Application And Research Of Personalized Recommender Systems

Internet has provided people with an extremely rich source of information. In these massive, heterogeneous resources on the web, there exists information and knowledge with great potential value. However, due to the huge amount of resources the phenomenon of"information overload"emergences. Recommender systems aim to provide people with items that they will appreciate by exploiting their previous preferences, the contents of items, and the demographic information, etc. In fact, recommender system has been one of the most effective tools to solve the problem of information overload. However, the real world is dynamic and things are always changing beyond peoples expectations. Indeed, volatile user interest drifts have been a major hinder to the applications of recommender systems. To this end, we design a density based segmentation method to identify user's interest patterns and detect user interest drifts. This, in turn, helps us to develop a general strategy to improve the performances of recommender systems. The main content of this paper is as follow:1) We provide an organized study of how to improve recommender systems under volatile user interest shifts. Along this line, we first analyze time-related user rating data and construct a rated item graph to track user interests according to the item similarities. Then, we reveal four types of user interest patterns. To identify these four types of patterns, we design a density-based segmentation method to detect user interest drifts and extract patterns by exploiting the rated item graph.2) To improve the performances of complex recommender systems, we remove both noise user and interest drift user which are identified by our method, and produce recommendations. The results show that the performance improvement is significant in terms of the Hit Ratio and Macro-DOA metrics.3) Then an application of website based on this system is implemented to recommend science and technology papers. We implemente three paper recommend methods: keywords based, taxonomy and keywords based, and three-level network based recommenders. The results show that our methods recommend papers which users appreciate.