Filtering as Rewiring for Bias Mitigation on Graphs

Machine learning over graphs (MLoG) has attracted growing attention due to its effectiveness in processing relational data from complex systems such as social networks, financial markets, and the brain. However, MLoG algorithms that use the graph topology for information aggregation have been shown to amplify the already existing bias towards certain under-represented groups, often leading to discriminatory results in downstream tasks. In this context, here we consider the prob-lem of topology-induced algorithmic bias mitigation by cross-pollinating tools from MLoG and graph signal processing. Specifi-cally, we argue that application of a tunable debiasing graph filter can be reinterpreted as a graph rewiring process, thus offering an explicit handle to manipulate the utility versus topological bias tradeoff. Building on this insight, we formulate a fairness-aware network topology inference problem to obtain a rewired graph minimizing a correlation-based, unsupervised bias metric. Node classification experiments on several real-world datasets demonstrate that the proposed approach typically outperforms state-of-the-art baselines in terms of fairness metrics, and without a degradation in classification accuracy.