Professional Documents
Culture Documents
Proceedings of the 29th Annual International Computer Software and Applications Conference (COMPSAC’05)
0730-3157/05 $20.00 © 2005 IEEE
marketing personnel or corporate decision makers can understand customer needs and adjust their marketing strategy
formulate more appealing marketing plan or operational rules accordingly to seek maximum return under limited resources.
and actively offer products that might interest the customers. 6.1. Classification analysis
Thus clusters may be used as basis for segmenting marketing We clustered the 2003 customers using the RFM model.
targets. In consideration that customer’s spending behavior Given that “most recent purchase date (Recent)” bears no
might change with time, we will also explore such time-varying significant influence in the analysis of customer value in credit
and dynamically adjusted spending patterns. card transactions, we considered only two factors -frequency
4. System analysis and design and monetary value (M) in clustering analysis. Cluster 0 and
4.1. System framework Cluster 2, which collectively accounted for 80% of bank
As shown in Figure 2, the system framework for the study customers, but had brought little profit to the bank (spending
contains customer clustering, creation of association rules, and less than $4500,000 per person each year), banks are advised
mining change of rules at different time. against spending too much energy and costs on these types of
customers. Cluster 3 and Cluster 4 were more loyal, but did not
4.2. Relationship between rules at different time generate high profit for the bank. How to enhance their value to
periods the bank should be the focus of marketing efforts. Cluster 1
We identify higher profit customers using clustering accounted for 3.3% of credit card customers and they were the
algorithm, select datasets of two different time periods, and gold customers with high loyalty and high profit, and hence one
generate the association rules in those two periods using of the bank’s most important assets. Losing such customers
association rules algorithm. We first measure the similarity and will result in substantial loss to the bank. Thus the bank should
difference between the rules, then use rule matching threshold endeavor to provide better services and greater added value to
to decide the type of the rule, and finally evaluate the degree of retain them. A business only stands to win if it is able to use
change to identify a significant trend of spending behavioral information and professional knowledge effectively to create
change. We use the term “similar” and “different” to depict the value for its customers and solidify its relationship with them.
disparity between rules. Here we use a rule matching threshold At 5.5% support and 75% confidence level, we identified
(RMT) [8] to decide the measure of difference between rules, 641 association rules in 2003 dataset and 578 association rules
shown in Figure 3. in 2004 dataset using the Apriori algorithm provided in Weka.
7. Conclusion
We used association rules algorithm to mine the data to
detect the association rules in each time period. We measured
the similarity, difference and modified difference of the mined
Figure 3 Different Types of Changes association rules based on three definitions and used the self-set
rule matching threshold (RMT) to find all types of rules.
5. System implementation
5.1. Data description References
The data for the empirical study are credit card customer [1] Alex Sbesbunoff,”Winning CRM Strategies”, ABA Banking
data and spending records in 2003 and 2004 from banks. In Journal,1999.
light that the study focuses on the spending behavior of [2] Vince Kellen, “CRM Measurement Frameworks”, Blue
customers, all purchases of the same customer using more than Wolf White Paper, p.4, 2002.
one credit card were combined under said customer. There were [3] Ott, J., “Successfully development and Implementing
altogether 10304500 and 1146200 pieces of purchase data for Continuous relationship management,” eBusiness executive
2003 and 2004 respectively, which were matched using a report, 2000, 26-30.
program, and data of the same customer were kept. Finally, [4] Adriaans P. and Zantinge D., “Data Mining,”
1063000 pieces of data for each year were used as dataset for Addson-Wesley, Harlow, 1996.
the study. [5] A. Berson, K. Thearling and S. Smith, “Building Data
Mining Applications for CRM,” McGraw-Hill, 2000.
5.2. Weka data mining tool [6] Jill Dyché, “The CRM Handbook: A Business Guide to
Weka (Waikato Environment for Knowledge Analysis) is Customer Relationship Management”, Addison Wesley
a Java-based data mining tool developed by Waikato Professional, 2002.
University. After loading the dataset into it, the preprocess [7] Swift, Ronald S., “Accelerating Customer Relationships
function of Weka allows the user to input undesired attributes Using CRM and Relationship Technologies,” Prentice Hall
to prevent them from affecting the quality of extracted PTR,2001
knowledge. Next, the user can use one of the three algorithms [8] Song H.S., Kim J.K. and Kim S.H. “Mining the change of
to mine the data: Classification, Clustering, and Association customer behavior in an internet shopping mall” Expert
Rule. Systems with Applications 2001.
6. Results [9] Shaw M.J., Subramaniam C., Tan G.W. and Welge
In this section, we used Weka to run the model. First, we M.E.,”Knowledge management and data mining for
clustered the 2003 customers to identify the most influential marketing”, Decision Support Systems 2001.
group in term of business profit and find their behavioral [10] Grupe, G. H., Owrang, M. M., “Database Mining
pattern. The 2004 dataset was used to observe change of rules Discovering New Knowledge and Cooperative
in different time periods. These steps can help decision makers Advantage,” 12, 1995, 26-31.
Proceedings of the 29th Annual International Computer Software and Applications Conference (COMPSAC’05)
0730-3157/05 $20.00 © 2005 IEEE