Research On Social Recommendation System Based On Graph Neural Networks

With the development of the big data era,information resources are increasingly abundant and the phenomenon of information overload is severe.Driven by the background of the times,recommendation system came into being,traditional recommendation algorithms have shortcomings in solving problems such as data sparsity and cold start.Due to its rich data resources,social networks combined with recommendation system can effectively help improve recommendation performance.Hence,research based on social recommendation has gradually become a hot spot.Although the existing methods have made some progress,there are still many deficiencies in making full use of social network information.Most models usually only consider the local neighborhood influence of the user’s social network,or ignore the heterogeneous association of the user’s social interaction,and shallow linear interaction functions are often used to learn the complex interaction between user and item,resulting in unsatisfactory recommendations.In recent years,graph neural network has made remarkable achievements in many research fields because of its excellent graph structure data processing ability.Therefore,this paper combines graph neural network technology with social recommendation tasks,puts forward two kinds of social recommendation models based on graph neural network to predicts users’ preferences for items and improves the accuracy of recommendation.The research contents and innovations of this paper are summarized as follows:(1)The principle of social recommendation based on graph neural network is deeply studied.Firstly,this paper introduces the relevant theories of recommendation system and social recommendation.In view of the current shortcomings of traditional recommendation algorithms,this paper puts forward a solution to build social recommendation system based on graph neural network,and outlines the related technologies involved in the model construction,which lays a theoretical basis for the later work.(2)A social recommendation model based on neighborhood aggregation(DAN)is proposed.The existing social recommendation models generally consider only local neighborhood information and ignore the global dynamic impact from social interaction,which leads to the inability to learn user depth feature representation,this paper presents a social recommendation model based on neighborhood aggregation(DAN).The model uses graph convolution neural network to simulate the dynamic influence of user’s interest preference by neighborhood interactions,and obtains the representation of user’s depth feature by aggregating user’s own features and neighborhood interactions.Then,the deep neural network multi-layer perceptron captures the complex interaction between the feature of the user and the item,and finally the user’s preference prediction for item is obtained.Experiments show that the DAN model can effectively alleviate the problems of traditional recommendation algorithms and significantly improve the recommendation performance.(3)A social recommendation model based on heterogeneous neighborhood interaction(SAN)is proposed.At present,most social recommendation models ignore the differences of neighborhood social interaction,and simply treat neighborhood interaction in a homogeneous way,which makes it difficult to obtain a more accurate representation of user features.Considering the depth and intensity of social interaction,this paper presents a social recommendation model based on heterogeneous neighborhood interaction(SAN).The model uses the graph attention network to learn the heterogeneous association of social interaction,realizes the adaptive allocation of weights of different neighborhood nodes,and obtains a more accurate representation of user features by aggregating the user’s own features and the heterogeneous influence of neighborhood interaction.Experiments show that the SAN model can further improve the recommendation performance and improve the accuracy of the results.