See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/254014721

Customer segmentation for a mobile telecommunications company based on

service usage behavior

Article · January 2011

CITATIONS READS

4 6,545

1 author:

Shohin Aheleroff
University of Auckland
4 PUBLICATIONS 4 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Customer Segmentation View project

All content following this page was uploaded by Shohin Aheleroff on 24 April 2014.

The user has requested enhancement of the downloaded file.

 Customer Segmentation
For a Mobile Telecommunications Company
Based on Service Usage Behavior
By:
Shohin Aheleroff

Advisor:
Dr. Gholamian

Jun 2011

Abstract: Competition between the mobile operators is II. Research methodology

becoming more based on subscriber’s behavior. In order
to improve mobile operator’s competitiveness and This study is designed to discover the patterns of use for
customer value, several data mining technologies can be mobile services based call/event detailed records including
used. One of the most important data mining technologies major service usage information and not personal information.
is customer clustering and segmentation. This targeting There are many clustering method, for example, fuzzy
practice has been proven manageable and effective for clustering method, system clustering method, dynamic
mobile telecommunications industry. Most clustering method and K-means clustering method. But the K-
telecommunications carriers cluster their mobile means method of cluster detection is most commonly used in
customers by billing system data. This paper discusses how practice that the number of clusters is an input. Based on
to cluster mobile customers based on their call detail business and infrastructure constrain generally operators are
records and analyze their consumer behaviors. Finally, the comfortable to have as few as five unique segments, while
subscribers categorized in four loyal groups and the other require as many as twenty segments to satisfy their data-
strategy to apply has been suggested in a specific life cycle. driven marketing needs. The decision of how many customer
segments a company should create is largely dictated by the
I. INTRODUCTION particular make-up of their customer base and the
organizations ability to develop and deliver unique segment
specific marketing treatments.
Mobile operator’s profits and ARPU (Average Revenue Per
User) are facing great challenges. Customer’s demand and
requirements of services has been changed. In order to
improve mobile operator’s competitiveness and customer
value, several data mining technologies can be used. One of
most important data mining technologies is customer
clustering analysis to categorize potential customers into
distinct groups for distinctive contact strategies. With the rapid
growing marketing business, data mining technology is imply
more important role in the demands of analyzing and utilizing
the large scale information gathered from customers especially
large amount call detailed record of mobile customers.
Information about customer’s behavior is required to segment
and personalize products and services along with business
strategy and planning.
But most of them segment customers only by personal
information such as age, gender and address from special
points, rather than from their actual behavior. Furthermore, Fig.1 Data flow from Network Elements to IS downstream systems
one of the key purposes of customer segmentation is customer
retention to increase the loyalty and avoid churning. This III. Prepare the data for clustering
paper focused on proposing a customer segmentation
framework based on actual customer behavior.
 I found a set of valuable information to identify core needs of The number of data in GSM is a barrier to analysis
subscribers based on their call detail record instead of their customer‘s behavior, so almost there is a limitation to
personal information such as gender, address and income. A analysis the whole data.
Call Detail/Data Record contains at a minimum the following:
 The number making the call (A number)
 The number receiving the call (B number)
 When the call started (date and time)
 How long the call was (duration)
 Call Type :
o Voice call
o SMS
o Data (content)
o GPRS/MMS
 Balance before & after.
 Location of mobile generator & terminator.
 Incoming and outgoing Voice
 Incoming and outgoing SMS
 Different type of content

In addition to CDR we also should consider subscribers

interest to active or change any service by capturing their
action via analyzing Event Detail Records. By getting CDRs
from different sources, we would be able to make sure that
customers’ behavior captured and we can evaluate what they Fig.2 Customer Segmentation Model
are interested more and less. Any call or event from network
elements such as IN / CCN and MSC will pass through ODS IV. CASE STUDY:
via mediation (Fig.1 shows the principle of data flow from A MOBILE OPERATOR’S CUSTOMERS
N.Es to IS downstream systems ) , so by accessing to the pull CLUSTERING ANALYSIS
of xDRs and defining proper characters ,we build our model of
customer segmentation based on their call/event detailed According to CDR (call detail records) analysis of a mobile
record. operator, located in the Middle East that has about 35 million
Considering the bellow steps (as illustrated in the Fig.2,) we mobile subscribers, in a normal day and also holiday, the trend
need enterprise hardware and software environment to deal shows that, the SMS usage is quite more than Call. As
with huge amount of data (generated CDRs & EDRs) but we Customer Segmentation is the process of splitting a customer
also can consider sample of data to evaluate customer habits database into distinct, meaningful, and groups, the major
and behaviors. Many segmentation algorithms and software parameters such as call duration, balance, call type, tariff plan
applications such as SAS and SPSS are already developed but and call time needs to be considered.
the most important is to follow bellow steps: As mentioned earlier in many methods number k of clusters to
construct is an input user parameter. Running an algorithm
 CDRs & EDRs collected (by push or pull mechanism several times leads to a sequence of clustering systems.
into a data warehouse) Selection of the number of clusters (e.g. 10, 5, and 3) before
 Different services selected: K-means implementation is required. However to achieve the
o GPRS / MMS optimum number of clusters using the data histogram (Fig.3)
o SMS will help to make a decision as a practical solution.
o Content e.g. RBT, Wallpaper, Java The majority of subscribers use to have less call duration and
application, Push mail, Music, Clip. only few of them have calls up to 200 seconds.
o Voice
 Specific factors selected as the core items to monitor
customer‘s behavior.
 Apply k-means as a well-known segmentation
algorithm.
 Customer segmentation as per each recognized
factors to generate a matrix of segments.
 Using Segmentation output for loyalty and customer
churn application.

Fig.3 Call Duration vs the number of Subscribers

By focusing on Call Duration and using K-means with the five
Final Cluster Centers
numbers of clusters (Table.1.), the center of selected clusters
areas after 19 number of iteration has been changed. The Cluster
minimum distance between initial centers is 165.000 (between
5th & 3rd segments) and the maximum is 200(between 2nd & 1 2 3 4 5
4th segments).The final center of clusters versus initial cluster
center has been improved as resulted in Table2.As we CALL_DURATION 421 139 59 257 10
expected the number of cases in each cluster is not close to
Initial Cluster Centers
each other. According to call duration histogram, the people
who their call duration (15085 cases in 5th segment) is quite Cluster
short are more than the others.
In addition to call duration exercise, the same practice is 1 2 3 4 5
applicable on the other CDRs parameters such as balance
before / after, SMS, MMS, GPRS usage and other content CALL_DURATION 722 337 166 537 1
based parameter for customer segmentation purpose.
Distances between Final Cluster Centers
Furthermore, it’s highly suggested to have a matrix of major
parameters to come up with a unique plan instead of each Cluster 1 2 3 4 5
individual service segments.
1 282.059 361.953 163.896 411.675

Iteration History 2 282.059 79.894 118.163 129.616

Change in Cluster Centers 3 361.953 79.894 198.057 49.722

Iteration 1 2 3 4 5 4 163.896 118.163 198.057 247.779

5 411.675 129.616 49.722 247.779
1 .000 19.596 28.955 43.135 15.805
2 51.500 27.633 12.061 30.882 .747 Table.2 Initial, Final and Distance between Cluster Centers

3 28.833 19.922 9.615 25.645 .756 V. Customer type definition:

4 53.381 16.630 7.859 20.455 .638
According to the definition of customer’s behaviors and due to
5 30.357 13.194 6.555 19.908 .565 the behavior of subscribers it’s clearly shows where they are.
6 22.833 11.932 5.573 19.376 .540
o Plain Loyal:
7 20.995 9.350 4.950 16.043 .562 A customer that has always been Active (never went into
Dormancy or Churn status).
8 14.232 8.933 4.130 13.544 .458 o Not Dependable:
9 11.224 9.720 3.325 14.647 .344 A customer has reached the Churn status for the first time. He
may in the future either stay in Churn status or return to
10 14.287 8.156 3.379 12.287 .381 Active (he will then be labeled ‘Loyal under Incentive’ from
now on until he reaches again and for good the Churn status).
11 7.786 7.666 2.777 10.875 .324 o Fence Seated:
12 5.498 7.140 2.288 6.635 .174 A customer has moved out of Active into Dormancy for the
first time. She/he may either fall into the Churn status, remain
13 7.047 8.058 3.272 8.796 .350 Dormant, are become Active again.
o Loyal under Incentive:
14 11.129 7.275 2.209 10.422 .178 A customer that has moved (once or several time) out of
15 8.958 6.040 2.922 7.709 .342 Dormancy or Churn and back into Active status.

16 6.050 5.457 1.844 7.140 .163 Based on the level of loyalty of customers during their life
cycle (Fig.3) we really need to keep the plain loyal motivated
17 5.223 4.057 1.790 5.896 .230 and also provide proper motivation and package to improve
18 .633 3.315 1.584 3.267 .211 their loyalty.

Table.1 implementation of K-means (k=5) for call duration

 o What happened to the ‘Fence Seated’ customers?
(7% of base)
o What happened to the ‘Not Dependable’ customers?
(16% of base)
o What happened to the ‘Loyal’ customers? (57% of
base)
o Where do each ‘’Loyalty type’ sit and what strategy
to apply?

Distinguish between groups, is the key milestone to make

distinguish promotion and motivation for each individual
groups. By the same way, marketing managements can design
Fig.3 Customer Life Cycle
more suitable marketing strategy.
The behavior of Fence seated subscribers shows that all the
In order to get a full picture of customer behavior in the existing promotions and various tariff plan have not any
network and to realize their interest and also to predict their impact on this group, so as per detailed graph (Fig.5) after one
behavior based on historical call detail record, we will analysis month only 41% of this subscribers remained in the same
mentioned scenarios in detail to identify situation while the 60% divided into two equal parts and joint
into “Not dependable” and “ Loyal under inventive” groups.
o Which segment he/she falls into and what are the It’s so interesting that the after about four month the loyal
characteristics of this segment including revenue value to under incentive (66%) group members increased to double
the business. compare to the Not dependable (33.5%) group members.
o What is the risk of this customer to leave the network or By focusing on not dependable subscribers which are 16% of
remain inactive total subscribers, it’s illustrated that only 10% of these people
moved to loyal under incentive while the rate of movement
VI. Customers’ Behavioral Evolution increased up to 37% after four month. I would like to highlight
During their Life Cycle: that not dependable staff only moved to loyal under incentive
group and not to any other group. this is a good message to
Based on six month historical data from enterprise data keep continue and boos the existing marketing plan to increase
warehouse, the statistics report of customers life cycle shows the number of loyal under incentive staff at the first step.
(Fig.4) that more than 50 percent of plain loyal subscribers are The goal is to make a plan to have specific motivation to move
significantly decreasing while the other there type of three segments into the plain loyal group. The selected period
subscribers are not in the same level or even increasing such of subscribers’ life cycle is totally in a green status as more
as loyal under incentive subscribers. than half out of total subscribers is plain total. There is a big
I would like to highlight that due to behavior of subscribers risk of churning to other operators because during the last six
during the selected period of customers’ life cycle, we can month the trend of loyal subscribers is not fluctuated and
predict that both loyal plain and loyal under incentive decreased to 49% that shows 8 % drop down to other groups.
subscribers will reach to a single point. In this situation the Besides not dependable and fence seated groups, the two other
two various segments will merge to a unique group with loyal groups have high rate of upward (Loyal under Incentive
population of 40 present of total subscribers. from 19% to 33%) and downward (Plain Loyal from 57% to
The volume of subscribers (Fig.4) shows that the operator is 49%) changes.
quite a bit in a safe side at this stage; however the monitoring
of customers behavior illustrated that we will face high rate of
changes in near future.
All the subscribers are active in the network till their status
will be changed to dormant and churned status. As soon as
they become a dormant subscriber, the risk will be remained to
leave the network and churn. The highest challenge is related
to the fence seated group because they have potential capacity
to change their status into loyal incentive in a very optimistic
view or they will leave the network for ever (pessimistic).
Besides the total number of subscribers in each segment, we
need to more focus on the detailed information and track the
dynamic behavior of subscribers in each group and find a
proper answer for the following questions:
Fig.4 Customer Life Cycle
 VII. Where do each ‘’Loyalty Type’ sit
What strategy to apply?

Any operator required to build strong, profitable customer

relationships with solutions that increase average revenue per
user, reduce subscriber churn and enhance brand loyalty.
In order to utilize the best criteria, two important parameters
selected to identify level of loyalty based on monthly recharge
and active on net (AON) for each segment of subscribers.
The propagation of subscribers and the location of each
loyalty type will lead the business to make an adequate plan,
so based on two mentioned items we mapped (Fig.6) four
loyalty groups into a matrix of duration (10 to 16 month) and
recharge (from 60$ to 100$) monthly basis. The margin for
AON is 13month and the recharge is 80$ monthly basis. It
means that if a subscriber is active more that 13 month on
network, then its ether plain or under incentive loyal
subscribers depend on the volume of recharge per month.
By the same way if they are less that 13 month active on
network then they are either fence seated or Not dependent ,
so it shows that we need development plan to keep them more
active on net by offering new packages as motivations.
As mentioned the boundary for monthly revenue is 80$, so it’s
quite important that even the subscribers who charge more
than specified range (80$) they are not loyal.
In addition to mentioned parameters to identify behaviors and
level of loyalty, we need to consider the service usage and also
the dependency or relation between services as part of cross
functional management to improve customer segmentation.
After we recognized the location of loyal subscribers, it’s time
to plan the right strategy for each specific segment.
The size and the status of each loyal group have been
illustrated in the Fig.6 based on their monthly recharge and
active time on the network. Both fence seated and not
dependable subscribers have less size than other two groups
and they stayed around 12 month on the network as an active
subscriber, while fence seated staff pied more than boundary.
On the other hand, two plain and under incentive loyal groups
are almost 15 month active on the net with different monthly
recharge payment. It’s clearly advised to motivate the loyal
under incentive subscribers to buy more vouchers (for
prepaid) or bill payment (postpaid) as the right strategy to
make then plain loyal as they are 33% of all subscribers.
As a general concern and risk, we might loss the entire 14% of
not dependable subscribers if they refuse to do payment.
By considering the size of each loyal group , the average
revenue per user and the location of group in the matrix
(Fig.6) , at least we would be able to initiated a strategic plan
because sometime it’s very costly to motivate this group than
put more effort and cost on the loyal under incentive or fence
seated groups to make them plain loyal on the network, so
depend on the constraints such as budget , priority and the
number of subscribers in each group, management will be able
to make a decision accordingly.

Fig.5 Detailed Behavior of Customer Life Cycle

 XI. REFERENCES

[01] Ngai, E.W.T., Xiu, Li., Chau, D.C.K. (2008).

“Application of Data Mining Techniques in Customer
Relationship Management.” Expert Systems with
Applications, No.8:003-124.

[02] Kim, Su-Yeon. , Jung, Tae-Soo. , Suh, Eui-Ho., Hwang,

Hyun-Seok. (2006).“Customer segmentation and strategy
development based on customer lifetime value.” Expert
Systems with Applications, Vol.31:101–107.

[03] So Young, Sohn. Kim, Yoonseong. (2008). “Searching

customer patterns of mobile service using clustering and
Fig.6 Loyalty type subscribers based on their monthly recharge and active on quantitative association rule.” Expert Systems with
net duration every month Applications, Vol.34:1070–1077.

VIII. CONCLUSION [04] McCarty, John A., Hastak, Manoj. (2008). “Segmentation
approaches in data-mining: A comparison of RFM, CHAID,
The mobile telecommunication marketplace is highly and logistic regression.“Journal of Business Research, Vol.60,
competitive. The operators often need to design No. 6:656-662.
distinguishable marketing strategy based on different behavior
of their mobile subscribers in order to improve their marketing [05] Hung, Shin-Yuan. , Yen, David C., Wang, Hsiu-Yu.
result and revenue. Call Detail Records describe customer (2008). “Applying data mining to telecom churn
utilization behavior. They have more information to describe management.”Expert Systems with Applications, Vol.31:515–
customer behavior than billing system data. Clustering 524.
analysis based on call detail records can give more
information than other clustering analysis for marketing [06] Charles Dennis, David Marsland, Tony Cockett. (2001).
management. We suggested a customer life cycle model “Data mining for shopping centers - Customer knowledge
considering the past contribution, potential value, and churn management framework.” Journal of Knowledge
probability at the same time. The model used for customer Management, Vol.5, No.4:368-374.
segmentation. Three perspectives on customer value (current
value, potential value, and customer loyalty) assist marketing [07] Leea, Jang Hee., Park, Sang Chan. (2005).”Intelligent
managers in identifying customer’s segmentation with more profitable customer segmentation system based on business
balanced viewpoints. After identification of subscriber’s intelligence tools.” Expert Systems with Applications,
behavior and identification of loyal groups, it’s feasible and Vol.29:145–152.
possible to put mobile customer clusters in place and make an
applicable strategic plan for each group to achieve customer [08] Jain, D., & Singh, S. S. (2002).”Customer lifetime value
satisfaction. research in marketing: a review and future directions.”Journal
of Interactive
IX. ACKNOWLEDGMENT Marketing, Vol.16, No.2, 34–45.

The author would like to thank Dr. Gholamian for his support [09] Mali, K. (2003).“Clustering and its validation in a
in the direction of the thesis in Shiraz University. During symbolic framework.” Expert Systems with Applications, Vol.
working on it, I got a lot of help from both supervisor and 24:2367-2376.
colleagues within MTN Iran. My supervisor always tracks the
work to make sure there is no problem, and if there is, he [10] Han, J., Kamber, M. (2006), Data mining: Concepts and
would give immediate help to solve the problem. And he also Techniques, USA, Morgan Kaufmann Publishers.
gave me a good idea on how to write the thesis and what is the
process. Thanks to my colleagues, they gave me lots of
encouragement and help on both studies and work. Thanks to
my wife, she always let me know she love me which gave me
motivation. At last, I want to thank this program for the
opportunity to grow and share knowledge.

View publication stats