DFGNN: Dual-frequency Graph Neural Network for Sign-aware Feedback

TL;DR

This work tackles the underexplored problem of leveraging negative feedback in graph-based recommender systems. By analyzing feedback signals in the graph frequency domain, the authors identify positive feedback as predominantly low-frequency and negative feedback as high-frequency, which challenges the common low-pass bias of many GNNs. They propose DFGNN, a dual-frequency graph neural network that uses a low-pass filter for positive feedback and a high-pass filter for negative feedback, paired with signed graph regularization to prevent representation degeneration. Across multiple real-world datasets, DFGNN achieves superior performance in both recommendation and feedback-type recognition, and analysis shows the embeddings become more uniform and expressive, mitigating over-smoothing. The approach offers a principled, frequency-aware framework for sign-aware recommendation with practical implications for improving user experience by better handling negative feedback.

Abstract

The graph-based recommendation has achieved great success in recent years. However, most existing graph-based recommendations focus on capturing user preference based on positive edges/feedback, while ignoring negative edges/feedback (e.g., dislike, low rating) that widely exist in real-world recommender systems. How to utilize negative feedback in graph-based recommendations still remains underexplored. In this study, we first conducted a comprehensive experimental analysis and found that (1) existing graph neural networks are not well-suited for modeling negative feedback, which acts as a high-frequency signal in a user-item graph. (2) The graph-based recommendation suffers from the representation degeneration problem. Based on the two observations, we propose a novel model that models positive and negative feedback from a frequency filter perspective called Dual-frequency Graph Neural Network for Sign-aware Recommendation (DFGNN). Specifically, in DFGNN, the designed dual-frequency graph filter (DGF) captures both low-frequency and high-frequency signals that contain positive and negative feedback. Furthermore, the proposed signed graph regularization is applied to maintain the user/item embedding uniform in the embedding space to alleviate the representation degeneration problem. Additionally, we conduct extensive experiments on real-world datasets and demonstrate the effectiveness of the proposed model. Codes of our model will be released upon acceptance.

Figures (8)

• Figure 1: Different frequency signals in the time domain and frequency domain.
• Figure 2: The distribution of f($\lambda$) calulated based on normalzied Laplacian matrix. (a) It is the distribution of only negative edges graph. (b) It is the distribution of only positive edges graph. We evenly split the frequency $\lambda$ into ten buckets and calculate the normalized f($\lambda$).
• Figure 3: The visualized embedding and singular values. (a) is the embedding learned by NCF. (b) is the embedding learned by GCN. (c) is the normalized singular values of GCN and NCF.
• Figure 4: The graph filter kernel function of LGF and HGF with layer stack. The left is the kernel (1-$\lambda$) and the right is$\lambda$
• Figure 5: Ablation study of feedback type recognition task on Arts and GFood dataset