Research Of Recommendation Algorithms Based On Collaborative Filtering

With the rapid development of the Internet, people can enjoy the convenience brought by information technology, but at the same time, they often feel lost in the sea of information. That's because there is not an effective bridge between the people and the data. Information filtering technology try to help people filter out useless information, find the information they need.Personalized recommendation system is an application of information filtering, it try to conjecture the latent interests of users by investigate the features of the users or content and the behavior of the uses, and recommend some items to current user which may satisfy the user. So the recommendation problem can be comprehended as a process of behavior analysing, data mining and machine learning. Collaborative filtering is a popular technique of information filtering, has been successfully applied to many commercial recommendation system. Though collaborative filtering has been proved can providing good recommendation, but there are still some disadvantages. So, concerned the sparsity, cold start and the expansibility, we have done a detail study of collaborative filtering on the MoiveLens dataset which was supplied by GroupLens.Combining content-based information filtering and collaborative filtering method have been proposed to solve sparsity and cold start problem. In this paper, we introduced Bayesian Network to apply a content-based filtering model, and NMF(Non-negative Matrix Factorization) to solve the sparsity, then combining the NMF collaborative filtering and the content-based filtering. And the experiments show that our approach improved the performance on rating predication.We concern another problem which was ignored by most research on collaborative filtering, the drift of user's interest. We introduced a gradual forgetting mechanism to adjust the weight of user's rating, the newer the rating occur, the more important it is for describing the user's interest. And the experiments show that our approach improved the performance on Top-N recommendation.