This paper presents MAPPING, a model-agnostic framework for debiasing Graph Neural Networks (GNNs) to ensure fair node classification and reduce leakage of sensitive information. Despite the success of GNNs in diverse applications, they can exacerbate historical discrimination and social stereotypes. Current fairness research, primarily based on i.i.d data, cannot be easily applied to non-i.i.d graph structures with topological dependencies among samples. MAPPING uses distance covariance-based fairness constraints to reduce feature and topology biases in any dimensions, combined with adversarial debiasing to limit the risks of attribute inference attacks. The study shows MAPPING can balance utility and fairness, and mitigate privacy risks.
Publication date: 24 Jan 2024
Project Page: https://doi.org/XXXXXXX.XXXXXXX
Paper: https://arxiv.org/pdf/2401.12824
