Discreetly Exploiting Inter-Session Information for Session-Based Recommendation

Limited intra-session information has constrained the performance of early Graph Neural Network (GNN)-based session-based recommendation (SBR) models. To address this issue, researchers have increasingly incorporated inter-session information to improve next-item prediction. Although such additional information can provide valuable cues, it may also introduce undesirable interference. However, existing approaches often fail to effectively extract relevant information while filtering out extraneous interference, largely owing to incomplete modeling of inter-session dependencies. Specifically, inter-session connections have been examined merely at the item level, overlooking more intricate distinctions and associations between sessions. Moreover, current methods rely exclusively on similarity as the metric for evaluating inter-session dependencies, overlooking their multidimensional complexity. In this work, we propose a GNN-based SBR model that discreetly exploits inter-session information (DEISI). To overcome the first limitation, we introduce factor-level inter-session dependencies using disentangled representation learning, enabling the capture of finer-grained interactions. To address the second limitation, we design a novel metric named “stability” to complement similarity, providing an additional perspective on inter-session relationships. Consequently, DEISI constructs more detailed and reliable inter-session dependencies. Extensive experiments on three benchmark datasets demonstrate the superior performance of DEISI over state-of-the-art models.