Spatio-Temporal Contrastive Heterogeneous Graph Attention Networks for Session-Based Recommendation

Abstract

The main goal of session-based recommendation (SBR) is to analyze the list of possible next interaction items through the user’s historical interaction sequence. The existing session recommendation models directly model the session sequence as a graph, and only consider the aggregation of neighbor items based on spatial structure information, ignoring the time information of items. The sparsity of interaction sequences also affects the accuracy of recommendation. This paper proposes a spatio-temporal contrastive heterogeneous graph attention network model (STC-HGAT). The session sequence is built as a spatial heterogeneous hypergraph, a latent Dirichlet allocation (LDA) algorithm is used to construct the category nodes of the items to enhance the contextual semantic information of the hypergraph, and the hypergraph attention network is employed to capture the spatial structure information of the session. The temporal heterogeneous graph is constructed to aggregate the temporal information of the item. Then, the spatial and temporal information are fused by sumpooling. Meanwhile, a modulation factor is added to the cross-entropy loss function to construct the adaptive weight (AW) loss function. Contrastive learning (CL) is used as an auxiliary task to further enhance the modeling, so as to alleviate the sparsity of data. A large number of experiments on real public datasets show that the STC-HGAT model proposed in this paper is superior to the baseline models in metrics such as P@20 and MRR@20, improving the recommendation performance to a certain extent.