Grad2Fair: A Gradient-driven Approach for Graph Fairness without Demographics

Graph neural networks (GNNs) often exhibit group fairness issues, producing biased predictions based on sensitive attributes like race or gender. While many existing solutions attempt to mitigate this, they typically require access to demographic information, which is often unavailable due to privacy regulations. Recent attempts to use predicted demographics as proxies often fail due to prediction inaccuracies. This paper asks: can we achieve group fairness in GNNs without any demographic information or demographic prediction?

The authors observe that the gradients of misclassified nodes during training contain implicit information about demographic groups. They first propose GradDist, a metric that quantifies bias by measuring the distance between the local modes of gradient distributions for misclassified nodes. Building on this, they introduce Grad2Fair, a two-stage training framework. First, the model undergoes bias amplification to make inherent model biases more pronounced. Second, the model uses these amplified gradient signals to perform adaptive reweighting, prioritizing challenging samples to debias the training process without ever needing explicit demographic labels.

The study demonstrates that gradient distributions of misclassified nodes are highly correlated with the degree of unfairness in a dataset. By leveraging these gradients, Grad2Fair successfully mitigates bias across several real-world datasets. Experimental results show that Grad2Fair consistently outperforms existing state-of-the-art baselines that either rely on demographic proxies or other fairness-without-demographics approaches, while maintaining high computational efficiency.

This research provides a principled, privacy-preserving way to address algorithmic bias in high-stakes graph applications. By eliminating the need for sensitive demographic data, Grad2Fair offers a robust solution for real-world scenarios where privacy regulations like GDPR strictly limit the collection and use of personal attributes, ensuring that fairness can be pursued without compromising data privacy or relying on noisy proxy predictions.