Revisiting (Un)Fairness in Recourse by Minimizing Worst-Case Social Burden

TL;DR

This work studies fairness in algorithmic recourse by linking prediction quality to recourse burden across groups. It critiques traditional equal-cost metrics and introduces a social-burden framework, culminating in MISOB, a minimax-based training method that reduces the worst-group burden while preserving overall accuracy and without accessing sensitive attributes during training. The authors establish theoretical connections between classifier decisions and recourse costs, and empirically validate MISOB on real datasets against multiple recourse methods, showing robust improvements in fairness across groups. The approach offers a practical, pipeline-aware tool for enhancing recourse fairness with broad applicability and no requirement for sensitive attribute data in training or deployment, signaling a meaningful step toward fairer decision-making systems.

Abstract

Machine learning based predictions are increasingly used in sensitive decision-making applications that directly affect our lives. This has led to extensive research into ensuring the fairness of classifiers. Beyond just fair classification, emerging legislation now mandates that when a classifier delivers a negative decision, it must also offer actionable steps an individual can take to reverse that outcome. This concept is known as algorithmic recourse. Nevertheless, many researchers have expressed concerns about the fairness guarantees within the recourse process itself. In this work, we provide a holistic theoretical characterization of unfairness in algorithmic recourse, formally linking fairness guarantees in recourse and classification, and highlighting limitations of the standard equal cost paradigm. We then introduce a novel fairness framework based on social burden, along with a practical algorithm (MISOB), broadly applicable under real-world conditions. Empirical results on real-world datasets show that MISOB reduces the social burden across all groups without compromising overall classifier accuracy.

Figures (2)

• Figure 1: Conventional metrics to measure fairness in recourse hide disparities. (a) Given a population balanced by sensitive group membership and ground‑truth outcomes (e.g., true ability to repay a loan), treating prediction and recourse fairness separately, as in the majority of prior work, can mask disparities: if positive classifications are unevenly distributed, one group bears recourse burdens more frequently even when per‑instance recourse costs are equal; misclassification (e.g., false negatives) further amplifies this effect. (b) Conventional cost‑parity metrics suggest balanced recourse costs across groups $S \in \{0,1\}$ (left). Our metric (Eqn. \ref{['eq:cost_group']}) reveals substantial disparity once group differences in decision and error rates are taken into account, even with "fair" classifiers, FairSVM gupta2019equalizing and CAU-FairSVM von2022fairness (right).
• Figure 2: Empirical evaluation of MISOB for different$\alpha \in \{0.1, \ldots, 1.0\}$. We report the average results and standard deviation (shaded area) over 10 runs, with WT wachter2017counterfactual and GS laugel2017inverse, on the Adult dataset with race and gender as the sensitive attributes. The worst burden and cost (Eqn. \ref{['eqn:eval_metrics']}) are in log scale. In brief, increasing $\alpha$, which favors high-burden instances, improves fairness guarantees, but large $\alpha$ values may hurt overall accuracy.

Theorems & Definitions (2)

• Definition 3.1
• Definition 3.2