Causality-guided Graph Learning for Session-based Recommendation

Dianer Yu, Qian Li, Hongzhi Yin, Guandong Xu

Abstract: Session-based recommendation systems (SBRs) aim to capture user preferences over time by taking into account the sequential order of interactions within sessions. One promising approach within this domain is session graph-based recommendation, which leverages graph-based models to represent and analyze user sessions. However, current graph-based methods for SBRs mainly rely on attention or pooling mechanisms that are prone to exploiting shortcut paths and thus lead to suboptimal recommendations. To address this issue, we propose Causality-guided Graph Learning for Session-based Recommendation (CGSR) that is capable of blocking shortcut paths on the session graph and exploring robust causal connections capturing users' true preferences. Specifically, by employing back-door adjustment of causality, we can generate a distilled causal session graph capturing causal relations among items. CGSR then performs high-order aggregation on the distilled graph, incorporating information from various edge types, to estimate the session preference of the user. This enables us to provide more accurate recommendations grounded in causality while offering fine-grained interaction explanations by highlighting influential items in the graph. Extensive experiments on three datasets show the superior performance of CGSR compared to state-of-the-art SBRs.