A Study Of Designing And Applying Recommenders Based On User Interests Modeling

One of the most important challenges in the era of information explosion is the way to filter out the disordered mixed information and then select the part that a given user may like. Along this line, recommender systems were proposed. Unlike search engines, which extract and return information according to user queries, recommender systems could automatically make item recommendations by mining user latent interests without user interventions. Thus, this technique has been successfully applied for improving the quality of services in a number of fields. However, existing recommendation algorithms often suffer from user interest overspecialization problem and cold-start problem. To that end, in this paper, we make a focused study of designing and applying recommenders based on user interests modeling for further enhancing both user experience and profit of the system. Our contributions could be summarized as:Firstly, we propose a method to represent user interests in collaborative filter-ing, and then describe a user interests expansion model by personalized ranking. Existing collaborative filtering based recommender systems usually focus on ex-ploiting the information about the user’s interaction with the systems, and the information about latent user interests is largely under-explored. Since learning to leverage the information about user interests is often critical for making better recommendations, we introduce a three-layer, user-interests-item, representation scheme. Specifically, after a topic model based method is used to capture each user’s interests, personalized ranking is developed for predicting user’s possible interests expansion. Moreover, a diverse recommendation list is generated by us-ing user latent interests as an intermediate layer between users and items. This recommendation strategy is summarized into an item-oriented model-based col-laborative framework, named iExpand. The extensive experimental results show that iExpand can lead to better ranking performance than state-of-the-arts.Secondly, we propose a way for context-rich tourist interest modeling, and design a cocktail approach (Cocktail) on personalized travel package recommen-dation. To deal with the technical and domain challenges in travel recommenda-tion, we first analyze the unique characteristics of travel packages and discover the spatial-temporal autocorrelations among their landscapes. Then develop the TAST model for representing the interests of the tourists, which can extract the interests conditioned on both the tourists and the intrinsic features (i.e., loca-tions, travel seasons) of the landscapes. Based on this TAST model, Cocktail is developed for personalized travel package recommendation by considering some additional factors. Finally, we evaluate TAST model and Cocktail on real-world travel package data (last for ten years) provided by a travel company in Chi-na. The experimental results show that the TAST model can effectively capture the unique characteristics of the travel data and the Cocktail is thus much more effective than traditional methods for travel package recommendation.Lastly, we provide an idea of recommending seed items to cold-start users for user interest elicitation. For choosing and evaluating seed items from a systematic perspective, we propose to identify influential seed nodes(items) from the item consumption network. Specifically, we first provide several influential seed items selection heuristics, which recommend the most influential items that can bring in more consumptions; Then, we present a seed items selection algorithm based on a linear information propagation model similar to PageRank. Different from the previous heuristics which do not consider the possible influence overlaps between seed items and select all the items simultaneously, the linear algorithm identifies the independent influence of each item and selects the seeds one by one following a greedy strategy. Finally, we show the effectiveness of these methods by the expected number of consumptions of cold-start users.