Research On Additive Co-clustering Recommendation Algorithm With Social Influence

With the rapid development of social networks,rich social information has been applied into recommendation,and personalized recommendation is facing great changes.In recent years,how to use social information to solve the user's cold start problem and improve the accuracy of recommendation results has become a hot issue for researchers.Influenced by the type and scale of the datasets,the traditional matrix factorization model has some problems,such as low recommendation accuracy and low computing efficiency.The additive co-clustering matrix approximation algorithm has attracted the author's attention because of its small prediction error and high computing efficiency.However,only the rating information was used in this model,and the recommendation quality on cold start users is not very perfect.Therefore,in order to improve the efficiency of the algorithm and the recommendation quality on the cold start users,this thesis has completed the following discussion and research.Firstly,we modeled social information based on a simple connection between users,proposed a graph-regularized Fuzzy C-Means algorithm(gFCM)for users based on the thought of the fuzzy partition.Combing the trust information makes it easy to recognize the user partial blocks even though the rating information is sparse.Secondly,with the depth study of social relations theory,we analyzed the social relations between users from multi-perspective.Social relation strength was captured from a local perspective and the reputation of users in the whole social networks was calculated from a global perspectives.Exploiting local and global social influence to recognize the user partial blocks,we develop a graph-regularized weighted-Fuzzy C-Means algorithm(gwFCM)to cluster users precisely.Then,an additive co-clustering recommendation model with social influence was proposed in this thesis.Without considering the users' social information,the traditional co-clustering matrix approximation algorithm only uses the rating matrix.So this thesis clustered the users according to the above two algorithms and then clustered the items of the rating matrix.It expected to get general and specific categories by generating the user and item additive co-clustering results in an iterative method and predicted the missing data in the rating matrix.The effective combination of social information improves the recommendation quality,and also has a higher performance improvement for cold start users.Finally,recommendation quality of two algorithms which based on the additive co-clustering recommendation model with social influence was verified and analyzed on three standard datasets in this chapter.The experimental results show that the proposed model has better recommendation quality.The two recommendation algorithms with different combination of social information improve the accuracy of rating prediction.At the same time,the prediction error of the proposed recommendation algorithm that exploits local and global social information is lower than that based on trust relation,which indicates that the comprehensive combination of social information can improve the recommendation quality effectively.In addition,the experimental results on cold start users show that the recommendation model represented in this thesis can effectively solve the user's cold start problem.