Reproduce “Avoiding Discrimination through Causal Reasoning” on OpenResearch

About this paper

Niki Kilbertus, Mateo Rojas-Carulla, Giambattista Parascandolo, Moritz Hardt, Dominik Janzing, Bernhard Schölkopf

Researchers from Max Planck Institute for Intelligent Systems and UC Berkeley propose a causal reasoning framework to address limitations of purely observational fairness criteria in machine learning. Their work introduces interventional definitions of discrimination and outlines algorithmic procedures for constructing predictors that mitigate bias based on causal relationships in data.

Key insights

• Observational fairness criteria are inherently limited, as formally shown by Theorem 1, because distinct causal graphs—some discriminatory, some not—can produce identical observed joint distributions, making it impossible to infer discrimination from statistics alone.
• A crucial conceptual distinction is established between the unmanipulable protected attribute and its observable, intervenable proxies, which provides a practical handle for applying intervention-based fairness criteria.
• The paper introduces 'Unresolved Discrimination' and 'Proxy Discrimination' as precise, causally-grounded definitions that enable a more nuanced understanding of different forms of algorithmic unfairness through the lens of interventions.

Summary by alphaXiv.