Towards a Unified Framework for Fair and Stable Graph Representation\n Learning

Abstract

As the representations output by Graph Neural Networks (GNNs) are\nincreasingly employed in real-world applications, it becomes important to\nensure that these representations are fair and stable. In this work, we\nestablish a key connection between counterfactual fairness and stability and\nleverage it to propose a novel framework, NIFTY (uNIfying Fairness and\nstabiliTY), which can be used with any GNN to learn fair and stable\nrepresentations. We introduce a novel objective function that simultaneously\naccounts for fairness and stability and develop a layer-wise weight\nnormalization using the Lipschitz constant to enhance neural message passing in\nGNNs. In doing so, we enforce fairness and stability both in the objective\nfunction as well as in the GNN architecture. Further, we show theoretically\nthat our layer-wise weight normalization promotes counterfactual fairness and\nstability in the resulting representations. We introduce three new graph\ndatasets comprising of high-stakes decisions in criminal justice and financial\nlending domains. Extensive experimentation with the above datasets demonstrates\nthe efficacy of our framework.\n