Research On Recommendation Algorithm Based On Social Network

Recommender system is an important technology in the field of information filtering. With the development of e-commerce, more and more users and items add to the network. Faced with such a large group of items, users usually give ratings to a small part of items, and more items have no ratings usually, so the data sparsity will become serious increasingly. In this condition, the recommendation accuracy of recommendation algorithm is relatively low. With the appearance of social networks, people find that incorporating trust information into the recommender system can improve the recommendation accuracy, and recommender systems based on social network are more realistic. But the social network information is also sparse. This paper aims to alleviate the data sparsity and improve the accuracy of recommender system.Firstly, to solve the problem of low predict accuracy, which most recommendation algorithms have in the condition of data sparsity, we propose a probabilistic matrix factorization algorithm, which bases on weighted trust. We first adopt an improved Jaccard similarity method to calculate the similarity value between any two trust users, and regard this value as a weighted factor of the original trust value, and obtain weighted trust. Then we construct basic bias, fusing weighted trust, and construct conditional distributions on user feature matrix and user-item ratings, which both fuse weighted trust. At last, we construct the extended probabilistic matrix factorization model. We use two important factors in the model to control the proportion of trust information during the recommended process.Secondly, to alleviate trust data sparsity in the social network, we propose a collaborative recommendation algorithm, based on similarity and filled trust matrix. We calculate the similarity of trust users based on original trust matrix. Then, we measure the trust degree of trust users, and make use of the trust transition mechanism to predict new trust value, and use new predicted trust values to fill the original trust matrix. Then we adopt the improved Jaccard similarity method to obtain similarity neighbors of the target user. In the end, we use the trust users and similarity neighbors of the target user to do recommendation.Finally, according to the proposed models, we write recommendation algorithm. We conduct the experiments on Epinions dataset, and compare our algorithm with other relevant algorithms and carry on analysis.