AN ADAPTIVE RECOMMENDATION SYSTEM 177 3. Overview The primary objective of a web-based recommendation system can be stated as follows Problem 1. Suppose the item-set /=[i l i is an item presented in a web-site) and u is a user interactively navigating the Web-site. The recommendation problem is to obtain the ut's wish-list I, e I. which is a list of items that are ranked based on us interests In general, to acquire a wish-list for a user, a recommendation process goes through three 1. Obtaining user perceptions: Data about user perceptions such as navigation behaviors are collected. In some systems [9, 22], these data need further processing for inferring information which is used in the later phases 2. Ranking the items: The inferred user interests are utilized to provide the predicted user wish-list 3. Adjusting user settings: The system acquires relevance feedback(or follow-up navigation behaviors)from the user and employs it to refine the user settings, which represent the user perceptions. On occasion, this phase is integrated into phase one Figure 1 illustrates the processing flow of Yoda. Suppose that music CDs are the recom- mending items in the Yoda web-site. The objective of Yoda is to provide each active user. who is using the system, a satisfactory-and-customized recommendation list of CDs by Experts wish-lists Other Experts' wish-lists Cluste Aggregat⊥on Learning List Cluster Centroi Update Confidence values User Navigation Figure 1. Processing flow of Yoda.AN ADAPTIVE RECOMMENDATION SYSTEM 177 3. Overview The primary objective of a web-based recommendation system can be stated as follows: Problem 1. Suppose the item-set I = {i | i is an item presented in a web-site} and u is a user interactively navigating the Web-site. The recommendation problem is to obtain the u’s wish-list Iu ∈ I, which is a list of items that are ranked based on u’s interests. In general, to acquire a wish-list for a user, a recommendation process goes through three steps/phases: 1. Obtaining user perceptions: Data about user perceptions such as navigation behaviors are collected. In some systems [9, 22], these data need further processing for inferring information which is used in the later phases. 2. Ranking the items: The inferred user interests are utilized to provide the predicted user wish-list. 3. Adjusting user settings: The system acquires relevance feedback (or follow-up navigation behaviors) from the user and employs it to refine the user settings, which represent the user perceptions. On occasion, this phase is integrated into phase one. Figure 1 illustrates the processing flow of Yoda. Suppose that music CDs are the recom￾mending items in the Yoda web-site. The objective of Yoda is to provide each active user, who is using the system, a satisfactory-and-customized recommendation list of CDs by Figure 1. Processing flow of Yoda