State space models with gated spatial kernels for sequential recommendation

Abstract Sequential recommendation aims to predict a user’s next interaction by modeling the sequential dependencies in historical behaviors. However, Transformer-based methods incur quadratic computational cost on long sequences and often struggle to capture behavioral patterns at different temporal scales, particularly short-term interest shifts. To address these limitations, we propose SEMamba, a sequential recommendation model based on state space models. SEMamba employs a Mamba encoder to model global dependencies and capture long-term user preferences. It further introduces a local spatial enhancement module to extract short-term interest patterns from neighboring interactions. A position-wise gating mechanism adaptively modulates the locally aggregated features before integrating them with the Mamba output. Experiments on the ML-1M, KuaiRand, and Mind datasets show that SEMamba improves Hit Rate by 1.60%, NDCG by 2.01%, and MRR by 2.23% on average over strong baselines. The results also demonstrate improved training stability and better generalization across datasets.