A1472
Title: Toward interventionally fair decision-making under partial causal graph knowledge
Authors: Yoichi Chikahara - NTT Communication Science Laboratories (Japan) [presenting]
Abstract: Machine learning is increasingly used to support high-stakes decisions such as loan approval, hiring, and abuse detection, but its use raises serious concerns about discriminatory predictions based on sensitive attributes such as gender and race. Because discrimination is often legally assessed in terms of causal effects, recent studies have used causal inference to ensure fairness in predictions. However, most existing methods assume that the true causal graph among variables is fully known. This assumption limits practical deployment because it requires solving the highly challenging problem of perfectly identifying the causal graph from observational data. A method is proposed for learning predictors that satisfy interventional fairness using a coarse causal representation called a cluster causal graph, which describes causal relationships among user-specified groups of variables. Although cluster causal graphs are easier to estimate in high-dimensional settings, they introduce graph uncertainty because the internal causal structure within each cluster is unknown. To address this challenge, (1) a graph algorithm that enumerates adjustment sets for identifying interventional distributions and (2) a barycenter kernel MMD constraint that efficiently measures distances between interventional distributions using multiple adjustment sets are developed. Experiments on synthetic and real-world data show improved accuracy-fairness trade-offs under graph uncertainty.
