Adaptive multi-embedding fusion via denoising bidirectional mamba for sequential recommendation

Sequential recommendation aims to predict users’ next interactions based on their historical behavior. However, existing models suffer from three major limitations: (1) reliance solely on collaborative filtering signals while overlooking item semantic information; (2) vulnerability to high-frequency noise in behavior sequences, which interferes with representation learning; and (3) computational inefficiency in Transformer-based bidirectional architectures. To address these challenges, we propose AFNBM (Adaptive Fusion via Noise-filtered Bidirectional Mamba), a novel framework that introduces three key innovations. First, we develop an adaptive semantic-collaborative fusion module that leverages pre-trained language models to extract item semantic embeddings and integrates them with collaborative embeddings through learnable parameters and. This mechanism effectively mitigates data sparsity and cold-start problems. Second, we integrate selective frequency modulation (SFM) into a bidirectional Mamba encoder, making this the first work to combine frequency-domain adaptive denoising with state space models (SSMs) for sequential recommendation. The bidirectional encoder captures temporal dependencies in both directions during training while the Mamba blocks themselves maintain linear complexity O (L). Each Mamba block incorporates an SFM layer with learnable parameter, enabling adaptive control over high-frequency component retention. The SFM operation introduces an O (L L) factor, making the overall framework complexity O (L L), which remains significantly more efficient than Transformer-based models with O (L^2) complexity while simultaneously improving computational efficiency and noise robustness. Third, we construct a comprehensive end-to-end framework built upon the adaptive fusion and bidirectional Mamba-SFM architecture. For prediction, we employ a gated linear unit (GLU) to selectively enhance salient features, followed by a masked-terminal prediction layer. Extensive experiments validate the effectiveness of our design: ablation studies quantify each component’s contribution (e. g. , GLU yields +2. 2% to +5. 2% gains across datasets), while AFNBM achieves 1. 02%-11. 76% improvements in Hit@10 and 2. 84%-7. 57% improvements in NDCG@10 over state-of-the-art baselines across four benchmark datasets, with particularly strong results on sparse and cold-start scenarios.