Attention Model Based Social Recommmendation

The rapid development of the Internet has triggered the explosive growth of information,making the problem of information overloading increasingly serious.As an effective information filtering technology,the recommender system greatly improve the efficiency of people browsing information.One of the most popular method is Collaborative Filtering(CF).However,due to the issues of data sparsity and cold start,these traditional recommendation methods achieve poor performance in many application scenarios.In recent years,with the rise of social platforms,social recommendation has become a widely applied recommendation method.Based on the social influence theory,these methods explore the role of social network in the recommender systems,thus alleviating the data sparsity issue in personalized recommendations.Therefore,in the social recommender systems,how to accurately model the social influence has become an important direction.However,most of previous research methods simply model the social influence,ignoring the role of social influence strength in recommendation.In addition,the tide of deep learning technology gives the recommender systems more opportunities for improvement.Benefiting from its powerful ability of feature representation,many neural network frameworks are applied in recommender systems to help models better extract relevant features from original data.This article uses some of the techniques commonly used in deep learning to learn the social information from social network,and combines with the idea of attention mechanism to model the social influence strength in social recommendation model,aiming at applying the related structure of neural network to catch more reasonable social influence information.The contribution of this article is listed as follows:(1)Beginning with the effective modeling of the social influence,we attempt to use the attention mechanism to learn social context information in social networks,and it is regarded as another part of factors which influence users' preferences.Then,combining with the users' own potential features of preference,the objective prediction function is constituted.Since our work focuses on the recommendation with implicit feedback,the datasets are also processed into corresponding forms.(2)In order to improve the effect of recommendation and the efficiency of model training,we adopt an autoencoder to learn the implicit structural information in thesocial matrix,and the parameters of users' potential feature vectors in the two parts of the original model are set to be independent.Several comparative experiments show the performance of our model on social recommendation and verify the feasibility of using attention mechanism to model social influence.