Research And Application Of Collaborative Filtering Recommendation Algorithm

With the application and popularization of Internet technology, especially the rapid development of E-commerce, Web has become one of the most important ways to access information. However, due to the resulted "information explosion" or "information overload", we may spend a lot of time in finding the target information. Recommendation system, as a new information filtering technology, emerges in response to the above problem and has become a hotspot in the research community.Collaborative filtering recommendation system appears earliest. Its basic idea is that users with similar interest may purchase same commodities. Collaborative filtering recommendation system is one of the most successful recommendation systems and has been widely used in many fields. Although, it still has two big shortcomings:the result of similarity calculation is not very accurate and the rating data is extremely sparse.In order to solve the above two problems, we propose two improved collaborative filtering recommendation algorithms in this thesis. The specific research topics are as follows:1. Collaborative filtering algorithm based on weight. It is the ratings that we use to analysis users' interest. The ratings of the active user on different items have different capacities to reflect his interest. It is apparent that the more peculiar a rating is, the more value it has. We call the ratings which are different from the common ratings on the same item as personalized ratings of the active user. So we divide the rating matrix into personalized and common matrix and calculate the similarities respectively. Then get the combined one by weighting. At the same time, different user has different influence. So at last, we amend the final similarity by neighbors'influential. The improved similarity calculation method can effectively increase the accuracy.2. Collaborative filtering algorithm with stepwise prediction. If arrange the predicting order in a reasonable way, we can effectively alleviate the data sparseness problem. So we put forward a new algorithm with stepwise prediction. It preprocessed the rating matrix first:rearranged the location of the matrix elements to concentrate the ratings to the left upper corner and filled part of user's missing data when he scored too few projects. Then it extracted a subsystem with high data density from rating matrix and filled the missing values by trust-based collaborative filtering algorithm. Finally it achieved stepwise prediction by constantly adding "new user" or "new project". This method can ensure that each calculation has high data density.