Supervised Reinforcement Session Recommendation Model Based on Dual-Graph Convolution

In the field of session-based recommendation by anonymous sessions, the commonly used supervised learning modeling method has the problem of sub-optimal recommendation. The supervised reinforcement learning (SRL) recommendation framework can be used to solve this problem, but there is currently a lack of research on graph state representations of anonymous sessions in this field. In this regard, we propose a supervised reinforcement session recommendation model, HG-SRL, based on global hypergraphs and local session graphs. In this model, we propose, for the first time, a state representation method based on hypergraph neural networks and graph neural networks to bridge the gap in SRL in graph state construction methods. To fully utilize graph information at both global and local levels, we propose a self-matching attention fusion mechanism, which fuses the different levels of information contained in the two graphs through cross-calculation and then embeds them in the final graph state representation. To make the state more comprehensive, we also mine the neighboring sessions of each anonymous session on the global session hypergraph to supplement the neighboring session information of the anonymous session. Experimental tests were conducted on three real-world datasets, which have shown that HG-SRL can effectively improve the accuracy of session recommendations.