Fair Risk Minimization under Causal Path-Specific Effect Constraints

Razieh Nabi, David Benkeser

Keywords: Constrained learning, Algorithmic fairness, Machine learning, Counterfactual fairness, Causal inference

Abstract: This paper introduces a comprehensive framework for deriving and estimating fair optimal predictions in machine learning, grounded in causal and counterfactual path-specific effects as constraints. We detail the theoretical foundations of our approach, and provide closed-form solutions for constrained optimization within prevalent risk frameworks, including mean squared error and cross-entropy risks. These solutions conceptualize the fair risk minimizer as a nuanced adjustment to the unconstrained minimizer, influenced by the magnitude of the constraint, its canonical gradient, and the variance of this gradient. Additionally, we propose flexible semiparametric estimation strategies for these nuisance components, tailored to diverse model specifications. Such flexibility is essential for accurately implementing fairness adjustments across varied contexts. This work advances the discourse on algorithmic fairness by seamlessly integrating complex causal considerations into model training, thus providing optimal strategies for implementing fair models in real-world applications. The full paper is on arXiv under the same title.

Submission Number: 33