Research On Collaborative Filtering Recommendation

With the rapid development of Internet,massive information has been created and stored in Internet.It's becoming difficult for users to access target information quickly and efficiently,and the "information overload"problem comes into being.Proposed to deal with this problem,recommender systems which can automatically analyze users' preference and push personalized information to them have been intensively studied and widely applied.Kinds of recommendation algorithms are applied to recommender systems in which Collaborative Filtering is the most widely applied one.But there are still some problems such as "Data Sparsity" and "Cold Start" with it.Besides,with the in-depth use of recommender systems,users want more diverse,novel recommendation and put forward some other higher requirements.This thesis aims to analyze collaborative filtering recommender systems and its problems,and with the mining and utilizing of the characteristics of user and item,and leveraging relative data mining techniques,proposes novel recommendation algorithms with better integrate performance.The main results are as follows:First,the thesis proposes an enhanced collaborative filtering with user interest and item popularity control.The quantity that a user gives thumb-ups to items can reflect some characteristics of participation,breadth of interest and criterion of the user.And the quantity that an item owns thumb-ups indicates its popularity.The thesis analyzes the way that user interest characteristics impact on item similarity and the way that item popularity impact the item-based collaborative filtering.And by considering user and item characteristics,the thesis proposes an enhanced collaborative filtering improving the integrate performance.Second,the thesis proposes an item clustering based collaborative filtering.Users' preference involves kinds of items with different degrees.Traditional item-based collaborative filtering tends to over-recommend items belonging to target-user's favorite class omitting the others making the,based on which the proposed collaborative filtering recommenrecommendation lacking diversity.Cluster techniques can classify items automaticallyds items from different classes independently making the recommendation more diverse.