kraljević, I. Krasić, S. Gotovac: „Data Mining Techniques and Applications in Telecommunication Systems“,
MIPRO 2007., BIS – Business Intelligence Systems, pp.
257–261, 2007.
[15] Marko Ferišak: „Intelligent Automata Models in Data Mining Processes for Telecommunication Service Forecasting“,
PhD Thesis, FER, Zagreb, 2005.
[16] S.M.H. Jansen: „Customer Segmentation and Customer
Profiling for a Mobile Telecommunications Company
Based on Usage Behavior“, A Vodafone Case Study July,
2007.
[17] E.W.T. Ngai , Li Xiu , D.C.K. Chau: „Application of
Data Mining techniques in customer relationship management: A literature review and classification“, Expert Systems with Applications: An International Journal, v.36 n.2,
pp. 2592–2602, 2009.
AUTOMATIKA 51(2010) 3, 275–283
G. Kraljević, S. Gotovac
[18] A.T.Jahromi, M.Moeini, I.Akbari, A.Akbarzadeh: „A dualstep multi-algorithm approach for churn prediction in Prepaid telecommunications service providers”, ICIM 2009,
Sao Paulo, Brazil, Dec.2009.
[19] CRISP-DM: Cross Industry Standard Process Model for
Data Mining (http://www.crisp-dm.org), 2010.
[20] WEKA (Open source,
Data Mining software
in Java), University of Waikato, New Zealand
(http://www.cs.waikato.ac.nz/ml/weka),
Weka 3.7.0.,
2010.
Goran Kraljević received B.Sc. degree in 2000.
in Computing from the Faculty of Electrical Engineering and Computing, University of Zagreb,
Croatia (FER Zagreb). Since 2003. he has been
employed in company HT Mostar. Currently he
is Head of Department for Business Information
Systems. Since 2004. he has worked as an external associate - assistant at the Faculty of Mechanical Engineering and Computing, University
of Mostar (FSR Mostar). His research areas of
interest include data mining and knowledge discovery, data warehousing and business intelligence systems.
Sven Gotovac received B.Sc. degree in 1982. in
Electrical Engineering from the Faculty of Electrical Engineering, Mechanical Engineering and
Naval Architecture University of Split, Croatia.
In 1988. he received M.Sc. degree in Computing
Science from the Faculty of Electrical Engineering and Computing, University of Zagreb, Croatia. From Technical University Berlin, Germany
in 1994. he received Ph.D. degree in Electrical
Engineering. Since 1983. he has been employed
at the Faculty of Electrical Engineering, Mechanical Engineering and Naval Architecture, University of Split, Croatia.
Now he is Full Professor, and the head of the Department for Computer
Architecture and Operating Systems. His research interest includes embedded systems and information system design and analysis.
AUTHORS’ ADDRESSES
Goran Kraljević
HT Mostar
Kneza Branimira bb, 88000 Mostar, B&H
[email protected]
Sven Gotovac
Faculty of Electrical Engineering, Mechanical Engineering
and Naval Architecture
Ružera Boškovića bb, 21000 Split, Croatia
[email protected]
Received: 2010-07-12
Accepted: 2010-11-13
283

Similar documents

×

Report this document