Social Recommendation Algorithm Integrating User-Group Relationships

With the advent of the Web 2.0 era,the amount of information on the Internet has grown exponentially,and the problem of information overload has become increasingly serious.Personalized recommender systems have been widely recognized as an effective tool to solve the problem.In recommendation tasks,interaction data sparsity is a common and critical issue.Therefore,social recommendation is proposed to mitigate the problem and improve the recommendation performance by introducing social data of users.Existing social recommendation studies primarily focus on direct connections between users,such as friendship or correlation between users.However,in fact,there is also a serious data sparsity issue in above social data,which limits the performance of those models that utilize this kind of social data as side information.On the contrary,the usergroup relationship,another type of valuable information in social networks,hasn’t gotten enough attention.It refers to the connections formed between users and groups as a result of users joining the groups in which they are interested.In this paper,we study this not widely used relationship,thoroughly analyze its characteristics,and attempt to integrate it into our recommendation model in a reasonable way to further improve the accuracy of recommendation.The accomplishments of our research are listed as follows:(1)A social recommendation method integrating user correlation information is proposed.Through qualitative and quantitative analysis,we discover that compared to direct social connections between users,user-group relationships may have better performance in alleviating the sparsity of social data and could provide more information for inferring users’ preferences.Existing social recommender systems that take usergroup relationships into account are mostly based on the traditional models of matrix factorization or meta-path based representation learning on the heterogeneous graphs.In these models,the collaborative interaction interests and social influence of users have not been modeled in depth.Graph Neural Network(GNN)could more deeply model the correlation between nodes,so we propose to apply stronger GNN-based model to our task of integrating user-group relationships for better social recommendation.Specifically,we first transform user-group relationships into the correlation relationships between users.And then adopt a Graph Convolution Network(GCN)-based model called IGRec-Trans to learn user and item representations from the generated user-user graph and user-item graph,combining a multi-layer attention mechanism composed of node-level attention and graph-level attention to capture valuable information in graphs.Finally,rating prediction could be modeled as the inner product between the user and item embedding for each user-item pair.Experimental results on datasets Douban-Book and DoubanMovie show that our proposed model outperforms previous social recommendation models and is capable of alleviating the cold-start problem to some extent.(2)A social recommendation method integrating user-group relationships is proposed.The method proposed before of leveraging user-group relationships in an indirect way by transforming user-group relationships into the correlation between users is not ideal enough.There would be some lost in the accuracy and integrity of social information after transformation.Therefore,we propose IGRec-Direc,a novel model to learn user and item representations directly from the user-group bipartite graph and the user-item bipartite graph for social recommendation.Furthermore,we notice that due to the high complexity of user-group networks,interests of some users in the same group may be far different,especially for those groups with a large number of users.The indiscriminate use of high-order neighbors’ information in user-group graph may result in the introduction of negative information during the embedding propagation.As a result,in our model,we propose to constrain the graph convolution operations at the social side inside subgraphs,which are composed of users with similar interests and the groups they have joined.Experimental results show that our proposed new model performs best on both two real datasets.Additionally,it does admirably well in terms of resolving the coldstart issue.