Research On The Collaborative Filtering Algorithm Based On HITS And Clustering

Personalized recommendation system,as one of the technologies to solve the problem of information overload,analyzes different sorts of past user behavior,such as purchase history,rating behavior,social relationships,context information,in order to identify new user-item associations.As e-commerce is growing in popularity,recommendation technology has become an indispensable part of ecommerce markets.However,many problems are still not well solved,such as sparsity of data sets and cold-start problem.In order to solve those problems,we proposed some approaches combining theory and practice:1.Applied the hybrid filtering technique based on the proposal Timed-HITS and collaborative filtering.Personalized recommendation system makes recommendations by applying the interaction information between users and items.However,traditional collaborative filtering suffers sparisity and cold start problems,which result in poor quality in recommendation systems.To overcome such problems,we propose a approach that we divided users into positive and negative users.And mining implicit data is also an effective method and inevitable trend to overcome the problem of sparseness.In this paper,we rely on logical regression algorithm to estimate the preference matrix given by users on items.And the HITS algorithm by employing the trust relationship among users,the preference for goods,and the time factor is improved.Finally,recommendation model is proposed by combining the improved HITS algorithm.With respect to other methods,our algorithm could generate better recommendation result in sparse data sets and cold-start situation.We apply the proposed recommendation algorithm on Movielens dataset and the experimental results show the accuracy of our method can be improved.2.Puted forward the hybrid recommendation method(APCF)combining the affinity propagation clustering technique and collaborative filtering.APCF focuses on sparsity problem of data sets.The algorithm is divided into two processes:(1)clustering process:Propose a twosteps clustering method for users that divide users into positive and negative users.For the first step,The affinity propagation clustering algorithm is used to cluster the active user.For the second step,On the base of clustering results from previous method,computing similarity between user and cluster center based on item's content vector.(2)recommendation process:Make improvements in two aspects that k-nearest neighborhood and similarity computation.The neighborhood is determined by the users in the same cluster.It takes into account not only the rating matrix but also the trust relationship and content information to improve predictive accuracy.In this experience,the experience compares the proposed method and other recommendation methods on Movielens data.And the RMSE and MAE is selected as an index to measure the recommendation quality.The results show that recommendation method proposed in this paper has a better performance.