Research On Recommendation Method Based On Graph Neural Network

With the advent of the intelligent era,the problem of information overload has seriously affected the user ’s network experience.The recommendation system came into being to provide users with personalized recommendations.In addition to using the interactive information of user items in the recommendation system,in fact,there are many auxiliary information such as comment information,social information,knowledge graph and so on,which are worth exploring and utilizing.The integration of these auxiliary information can effectively alleviate the problem of data sparsity and improve the accuracy and interpretability of recommendation.Graph structure data widely exists in recommendation systems.Graph neural network technology has been proved to be very powerful in learning graph structure data representation in various fields.However,the auxiliary information mining of users and items in the recommendation system based on graph neural network is not sufficient,the information fusion of multiple nodes is difficult,the mining of high-order connectivity information is not easy,and the interpretability is not satisfactory.This paper mainly studies the recommendation method based on graph neural network,and improves the problem of insufficient auxiliary information mining in the recommendation system.The main research is as follows :(1)Aiming at the sparse user-item interaction relationship and lack of interpretability,this paper proposes a general recommendation framework that integrates user,entity attributes,and item relationships.It aims to use the entity attributes and their emotional information in the comment text to improve the performance of the model while enhancing the interpretability of the recommendation results.Firstly,the entity aspectsentiment pair extraction method is used to annotate the comments.Based on the previous unified sequence labeling strategy,the method uses the transition matrix to fuse the boundary information of the entity attributes,and gives the prediction results of the two methods.Secondly,the weight measurement of entity attributes is defined,and the userentity attribute graph and entity attribute-project graph are constructed to describe user preferences and project attributes from the perspective of entity attributes,which improves the interpretability of the model.Experimental results on two datasets,Home＿and＿Kitchen and Electronics,demonstrate the effectiveness of the proposed model.(2)Aiming at the problem of insufficient mining of project and its attribute information in the recommendation system,a graph attention neural network recommendation model based on knowledge graph is proposed.The knowledge graph is used to expand the project attributes,explicitly model the high-order connectivity of users and projects in the recommendation system,and explore the high-order connections with semantic relationships in the recommendation system.Two corresponding attention components are proposed to fuse different nodes in the graph.Two types of attention mechanisms are used to embed neighbor nodes and update the node representation.The attention mechanism is used to filter the information contained in different nodes,and the heterogeneous node information in the graph structure is effectively fused.The model is compared on the three data sets of Last-FM,Yelp2018 and Amazon-book.The results show that the proposed method is improved on the two evaluation indicators of Recall and NDCG.