486 Section: Service

Data Mining in the Telecommunications

Industry
Gary M. Weiss
Fordham University, USA

INTRODUCTION Telecommunication companies maintain data about

the phone calls that traverse their networks in the form of
The telecommunications industry was one of the first call detail records, which contain descriptive informa-
to adopt data mining technology. This is most likely tion for each phone call. In 2001, AT&T long distance
because telecommunication companies routinely gener- customers generated over 300 million call detail records
ate and store enormous amounts of high-quality data, per day (Cortes & Pregibon, 2001) and, because call
have a very large customer base, and operate in a detail records are kept online for several months, this
rapidly changing and highly competitive environment. meant that billions of call detail records were readily
Telecommunication companies utilize data mining to available for data mining. Call detail data is useful for
improve their marketing efforts, identify fraud, and marketing and fraud detection applications.
better manage their telecommunication networks. Telecommunication companies also maintain exten-
However, these companies also face a number of data sive customer information, such as billing information,
mining challenges due to the enormous size of their as well as information obtained from outside parties,
data sets, the sequential and temporal aspects of their such as credit score information. This information
data, and the need to predict very rare eventssuch as can be quite useful and often is combined with tele-
customer fraud and network failuresin real-time. communication-specific data to improve the results
The popularity of data mining in the telecommunica- of data mining. For example, while call detail data
tions industry can be viewed as an extension of the use can be used to identify suspicious calling patterns, a
of expert systems in the telecommunications industry customers credit score is often incorporated into the
(Liebowitz, 1988). These systems were developed to analysis before determining the likelihood that fraud
address the complexity associated with maintaining a is actually taking place.
huge network infrastructure and the need to maximize Telecommunications companies also generate and
network reliability while minimizing labor costs. The store an extensive amount of data related to the opera-
problem with these expert systems is that they are ex- tion of their networks. This is because the network
pensive to develop because it is both difficult and time- elements in these large telecommunication networks
consuming to elicit the requisite domain knowledge have some self-diagnostic capabilities that permit them
from experts. Data mining can be viewed as a means to generate both status and alarm messages. These
of automatically generating some of this knowledge streams of messages can be mined in order to support
directly from the data. network management functions, namely fault isolation
and prediction.
The telecommunication industry faces a number of
BACKGROUND data mining challenges. According to a Winter Cor-
poration survey (2003), the three largest databases all
The data mining applications for any industry depend belong to telecommunication companies, with France
on two factors: the data that are available and the busi- Telecom, AT&T, and SBC having databases with 29, 26,
ness problems facing the industry. This section provides and 25 Terabytes, respectively. Thus, the scalability of
background information about the data maintained by data mining methods is a key concern. A second issue
telecommunications companies. The challenges as- is that telecommunication data is often in the form of
sociated with mining telecommunication data are also transactions/events and is not at the proper semantic
described in this section. level for data mining. For example, one typically wants
to mine call detail data at the customer (i.e., phone-

Copyright 2009, IGI Global, distributing in print or electronic forms without written permission of IGI Global is prohibited.
Data Mining in the Telecommunications Industry

line) level but the raw data represents individual phone was MCIs Friends and Family program. This program,
calls. Thus it is often necessary to aggregate data to long since retired, began after marketing researchers D
the appropriate semantic level (Sasisekharan, Seshadri identified many small but well connected subgraphs
& Weiss, 1996) before mining the data. An alternative in the graphs of calling activity (Han, Altman, Kumar,
is to utilize a data mining method that can operate on Mannila & Pregibon, 2002). By offering reduced rates
the transactional data directly and extract sequential or to customers in ones calling circle, this marketing strat-
temporal patterns (Klemettinen, Mannila & Toivonen, egy enabled the company to use their own customers
1999; Weiss & Hirsh, 1998). as salesmen. This work can be considered an early use
Another issue arises because much of the telecom- of social-network analysis and link mining (Getoor &
munications data is generated in real-time and many Diehl, 2005). A more recent example uses the interac-
telecommunication applications, such as fraud identi- tions between consumers to identify those customers
fication and network fault detection, need to operate in likely to adopt new telecommunication services (Hill,
real-time. Because of its efforts to address this issue, Provost & Volinsky, 2006). A more traditional approach
the telecommunications industry has been a leader in involves generating customer profiles (i.e., signatures)
the research area of mining data streams (Aggarwal, from call detail records and then mining these profiles
2007). One way to handle data streams is to maintain a for marketing purposes. This approach has been used
signature of the data, which is a summary description of to identify whether a phone line is being used for voice
the data that can be updated quickly and incrementally. or fax (Kaplan, Strauss & Szegedy, 1999) and to clas-
Cortes and Pregibon (2001) developed signature-based sify a phone line as belonging to a either business or
methods and applied them to data streams of call detail residential customer (Cortes & Pregibon, 1998).
records. A final issue with telecommunication data Over the past few years, the emphasis of marketing
and the associated applications involves rarity. For applications in the telecommunications industry has
example, both telecommunication fraud and network shifted from identifying new customers to measuring
equipment failures are relatively rare. Predicting and customer value and then taking steps to retain the most
identifying rare events has been shown to be quite dif- profitable customers. This shift has occurred because it
ficult for many data mining algorithms (Weiss, 2004) is much more expensive to acquire new telecommunica-
and therefore this issue must be handled carefully in tion customers than retain existing ones. Thus it is useful
order to ensure reasonably good results. to know the total lifetime value of a customer, which
is the total net income a company can expect from that
customer over time. A variety of data mining methods
MAIN FOCUS are being used to model customer lifetime value for
telecommunication customers (Rosset, Neumann, Eick
Numerous data mining applications have been de- & Vatnik, 2003; Freeman & Melli, 2006).
ployed in the telecommunications industry. However, A key component of modeling a telecommunica-
most applications fall into one of the following three tion customers value is estimating how long they
categories: marketing, fraud detection, and network will remain with their current carrier. This problem
fault isolation and prediction. is of interest in its own right since if a company can
predict when a customer is likely to leave, it can take
Telecommunications Marketing proactive steps to retain the customer. The process of
a customer leaving a company is referred to as churn,
Telecommunication companies maintain an enormous and churn analysis involves building a model of
amount of information about their customers and, due customer attrition. Customer churn is a huge issue in
to an extremely competitive environment, have great the telecommunication industry where, until recently,
motivation for exploiting this information. For these telecommunication companies routinely offered large
reasons the telecommunications industry has been a cash incentives for customers to switch carriers. Nu-
leader in the use of data mining to identify customers, merous systems and methods have been developed to
retain customers, and maximize the profit obtained from predict customer churn (Wei & Chin, 2002; Mozer,
each customer. Perhaps the most famous use of data Wolniewicz, Grimes, Johnson, & Kaushansky, 2000;
mining to acquire new telecommunications customers Mani, Drew, Betz & Datta, 1999; Masand, Datta, Mani,

487
 Data Mining in the Telecommunications Industry

& Li, 1999). These systems almost always utilize call was augmented by comparing the new calling be-
detail and contract data, but also often use other data havior to profiles of generic fraudand fraud is only
about the customer (credit score, complaint history, signaled if the behavior matches one of these profiles.
etc.) in order to improve performance. Churn predic- Customer level data can also aid in identifying fraud.
tion is fundamentally a very difficult problem and, For example, price plan and credit rating information
consequently, systems for predicting churn have been can be incorporated into the fraud analysis (Rosset,
only moderately effectiveonly demonstrating the Murad, Neumann, Idan, & Pinkas, 1999). More recent
ability to identify some of the customers most likely work using signatures has employed dynamic clustering
to churn (Masand et al., 1999). as well as deviation detection to detect fraud (Alves
et al., 2006). In this work, each signature was placed
Telecommunications Fraud Detection within a cluster and a change in cluster membership
was viewed as a potential indicator of fraud.
Fraud is very serious problem for telecommunica- There are some methods for identifying fraud that
tion companies, resulting in billions of dollars of lost do not involve comparing new behavior against a
revenue each year. Fraud can be divided into two profile of old behavior. Perpetrators of fraud rarely
categories: subscription fraud and superimposition work alone. For example, perpetrators of fraud often
fraud (Fawcett and Provost, 2002). Subscription fraud act as brokers and sell illicit service to othersand the
occurs when a customer opens an account with the illegal buyers will often use different accounts to call
intention of never paying the account and superim- the same phone number again and again. Cortes and
position fraud occurs when a perpetrator gains illicit Pregibon (2001) exploited this behavior by recognizing
access to the account of a legitimate customer. In this that certain phone numbers are repeatedly called from
latter case, the fraudulent behavior will often occur compromised accounts and that calls to these numbers
in parallel with legitimate customer behavior (i.e., is are a strong indicator that the current account may
superimposed on it). Superimposition fraud has been be compromised. A final method for detecting fraud
a much more significant problem for telecommunica- exploits human pattern recognition skills. Cox, Eick
tion companies than subscription fraud. Ideally, both & Wills (1997) built a suite of tools for visualizing
subscription fraud and superimposition fraud should data that was tailored to show calling activity in such
be detected immediately and the associated customer a way that unusual patterns are easily detected by us-
account deactivated or suspended. However, because it ers. These tools were then used to identify international
is often difficult to distinguish between legitimate and calling fraud.
illicit use with limited data, it is not always feasible
to detect fraud as soon as it begins. This problem is Telecommunication Network Fault
compounded by the fact that there are substantial costs Isolation and Prediction
associated with investigating fraud, as well as costs if
usage is mistakenly classified as fraudulent (e.g., an Monitoring and maintaining telecommunication net-
annoyed customer). works is an important task. As these networks became
The most common technique for identifying super- increasingly complex, expert systems were developed
imposition fraud is to compare the customers current to handle the alarms generated by the network elements
calling behavior with a profile of his past usage, using (Weiss, Ros & Singhal, 1998). However, because these
deviation detection and anomaly detection techniques. systems are expensive to develop and keep current, data
The profile must be able to be quickly updated because mining applications have been developed to identify
of the volume of call detail records and the need to and predict network faults. Fault identification can be
identify fraud in a timely manner. Cortes and Pregi- quite difficult because a single fault may result in a
bon (2001) generated a signature from a data stream cascade of alarmsmany of which are not associated
of call-detail records to concisely describe the calling with the root cause of the problem. Thus an important
behavior of customers and then they used anomaly part of fault identification is alarm correlation, which
detection to measure the unusualness of a new call enables multiple alarms to be recognized as being
relative to a particular account. Because new behavior related to a single fault.
does not necessarily imply fraud, this basic approach

488
Data Mining in the Telecommunications Industry

The Telecommunication Alarm Sequence Analyzer the telecommunications industry, but as telecommuni-
(TASA) is a data mining tool that aids with fault iden- cation companies start providing television service to D
tification by looking for frequently occurring temporal the home and more sophisticated cell phone services
patterns of alarms (Klemettinen, Mannila & Toivonen, become available (e.g., music, video, etc.), it is clear
1999). Patterns detected by this tool were then used to that new data mining applications, such as recommender
help construct a rule-based alarm correlation system. systems, will be developed and deployed.
Another effort, used to predict telecommunication Unfortunately, there is also one troubling trend that
switch failures, employed a genetic algorithm to mine has developed in recent years. This concerns the in-
historical alarm logs looking for predictive sequential creasing belief that U.S. telecommunication companies
and temporal patterns (Weiss & Hirsh, 1998). One are too readily sharing customer records with govern-
limitation with the approaches just described is that mental agencies. This concern arose in 2006 due to
they ignore the structural information about the under- revelationsmade public in numerous newspaper and
lying network. The quality of the mined sequences can magazine articlesthat telecommunications companies
be improved if topological proximity constraints are were turning over information on calling patterns to
considered in the data mining process (Devitt, Duffin the National Security Agency (NSA) for purposes of
and Moloney, 2005) or if substructures in the telecom- data mining (Krikke, 2006). If this concern continues to
munication data can be identified and exploited to allow grow unchecked, it could lead to restrictions that limit
simpler, more useful, patterns to be learned (Baritchi, the use of data mining for legitimate purposes.
Cook, & Lawrence, 2000). Another approach is to use
Bayesian Belief Networks to identify faults, since they
can reason about causes and effects (Sterritt, Adamson, CONCLUSION
Shapcott & Curran, 2000).
The telecommunications industry has been one of the
early adopters of data mining and has deployed numer-
FUTURE TRENDS ous data mining applications. The primary applications
relate to marketing, fraud detection, and network
Data mining should play an important and increasing monitoring. Data mining in the telecommunications
role in the telecommunications industry due to the industry faces several challenges, due to the size of
large amounts of high quality data available, the com- the data sets, the sequential and temporal nature of the
petitive nature of the industry and the advances being data, and the real-time requirements of many of the
made in data mining. In particular, advances in mining applications. New methods have been developed and
data streams, mining sequential and temporal data, existing methods have been enhanced to respond to
and predicting/classifying rare events should benefit these challenges. The competitive and changing nature
the telecommunications industry. As these and other of the industry, combined with the fact that the industry
advances are made, more reliance will be placed on the generates enormous amounts of data, ensures that data
knowledge acquired through data mining and less on the mining will play an important role in the future of the
knowledge acquired through the time-intensive process telecommunications industry.
of eliciting domain knowledge from expertsalthough
we expect human experts will continue to play an
important role for some time to come. REFERENCES
Changes in the nature of the telecommunications
industry will also lead to the development of new ap- Aggarwal, C. (Ed.). (2007). Data Streams: Models and
plications and the demise of some current applications. Algorithms. New York: Springer.
As an example, the main application of fraud detection
in the telecommunications industry used to be in cellular Alves, R., Ferreira, P., Belo, O., Lopes, J., Ribeiro, J.,
cloning fraud, but this is no longer the case because the Cortesao, L., & Martins, F. (2006). Discovering telecom
problem has been largely eliminated due to technologi- fraud situations through mining anomalous behavior
cal advances in the cell phone authentication process. patterns. Proceedings of the ACM SIGKDD Workshop
It is difficult to predict what future changes will face on Data Mining for Business Applications (pp. 1-7).
New York: ACM Press.

489
 Data Mining in the Telecommunications Industry

Baritchi, A., Cook, D., & Holder, L. (2000). Discov- Liebowitz, J. (1988). Expert System Applications to
ering structural patterns in telecommunications data. Telecommunications. New York, NY: John Wiley &
Proceedings of the Thirteenth Annual Florida AI Re- Sons.
search Symposium (pp. 82-85).
Mani, D., Drew, J., Betz, A., & Datta, P (1999). Statistics
Cortes, C., & Pregibon, D (1998). Giga-mining. Pro- and data mining techniques for lifetime value modeling.
ceedings of the Fourth International Conference on Proceedings of the Fifth ACM SIGKDD International
Knowledge Discovery and Data Mining (pp. 174-178). Conference on Knowledge Discovery and Data Mining
New York, NY: AAAI Press. (pp. 94-103). New York, NY: ACM Press.
Cortes, C., & Pregibon, D. (2001). Signature-based Masand, B., Datta, P., Mani, D., & Li, B. (1999).
methods for data streams. Data Mining and Knowledge CHAMP: A prototype for automated cellular churn
Discovery, 5(3), 167-182. prediction. Data Mining and Knowledge Discovery,
3(2), 219-225.
Cox, K., Eick, S., & Wills, G. (1997). Visual data min-
ing: Recognizing telephone calling fraud. Data Mining Mozer, M., Wolniewicz, R., Grimes, D., Johnson, E.,
and Knowledge Discovery, 1(2), 225-231. & Kaushansky, H. (2000). Predicting subscriber dis-
satisfaction and improving retention in the wireless
Devitt, A., Duffin, J., & Moloney, R. (2005). Topo-
telecommunication industry. IEEE Transactions on
graphical proximity for mining network alarm data.
Neural Networks, 11, 690-696.
Proceedings of the 2005 ACM SIGCOMM Workshop
on Mining Network Data (pp. 179-184). New York: Rosset, S., Murad, U., Neumann, E., Idan, Y., & Gadi,
ACM Press. P. (1999). Discovery of fraud rules for telecommu-
nicationschallenges and solutions. Proceedings of
Fawcett, T., & Provost, F. (2002). Fraud Detection. In
the Fifth ACM SIGKDD International Conference on
W. Klosgen & J. Zytkow (Eds.), Handbook of Data
Knowledge Discovery and Data Mining (pp. 409-413).
Mining and Knowledge Discovery (pp. 726-731). New
New York: ACM Press.
York: Oxford University Press.
Rosset, S., Neumann, E., Eick, U., & Vatnik (2003).
Freeman, E., & Melli, G. (2006). Championing of
Customer lifetime value models for decision support.
an LTV model at LTC. SIGKDD Explorations, 8(1),
Data Mining and Knowledge Discovery, 7(3), 321-
27-32.
339.
Getoor, L., & Diehl, C.P. (2005). Link mining: A survey.
Sasisekharan, R., Seshadri, V., & Weiss, S (1996). Data
SIGKDD Explorations, 7(2), 3-12.
mining and forecasting in large-scale telecommunica-
Hill, S., Provost, F., & Volinsky, C. (2006). Network- tion networks. IEEE Expert, 11(1), 37-43.
based marketing: Identifying likely adopters via con-
Sterritt, R., Adamson, K., Shapcott, C., & Curran, E.
sumer networks. Statistical Science, 21(2), 256-276.
(2000). Parallel data mining of Bayesian networks from
Kaplan, H., Strauss, M., & Szegedy, M. (1999). Just telecommunication network data. Proceedings of the
the faxdifferentiating voice and fax phone lines us- 14th International Parallel and Distributed Processing
ing call billing data. Proceedings of the Tenth Annual Symposium, IEEE Computer Society.
ACM-SIAM Symposium on Discrete Algorithms (pp.
Wei, C., & Chiu, I (2002). Turning telecommunications
935-936). Philadelphia, PA: Society for Industrial and
call details to churn prediction: A data mining approach.
Applied Mathematics.
Expert Systems with Applications, 23(2), 103-112.
Klemettinen, M., Mannila, H., & Toivonen, H. (1999).
Weiss, G., & Hirsh, H (1998). Learning to predict rare
Rule discovery in telecommunication alarm data.
events in event sequences. In R. Agrawal & P. Stolorz
Journal of Network and Systems Management, 7(4),
(Eds.), Proceedings of the Fourth International Confer-
395-423.
ence on Knowledge Discovery and Data Mining (pp.
Krikke, J. (2006). Intelligent surveillance empowers 359-363). Menlo Park, CA: AAAI Press.
security analysts. IEEE Intelligent Systems, 21(3),
102-104.
490
Data Mining in the Telecommunications Industry

Weiss, G., Ros, J., & Singhal, A. (1998). ANSWER: Call Detail Record: Contains the descriptive in-
Network monitoring using object-oriented rule. Pro- formation associated with a single phone call. D
ceedings of the Tenth Conference on Innovative Ap-
Churn: Customer attrition. Churn prediction refers
plications of Artificial Intelligence (pp. 1087-1093).
to the ability to predict that a customer will leave a
Menlo Park: AAAI Press.
company before the change actually occurs.
Weiss, G. (2004). Mining with rarity: A unifying frame-
Signature: A summary description of a subset of
work. SIGKDD Explorations, 6(1), 7-19.
data from a data stream that can be updated incremen-
Winter Corporation (2003). 2003 Top 10 Award Win- tally and quickly.
ners. Retrieved October 8, 2005, from http://www.
Subscription Fraud: Occurs when a perpetrator
wintercorp.com/VLDB/2003_TopTen_Survey/Top-
opens an account with no intention of ever paying for
Tenwinners.asp
the services incurred.
Superimposition Fraud: Occurs when a perpetra-
tor gains illicit access to an account being used by a
KEY TERMS legitimate customer and where fraudulent use is su-
perimposed on top of legitimate use.
Bayesian Belief Network: A model that predicts
the probability of events or conditions based on causal Total Lifetime Value: The total net income a com-
relations. pany can expect from a customer over time.

491