A Two-stage Collaborative Filtering Algorithm Based On Random Walking And Cluster-based Smoothing

With the rapid development of Internet, the information of Internet is growing exponentially. The users are difficult to find the products or information that they need in a short time. To solve the "information overload problem", personalized recommendation system has emerged as the times demand, it recommends the information or goods that the users are interested in according to the users'characteristics of interest or buying behavior, and it is a personalized service system.Collaborative filtering is a personalized recommendation technology which is used widely, it analyses user preferences according to the users'evaluation information or purchase history, and then it recommends the products according to the users'interest. As the result of the increasing number of the users and the items and the scoring information of products that the users rating is very limited, the user - item rating matrix is extremely sparse, so the performance of the recommend system and the quality of recommendations has been seriously affected.In this paper, we propose a two-stage collaborative filtering algorithm based on random walking and cluster-based smoothing. Off-line stage: calculate the correlation between items, the common methods calculate the statistical correlation between items, such as cosine similarity, however, these methods do not work well under the sparse data. This paper proposed a new method which describes the correlation between items by cumulating weighted transition probability of each step; Cluster items according to the item correlation matrix, then smooth the unrated data by using clustering information. On-line stage: search the target item's neighbors according to the correlation between items which is cumulated during the off-line, and then predict the target user's ratings. This method can enhance the description of the correlation between items. When training set is considerable sparse, traditional similarity measures fail to describe actual relationships between items, but this method can work well .The result of experiments illustrate that searching neighbors according to the item correlation matrix which obtained by this method will become more accurate, this method can effectively relieve the impact of sparse data and improve the quality of recommendation.