Reweighted Spectral Graph Filter for Collaborative Filtering

Recent graph signal processing (GSP)-based collaborative filtering (CF) methods have demonstrated superior performance. However, these methods tend to overemphasize low-frequency signals in user-item interactions, while neglecting the significance of high-frequency signals, leading to excessive smoothing. To address this issue, we propose a reweighted spectral graph filter method for collaborative filtering(RSGF-CF). First, we construct a generalized normalized user-item interaction graph using asymmetric edge weight regularization. Next, we design a reweighted spectral band-pass filter that reweights both high-frequency and low-frequency signals to effectively capture valuable information. Finally, we combine this band-pass filter with a linear filter to create a strong baseline. Compared to state-of-the-art GCN-based and GSP-based CF methods, our model achieves superior performance on three real-world recommendation datasets.