DWTRec: An Efficient Wavelet Enhanced Model for Sequential Recommendation

Transformer-based sequential recommender systems have achieved notable successes. Despite their effectiveness, recent studies reveal that self-attention mechanism used in current Transformer-based sequential recommendation models is constantly a low-pass filter which results in several problems. These problems include being incompetent in learning evolving user patterns and capturing users' abrupt interests. To integrate low and high frequency information effectively and overcome these problems, we propose a novel Wavelet enhanced hybrid attention network called DWTRec for sequential recommendation. Due to excellent characteristics of Wavelet Transform in signal processing, we can perform fine-grained signals decomposition while removing noise and capturing trending signals behind non-stationary sequences of user-item interactions. To address the bound distortion problem of the reconstructed signal of Discrete Wavelet Transform, we design a learning algorithm which adapts to different datasets. Our findings suggest a promising approach to improve the modeling power of self-attention. We tested our method on six datasets with different domains, different sparsity levels, and different average sequence lengths. Experiments show that our method outperforms all eight baseline models in recommendation performance on all datasets.