Research On Personalized Recommendation Algorithm Based On Collaborative Filtering

It has become more and more difficult for us to get information that we are interested in from the web owing to the tremendous amount of information available on it today. As the most successful technique, collaborative filtering has been widely used in personalized recommendation system. Through analyzing users' activities instead of the contents of information, it gathers ratings from people of the same interest with the target user and then creates recommendation. However, application of conventional algorithm of recommendation is hindered by the inaccurate calculation in similarity and problems of new-item and scalability.This dissertation introduces the basic theory of collaborative filtering technology and the steps of the implementation of CF-based algorithm, especial user-based CF algorithm and Item-based CF algorithm. Then it focuses on the calculation of the similarity, and analysis the difference among cosine-based similarity, correlation-based similarity and adjusted cosine-based similarity through experiments. The results show that the cosine-based similarity is better than the others in high-dimensional and sparse matrix. In order to solve the sparse problem of the traditional CF-based algorithm, the research presents an improved recommendation algorithm based on the prediction of the weighted mean of items' ratings. The experiments show that the improved algorithm is more accuracy in the similarity calculation than traditional methods. With the gradual increase of the users and items, problem of expansibility of recommendation system becomes serious, so the real-time response needs to be improved. The research also presents a frame for cluster-based recommendation and an improved k-means based algorithm named GKCF. The cluster-based recommendation firstly performs the improved GKCF algorithm to cluster the original user space to a new user space in offline mode, then the active user's neighbors are found through calculating the similarity between active user and the other virtual users in the new user space in online mode, and the items' ratings of the active user are predicted with these virtual neighbors' ratings. At last, the high ratings of items that predicted are recommended to the target user. For the cluster algorithm is performed in offline mode, the efficiency and speed of the real-time response of the recommendation system can be improved.