Research On Personalized Recommendation Based On User-item Co-Clustering

In the background of big data in recent years,the recommender system is a research hotspot in data mining and business intelligence,which can alleviate the information overload problem to some extent.Clustering techniques have been proved effective to deal with the sparsity and scalability problems in collaborative filtering recommender systems.They aim to identify groups of users having similar preferences or items sharing similar topics.However,users having similar tastes on one item subset may have totally different tastes on another set.Users have multiple preferences,and the preference of a user group may disperse on several item topics with different degree.In this study,we propose an integrated recommendation framework to exploit the cluster-level preference patterns.A matrix tri-factorization method is applied firstly to cluster users and items simultaneously,discovering cluster users and items into multiple groups.Posterior probability is used to describe cluster membership of users and items.A pair of strongly related user group and item group forms a submatrix.Then some traditional collaborative filtering technique is executed in every submatrix.Finally,final rating predictions are generated by aggregating results from all the submatrices and the items are recommended with a top-N strategy.We propose several meaningful aggregation methods.Experimental results show that the proposed framework significantly improves the recommendation accuracy of several state-of-the-art collaborative filtering methods.Our framework also relieves data sparsity problem.The computational efficiency can be improved the CF calculations can be parallelized,improving scalability of the recommendation system.