Fairness under Graph Uncertainty: Achieving Interventional Fairness with Partially Known Causal Graphs over Clusters of Variables

A new research paper by Yoichi Chikahara introduces a practical framework for achieving interventional fairness in AI systems when detailed causal knowledge is unavailable. Current causal fairness methods typically require a complete variable-level causal graph—a demanding assumption that limits real-world application. Chikahara's innovation addresses this by enabling fairness guarantees using only a causal graph over clusters of variables, which is substantially easier to estimate from domain knowledge or data. The framework identifies possible adjustment cluster sets from this higher-level graph and trains prediction models to minimize the worst-case discrepancy between interventional distributions across these sets.

The technical core involves a computationally efficient barycenter kernel maximum mean discrepancy (MMD) that scales favorably with the number of sensitive attribute values (like gender or race categories). Extensive experiments across 26 pages and 9 figures demonstrate that this cluster-based approach strikes a superior balance between fairness and accuracy compared to existing methods that either assume full causal knowledge or use non-causal fairness constraints. This work, published on arXiv, represents a significant step toward deploying legally-aligned, causally-fair algorithms in domains like hiring, lending, and healthcare, where perfect causal models are rare but fairness is critical.