Research And Implementation On Recommendation System Based On Graph Learning

With the development of the graph neural network,more and more graph neural network models are used in recommender systems to learn information of nodes from the user-item interaction bipartite graph.However,there are still some deficiencies in the current mainstream recommendation algorithm research.First of all,many current studies mainly focus on extracting explicit interaction from the user-item interaction bipartite graph,however,it is not enough to mine implicit relationships between nodes.Secondly,most graph neural network models are still based on supervised learning,so they depend on high-quality data input.In addition,interaction noise is inevitable in the acquired interaction data.Therefore,the thesis proposed an improved method for the above problems,and verifies the effectiveness of the improved method on public datasets.The main work and innovation of the thesis are as follows:1.The thesis proposes a relation enhanced neural attention network model named REGAT.The model defines implicit relationship graph of the nodes,as well as the construction details of the implicit relationship graph.The embedding vector representation of nodes in the implicit relation graph is learned by stacking multi-layer attention networks,and fused with the node vectors learned in the user-item interaction bipartite graph to obtain the final node vector representation.Experimental verification was carried out on three public datasets.Compared with the optimal baseline model,the index Recall@20 increased by 1.25%,1.59% and 1.94% respectively,and the index NDCG@20 increased by 0.46%,2.39% and 1.16% respectively.2.The thesis proposes a graph augmentation model named REGAT-DA.Firstly,the model adds edges to the long-tail nodes in the graph.Secondly,the thesis comes up with the concept of scoring centrality,which helps define the probability of nodes unremoved,and finally generate two different sub-views of bipartite graphs.With the help of contrastive learning,node vector representations can be acquired.Experiments were carried out on three public datasets.Compared with the REGAT model,the REGAT-DA model increased by 0.76%,1.97% and 1.78% in the index Recall@20,and the index NDCG@20 increased by 1.04%,2.07% and 1.90% respectively.3.According to the model proposed,a system for recommending movies is created and put into operation,which realizes the function of the user obtaining the recommended movie list and the administrator managing the recommendation model.