CHSR: Cross-view Learning from Heterogeneous Graph for Session-Based Recommendation

Junchen Wang, Lei Duan, Runze Ma, Yidan Zhang, Zhaohang Luo

Abstract: Session-based recommendation (SBR) aims to predict the next item based on short behavior sequences for anonymous users. Most of the current SBR methods consider the scenario that a session just consists of a series of items. However, the multiple item attributes can also reflect user behaviors and provide information for recommendation. In other words, a session in the real world should consist of items and multiple item attributes, which means that the session is heterogeneous. In this paper, we propose a novel method for the anonymous recommendation with heterogeneous item attributes, named CHSR. Firstly, we construct homogeneous session graph and heterogeneous global graph for heterogeneous sessions to map the relationships among different item attributes. Secondly, homo-view and hetero-view of these two kinds of graph encoders are proposed to capture both intra and inter patterns of heterogeneous sessions. Thirdly, a cross-view fusion strategy with consistency loss is introduced to integrate the heterogeneous attribute information by fusing the representations from the two-type views. Finally, the interest preference of anonymous users is represented from the above steps. Extensive experiments conducted on three large-scale real-world datasets demonstrate the superior performance of CHSR over the state-of-the-art methods.