Research On Graph Neural Network Recommendation Method Based On User Behavior

In recent years,Graph Neural Network(Graph Neural Network,GNN)related research fields have gradually become hot,and the combination of deep learning has brought tremendous improvements to reasonability,interpretability and model effects.The recommendation field has gradually begun to adopt GNN-like methods to solve the challenges it faces.GNN can naturally integrate node information and topology,and it has been proved to have a strong learning ability on the graph.In the e-commerce scenario,users,commodities,and the behavior between the two can be represented by a bipartite graph.Predicting the user's future behavior is transformed into predicting the probability of the user-commodity edge in the two-part graph,which has better interpretability and reasonability.Aiming at the problem that most existing GNN e-commerce recommendation models are based on homogenous graphs or single heterogeneous graphs,heterogeneous edges generated based on different behaviors of users are proposed,combined with graph convolution network(Graph Convolution Network,GCN)and collaboration The idea of filtering(Collaborative Filtering,CF)completes the modeling of heterogeneous graphs,making full use of the preference information implied by user behavior.For the cold start problem,with the help of the idea of inductive learning,train the function of mapping node embedding vectors,and use high-order adjacency and heterogeneous edges,even if the nodes have little interaction,this method can also be used to mine collaborative information to enrich the embedding of nodes Vector representation to overcome the poor prediction effect caused by insufficient sample size.Finally,a heterogeneous graph convolution collaborative filtering model(HGCF)is proposed.The Top-N recommendation experiment on the pre-processed Alibaba real e-commerce data shows that compared with the traditional Matrix Factorization(MF)model and Neural Graph Collaborative Filtering(NGCF)model,HGCF has higher recommendation accuracy and ability to deal with cold start.In addition,the HGCF superparameter experiment was designed to explore the performance of HGCF under different number of propagation learning layers,different normalization coefficients p_rui,and different output dimensions of the embedded propagation layer.