Research On Singularity-based And Reconstrcuting Trust Matrix Collaborative Filtering Recommendation Algorithm

With the vigorous development of e-commerce, recommendation system is gradually permeating people’s daily life, and constantly improving the users’ experience of online. As the most successful and widely used technique, collaborative filtering has been focused and researched comprehensively. Based on low prediction accuracy and cold start problem which are caused by data sparsity in collaborative filtering, we put forward the corresponding solutions. The specific content is shown as below.This paper first analyzes the present situation of collaborative filtering, then the classification and methods of collaborative filtering recommendation system are dicussed in detail. In the view of the problems of collaborative filtering, we summarize the corresponding strategies as well as the disadvantages.Secondly, aiming at the low prediction accuracy, the concept of singular value is introduced through analyzing the context of user-item rating. Considering the ratio of the number that item rated and the propotion of common rating items, the traditional Pearson Correlation Coefficient(PCC) and Jaccard algorithm are improved, and the two improved methods are combined in different two ways, in order to get better recommendation algorithm.Thirdly, to counter the problem of cold start, the trust information is incorporated. We reset the trust value in trust matrix based on user similarity, namely remove the trust value when the user similarity is lower than a certain threshold and add the trust statement into trust matrix when the user similarity is higher than another certain threshold. Then use weighted trust propagation to find more trust neighbors and distinguish different diatance trust neighbors, in order to solve cold start problems and improve prediction accuracy by this means.Finally, the two methods proposed in this paper, namely singular value based collaborative filtering recommendation algorithm and reconstructing trust matrix collaborative filtering recommendation algorithm, are varified on two different data sets respectively. The experimental results show that our methods can improve prediction accuracy obviously and resolve cold start problem effectively.