Toward System-Level Algorithmic Recourse: Capacity, Competition, and Evaluation
Loading...
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Waterloo
Abstract
Decision makers are increasingly relying on machine learning in sensitive, high-stakes settings. In such contexts, algorithmic recourse aims to provide individuals with actionable and minimally costly steps to reverse unfavorable AI-driven decisions. While existing research predominantly focuses on single-individual (i.e., seeker) and single-model (i.e., provider) scenarios, real-world applications involve multiple interacting stakeholders operating under shared resource constraints. Optimizing outcomes under an individual welfare perspective overlooks this inherently multi-agent nature, where individuals interact and compete for limited resources. Individually computed recourse recommendations no longer remain valid in such settings: when multiple seekers simultaneously compete for limited provider capacity, recommendations that appear optimal in isolation can become infeasible at the system level, and stakeholder interactions directly affect recourse validity.
To address this limitation, we introduce a novel framework for multi-agent algorithmic recourse that explicitly models system-level interactions among multiple seekers and providers. We formulate this many-to-many setting as a capacitated weighted bipartite matching problem, where matches are determined by recourse costs and constrained by provider capacities. Edge weights, derived from recourse costs, enable the system to optimize social welfare while quantifying the welfare gap between individually optimal outcomes and those achievable under resource constraints.
Building on this formulation, we propose a three-layer optimization framework. First, we compute welfare-maximizing assignments under fixed capacity constraints via capacitated matching. Second, we optimize the distribution of provider capacities to minimize the welfare gap, aligning system resources with seeker preferences. Third, we incorporate cost-aware adjustments by penalizing deviations from existing capacity configurations, enabling practical deployment under real-world operational constraints.
We further extend the framework to support inequality-averse objectives through a concave social welfare formulation that prioritizes the most disadvantaged seekers, allowing a principled trade-off between efficiency and equity. In addition, we present a unified perspective and reproducible benchmarking framework for evaluating recourse methods, enabling consistent comparison across system-level and individual-level criteria.
Experimental results on synthetic and real-world datasets demonstrate that our framework achieves near-optimal social welfare with minimal modifications to system configurations. The results also highlight an inherent efficiency–fairness trade-off, showing how recourse systems can be designed to balance aggregate welfare with distributive considerations in resource-constrained environments.
Overall, this work extends algorithmic recourse from isolated individual recommendations to system-level design, providing a tractable and principled approach for improving collective outcomes while preserving individual actionability.