Research On Rating Propogation Based Collaborative Filtering For Recommendation

From the perspective of information retrieval, the development of the Internet has experienced web portal, search engine, and personalized recommendation stages. According to user’s interests, personalized recommendation can provide users with personalized service automatically, which makes it enjoy a more broad application prospect. As one of the most successful recommendation technologies, collaborative filtering method is also faced with accuracy and sparsity problems. Aiming at alleviating these problems, the rating propogation based collaborative filtering for recommendation was studied in this paper. The main research results are as follows:1. Based on traditional user-based graph model and item-based graph model, an improved "user-item" graph propogation model for Top-N recommendation is proposed. By introducing the local density sensitive factor, the dense nodes will receive much more gains compared with those sparse nodes. Thus, the performance of graph based collaborative filtering recommendation can be effectively improved.2. To overcome the deficiency that both user-based graph and item-based graph can not make full use of the user and item cues simultaneously, this paper proposes a rating score diffusion based graph propogation model. In this graph model, each rating entry of the rating matrix is used to build the graph node directly. Meanwhile, by jointly considering user-item information when constructing the relationship between nodes, a more reliable rating prediction can be obtained via graph model propogation.3. This paper proposes a Top-N recommendation framework based on multi-view regression ensemble learning. Under this framework, multiple preliminary rating predictions are obtained by using PLS to make multi-view interactive regression. On this basis, a modified adaptive iterative threshold involved in AdaBoost ensemble learning is proposed to learn weak predictors dynamically. Experimental results show that the proposed ensemble method for Top-N recommendation achieves much more promising performance.