MAPLE: Mamba-Based Adaptive Prototype Learning for Intent Disentanglement in Sequential Recommendation

Abstract

Sequential recommender systems struggle to simultaneously capture persistent long-term preferences and volatile short-term intentions, while existing dual-stream architectures are computationally inefficient when handling long sequences and suffer from feature entanglement between the two interest streams. To address these issues, we propose a Mamba-based Adaptive Prototype LEarning and cross-intent contrastive alignment framework (MAPLE). MAPLE employs Mamba as the long-term encoder; its selective state space mechanism selectively retains important historical signals across the entire interaction sequence, making it well-suited for capturing stable long-term interest features with linear computational complexity. Meanwhile, the GRU-based short-term encoder focuses on a local tail window to model recent preference shifts. To decouple these two interest streams, we maintain a set of global semantic prototypes for each interest type and extract soft proxies as personalized semantic anchors through attention-weighted aggregation, thereby leveraging global intent consensus to enrich individual user representations. These proxies are further applied as contrastive anchors in a novel cross-intent InfoNCE loss, which treats proxies of the complementary intent stream as cross-intent contrastive anchors to promote semantic differentiation between long and short-term representations. Finally, we introduce an asymmetric dynamic gated fusion module that prioritizes long-term features as the main component and incorporates short-term features as the auxiliary component for next-item prediction. Experiments on four real-world benchmark datasets demonstrate that MAPLE achieves highly competitive performance, significantly outperforming state-of-the-art models on rank-aware metrics.