Better Fair Than Sorry: Adversarial Missing Data Imputation for Fair GNNs

This paper addresses the problem of learning fair Graph Neural Networks(GNNs) under missing protected attributes. GNNs have achieved state-of-the-artresults in many relevant tasks where decisions might disproportionately impactspecific communities. However, existing work on fair GNNs assumes that eitherprotected attributes are fully-observed or that the missing data imputation isfair. In practice, biases in the imputation will be propagated to the modeloutcomes, leading them to overestimate the fairness of their predictions. Weaddress this challenge by proposing Better Fair than Sorry (BFtS), a fairmissing data imputation model for protected attributes used by fair GNNs. Thekey design principle behind BFtS is that imputations should approximate theworst-case scenario for the fair GNN – i.e. when optimizing fairness is thehardest. We implement this idea using a 3-player adversarial scheme where twoadversaries collaborate against the fair GNN. Experiments using synthetic andreal datasets show that BFtS often achieves a better fairness × accuracytrade-off than existing alternatives.