Research On Personalized Recommendation Algorithm Based On User Interest Model

In the era of information explosion, how to filter out useless information quickly and find users’favorite items from the flood of information was one of the most challenging problems to be solved in recent years. In this context, the recommendation system as a powerful filtering tool emerged. Recommended system could push services to users initiatively based on the potential interests of them, and thus improving the efficiency and quality of services. As the traditional collaborative filtering algorithms were short of good understanding of the target users’interests, this paper attached importance to building effective user interest models. After gaining users’interests accurately, we could then design efficient recommendation algorithms.The main research aspects are as follows:Firstly, the development of Internet and mobile Internet was elaborated and the opportunities and challenges of recommendation system were studied. Besides, the methods of constructing and updating interest model in recommendation system were introduced exactly. After that, the common recommended algorithms were introduced.Secondly, as the method of using the probability model to extract interest directly only focused on the users’consumption number and ignored the other important feedback information, such as consumption time and ratings on items and so on, a method was proposed to add the feedback information to the probability topic model. Besides, users’ interests were expanded with the aid of PageRank algorithm in random walk way. Then, a user -interest-item model was designed to recommend items or predict users’ratings.Thirdly, because the above algorithm did not consider the situation where context information affected the users’behavior seriously, a recommendation algorithm with context was put forward and it was applied to food recommendation. First of all, the context was added to the traditional user-item model expressed as a vector consisting of several attributes, resulting in a User-Interest-Context interest model. Then Sub-Users were created according to different contexts from every user, thus obtaining a new user-item ratings matrix in certain context. Because of the data sparseness problem in this approach, an improved Slope One algorithm was designed to predict unknown ratings. Besides, similarity formula was optimized in order to provide more effective recommendation service.Finally, experiments were done to verify the contents this paper proposed and expectation for further research was bring forward.