Abstract
This project experiments with the designing of a Web site that has the self-adaptive feature of generating and adapting the site contents dynamically to match visitors’ tastes based on their activities on the site. No explicit inputs are required from visitors. Instead a visitor’s clickstream on the site will be implicitly monitored, logged, and analyzed. Based on the information gathered, the Web site would then generate Web contents that contain items that have certain relatedness to items that were previously browsed by the visitor. The relatedness rules will have multidimensional aspects in order to produce cross-mapping between items. The Internet has become a place where a vast amount of information can be deposited and also retrieved by hundreds of millions of people scattered around the globe. With such an ability to reach out to this large pool of people, we have seen the expulsion of companies plunging into conducting business over the Internet (e-commerce). This has made the competition for consumers’ dollars fiercely stiff. It is now insufficient to just place information of products onto the Internet and expect customers to browse through the Web pages. Instead, e-commerce Web site designing is undergoing a significant revolution. It has become an important strategy to design Web sites that are able to generate contents that are matched to the customer’s taste or preference. In fact a survey done in 1998 (GVU, 1998) shows that around 23% of online shoppers actually reported a dissatisfying experience with Web sites that are confusing or disorganized. Personalization features on the Web would likely reverse this dissatisfaction and increase the likelihood of attracting and retaining visitors. Having personalization or an adaptive site can bring the following benefits: 1. Attract and maintain visitors with adaptive contents that are tailored to their taste. 2. Target Web contents correspondingly to their respective audience, thus reducing information that is of no interest to the audience. 3. Advertise and promote products through marketing campaigns targeting the correct audience. 4. Enable the site to intelligently direct information to a selective or respective audience. Currently, most Web personalization or adaptive features employ data mining or collaborative filtering techniques (Herlocker, Konstan, Borchers, & Riedl, 1999; Mobasher, Cooley, & Srivastava, 1999; Mobasher, Jain, Han, & Srivastava, 1997; Spiliopoulou, Faulstich, & Winkler, 1999) which often use past historical (static) data (e.g., previous purchases or server logs). The deployment of data mining often involves significant resources (large storage space and computing power) and complicated rules or algorithms. A vast amount of data is required in order to be able to form recommendations that made sense and are meaningful in general (Claypool et al., 1999; Basu, Hirsh, & Cohen, 1998). While the main idea of Web personalization is to increase the ‘stickiness’ of a portal, with the proven presumption that the number of times a shopper returns to a shop has a direct relationship to the likelihood of resulting in business transactions, the method of achieving the goal varies. The methods range from user clustering and time framed navigation sessions analysis (Kim et al., 2005; Wang & Shao, 2004), analyzing relationship between customers and products (Wang, Chuang, Hsu, & Keh, 2004), performing collaborative filtering and data mining on transaction data (Cho & Kim, 2002, 2004; Uchyigit & Clark, 2002; Jung, Jung, & Lee, 2003), deploying statistical methods for finding relationships (Kim & Yum, 2005), and performing recommendations bases on similarity with known user groups (Yu, Liu, & Li, 2005), to tracking shopping behavior over time as well as over the taxonomy of products. Our implementation works on the premise that each user has his own preferences and needs, and these interests drift over time (Cho, Cho, & Kim, 2005). Therefore, besides identifying users’ needs, the system should also be sensitive to changes in tastes. Finally, a truly useful system should not only be recommending items in which a user had shown interest, but also related items that may be of relevance to the user (e.g., buying a pet => recommend some suitable pet foods for the pet, as well as suggesting some accessories that may be useful, such as fur brush, nail clipper, etc.). In this aspect, we borrow the concept of ‘category management’ use in the retailing industry to perform classification as well as linking the categories using shared characteristics. These linkages provide the bridge for cross-category recommendations.