Research And Application On Collaborative Filtering Algorithm In Friend Recommendation

Collaborative filtering algorithm is widely used in recommendation system, the key part of collaborative filtering recommendation algorithm is similarity computing method. Although there are many similarity computing methods, collaborative filtering recommendation algorithm can not work well in data sparsity conditions. "cold-start" problem is another problem of collaborative filtering recommendation algorithm. In order to solve these problems, this article proposed two methods, these two method can work well actually.First, we studied nine existing similarity methods, and proposed the design idea of our new similarity computing method PSJ (Proximity-Significance-Jaccard), it regards the variance value of user ratings, user global rating preferences and the number of common rating items; At the same time its Proximity factor used exponential function to describe how the variance value of user ratings affects user similarity computing and this design also avoid division by zero. The Significance factor and the URP factor ofNHSM (new heuristic similarity model) method were merged to build the Significance factor of the new method and this makes the computational complexity of the new method lower than NHSM method. For improving the recommendation performance in data sparsity conditions, the new method considers both the variance value of user ratings and user global rating preferences. The result of the experiment our PSJ method can improve recommendation performance than the other methods.Second, because of the performance of PSJ method is not satisfied when data sparsity reached to 99.99%, and it still exist "cold-start" problem. Due to the Matrix Factorization method can solve these two problems, we proposed PSJ Matrix Factorization method. The new method computes user similarity using PSJ method at first, secondly it use matrix factorization method to process user-user similarity matrix. We finally get recommendation result from user-user similarity matrix. The result of the experiment showed that the new method worked well than PSJ method.Finally, we used our two methods onto www.woao.com. www.woao.com is a basketball social network site, it contains the career stats of the user and game stats of the league. Our recommendation method use these data to compute similarity between users and recommended similarity users to the target user, it helps the target user to find friend who has similarity stats with him.