Research On Collaborative Filtering Recommendation System Based On Social Trust Network

With the rapid development of information technology and computer science, Internet is widely used in the real world, and consequently the amount of information on the web is dramatically increased. Users can acquire more and more information such that their requirements can be satisfied. But the process time to disvoer satisfied information with a limited time is too long due to the huge amount of information. Apparently, the traditional collaborative filtering systemdoes not work well in this situation. This leads to the emerging of personalized recommendation system. Through vigorous research efforts, theoretical foundation of novel recommendation system has been developed to some extent.Recommendation algorithm is crucial to the recommendation system, Collaborative filtering is facing some challenging problems such as sparsity, cold start problem. So the recommendation quality and accuracy of collaborative filtering alogrithm is not satisfying. In order to solve this problem, this thesispropose a novel collaborative filtering recommendation based on the trust of social netwok. In social network, user tend to buy products recommended by friends rather than those of recommended by strangers. Due to the sparsity and cold start problems, the collaborative filtering recommendation does not work well. However, collaborative filtering recommendation based on social netwok can achieve better performance.The main difference between our proprosed method and previous recommendation algorithm are summarized. Usually, the trust matrix of social network is very sparse. In order to solve this problem, this thesis combined local trust with global trust. This method can propagate the social network and make the trust matrix dense. This thesis used LDA to generate user's topic distribution. Then we can divide social users into different topic groups. Topic reflect the interest of user. The ratings of users who have similar interest with the target user will be assigned a high weight. Using this method, every trusted user's final trust value consists of two parts, which is original trust and topic similarity. At last, we can predict the user's rating matrix based on probability matrix factorization. Combining user own ratings with trust user' ratings comprehensively, we make the final prediction of the score. This method can not only consider user's own interest but also can consider the users who share the similr interests with the objective user, which can make the recommendation more accurate.In this paper's algorithm, propagating the trust network and using LDA to group users is our main innovative idea. The main work of this thesis have great effort on improving the accuracy of predicting ratings. Every step of our improvement of algorithm is reasonable. Through the experiment, we can see that our algorithm can improve the recommendation's accuracy.