Leveraging User And Product Relationships To Enhance Recommendation Algorithms

With the exponentially increased data on the Internet,the customers suffer from heavy information overload problem.Therefore,recommender systems have emerged to satisfy the increasing personalization demand,so as to help customers efficiently make better decision.Collaborative filtering has been of the most successful and widely used recommendation algorithms.However,traditional collaborative filtering algorithms heavily suffer from data sparsity problem.In order to improve the recommendation accuracy,a number of user relationship based collaborative filtering algorithms have been proposed,such as social network based approaches.Similarly,product relationship based methods have been widely applied.However,most researchers ignore the integration of both user and product relationships,thus to further improve the recommendation performance.This paper aims to incorporate user and product relationships into traditional collaborative filtering algorithms to further boost recommendation performance.First,we fuse implicit user and product relationships into matrix factorization model,which is considered as regularization terms to restrict the learning process of user and product latent vectors,so as to investigate the influence on improving recommendation performance.Implicit relationships refer to the nearest neighbors that are discovered by the traditional similarity methods.In the meanwhile,we study the impact of number of nearest neighbors and different similarity methods on recommendation.Then,we integrate explicit user and item relationships into matrix factorization model,and regard them as regularization terms to constrain the learning of user and item latent factors,in order to further investigate the effect of explicit relationships on recommendation.Explicit relationships refer to the nearest neighbors that are discovered by auxiliary information(e.g.,social network,product category,etc.).Meanwhile,by comparing the experimental results of fusing implicit and explicit relationships,we analyze the significant of their respective impacts on recommendation.Lastly,we conduct extensive and comprehensive experiments on two real-worlddatasets to evaluate the effectiveness of the proposed model.