Abstract: Collaborative filtering is one of the key technologies widely applied in personalized recommendation system with great success.The critical step of collaborative filtering is to get k nearest neighbors (kNNs),which is utilized to predict user ratings and recommend.In order to improve the recommendation quality which is affected by the matter that rating data is characterized by its large scalability,high dimensionality,extreme sparsity,and the lower real-time ability by direct similarity measuring method in finding the nearest neighbors,we proposed a collaborative filtering re-commendation algorithm based on locality-sensitive hashing,and improved it.The algorithm applies locality-sensitive ha-shing technology based on p-state distribution to get lower dimensionality and index for large rating data.Then a multi-probe mechanism is utilized to improve the algorithm with great efficiency in obtaining the approximate nearest users of target user.Then,a weighted method is used to predict the user ratings,and finally perform collaborative filtering recommendation.Experiment results on typical dataset show that the proposed algorithm can overcome the limitation of high dimensionality and sparsity in some degree,and has good recommendation performance,high efficiency and less memory consumption.

Key words: Recommendation system,Approximate nearest neighbor,Collaborative filtering,Similarity measuring,Locali-ty-sensitive hashing