ScholarWorks@성균관대학교

초록

Recently, sequential recommendation models leverage deep neural networks, including RNN, CNN, and Transformer, to effectively capture user preferences from behavioral data. User behavior sequences inherently contain noise, which is often addressed through the use of filtering algorithms to mitigate its effects. Generally, filtering algorithms utilize the Fourier transform for processing. However, the Fourier transform, which relies on combinations of sine and cosine waves, is not entirely effective in handling diverse real-world user sequences. To address this limitation, we propose a method called Wavelet transform for sequential Recommendation (WaveRec), which takes advantage of wavelet transform as an alternative approach. Wavelet transform allows for detailed frequency decomposition by using various filters. We conduct experiments on four real-world benchmark datasets to demonstrate that wavelet transform better captures complex user behavior patterns. The experimental results show that our model outperforms four baseline methods. Furthermore, when we replace the Fourier transform component in existing state-of-the-art models with wavelet transform, we observe additional performance improvements, underscoring the effectiveness of our approach. We also discover that the appropriate wavelet filter varies for each dataset. Our code is available at https://github.com/Byungmoon-Heo/WaveRec.

키워드