Research On Recommendation System Based On Graph Neural Network

In recent years,the recommendation system based on the graph neural network has become a research hotspot in the industry.The algorithm is the core of the recommendation system.Learning representations of items and sessions through the graph neural network obtained more researchers’ attention.Most algorithms ignored the global relationship of items.Therefore,this thesis studies the recommendation system based on graph neural network from two aspects.(1)This thesis proposes a graph neural network recommendation algorithm GS-GNN based on global graph and session graph.The algorithm models the user’s historical behavior sequence from two perspectives,and uses graph neural networks to learn the global and local representations of items.In comparison,the innovation of this algorithm is that it takes the influence of the global features of the item on the recommendation effect into account.The experimental results also show that the algorithm can improve the recommendation effect.(2)This thesis also proposes a graph neural network recommendation algorithm GSS-GNN based on the global item graph,session item graph and session similarity graph,whose application scenario is different from GS-GNN’s.The algorithm starts from the perspective that there is also mutual influence between sessions,and further considers the impact of the similarity between sessions on the recommendation effect on the basis of GS-GNN.In comparison,the features considered by GSS-GNN are more comprehensive.Compared with the GS-GNN algorithm,the larger number of sessions in dataset leads to GSS-GNN’s better performs.