Fairness in Algorithmic Recourse Through the Lens of Substantive Equality of Opportunity

TL;DR

This work targets fairness in algorithmic recourse by aligning recourse fairness with substantive equality of opportunity and explicitly incorporating time. It introduces two metrics—effort-to-recourse (ETR) and time-to-recourse (TTR)—and analyzes their disparities between advantaged and disadvantaged groups using an agent-based simulation of time-evolving recourse. The authors propose Circumstance-Normalized Selection (CNS) as a post-hoc intervention and compare it to Counterfactual Data Augmentation (CDA), showing CNS generally reduces disparities and that combining CNS with CDA yields the best performance. The study highlights the critical role of time and initial circumstances in recourse fairness, discusses trade-offs with overall utility, and calls for real-world datasets to validate these insights and guide practical deployments.

Abstract

Algorithmic recourse -- providing recommendations to those affected negatively by the outcome of an algorithmic system on how they can take action and change that outcome -- has gained attention as a means of giving persons agency in their interactions with artificial intelligence (AI) systems. Recent work has shown that even if an AI decision-making classifier is ``fair'' (according to some reasonable criteria), recourse itself may be unfair due to differences in the initial circumstances of individuals, compounding disparities for marginalized populations and requiring them to exert more effort than others. There is a need to define more methods and metrics for evaluating fairness in recourse that span a range of normative views of the world, and specifically those that take into account time. Time is a critical element in recourse because the longer it takes an individual to act, the more the setting may change due to model or data drift. This paper seeks to close this research gap by proposing two notions of fairness in recourse that are in normative alignment with substantive equality of opportunity, and that consider time. The first considers the (often repeated) effort individuals exert per successful recourse event, and the second considers time per successful recourse event. Building upon an agent-based framework for simulating recourse, this paper demonstrates how much effort is needed to overcome disparities in initial circumstances. We then proposes an intervention to improve the fairness of recourse by rewarding effort, and compare it to existing strategies.

Figures (10)

• Figure 1: Summary of the normative dimensions of the fairness formalization and fairness-enhancing intervention used in this paper, according to the classification of zehlike2022fairness.
• Figure 2: The effort-to-recourse, or the total amount of effort exerted per successful recourse event, by advantaged and disadvantaged groups (in terms of initial circumstances) over 20 time-steps.
• Figure 3: Reproduced with permission from DBLP:conf/eaamo/FonsecaBABS23.The x-axis shows time-steps $t$, and the y-axis shows agent scores $f(x_t)$. In this example, there are $k = 3$ positive outcomes available at each time-step. At $t = 0$, green agents receive a positive outcome ($f(x_0) \geq s_0$, where $s_0$ is represented by the horizontal line), and blue agents receive a negative outcome along with a recommendation $x'$ on how to change their features to receive a positive outcome. At time $t = 1$, new agents $N_1$, shown in black, enter the environment. Grey arrows show recourse actions. The agent shown in red acted on the recourse recommendation as directed, but (disappointingly) its effort turned out to be insufficient because competition from other agents "raised the bar" for acceptance.
• Figure 4: Illustration of score distributions that can result in fair decision-making but unfair recourse. Let individuals to the right of the vertical dashed grey line receive a positive outcome; then in all cases, decision-making is fair with respect to Demographic Parity between the advantaged (blue) and disadvantaged (orange) groups, but recourse is unfair. In our experiments, we assume that features are generated according to (a), where $\mu$ is the mean value for the high-performing agents from both populations, and $\mu_a$ and $\mu_d$ are the means of lower-performing agents of the advantaged and disadvantaged populations, respectively.
• Figure 5: The figure shows agents' feature values and the decision boundary for a positive outcome (the dashed lines) for 3 time-steps from a single run of the simulation, using the in-processing Group Recourse Regularization method by gupta2019equalizing. Note that faded-color points show agents' feature values at $t=0$, while full-color points show feature values at $t=15$.

Theorems & Definitions (2)

• Proposition 1
• proof