Gated-Slot Linear Attention Graph Neural Networks for Session-based Recommendation

Session-based Recommendation (SBR) systems aim to predict the next item which a user is likely to interact with, based on behavior sequence within the current session. Recent approaches leveraging the Transformer attention mechanism (e.g.,GC-SAN) have demonstrated superior performance over traditional methods; however, they suffer from heavily computation workload and space complexity due to the quadratic cost of self-attention. To address these limitations, a novel model termed GSLA4Rec (gated‑slot linear attention graph neural networks for session‑based recommendation) is proposed, which integrates linear attention with a learnable gating mechanism and a slot‑based memory update strategy inspired by gated‑slot attention. Specifically, GSLA4Rec first constructs two graphs: session graph to capture intra-session item transitions and neighbor global graph to encode cross-session collaborative signals, thereby modeling both local and global context information. The proposed approach centers on the gated-slot linear attention mechanism, which (1) employs gating units to suppress noise from irrelevant or redundant interactions in long sessions, (2) utilizes slot units to dynamically store and update compact session representations-reducing memory overhead, and (3) replaces self-attention with linear attention to achieve linear-time complexity. Together, these components significantly enhance the model's predictive accuracy and efficiency in capturing evolving users' interests. Extensive experiments on multiple public benchmark datasets, including Diginetica, Tmall, and NowPlaying, demonstrate that GSLA4Rec consistently outperforms SOTA SBR methods in recommendation performance, which validates the effectiveness of GSLA mechanism.