Research On Social Recommendation Model Based On Graph Neural Network

In recent years,with the advancement of technology and the development of social networks,social recommendation technology has attracted the attention of many scholars and become the focus of research in recommendation systems,which are widely used in online social networking and personalized advertising.A series of achievements have been made in the efficiency and accuracy of social recommendation,but there is still room for further improvement in the construction of item features,integration of social user preferences and the use of higher-order neighborhood features,etc.Therefore,this paper proposes corresponding improvements to address the above issues,mainly as follows.(1)To address the problem that existing social recommendation models do not consider the influence of similar items on item features and ignore the correlation between feature vector dimensions when using element inner product to model interactions.In this paper,we propose a graph attention social recommendation model based on similar item graph(SIG-GATSR),which consists of user feature modeling,item feature modeling and rating prediction modules,firstly,we construct user and item feature vectors from useritem interaction graph,social relationship graph and similar item graph by improved graph attention networks,then we obtain user and item feature mappings by using the outer product operation and finally apply multilayer convolutional neural network to model interactions to obtain predicted scores.(2)Further,to address the problem of not utilizing higher-order neighborhood features in interaction graphs and social graphs to enrich user and item representations,while ignoring social neighborhood interest preferences.In this paper,we propose a graph neural network social recommendation model that fuses higher-order preferences(So PGNNHOR),which mainly consists of an embedding fusion layer,an embedding propagation layer,and a scoring prediction layer.Firstly,the initial embedding vectors are fused in the embedding fusion layer,then the user features in different graphs are aggregated in the embedding propagation layer using the graph-level attention mechanism,the item features are obtained by an improved graph attention network,the higher-order neighbor representations of users and items are obtained by stacking multilayer graph neural networks,and finally the final user and item feature vectors are obtained in the scoring prediction layer using the layer-level attention mechanism,and the scoring prediction is completed.(3)Detailed experimental results on datasets Ciao,Epinions,and Flixster show that the SIG-GATSR model is able to deliver performance improvements of 0.81%-9.66% in several metrics,including prediction accuracy and normalized discounted cumulative gain(NDCG).By modeling higher-order neighborhood features and incorporating social preferences,So P-GNNHOR is able to further improve the recommendation performance of the model and validate its effectiveness.Figure [43] table [9] reference [66]...