Research Of Recommendation System Based On Collaborative Filtering

With the vigorous development of Internet, we can obtain rich information easily which, however, caused a very serious problem of information overload. Personalized recommendation system, since proposed, has got rapid development and being applied for its powerful capabilities and adaptability. Collaborative filtering is the most widely used and most successful among all recommendation algorithms. Based on the historical ratings collected from a specific domain, Collaborative filtering recommendation system provides a user with goods which he may be interested. However, in real world applications, these algorithms are unsatisfactory because of the data sparsity and the “cold start” problems. In this paper, we focus on two types of data in recommendation system, and use multi-source transfer learning technology and social trust relation among users to alleviate the impact of the above problems to some extent. Here are the main focus of this work and our improvements:1.We have designed a novel shared cluster-level factor model based on multi-source transfer learning. It solves the problem of single-source transfer learning, which cannot cover the entire target domain because of the limited number of users and items. The new model can learn the correlation coefficients between the different source domains and the target domain. Experiments results demonstrate that the improved algorithm can enhance the algorithm's prediction correct rate, while alleviating the impact of the data sparsity.2.We have studied the trust relationship between users in social networks, and taken into account the user's two roles, truster and trustee, in the trust network when affected by different users. Based on singular value decomposition model and starting from the two roles of the users, we built a new Dual-Trust SVD model by incorporating the explicit and implicit influence of both ratings and trust information. Then, combined with some points from nearest neighbor models, we design an improved Dual-Trust SVDNN model by introducing the similarity between items. The experiments later also proved that the new algorithms not only can improve the prediction effect, but also are more suitable for the “cold start” conditions.3.We analyze the impact of trust relationships on the problem of implicit feedback item ranking, and use the trust information to divide the items between user's explicit feedbacks, solving the uncertainty problem of the user preference between the explicit feedbacks. Then we use adaptive sampling to solve the drawback of the slow convergence speed of the BPR method. Several experiments on public data sets show that ATBPR model works better than the original algorithm on multiple evaluation criterions and also significantly speeds up the convergence rate.