Abstract
Graph Neural Networks (GNNs) often inherit and amplify social biases present in graph structures due to homophily. This paper introduces a fairness-aware graph filter ($h^{fair}$) that leverages Graph Signal Processing (GSP) to identify and dampen biased spectral components. By selectively filtering frequencies where sensitive attributes are disproportionately represented, the proposed method achieves significant bias reduction while maintaining high predictive accuracy across real-world social network datasets.
Introduction
Machine learning on graphs is integral to applications ranging from social recommendations to fraud detection. However, these models are prone to "bias amplification." Because of homophily—the tendency for similar individuals to connect—graph topologies often mirror historical social biases. During the message-passing process in GNNs, these structural biases are absorbed and magnified. Existing mitigation strategies, such as adversarial training or random edge pruning, often lack a principled approach to target the specific structural components driving bias.
Methodology: Spectral Scrubbing
The authors propose a signal-processing approach to fairness. By applying the Graph Fourier Transform (GFT), graph signals (features and sensitive attributes) are decomposed into the spectral domain.
- Bias Analysis via GFT: The graph topology and sensitive attributes are treated as signals. Analysis reveals that while both labels ($y$) and sensitive attributes ($s$) are often "smooth" (low-frequency), their distributions differ across the spectrum.
- The Fair Filter ($h^{fair}$): Unlike standard low-pass filters, $h^{fair}$ is a "matched" filter designed to identify frequencies where the sensitive attribute signal is significantly stronger than the label signal.
- Selective Dampening: The filter reduces coefficients at these "bias-heavy" frequencies. A hyperparameter $\tau$ controls the trade-off between fairness and utility.
Experimental Evaluation
The method was evaluated on two sampled versions of the Pokec social network (Pokec-z and Pokec-n), where the task was to predict a user's working field without being influenced by their region (the sensitive attribute).
Key Results
The proposed $h^{fair} + GNN$ framework was compared against several baselines, including standard GNNs, adversarial training, and FairDrop.
As shown in the results table, the proposed method significantly reduces bias. On Pokec-z, Statistical Parity ($\Delta_{SP}$) dropped from 6.79% in a standard GNN to 1.24%. Crucially, accuracy remained high (66.29% vs. 66.52%), demonstrating a superior fairness-utility trade-off compared to stochastic methods like FairDrop. The results also showed lower standard deviations, suggesting the spectral filter provides more consistent performance than stochastic alternatives.
Discussion and Conclusion
The core insight of this work is that bias is spectral. By visualizing the spectra of sensitive attributes versus labels, the researchers demonstrated that specific frequencies act as primary drivers of correlation with sensitive attributes. This provides a theoretical foundation for "spectral scrubbing."
While the current implementation requires a Laplacian eigendecomposition—limiting scalability to massive graphs—future work aims to develop computationally efficient polynomial approximations. This would enable the application of fairness-aware filtering to large-scale industrial graphs.
Summary updated