Intervention-based Causal Discrimination Discovery and Removal

Cong Su, Guoxian Yu, Jun Wang, Yongqing Zheng, Han Yu

Keywords: Fairness, Causal inference, Intervention-based metric

Abstract: Causal inference is a recent and widely adopted paradigm to deal with algorithmic discrimination. Building on Pearl's structure causal model, several causality-based fairness notions have been developed, which estimates the unfair causal effects from the sensitive attribute to the outcomes by incorporating the intervention or counterfactual operators. Among them, interventional fairness (i.e., $K$-Fair) stands out as the most fundamental and broadly applicable concept that is computable from observantional data. However, existing interventional fairness notions fail to accurately evaluate causal fairness, due to their following inherent limitations: (i) the causal effects evaluated by interventional fairness cannot be uniquely computed; (ii) the violation of interventional fairness being zero is not a sufficient condition for a causally fair model. To address these issues, we firstly propose a novel causality-based fairness notion called post-Intervention Cumulative Ratio Disparity (ICRD) to assess causal fairness of the decision models. Subsequently, we present a fairness framework (ICCFL) based on the proposed ICRD metric. ICCFL firstly generates interventional samples, and then computes the differentiable approximation of the ICRD to train a causally fair model. Both theoretical and empirical results demonstrate that the proposed ICRD effectively assesses causal fairness, and ICCFL can better balance accuracy and fairness.

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

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 10091