Research On Social Recommendation With Graph Neural Networks

With the continuous development of Internet and online services,we have witnessed the explosive growth of user information.To alleviate the information overload issue,recommender systems have attracted increasing attention from more and more researchers,and have been applied to many real-world applications,such as online shopping systems,online multimedia systems,search engines and so on.Social recommender systems,as one of the most important category in modern recommender systems,not only utilize the user-item interactions but also introduce the additional social relationship between users,to further improve the user preference prediction and alleviate the data sparsity issue.Although lots of work in recent years has shown the effectiveness of deep learning in social recommendation algorithms,there are still three significant challenges that have not been well resolved:(1)Existing methods only consider the social relationship of users and ignore the relationship between items;(2)Most of the existing methods aim at a single type of user-item interaction behavior,and cannot handle multiple types of interaction behavior;(3)Existing methods only model the similarity of local features and ignore the importance of global features.In order to solve the three aforementioned challenges,this paper proposes a social recommendation model based on the Graph Neural Networks.The model is mainly composed of two parts:(1)Multi-behavior feature coding module,which models the multi-behavior interaction between users and items with Graph Neural Networks,and uses the high-order connectivity of user-item interaction graphs to construct high-order representations for users and items;(2)Local feature constraint module,which restricts local embedding learning with the generated global features in the infomax manner.This paper proposes two different local feature constraint methods based on the deep graph infomax and the local affinity-aware ranking respectively,and applies them to user social graphs and item correlation graphs.So that the learned representations contain both local similarities and the global similarities.Finally,this paper compares the proposed model with the cutting-edge algorithms on four large-scale real-world datasets to verify the feasibility and effectiveness of the model.