Variational Perspective on Fair Edge Prediction

Algorithmic fairness has been of great interest in the machine learning community and more recently in the graph context. In this paper, we address the problem of dyadic fairness where the task at hand is edge prediction, and the population of interest (nodes) is divided into a protected and a non-protected group, e.g. men and women. The goal is then to ensure that there should be no statistically significant difference in the prediction outcomes between the two groups, after accounting for any relevant factors that may impact the outcome. To proceed, we design a novel loss based on the variational information bottleneck principle to learn individual node representation while controlling a given level of dyadic fairness. The optimization of the loss is done with a Graph Neural Network. Experiments carried out on several real-world datasets confirmed the capacity of the proposed method, to maintain high accuracy on the edge prediction task while significantly reducing potential bias.