Research On Personalized Recommendation System Based On Collaborative Filtering

With the development of science and technology and the emergence of the Internet,we entered the era of information explosion.Massive information into the people's lives,resulting in the problem of information overload.The recommendation system appears to solve this problem.Recommendation system is based on the user's interest and behavior characteristics and the characteristics of the information to recommend for the user,At present,the most widely used and the most effective algorithm is collaborative filtering algorithm.The collaborative filtering system recommends similar information to the target user about the preferences of the similar user or the information of the target user preference.However,collaborative filtering system is still affected by cold start and data sparse,so we combine implicit feedback,tag information and trust mechanism to improve it,to a certain extent,alleviate the above problems,Here are the main focus of this work and improvements:1.We have studied the influence of trust relationship among users in the recommendation system,solving the problem that the user's own information is insufficient and the positive feedback is uncertain by using the trust information,and we use the implicit feedback to alleviate the sparse data.We improved the PPMF(Pair Wise Probabilistic Matrix Factorization)model,giving a new Trust Pair Wise Probabilistic Matrix Factorization model(T-PPMF).Then,an improved NT-PPMF algorithm is designed by introducing the neighborhood relationship between trusted users.The experimental results show that the improved algorithm can get better accuracy and faster convergence rate than the original algorithm.2.We have studied "group shilling attack" in the recommendation system,first,we divide the attack user group and the attacked items into an "attack block",then by analyzing the behavior of "attack block",we propose two indicators: ADFI(Average Deviation of Attack User to Fill Items)and UARD(Under Attack Items Average Rate Deviation)to identify the real attack block.Then we refer to the frequent itemset mining algorithm to design a new pruning algorithm to dig out the "suspected attack block".The final experimental results also show that our algorithm can effectively identify the "attack block".3.We have designed a music group recommendation algorithm based on tripartite graph diffusion algorithm.We assume that the behavior of the user in the network is determined by its interests,and its interests is distributed in the item space.The group interests we designed have both considered the average situation of the group users and the differences between the users,then we make the target group users' interests spread in the user group-item-item label tripartite graph.Finally,we generate recommendation for the target group.The experimental results show that our algorithm can get better experimental results and alleviate the problem of data sparseness.