Intervention-based Cumulative Causal Fairness Learning

Cong Su, Guoxian Yu, Jun Wang, Han Yu, Dong Zhang

Keywords: Fairness, Causal inference, Intervention-based metric

Abstract: Causal inference has emerged as a powerful framework for addressing algorithmic discrimination, offering a principled approach to understand and mitigate unfairness in decision-making systems. Various causality-based fairness notions have been proposed to quantify unfair causal effects stemming from sensitive attributes by leveraging interventions or counterfactuals. Among these, intervention-based fairness has gained prominence as a foundational and widely applicable concept, computable from observational data. However, existing intervention-based fairness notions face critical limitations: (i) they fail to uniquely determine causal effects, and (ii) achieving zero interventional fairness does not guarantee causal fairness. To overcome these drawbacks, we introduce a novel intervention-based fairness metric, the post-Intervention Cumulative Ratio Disparity (ICRD), to rigorously assess causal fairness. Building on this metric, we propose the Intervention-based Cumulative Causal Fairness Learning (ICCFL) framework, which trains causally fair decision models by generating interventional samples and computing differentiable approximations of ICRD. Theoretical analysis and empirical evaluations demonstrate that ICRD provides a robust measure for causal fairness. Extensive experiments on four benchmark datasets demonstrate that ICCFL significantly outperforms six state-of-the-art methods, improving fairness (MMD metric) over 40% on average, and effectively balancing fairness and accuracy.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 24242