Community Structure Aware Social Recommendation Methods

With the rapid development of science and technology,massive data are flooding people's lives,and the problem of information overload is becoming increasingly serious.Recommendation system has gradually become a powerful tool for people to obtain personalized information,helping people to obtain useful information from massive information,and has been successfully applied to all walks of life.Recommendation methods based on collaborative filtering learn the feature representation of users and items by using the behavior data of users and items,and gradually become the mainstream recommendation method of recommender system.However,in the real world,the useritem rating matrix is highly sparse and unevenly distributed.When the recommendation system only considers the score matrix information between users and items,its performance is easily limited by sparse data and cold start problems,which makes users unable to get satisfactory recommendation content.We are now in the era of Web2.0,more and more users participate in online social activities,and a large number of social relationships are generated among users.These independent social data provide an opportunity to solve the problems of data sparsity and cold start inherent in collaborative filtering methods and improve the performance of recommendation systems.Some researchers begin to add social information to the traditional recommendation system to improve the performance of the model.In addition,there are a large number of community structures in the mobile social network between users,and it is particularly important to consider the social relations between users within the community.On the basis of the existing work,this paper mainly studies the method of social recommendation integrating community structure information.The main research contents and research results are summarized as follows:(1)A collaborative filtering method OCRIF which integrates overlapping community regularization and implicit feedback is proposed.This method firstly divides users into communities by community discovery algorithm and allows users to belong to different communities at the same time,so as to fully explore the potential community of common interest of users.Secondly,the OCRIF model incorporates the social implicit feedback within different communities and the conformance constraint of the implicit representation of goods.Finally,combined with the network representation learning model,we present a more effective user-community membership calculation method and a similarity calculation method between users,and extends the OCRIF model to the OCRIF+ model.Experiments on three public datasets show that the two proposed methods are superior to the previous similar recommendation algorithms.(2)A graph attention social recommendation model GASC,which integrates community influence diffusion and user interest diffusion,is proposed.In this model,the high-order social influence diffusion in the social network between users and the interest diffusion in the user-project interest network are modeled.The community discovery algorithm is used to divide users into their communities,and the finer-grained high-order social influence diffusion in the community is considered.In addition,a multi-layer attention network is designed to distinguish the different effects of different neighborhood nodes,different communities and different heterogeneous network graphs on the lowdimensional embedding representation of users and items.Finally,an attention network is introduced into the model to distinguish different order effects.Experiments on two publicly available datasets show that our proposed method is superior to the similar recommendation algorithms previously proposed.