Counterfactual Fairness With the Human in the Loop

Zhiqun Zuo, Tian Xie, Xuwei Tan, Xueru Zhang, Mohammad Mahdi Khalili

Primary Area: societal considerations including fairness, safety, privacy

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.

Keywords: Counterfactual Fairness, Strategic Behavior, Human in the Loop

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We proposed a definition of counterfactual fairness in the scenario where people would be impacted by the prediction and a kind of decision function that could improve counterfactual fairness in this scenario.

Abstract: Machine learning models have been increasingly used in human-related applications such as healthcare, lending, and college admissions. As a result, there are growing concerns about potential biases against certain demographic groups. To address the unfairness issue, various fairness notions have been introduced in the literature to measure and mitigate such biases. Among them, Counterfactual Fairness (CF) (Kusner $\textit{et al.}$) is a notion defined based on an underlying causal graph that requires the prediction perceived by an individual in the real world to remain the same as it would be in a counterfactual world, in which the individual belongs to a different demographic group. Unlike Kusner $\textit{et al.}$, this work studies the long-term impact of machine learning decisions using a causal inference framework where the individuals' future status may change based on the current predictions. We observe that imposing the original counterfactual fairness may not lead to a fair future outcome for the individuals. We thus introduce a fairness notion called $\textit{lookahead counterfactual fairness}$ (LCF), which accounts for the downstream effects of ML models and requires the individual $\textit{future status}$ to be counterfactually fair. We theoretically identify conditions under which LCF can be improved and propose an algorithm based on our theoretical results. Experiments on both synthetic and real data show the effectiveness of our method.

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: 6635