A Debiasing Framework for Graph Neural Networks Using Contrastive Learning

Graphs are a prevalent data structure used across various domains, but they often inherit human biases that can propagate into downstream systems. For instance, recommendation systems for social networks may amplify gender or ethnic homophily, leading to biased recommendations that influence critical decisions. The increasing complexity of graph neural networks (GNNs) has made identifying and mitigating such biases particularly challenging. This paper introduces a novel debiasing framework for GNNs, leveraging the widely adopted contrastive learning paradigm. In contrastive learning, GNNs are trained to differentiate between observed and randomly generated graphs, with bias often serving as a useful feature for this discrimination task. Our approach counteracts this by explicitly introducing bias into the random graphs, rendering it an ineffective feature for discrimination and preventing its propagation into the model. By making the bias model explicit, our framework enhances control and transparency in bias mitigation. Evaluations on link prediction tasks demonstrate that our framework significantly reduces bias while preserving overall predictive accuracy, advancing the fairness and reliability of GNN-based systems.