Fairness Under Uncertainty in Sequential Decisions
MCML Authors
Kirtan Padh
Niki Kilbertus
Prof. Dr.
Core PI
Abstract
Kirtan Padh
Niki Kilbertus
Prof. Dr.
Core PI
Abstract
Fair machine learning (ML) methods support the identification and mitigation of the risk that algorithms will encode and/or automate social injustices. While algorithmic approaches alone cannot resolve structural inequalities, these techniques can support socio-technical decision systems by surfacing unintended discriminatory biases, clarifying trade-offs, and enabling governance. Although fairness has been well studied in supervised learning, many real-life ML applications are online and sequential, with feedback from previous decisions informing future decisions. Each decision in such a setting is taken under uncertainty due to unobserved counterfactual outcomes and finite samples, with especially dire consequences for under-represented groups, who are often systematically under-observed due to historical exclusion and selective feedback. For example, a bank cannot know whether a denied loan would have been repaid, and it may have less data about previously marginalized and financially excluded populations. Towards this, this paper introduces a taxonomy of uncertainty in sequential decision-making—including model uncertainty, feedback uncertainty, and prediction uncertainty—to provide a shared vocabulary for assessing and governing sequential decision systems in which uncertainty is unevenly distributed across groups. We formalize the model uncertainty and feedback uncertainty using counterfactual logic and reinforcement learning techniques. We illustrate the potential harms for both the decision maker (unrealized gains and losses) and the decision subject (compounding exclusion and reduced access) of naïve policies that ignore the unobserved space. We illustrate our framework using simple algorithmic examples that demonstrate the possibility of simultaneously reducing the variance in outcomes for historically disadvantaged groups while preserving institutional objectives (e.g. expected utility) of the decision maker. Our experiments on data, simulated to include varying degrees of bias, illustrate how unequal uncertainty and selective feedback can produce disparities in sequential decision systems, and how uncertainty-aware exploration can alter observed fairness metrics. By providing a structured lens for identifying where and how uncertainty arises, this framework equips researchers and practitioners to better diagnose, audit, and govern fairness risks in sequential decision systems. In sequential and online systems in which uncertainty is a core driver of unfair outcomes, not merely incidental noise, our results highlight the importance of explicitly accounting for uncertainty in fair and effective decision-making.
inproceedings LPW+26
ACM FAccT 2026
9th ACM Conference on Fairness, Accountability, and Transparency. Montréal, Canada, Jun 25-28, 2026.Authors
M. S. A. Lee • K. Padh • D. Watson • N. Kilbertus • J. SinghLinks
DOI GitHubResearch Area
BibTeXKey: LPW+26