SHAHABI AND CHEN In an earlier work [24], we introduced a hybrid recommendation system--Yoda, which simultaneously utilizes the advantages of clustering, content analysis, and collaborate fil- tering(CF)approaches. Basically, Yoda is a two-step approach recommendation system. During the offline process, Yoda maintains numerous recommendation lists obtained from typical navigation patterns analyzed by clustering and content analysis techniques. During the online process, the confidence value of an active user to each navigation-pattern cluster is estimated using the PPED distance measure [25] by comparing the user's navigation attern with centroid of each cluster. Finally, Yoda generates customized recommendations for the user by aggregating across recommendation lists(one list for each cluster) using the confidence values as the weights To expedite the aggregation step, we proposed an optimized fuzzy aggregation function that reduces the time complexity of aggregation from O(N X E)to O(N), where N is the number of recommended items and e is the number of clusters. Besides the advantage of efficiency, our aggregation function(similar to FALCON [31] can manage disjunctive queries while the traditional weighted average method cannot. For example, assume that the system knows all the information about users and user U is interested in the items the list of items recommended by both expert A and B while our aggregation metho rate recommended by expert A or B. The traditional weighted average method can only gene retrieve the expected list. In sum, the time complexity is reduced through a model-based technique, a clustering approach, and the optimized aggregation method. Additionally, due to the utilization of con tent analysis techniques, Yoda can detect the latent association between items and therefore provides better recommendations. Moreover, Yoda is able to collect information about user interests from implicit web navigation behaviors while most other recommendation systems 3, 9, 20, 23, 28] do not have this ability and therefore require explicit rating information from users. Consequently, Yoda puts less overhead on the users. Since content analysis techniques only capture certain characteristics of products, some desired products might not be included in the recommendation lists produced by analyzing the content. For example, picking wines based on brands, years, and descriptors might not be adequate if"smell"and"taste"are more important characteristics. In order to remedy for this problem, we extend Yoda to incorporate more recommendation lists than just web navigation patterns. These recommendation lists can be obtained from various experts, such as human experts and clusters of user evaluations. Meanwhile, because PPED is specially designed for measuring the similarity between two web navigation patterns including related data such as browsed items, view time, and sequences information, it can only be used for estimating confidence values to navigation- pattern clusters. Therefore, a leaming mechanism is needed for obtaining the complete confidence values of an active user toward all experts. We propose a learning mechanism that utilizes users'relevance feedback to improve confidence values automatically using To the best of our knowledge, only a few studies[ 18, 29]incorporate Ga for improving the user profiles. In these studies, users are directly involved in the evolution processes. Because users have to enter data for each product inquiry, they are often frustrated with this method On the contrary, in our design, users are not required to offer additional data to improve174 SHAHABI AND CHEN In an earlier work [24], we introduced a hybrid recommendation system—Yoda, which simultaneously utilizes the advantages of clustering, content analysis, and collaborate fil￾tering (CF) approaches. Basically, Yoda is a two-step approach recommendation system. During the offline process, Yoda maintains numerous recommendation lists obtained from typical navigation patterns analyzed by clustering and content analysis techniques. During the online process, the confidence value of an active user to each navigation-pattern cluster is estimated using the PPED distance measure [25] by comparing the user’s navigation pattern with centroid of each cluster. Finally, Yoda generates customized recommendations for the user by aggregating across recommendation lists (one list for each cluster) using the confidence values as the weights. To expedite the aggregation step, we proposed an optimized fuzzy aggregation function that reduces the time complexity of aggregation from O(N × E) to O(N), where N is the number of recommended items and E is the number of clusters. Besides the advantage of efficiency, our aggregation function (similar to FALCON [31]) can manage disjunctive queries while the traditional weighted average method cannot. For example, assume that the system knows all the information about users and user U is interested in the items recommended by expert A or B. The traditional weighted average method can only generate the list of items recommended by both expert A and B while our aggregation method can retrieve the expected list. In sum, the time complexity is reduced through a model-based technique, a clustering approach, and the optimized aggregation method. Additionally, due to the utilization of con￾tent analysis techniques, Yoda can detect the latent association between items and therefore provides better recommendations. Moreover, Yoda is able to collect information about user interests from implicit web navigation behaviors while most other recommendation systems [3, 9, 20, 23, 28] do not have this ability and therefore require explicit rating information from users. Consequently, Yoda puts less overhead on the users. Since content analysis techniques only capture certain characteristics of products, some desired products might not be included in the recommendation lists produced by analyzing the content. For example, picking wines based on brands, years, and descriptors might not be adequate if “smell” and “taste” are more important characteristics. In order to remedy for this problem, we extend Yoda to incorporate more recommendation lists than just web navigation patterns. These recommendation lists can be obtained from various experts, such as human experts and clusters of user evaluations. Meanwhile, because PPED is specially designed for measuring the similarity between two web navigation patterns including related data such as browsed items, view time, and sequences information, it can only be used for estimating confidence values to navigation￾pattern clusters. Therefore, a learning mechanism is needed for obtaining the complete confidence values of an active user toward all experts. We propose a learning mechanism that utilizes users’ relevance feedback to improve confidence values automatically using genetic algorithms (GA) [7]. To the best of our knowledge, only a few studies [18, 29] incorporate GA for improving the user profiles. In these studies, users are directly involved in the evolution processes. Because users have to enter data for each product inquiry, they are often frustrated with this method. On the contrary, in our design, users are not required to offer additional data to improve