Learning a Wavelet Neural Filter with Mamba for Sequential Recommendation

Transformer-based models have achieved significant success in sequential recommendation tasks. However, they often suffer from over-smoothing and inference inefficiency when handling long user-interaction histories. In real-world scenarios, user behaviour patterns are complex and non-stationary, posing further challenges. To address these problems, we propose a Mamba-based model enhanced with a wavelet frequency rescaler, inspired by attentive inductive bias and the strengths of state-space models. The Mamba block ensures efficient inference, while the wavelet filter attenuates noise in the time-frequency domain. Our approach not only showcases the effectiveness of state-space models but also highlights the potential of wavelet transforms for managing non-stationary data with long interaction histories in sequential recommendation. Extensive experiments on two public datasets with long interaction histories demonstrate that our model outperforms CNN, RNN, attention-based, and pure state-space model baselines in recommendation performance.