MVC-HGAT: multi-view contrastive hypergraph attention network for session-based recommendation

Session-based recommendation (SBR) mainly analyzes user’s interaction sequences and recommends a list of items for the next potential interaction. The existing models of SBR mostly obtain different item features through dual-channel graph neural networks, but there are often many different views of high-order relationships hidden in the real session sequence. Especially, the sparsity of the interaction sequence affects the performance of the SBR model. To make the recommendation results more comprehensive and accurate, we propose a multi-view contrastive hypergraph attention network (MVC-HGAT) for session-based recommendation, which models the session sequence as multi-view hypergraphs from three different views: the context relationship of the interaction sequence, the click unit and the hidden similarity attribute of items. The multi-view feature information of items is captured by hypergraph attention network (HGAT) and fused by sum-pooling. Additionally, multi-view contrastive learning is employed to alleviate data sparsity in the hypergraph. To prevent fitting, label smoothing is introduced in the loss function. Extensive experiment results on selected real datasets, including Diginetica, Yoochoose and Retailrocket, demonstrate that our proposed MVC-HGAT has improved recommendation performance to some extent, and is better than the baselines for two metrics Prec@20 and MRR@20.