A CREDIT CARD USAGE BEHAVIOUR ANALYSIS
FRAMEWORK - A DATA MINING APPROACH
Chieh-Yuan Tsai
Department of Industrial Engineering and Management, Yuan Ze University
No. 135, Yuantung Rd., Chungli City, Taoyuan County, Taiwan
Keywords: e-Commerce, Customer Relationship Management, Data Mining, LabelSOM, Fuzzy Decision Tree.
Abstract: Credit card is one of the most popular e-payment approaches in current online e-commerce. To consolidate
valuable customers, card issuers invest a lot of money to maintain good relationship with their customers.
Although several efforts have been done in studying card usage motivation, few researches emphasize on
credit card usage behaviour analysis when time periods change from t to t+1. To address this issue, an
integrated data mining approach is proposed in this paper. First, the customer profile and their transaction
data at time period t are retrieved from databases. Second, a LabelSOM neural network groups customers
into segments and identify critical characteristics for each group. Third, a fuzzy decision tree algorithm is
used to construct usage behaviour rules of interesting customer groups. Finally, these rules are used to
analysis the behaviour changes between time periods t and t+1. An implementation case using a practical
credit card database provided by a commercial bank in Taiwan is illustrated to show the benefits of the
proposed framework.
1 INTRODUCTION
Companies across all sectors and of all sizes are now
benefiting from Internet e-commerce where services,
information and/or products are exchanged via the
web. To finalize the online transaction, providing
convenient e-payment approaches for consumers are
very important. Among several e-payment
mechanisms, credit card is one of the most welcome
approaches for online stores. Card issuers earn their
profit from the fee charged to the store that accepts
the credit card, charging interest on outstanding
balances, and fee charged to customers. The major
fees contributed from customers are (1) payments
received late (past the "grace period"); (2) charges
that result in exceeding the credit limit on the card
(whether done deliberately or by mistake); (3) cash
advances and convenience checks; and (4)
transactions in a foreign currency. However, raising
these fees or interesting rates to increase the profit of
card issuers could be very difficult in current
competitive credit card markets.
Another strategy to increase the profit of card
issuers is to forge closer and deeper relationships
with customers by emphasizing on customer
relationship management (CRM) (Giudici and
Passerone, 2002; Tsai and Chiu, 2004). CRM
focuses on customer’s need and regard customer life
cycle as important assets of an enterprise. Due to the
advance of the Information Technology (IT), it is
easy to discover the usage information of what their
customers purchase, when they use the card, and
how often they consume. When the usage
information is available, the card issuers can
encourage customers use their cards more frequently
through offering suitable products and services.
Data mining is the technique to discover
meaningful patterns (rules) from large databases.
Much of existing data mining researches in credit
card fields has focused on building accurate models
for risk and scoring analysis (Lee et al. 2006), cross
selling (Wu and Lin, 2005), and fraud detection
(Kou et al. 2004; Chen et al., 2005b). Relatively
little attention has been made to analyze pattern
changes in databases collected over time (Donato et
al., 1999). However, customer behaviour usually
changes over time. Some frequent patterns at one
time period may not be valid for another time period
(Chen et al., 2005a; Tsai et al., 2007). For example,
a group of customer has preference in shopping at
department stores this year and might change their
preference to doing outdoor activities and travelling
219
Tsai C. (2007).
A CREDIT CARD USAGE BEHAVIOUR ANALYSIS FRAMEWORK - A DATA MINING APPROACH.
In Proceedings of the Second International Conference on e-Business, pages 219-226
DOI: 10.5220/0002108102190226
Copyright
c
 SciTePress

in the following year. If issuers cannot capture the
behaviour change dynamically due to the time
difference, it will be hard to retain customers by
tailoring appropriate products and services to satisfy
their real needs.
This paper is organized as follows. Section 2
reviews the literatures related to the change analysis
models. Section 3 introduces the proposed credit
card usage analysis framework that adopts
LabelSOM algorithm and fuzzy decision tree
algorithm. Section 4 provides an implementation
case using a practical credit card database provided
by a commercial bank in Taiwan to demonstrate the
benefit of the proposed framework. A summary and
future works are concluded in Section 5.
2 LITERATURE REVIEW
Current businesses face the challenge of a constantly
evolving market where customer’s needs are
changing rapidly. Some researches applied customer
demographic variables such as recency, frequency,
and monetary (RFM) to analyze customer behavior
(Tsai and Chiu, 2004). Although RFM analysis can
effectively investigate customer values and segment
markets, it is not a suitable tool for detect the
customer behavior changes. Therefore, to better
understand customer behaviors, developing suitable
change detection models becomes an important
research topic in the financial business.
Except the studies of rule maintenance in the
changed database, some researches focus on
discovering emerging patterns. Emerging pattern
mining can be defined as the process to discover
significant changes or differences from one database
to another (Dong and Li, 1999). Emerging pattern
captures emerging trends in time stamped database.
Another related research trend is subjective
interestingness mining. Interestingness mining is to
find unexpected rules with respect to the user’s
existing knowledge. Unexpected changes compare
each newly generated rule with each existing rule to
find degree of difference (Liu and Hsu, 1996). Liu et
al. (Liu et al., 1999) proposed a DM- II (Data
Mining-Integration and Interestingness) system
which has classification and association rule mining
tasks to help users perform interestingness analysis
of the rules. Its analysis compares each newly
generated rule with each existing rule to find degree
of difference, which is useful and important for real-
life data mining applications. Han et al. (1999)
presented several algorithms for efficient mining of
partial periodic patterns, by exploring some
interesting properties related to partial periodicity.
The algorithms show that mining partial periodicity
needs only two scans over the time series database to
make efficient in mining long periodic patterns.
3 ANALYSIS FRAMEWORK
The proposed credit card usage behaviour analysis
framework consists of four major stages as shown in
Figure 1. The first stage is data extraction and pre-
processing. In this stage, the customer profile and
their transaction data at time period t are retrieved
from databases. The second stage is to conduct
customer segmentation using the LabelSOM neural
network. The LabelSOM adaptively cluster
customers into groups and automatically identifies
critical demographic features for each group. In the
third stage, the usage behaviour of the customers in
the interesting group is generated using fuzzy
decision tree (FDT) algorithm that represents usage
behaviour as a set of IF-THEN rules. After obtaining
the usage patterns of interesting customer group at
time period t, we can trace the behaviour changes of
these customers from time period t to t+1 when
retrieving their corresponding data at time period t+1.
The LabelSOM algorithm in the second stage and
the FDT the third stage are further introduced in the
following sub-sections.
Customer and
Transaction data at
time period t
Customer and
Transaction data at
time period t+1
Data Preprocessing
LabelSOM
Neural Network
Fuzzy Decision Tree
Chang Analysis
Select appropriate
customer groups
Generate usage
behavior rules
Marketing
Suggestions
Figure 1: The proposed credit card usage behaviour
analysis framework.
ICE-B 2007 - International Conference on e-Business
220

3.1 Label Self-Organizing Map
(LabelSOM)
Self-Organizing Map (SOM) neural network,
proposed by Kohonen (1990), is recognized as
one of the popular clustering methods. The SOM
employs a competitive unsupervised learning
technique to project high dimension data into a
two-dimensional grid without destroying data
topology. Data points that are close each other in
the input space are mapped to nearby output
neurons in the SOM. An input neuron represents
a data feature in N dimension and an output
neuron represents the clustering result in two-
dimensional space. The output neuron that has the
highest similarity with an input data point is
claimed as the winner (the best matching unit).
The weights of the winner node and its
neighbouring neurons are then adjusted
automatically to force the weights closer to the
input vector. After completing this learning
process, each neuron represents a set of data.
Self organizing maps are an unsupervised
neural network model which lends itself to the
cluster analysis of high dimensional input data.
However interpreting a trained map could be
difficult because the features responsible for a
specific cluster assignment are not evident from
the resulting map representation. To solve this
difficulty, the LabelSOM was developed to
automatically label every node of a trained SOM
(Rauber and Merkl, 1999). That is, LabelSOM
algorithm can not only conduct cluster operation
well but also distinguish the difference between
each cluster clearly. Therefore, the LabelSOM is
adopted for customer segmentation in our second
stage.
The operations of the LabelSOM algorithm
are summarized into the following five steps.
1. Input the primary parameters of the
LabelSOM. The primary parameters include the
number of input neurons, number of output
neurons, number of input data, learning iteration,
learning rate
α
, and radius
η
. In addition, initial
weight matrix W is set randomly.
2. Conduct the following three sub-steps for each
input vector
12
( , ,... )
iiiik
x
xx x= sequentially
where
1,2,...,im=
, m is the total number of input
data points and k is the number of neurons in the
input layer. Notes that k is also the number of
features for an input point.
 Calculate the Euclidean distance d
j
between
the input vector i and output neuron j. That
is,
jij
dxw=− where
12
( , ,..., )
j
jj jk
www w
=
, 1,2,...,jn= , and n is
the number of output neurons.
 Find the output neuron j* with minimum
Euclidean distance between output neurons
and the input vector. Mathematically, it is
represented as
min
1,2,...,
min
j
jn
dd
=
=
.
 Update the weights of output neurons j*
using
** *
()
new old old
i
jj j
ww xw ζα =+××− where
exp( / )R
ζ
η
=
− ,
α
is the learning rate,
η
is
the radius, and R is the closed distance.
3. Decrease the learning rate α and radius η and
repeat Step 2 until the stopping criteria of the
learning process are reached.
4. After the above learning process is completed,
the quality of this network is evaluated using an
average total distance measure which is defined
as:
p
j
N
d
G
n
=
where
2
()
p
jpj
p
dxw=−
∑
is
the distance between data p and jth output node,
and n is the number of data in cluster N.
5. Let
j
D
be the set of input vector
i
x mapped
onto node j. Summing up the distances for each
vector element over all the vectors
i
x
(
ij
x
D∈
)
yields a quantization error vector
q
j
for every
node. This can be represented as:
2
(),1...
ij
jl jl il
xD
qwxlk
∈
=−=
∑
3.2 Fuzzy Decision Tree (FDT)
Decision tree algorithms are supervised learning
models that express knowledge rules using a tree
structure. Among several decision tree algorithms,
the ID3 is one of the most popular algorithms since
it is efficient and easy to operate (Quinaln, 1986).
However, data in the input vector to the ID3 should
be in categorical format. All numerical data need to
be discretized into proper categorical data format in
advance. If the discretization process is not well
performed, the classification quality of the ID3 will
be erroneous. In addition, when a new input vector is
fed to the ID3, only one branch of the tree is
initialized and the end node of the active branch
returns the class label for the new input vector.
A CREDIT CARD USAGE BEHAVIOUR ANALYSIS FRAMEWORK - A DATA MINING APPROACH
221

Although this approach is straightforward, the class
information of other branches, which are similar to
the active branch, is not considered. This might
significantly increase the classification error rate. To
solve this difficulty, fuzzy set theory is integrated
with ID3 for generating customer behaviour rules in
the third stage, called the fuzzy decision tree (FDT).
The rules generated using the FDT are easier to be
interpreted by human beings. In addition,
discretization process is not required since fuzzy
membership functions will map the numerical data
into proper membership value. Moreover, more than
one branch of the fuzzy decision tree might be
initiated. Therefore, class labels suggested by
multiple branches will be fused by majority voting
into a more trustable class label.
Let
1
{ ( ,..., , )}
n
jj j j j
Egg g gC==
%%
be the set of our
interesting customer group decided in the last stage
where
i
j
g
represents the value of attribute i for
customer
j and
j
C
is the predefined class label for
customer
j.
1
{ ,..., }
n
A
aa= is the attribute set where
i
aA∈ with value [0,1].
1
{ ,..., }
i
ii ip
Da a=
denotes
the set of fuzzy linguistic terms for attribute
i
a
where
i
ip
a
can be described using membership
function
()
ip
i
a
x
μ
. Thus, ()
ip
i
i
aj
g
μ is the membership
value for attribute
i of vector j for fuzzy linguistic
term
ip
a . In addition, for each node N in the fuzzy
decision tree,
N
F
denotes the set of fuzzy
restrictions on the path leading to
N.
ip
Na denotes
the particular child of node
N created using
i
a to
split node
N and following the branch
i
ip
a
.
The training process for the FDT algorithm
contains five steps and is introduced as follows
(Janikow, 1998).
1. Set all input vector set of the interesting
customer group
E in the root node of the tree. At
node
N to be expanded, compute the number of
data included in the node which need to be
subdivided as:
1
C
NN
k
k
PP
=
=
∑
where
C is the number of predefined classes,
N
k
P
is the number of vectors belonging to kth
class and is computed as
{}
1
1
j
E
NN
kjCk
j
PX
=
=
=×
∑
where
{}
10
j
Ck=
=
if
j
Ck
≠
and
{}
11
j
Ck=
= if
j
Ck
=
.
2. Compute the information gain at node
N as
1
(log)
C
NN
N
kk
NN
k
PP
I
PP
=
=− ×
∑
where
i
a means all
clustering attributes which have not appeared at
path
N
F
.
3. Compute the extended attached nodes from the
above attributes
ip
Na . The information quantity
of
iip
Da
∈
is
ip
aN
I
.
4. Choose the attribute
*
i
a with largest
information gain. The formula to compute
information quantity is
i
Na
N
ai
GII=−
where
1
()
ip
ii
Na
Na D
pp
IwI
=
=×
∑
. The weight
p
w is the
proportion of examples belonging to
node
ip
Na and is shown as
ip ip
Na Na
p
p
wP P=
∑
／ .
5. Subdivide the node N again by clustering
attribute
*
i
a and delete the attached nodes with
few vectors.
4 A CASE STUDY
The proposed framework is implemented using the
database provided by a major credit card issuer in
Taiwan.
4.1 Data Extraction and Data
Preprocessing
In the database, there are 314,339 activate card users
who generated 2,153,062 transactions in year 2001
(time period t) and 2,561,202 transactions in year
2002 (time period t+1). Marketing managers want to
concentrate on the customer behaviour of their VIP
customers. The VIP selection criteria are based on
the corporation regulations, credit assessment
policies, and customer life value evaluations. “No
delayed payment is made in recent nine months” and
“lowest limit amount are paid in the past two
months” are two typical criteria they set. A serial of
COBOL (common business oriented language) and
ICE-B 2007 - International Conference on e-Business
222

JCL (job control language) programs are coded to
retrieve customer profiles and customer behaviour
data from the VSAM (virtual storage access method)
files in OS/390 operation system of an IBM
9121main frame computer. As a result, 9,086 VIP
customers are identified. In addition, these
customers made 354,063 transactions in year 2001
and 440,010 transactions in year 2002.
4.2 Customer Segmentation Using
LabelSOM
Based on the available data in the database,
demographic attributes of gender, age, marital
status, education, occupation, card holding
period, and credit limit are used to describe a VIP
customer profile. Therefore, seven input nodes
are required for the LabelSOM neural network. In
addition, a two-dimension rectangle topology is
selected as output layer. Since the clustering
quality of the LabelSOM might be affected by
different parameter settings, a number of
experiments are conducted based on literature
suggestion (Vesanto and Alhoniemi, 2000; Zhang
and Li, 1993) and our own experience. Table 1
shows the primary parameter settings in our
experiments.
Table 1: The parameters of LabelSOM.
Parameter Value
N
umber of Inpu
t
nodes 7
Topology Two-dimension rectangle
N
umber of output nodes 3~7 for each side
Learning coefficient 0.3~0.9
N
eighbourhood radius 1~35
Epoch 9,086
Training number 272,580
After a systematic experimental design, the best
clustering result with minimal distance 0.346 is
found. The best clustering result consists of five VIP
customer groups where the number of customers and
the average total distance for each group are shown
in Table 2. Table 3 shows the quantization error
vectors, introduced in Section 3.1, of all attributes
for the five groups. The smaller vector value
indicates that the attribute is more important for
distinguishing data among clusters.
Table 2: The grouping results for the best experiment trial.
Group No. Number of Customers Average Total Distance
1 1697 0.263
2 3279 0.429
3 1594 0.369
4 1379 0.379
5 1137 0.294
Table 3: The quantization error vectors for all attributes.
Attributes
Group
No.
GenderMartial
Status
Edu. Occp. Age Holding
period
Credit
Limit
1 0 0 0 6.376 6.929 8.385 1.380
2 0 0 18.140 8.852 9.894 13.439 2.698
3 0 0 11.354 5.714 5.558 7.639 1.141
4 0 0 11.030 5.739 4.911 7.114 1.209
5 0 0 3.945 5.464 5.314 7.915 1.307
4.3 Customer Behaviour Pattern
Generation Using FDT
After discussion with marketing managers, they are
interesting in the customers in VIP Group 2. Table 4
shows the comparison between the customers in
Group 2 and all customers.
Table 4: The comparisons between Group 4 and all VIP
customers.
Group 2 Average for all VIP
Gender Male: 100%
Female: 49%
Male: 51%
Martial
Status
Married: 100%
Married: 67%
Single: 33%
Edu.
High School: 8%
Undergraduate: 54%
Graduate: 38%
High School: 5%
Undergraduate: 47%
Graduate: 48%
Occup.
Self or Intl. Business:
88%
Finance or Service: 10%
Others: 2%
Self or Intl. Business: 88%
Finance or Service: 10%
Others: 3%
Age
20~29: 0.4%
30~39: 24%
40~49: 44%
50~59: 25%
20~29: 5%
30~39: 39%
40~49: 36%
50~59: 16%
Holding
Period
3~7 Year: 59% 3~7 Year: 65%
Credit
Limit
Below 100K: 10%
100K~200K: 27%
200K~300K: 37%
300K~400K: 14%
Below 100K: 17%
100K~200K: 37%
200K~300K: 31%
300K~400K: 9%
For managerial reasons, each customer was
classified as one of the four types (Type 1 to 4)
according to their RFM scores in year 2001 (time
A CREDIT CARD USAGE BEHAVIOUR ANALYSIS FRAMEWORK - A DATA MINING APPROACH
223

period t). Marketing managers want to know
whether this classification method is still valid in
year 2002 (time period t+1). Therefore, a
classification model for year 2001 needs to be
constructed first.
There are two factors that might affect the
inference result of the FDT algorithm. They are the
number of linguist terms for each variable, and the
shape of membership functions for each linguistic
term. To understand the influence, the following
experiments are conducted.
Assume that a trapezoid membership function,
which can be described as T(a, b, c, d, e), is to
represent an numeric interval in this study where a is
the left-bottom corner point, b is the left-top corner
point, c is the middle point between b and d, d is the
right-top corner point, and e is the right-bottom
corner point. If we adjust top and/or bottom widths
of T(a, b, c, d, e), the fuzzy degree will be different.
Therefore, an experiment Shape(x, y) denotes that a
trapezoid fuzzy number T(a, b, c, d, e) is modified as
T (a×(1-y), c-(c-b)×x, c, c
+(d-c) ×x, e×(1+y)). For
example, Figure 2(a) shows a linguistic term
“Medium” with a crisp membership function
T(210000, 210000, 250000, 290000, 290000) for a
“credit limit” attribute, while Figure 2(b) shows the
fuzzy membership functions T(199500, 240000,
250000, 260000, 304500) after Shape(25%, 5%) is
applied to T(210000, 210000, 250000, 290000,
290000).
μ (χ )
1
Credit Limit
199500 304500240000 250000 260000
(b)
μ (χ )
1
Credit Limit
210000 290000250000
(a)
Figure 2: Crisp and fuzzy membership functions.
Table 5 shows examples of our experiment result
when an attribute is represented using 3, 5, and 7
linguistic terms respectively. For each linguistic
term, one crisp and four fuzzy membership functions
with different bottom widths are experimented. It is
clear that the classification accuracy using fuzzy
membership functions are higher than the one using
crisp membership function in all cases. In addition,
the classification accuracy using 5 linguistic terms is
higher than ones using 3 or 7. It also indicates that
when the bottom with of the trapezoid fuzzy number
increases, a more accurate classification result can
be obtained. Table 6 shows the experiment results
when the top widths of fuzzy membership functions
change.
Table 5: The classification accuracy using different
number of linguist terms and membership functions (I).
Number of
Linguistic Terms
Membership Functions
Classification
Accuracy
crisp 59%
Shape(25%, 5%) 68%
Shape(25%, 10%) 69%
Shape(25%, 15%) 72%
3
Shape(25%, 20%) 72%
crisp 63%
Shape(25%, 5%) 70%
Shape(25%, 10%) 74%
Shape(25%, 15%) 74%
5
Shape(25%, 20%) 77%
crisp 62%
Shape(25%, 5%) 69%
Shape(25%, 10%) 69%
Shape(25%, 15%) 68%
7
Shape(25%, 20%) 68%
Table 6: The classification accuracy using different
number of linguist terms and membership functions (II).
Number of
Linguistic Terms
Membership Functions
Classification
Accuracy
Shape(20%, 10%) 68%
Shape(25%, 10%) 69%
Shape(30%, 10%) 68%
3
Shape(35%, 10%) 68%
Shape(20%, 10%) 72%
Shape(25%, 10%) 74%
Shape(30%, 10%) 72%
5
Shape(35%, 10%) 72%
Shape(20%, 10%) 68%
Shape(25%, 10%) 69%
Shape(30%, 10%) 68%
7
Shape(35%, 10%) 68%
4.4 Change Analysis
Based on the experiment result of Table 5 and Table
6, managers decide to use the FDT for the following
analysis where the number of linguist terms for each
variable is 5 and the shape of membership functions
ICE-B 2007 - International Conference on e-Business
224

for each linguistic term is fuzzy(25%, 20%). Under
these settings, the FDT generate 109 rules and has
the highest classification accuracy. Among these
rules, 16 rules are to identify Customer Type 1, 30
rules are to identify Customer Type 2, 38 rules are to
identify Customer Type 3, and 25 rules are to
identify Customer Type 4. Figure 3 shows some of
these rules.
Rule1:
Usage=Very Low (0~17) &
Interest Amount = Very Low（0~432）&
Expenditure Amount = Very Low (0~27423)
=>Customer Type 1
Rule 2:
Usage =Very Low (0~17) &
Interest Amount=High（8929~18251）&
Expenditure Amount=Very Low (0~27423) &
Credit Limit=Low (125000~250000) &
=> Customer Type 2
Rule 3:
Usage =High (37~81) &
Expenditure Amount=High (119140~393753) &
Interest Amount=Very Low (0~432) &
Credit Limit=High (255000~360000)
=> Customer Type 3
Rule 4:
Usage =Very High（63~317）&
Expenditure Amount=Very High (316215~4944133) &
Credit Limit=High (255000~360000) &
Interest Amount=Low (0~432) &
=> Customer Type 4
Figure 3: Example rules generated by the FDT algorithm.
For example, rule 4 indicates that, in year 2001 (time
period t), if a customer has the usage behaviour
such as “Usage = Very High (63~317) AND
Expenditure = Very High (316215~4944133) AND
Credit Limit = Very High (310000~3600000) AND
Interest Amount = Very Low (0~432),” then he/she
should be “Customer Type 4”. When we further
check the database, 143 customers in year 2001
(time period t) satisfy this rule. However, when this
rule applies to these 143 customers in year 2002
(time period t+1), only 107 customers still confirm
this rule. 24 customers change to Type 3, 11
customers change to Type 2, and 1 customers
change to Type 1. Table 7 summarizes basic
changing information. It is surprised that all
changing persons are male customers, married, and
own business. The company should note the changes,
since customer Type 4 is most valuable for the
company.
Table 7: The changing information for Type 4 customers.
Type (4 → 1)
(1 person)
Type (4→2)
(11 persons)
Type (4→3)
(24 persons)
Gender
Male Male Male
Martial
Status
Married Married Married
Edu.
High School
and Below
High School: 2
Undergraduate: 6
Graduate: 3
High School: 4
Undergraduate: 15
Graduate: 5
Occup.
Self or Intl.
Business
Self or Intl.
Business
Self or Intl.
Business
Age
46
30~39: 3
40~49: 4
50~59: 3
60~69: 1
30~39: 3
40~49: 8
50~59: 11
60~69: 2
Holding
period
8 Year 5~10 Years 4~12 Years
Credit
Limit
500K 325K~650K 320K~1850K
Interest
Amount
Y2001: None
Y2002: None
Y 2001: None
Y2002: 1
Y2001: 1
Y2002: 3
Average
Spending
Amount
Y2001: 362704
Y2002: 242075
Y2001: 688302
Y2002: 303967
Y2001: 499210
Y2002:: 354909
5 CONCLUSIONS
The magnificent increase in credit card markets for
e-commerce leads card issuers put more efforts to
understand their usage behaviour. In reality,
customer behaviours usually change over time.
Some frequent patterns at one time period may not
be valid for another time period. To fulfil this need,
this research proposes an integrated data mining
approach for credit card usage behaviour analysis.
The proposed credit card usage behaviour
analysis framework consists of four major stages.
The first stage is data extraction and pre-processing.
In this stage, the customer profile and their
transaction data at time period t are retrieved from
databases. The second stage is to conduct customer
segmentation using the LabelSOM neural network.
The LabelSOM adaptively cluster customers into
groups and automatically identifies critical
demographic features for each group. In the third
stage, the usage behaviour of the customers in the
interesting group is generated using fuzzy decision
tree (FDT) algorithm that represents usage
A CREDIT CARD USAGE BEHAVIOUR ANALYSIS FRAMEWORK - A DATA MINING APPROACH
225

behaviour as a set of IF-THEN rules After obtaining
the usage patterns of interesting customer group at
time period t, we can trace the behaviour changes of
these customers from time period t to t+1 when
retrieving their corresponding data at time period t+1.
The proposed model has been successfully
implemented using real credit card data provided by
a commercial bank in Taiwan. The provided analysis
procedure should provide card issuers a systematic
approach to set up marketing strategies for
interesting customer groups. However, there are still
some rooms for improvement in the future.
Currently, only the fuzzy number with trapezoid
shape is considered. It is suggested that automatic
membership function fitting algorithms can be
incorporated into the proposed framework. Besides,
it will be worthwhile to explore variant customer
groups and study what marketing strategies can
affect their behaviour.
REFERENCES
Chen, M.-C., Chiu, A.-L. and Chang, H.-H., 2005a.
Mining changes in customer behavior in retail
marketing, Expert Systems with Applications, Vol. 28,
No. 4, pp. 773-781.
Chen, R., Chen, T., Chien, Y., and Yang, Y., 2005b.
Novel questionnaire-responded transaction approach
with SVM for credit card fraud detection, Lecture
Notes in Computer Science (LNCS), Vol. 3497, pp.
916-921.
Donato, J. M., Schryver, J. C., Hinkel, G. C., Schmoyer, R.
L., Leuze, M. R., and Grandy, N. W., 1999. Mining
multi-dimensional data for decision support, IEEE
Future generation Computer Systems, Vol. 15, No. 3,
pp. 433-441.
Dong, G., Li, J., 1999. Efficient mining of emerging
patterns: discovering trends and differences, In
Proceedings of the Fifth International Conference on
Knowledge Discovery and Data Mining, pp. 43-52.
Giudici, P., Passerone, G., 2002. Data mining of
association structures to model consumer behavior,
Computational Statistics and Data Analysis, Vol. 38,
No. 4, pp. 533-541.
Han, J., Dong, G. and Yin, Y., 1999. Efficient mining of
partial periodic patterns in time series database. In
Proceedings of the Fifteenth International Conference
on Data Engineering, pp. 106-115.
Janikow, C. Z., 1998. Fuzzy decision trees: issues and
Methods, IEEE Transactions on Systems, Man, and
Cybernetics-Part B: Cybernetics, Vol. 28, No. 1, pp. 1-
14.
Lee, T.-S., Chiu, C.-C., Chou, Y.-C. and Lu, C.-J., 2006.
Mining the customer credit using classification and
regression tree and multivariate adaptive regression
splines, Computational Statistics & Data Analysis, Vol.
50, No. 4, pp. 1113-1130.
Kohonen, T., 1990. The self-organizing map, In
Proceedings of the IEEE, Vol. 78, No. 9, pp. 1464-
1480.
Kou, Y., Lu, C.-T., Sirwongwattana, S., Huang, Y.-P.,
2004. Survey of fraud detection techniques, In
Proceedings of IEEE International Conference on
Networking, Sensing and Control, Vol. 2, pp. 749- 754.
Liu, B., Hsu, W., 1996. Post-analysis of learned rules, In
Proceedings of the Thirteen National Conference on
Artificial Intelligence, pp. 220-232.
Liu, B., Hsu, W., Ma, Y. and Chen, S., 1999. Mining
interesting knowledge using DM-II, In Proceedings of
the Fifth International Conference on Knowledge
Discovery and Data Mining, pp. 430-434.
Quinaln, J. R., 1986. Induction of decision trees. Machine
Learning, Vol. 1, No. 1, pp. 81-106.
Rauber, A. and Merkl, D., 1999. The SOMLib digital
library system. In Proceedings of the Third European
Conference, ECDL'99, pp. 323.
Tsai, C.-Y., Chiu, C.-C., 2004. A purchase-based market
segmentation methodology, Expert Systems with
Applications, Vol. 27, No. 2, pp. 265-276.
Tsai, C.-Y., Wang, J.-C., Chen, C.-J., 2007. Mining usage
behavior change for credit card users, WSEAS
Transactions on Information Science and Applications,
Vol. 4, No. 3, pp. 529-536.
Vesanto, J., Alhoniemi, E., 2000. Clustering of the Self-
Organization Map, IEEE Transactions on Neural
Networks, Vol. 11, pp. 568-600.
Wu, J. and Lin, Z., 2005. Research on customer
segmentation model by clustering. In Proceedings of
the 7th international Conference on Electronic
Commerce (ICEC '05), pp. 316 – 318.
Zhang, X., Li, Y., 1993. Self-organizing map as a new
method for clustering and data analysis, In
Proceedings of International Joint Conference on
Neural Networks, pp. 2448-2451.
ICE-B 2007 - International Conference on e-Business
226