استخراج تغییر رفتار مشتری در مرکز خرید اینترنتی

Understanding and adapting to changes of customer behavior is an important aspect for a internet-based company to survive in a continuously changing environment. The aim of this paper is to develop a methodology which detects changes of customer behavior automatically from customer profiles and sales data at different time snapshots. For this purpose, we first define the three types of changes as emerging pattern, unexpected change and the added/perished rule, then, we develop similarity and difference measures for rule matching to detect all types of change. Finally, the degree of change is evaluated to detect significantly changed rules. Our proposed methodology can evaluate the degree of changes as well as detect all kinds of change automatically from different time snapshot data. A case study on an internet shopping mall for evaluation of this methodology is also provided.

Understanding and adapting to changes of customer behavior is an important aspect of surviving in a continuously changing environment. Especially for internet-based companies, knowing what is changing and how it has been changed is of crucial importance because it allows businesses to provide the right products and services to suit the changing market needs (Liu, Hsu, Han & Xia, 2000). More specifically, most decision makers in internet-based companies have a strong need to know and adapt to the answers to following questions about their customers. • Which customer group's sales are gradually increasing? • Which customer groups moved from product A to product B over the years? • Whether certain groups of customers had gradually emerged to be the major buyers over the years? Data mining is the process of exploration and analysis of large quantities of data in order to discover meaningful patterns and rules, but much of existing data mining research has focused on devising techniques to build accurate models and to discover rules. Relatively little attention has been paid to mining changes in databases collected over time (Liu et al., 2000). Most data mining techniques such as association rules, decision trees and neural networks cannot be applied alone to answer the above research questions, because they cannot handle dynamic situations well. Also, most data mining techniques usually ignore rare items which have a small frequency of occurrence, but rare items which have a large growth rate or decreasing rate may give very significant implications to managers in changing environment. These are the reasons why we are motivated to develop other methodology to detect change. Association rule mining finds interesting association relationships among a large set of data items (Agrawal et al., 1993b, Agrawal et al., 1993b and Agrawal and Srikant, 1994). With massive amounts of data continuously being collected and stored, many industries are becoming interested in mining association rules from their databases. Association rule mining is used as a basic mining methodology in our research. Detected changes can be usefully applied to plan various niche marketing campaigns. For example, in a shop, people used to buy beer and snacks together—now they still buy beer, but seldom buy snacks. The shop manager needs to know this information so that he/she can find out the reason for this and design some catalysts to attract customers to buy snacks again. As an another example, if a manager can find out that a certain customer's preference has moved from a medium-size car to a large-size car, then that manager can establish a trade-in plan for customers who have a medium-size car and have the intention of buying a large-size car for replacement. In this paper, we develop a methodology which detects changes automatically from customer profiles and sales data at different periods of time. The most common approach to discover changes between two datasets is to generate rules from each dataset and directly compare the rules by rule matching, but this is not a simple process because of the following reasons. First, some rules cannot be easily compared due to different rule structures. Second, even with matched rules, it is difficult to know what kind of change and how much change has occurred. To simplify these difficulties, we first define three types of changes as emerging pattern, unexpected change and the added/perished rule, then we develop similarity and difference measures for rule matching to detect all types of change. Finally, the degree of change is evaluated to detect significantly changed rules. The proposed methodology can evaluate degree of changes as well as detect all kinds of changes automatically from different time snapshot data. Furthermore, the methodology can also be applicable for discovery of different characteristics from different categorical data. We begin by reviewing the concept of association rules which are a prerequisite of our research and the discussion of related works in Section 2. We define types of change and change detection problems to clarify our objectives in Section 3. In Section 4, we provide our methodology. A case study for evaluation and its business implications are presented in Section 5 and Section 6. Finally we summarize our contributions and outline areas for further research in the Conclusion section.

In this paper, we developed a methodology which detects changes of customer behavior automatically from customer profiles and sales data at different time snapshots. For this purpose, we defined types of change as emerging pattern, unexpected change and the added/perished rule. We then developed similarity and difference measures for rule matching to detect all types of change in syntactic aspects. Additionally, the degree of change is evaluated to detect significantly changed rules in semantic aspects. We also suggested practical applications and opportunities to use for our methodology. As a further research area, we plan to extend our methodology to discover changes of a more general nature than association rules. It will be also promising to set up the campaign management planning based on our suggested methodology and it will be also interesting to check the effectiveness of the campaign. The contribution of this research is that the proposed methodology can evaluate the degree of changes as well as detecting all kinds of changes automatically from different time snapshot data. For such a purpose, we developed new measures for changes. We believe that the change detection problem will become more and more important as more data mining applications are implemented.