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

With the increasing popularity of machine learning models for high-stakes decision-making, it has become a consensus that (1) these models carry implicit biases [1, 2] and (2) these biases should be addressed to improve the fairness of algorithmic decisions [3, 4]. The disparate treatment of such models towards African Americans and women has been illustrated in the well-documented COMPAS [1] and Apple credit card [2] cases. While there has been extensive research on fair algorithms and fair machine learning in recent years, the proposed solutions have mostly disregarded important challenges that arise in real-world settings. Existing work in fair machine learning has focused on tabular, image, and text data [5, 6]. However, in several applications, data can be naturally modeled as graphs (or networks), representing different objects, their relationships, and attributes [7, 8]. Graph Neural Networks (GNNs) have achieved state-of-the-art results in many graph machine learning tasks, including node classification, link prediction, and graph classification [9, 10, 11, 12]. For instance, in the case of link prediction in a professional network, such as LinkedIn, which is a key recruiting and networking tool, link recommendations should not be biased against protected groups [13, 14, 15, 16, 17]. However, guaranteeing fairness in graph data is a challenge due to well-known correlations in the network caused by homophily and influence.