pith. sign in

arxiv: 2604.21711 · v1 · submitted 2026-04-23 · 💻 cs.LG · cs.AI

Fairness under uncertainty in sequential decisions

Michelle Seng Ah Lee , Kirtan Padh , David Watson , Niki Kilbertus , Jatinder Singh This is my paper

Pith reviewed 2026-05-09 22:20 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords fairnesssequential decision makinguncertaintycounterfactualsreinforcement learningselective feedbackcompounding exclusion
0
0 comments X

The pith

A taxonomy of model, feedback, and prediction uncertainty shows how uneven distributions across groups in sequential decisions produce compounding exclusion that uncertainty-aware policies can mitigate.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that fairness in online machine learning systems must address uncertainty that hits groups unequally because of historical exclusion and selective feedback loops. It introduces a three-part taxonomy—model uncertainty from limited data, feedback uncertainty from unobserved counterfactuals, and prediction uncertainty—to give practitioners a common language for diagnosing risks in sequential settings like lending or hiring. Formalizing the first two via counterfactual logic and reinforcement learning reveals concrete harms: institutions forgo gains while disadvantaged subjects face reduced access and self-reinforcing exclusion. Algorithmic constructions and bias-varied simulations then demonstrate that policies incorporating uncertainty-aware exploration can lower outcome variance for those groups without lowering institutional expected utility.

Core claim

Policies that ignore the unobserved space in sequential decisions under uneven uncertainty systematically disadvantage under-represented groups through compounding exclusion and reduced access while also imposing unrealized losses on decision makers; formalizing model and feedback uncertainty with counterfactuals and reinforcement learning, together with the additional category of prediction uncertainty, supplies the diagnostic tools needed to surface these effects, and uncertainty-aware exploration strategies can reduce outcome variance for disadvantaged groups while preserving expected utility.

What carries the argument

Three-category taxonomy of uncertainty (model, feedback, prediction) in sequential decisions, formalized via counterfactual logic and reinforcement learning to trace how selective feedback creates self-reinforcing disparities.

If this is right

  • Sequential systems can audit fairness risks by measuring how uncertainty is distributed across demographic groups rather than only tracking average outcomes.
  • Uncertainty-aware exploration changes fairness metrics in simulations with varying bias levels, showing that selective feedback is not incidental noise but a driver of disparity.
  • Decision makers can reduce unrealized gains and losses by explicitly modeling unobserved counterfactuals instead of relying on observed data alone.
  • The taxonomy equips governance processes to distinguish incidental noise from uncertainty-driven unfairness in online applications.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Extending the approach to real lending or hiring platforms would require integrating uncertainty estimates into existing reinforcement learning pipelines to test long-term access effects.
  • Neighbouring problems such as dynamic pricing or content recommendation could adopt the same taxonomy to check whether selective feedback creates parallel exclusion loops.
  • A testable extension is to apply the framework to recidivism or medical triage data and measure whether uncertainty-aware policies alter group-level outcome variance over multiple decision rounds.

Load-bearing premise

That the three uncertainty categories are exhaustive and non-overlapping, and that counterfactual logic plus reinforcement learning suffice to diagnose and correct the resulting harms without missing interactions.

What would settle it

A controlled simulation or real deployment in which uncertainty-aware exploration policies produce no reduction—or an increase—in outcome variance for disadvantaged groups relative to standard policies, or in which observed disparities are better explained by factors outside the three-category taxonomy.

Figures

Figures reproduced from arXiv: 2604.21711 by David Watson, Jatinder Singh, Kirtan Padh, Michelle Seng Ah Lee, Niki Kilbertus.

Figure 1
Figure 1. Figure 1: Uncertainty taxonomy throughout ML lifecycle [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of methods on performance and bias over time. The [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Sensitivity to 𝑌 measurement bias at end of Q10, keeping the other biases at 0. counterfactual utility maximization is much better at mitigating bias without a cost of utility (cumulative profit). While we cannot make generalized statements about its effectiveness across scenarios, this illustrates how accounting for unrealized outcomes can shift observed fairness metrics. Comparing the methods for differe… view at source ↗
Figure 4
Figure 4. Figure 4: Method rankings by four fairness metrics (lower disparity = better; rank 1 = best). No single method dominates all [PITH_FULL_IMAGE:figures/full_fig_p021_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Method ranking by cumulative profit (higher = better; rank 1 = most profitable). Prob-weighted exploration ranks first [PITH_FULL_IMAGE:figures/full_fig_p022_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Selection-rate difference (ΔSR) over decision rounds under each bias type at maximum severity. Counterfactual utility converges toward parity under label bias but diverges under proxy measurement bias (𝛽 (𝑚) 𝑅 ), where its utility estimate inherits the proxy corruption. Under interaction proxy bias (𝛽𝑌𝑏 ), all methods show persistent disparity. C.3 Worst-Case Distribution of Outcomes The violin plots in Fi… view at source ↗
Figure 7
Figure 7. Figure 7: False-positive-rate difference (ΔFPR) over decision rounds. Under label bias, counterfactual utility reduces ΔFPR; under proxy measurement bias, prob-weighted exploration achieves the smallest gap. Which method controls FPR best depends on the bias mechanism. Q2 Q3 Q4 Q5 Q6 Q7 Q8 Q9 Quarter 0.0 0.2 0.4 0.6 FNR diff. Label hist. bias ( y = 4) Q2 Q3 Q4 Q5 Q6 Q7 Q8 Q9 Quarter Label meas. bias ( m y = 4) Q2 Q3… view at source ↗
Figure 8
Figure 8. Figure 8: False-negative-rate difference (ΔFNR) over decision rounds. Trajectories are more unstable than for ΔSR or ΔFPR: methods that increase minority selection can initially reduce ΔFNR but overshoot in later quarters. Under interaction proxy bias, trajectories diverge rather than converge, illustrating the limits of uncertainty treatment under structural proxy corruption. Q2 Q3 Q4 Q5 Q6 Q7 Q8 Q9 Quarter 0.0 0.1… view at source ↗
Figure 9
Figure 9. Figure 9: Accuracy difference (ΔAcc) over decision rounds. Under label historical bias, the naive baseline maintains the smallest gap; under label measurement bias, counterfactual utility converges to near-zero ΔAcc. These opposing dynamics reflect that ΔAcc responds to changes in both decision policy and label quality. combinatorial sweep, providing a worst-case view: a method with a narrow, low-positioned violin i… view at source ↗
Figure 10
Figure 10. Figure 10: Per-quarter utility (profit) over decision rounds. Prob-weighted exploration maintains the highest or near-highest [PITH_FULL_IMAGE:figures/full_fig_p024_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Distribution of |ΔSR| at the final quarter across all bias conditions and random seeds. Uncertainty-aware and prob-weighted exploration show narrower, lower-positioned violins than the naive baseline. Counterfactual utility has a bimodal distribution: concentrated near zero under label bias but with a long upper tail under proxy bias [PITH_FULL_IMAGE:figures/full_fig_p024_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Distribution of |ΔFPR| across all bias conditions. Prob-weighted exploration produces the most compact violin. The naive baseline has a wider distribution, and the tails,not the medians,are where methods diverge most. Naive Naive exploration Uncertainty-aware Prob-weighted Counterfactual utility 0.0 0.2 0.4 0.6 0.8 |FNR difference| |FNR Difference| across all bias conditions [PITH_FULL_IMAGE:figures/full… view at source ↗
Figure 13
Figure 13. Figure 13: Distribution of [PITH_FULL_IMAGE:figures/full_fig_p025_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Distribution of |ΔAcc| across all bias conditions. Distributions are narrower than for FNR disparity, as accuracy aggregates false positives and false negatives that partially cancel. Uncertainty-aware and prob-weighted exploration show lower medians than the naive baseline; counterfactual utility shows the widest spread. Naive Naive exploration Uncertainty-aware Prob-weighted Counterfactual utility 0.0 0… view at source ↗
Figure 15
Figure 15. Figure 15: Distribution of normalised cumulative profit across all bias conditions. Prob-weighted exploration achieves the [PITH_FULL_IMAGE:figures/full_fig_p026_15.png] view at source ↗
read the original abstract

Fair machine learning (ML) methods help identify and mitigate the risk that algorithms encode or automate social injustices. Algorithmic approaches alone cannot resolve structural inequalities, but they can support socio-technical decision systems by surfacing discriminatory biases, clarifying trade-offs, and enabling governance. Although fairness is well studied in supervised learning, many real ML applications are online and sequential, with prior decisions informing future ones. Each decision is taken under uncertainty due to unobserved counterfactuals and finite samples, with dire consequences for under-represented groups, systematically under-observed due to historical exclusion and selective feedback. A bank cannot know whether a denied loan would have been repaid, and may have less data on marginalized populations. This paper introduces a taxonomy of uncertainty in sequential decision-making -- model, feedback, and prediction uncertainty -- providing shared vocabulary for assessing systems where uncertainty is unevenly distributed across groups. We formalize model and feedback uncertainty via counterfactual logic and reinforcement learning, and illustrate harms to decision makers (unrealized gains/losses) and subjects (compounding exclusion, reduced access) of policies that ignore the unobserved space. Algorithmic examples show it is possible to reduce outcome variance for disadvantaged groups while preserving institutional objectives (e.g. expected utility). Experiments on data simulated with varying bias show how unequal uncertainty and selective feedback produce disparities, and how uncertainty-aware exploration alters fairness metrics. The framework equips practitioners to diagnose, audit, and govern fairness risks. Where uncertainty drives unfairness rather than incidental noise, accounting for it is essential to fair and effective decision-making.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 4 minor

Summary. The manuscript introduces a taxonomy of uncertainty in sequential decision-making consisting of model, feedback, and prediction uncertainty. It formalizes model and feedback uncertainty via counterfactual logic and reinforcement learning, illustrates harms to decision makers and subjects from policies that ignore unobserved outcomes, provides algorithmic examples demonstrating that outcome variance for disadvantaged groups can be reduced while preserving institutional objectives such as expected utility, and reports experiments on simulated data with varying bias showing how unequal uncertainty and selective feedback produce disparities and how uncertainty-aware exploration alters fairness metrics. The framework is intended to equip practitioners to diagnose, audit, and govern fairness risks where uncertainty is unevenly distributed.

Significance. If the taxonomy and formalizations hold, the work supplies a shared vocabulary and diagnostic lens for fairness issues that arise specifically in online and sequential settings, an area where supervised-learning fairness methods are known to be insufficient. The constructed algorithmic examples are a strength, as they demonstrate in principle that variance reduction for disadvantaged groups need not trade off against expected utility. The simulations, though illustrative, highlight compounding effects of selective feedback. These elements together could support more targeted fairness audits in high-stakes sequential domains.

minor comments (4)
  1. [Abstract] Abstract: the claim that 'experiments on data simulated with varying bias show how unequal uncertainty and selective feedback produce disparities' is stated without any quantitative metrics, sample sizes, bias parameters, or specific fairness measures (e.g., demographic parity, equalized odds); adding these details would allow readers to assess the magnitude and robustness of the reported alterations.
  2. [Formalization] The formalization section references counterfactual logic and reinforcement learning for model and feedback uncertainty but does not display the key definitions or equations; including them (even if standard) would clarify how the taxonomy maps onto existing tools and avoid reliance on high-level description alone.
  3. [Algorithmic examples] Algorithmic examples section: the statement that variance reduction is achieved 'while preserving institutional objectives' would be strengthened by an explicit statement of the objective function or constraint used in each example, so that readers can verify the claimed lack of trade-off.
  4. [Taxonomy] The manuscript would benefit from a short discussion of potential overlaps or gaps between the three uncertainty categories (model, feedback, prediction), even if only to note that the taxonomy is intended as a practical rather than exhaustive partition.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive summary of the manuscript and the recommendation for minor revision. We are pleased that the taxonomy of uncertainty, the formalizations via counterfactual logic and reinforcement learning, the algorithmic examples showing variance reduction without sacrificing expected utility, and the simulations on selective feedback were viewed as strengths that could support targeted fairness audits in sequential domains.

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper presents a conceptual taxonomy of uncertainty (model, feedback, prediction) in sequential decisions, formalized using standard counterfactual logic and reinforcement learning concepts. It illustrates harms via examples and reports experiments on simulated data with varying bias. No load-bearing derivation, prediction, or result is shown to reduce by construction to fitted parameters, self-citations, or ansatzes that presuppose the target fairness outcomes. The central contribution is the taxonomy and diagnostic framing itself, which does not rely on equations that equate outputs to inputs by definition. This is a standard non-circular framework paper.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The paper rests on standard counterfactual and RL formalisms plus the assumption that the three uncertainty categories are the right partition. No new physical constants or fitted parameters are introduced in the abstract.

axioms (1)
  • domain assumption Counterfactual logic and reinforcement learning provide an adequate formalization of model and feedback uncertainty in sequential decisions.
    Stated in the abstract as the basis for formalization.
invented entities (1)
  • Taxonomy of model, feedback, and prediction uncertainty no independent evidence
    purpose: Shared vocabulary for assessing uneven uncertainty across groups
    New organizing structure introduced by the paper; no independent empirical test supplied in abstract.

pith-pipeline@v0.9.0 · 5585 in / 1323 out tokens · 25014 ms · 2026-05-09T22:20:22.126893+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

59 extracted references · 59 canonical work pages

  1. [1]

    Fair Housing Act

    1968. Fair Housing Act. https://uscode.house.gov/view.xhtml?edition=prelim&path=%2Fprelim%40title42%2Fchapter45 42 U.S.C. §§3601–3619

    work page 1968

  2. [2]

    Equal Credit Opportunity Act

    1974. Equal Credit Opportunity Act. https://uscode.house.gov/view.xhtml?edition=prelim&req=granuleid:USC-prelim-title15-chapter41- subchapter4 15 U.S.C. §§1691–1691f

    work page 1974

  3. [3]

    Equality Act 2010

    2010. Equality Act 2010. https://www.legislation.gov.uk/ukpga/2010/15/contents Accessed 2026-03-23

    work page 2010

  4. [4]

    UN General Assembly et al. 1948. Universal declaration of human rights. UN General Assembly 302, 2 (1948), 14–25

    work page 1948

  5. [5]

    Solon Barocas, Moritz Hardt, and Arvind Narayanan. 2019. Fairness and Machine Learning. fairmlbook.org. http://www.fairmlbook.org

    work page 2019

  6. [6]

    Joachim Baumann, Alessandro Castelnovo, Riccardo Crupi, Nicole Inverardi, and Daniele Regoli. 2023. Bias on demand: a modelling framework that generates synthetic data with bias. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency. 1002–1013

    work page 2023

  7. [7]

    Richard Berk, Hoda Heidari, Shahin Jabbari, Michael Kearns, and Aaron Roth. 2021. Fairness in Criminal Justice Risk Assessments: The State of the Art. Sociological Methods & Research 50, 1 (2021), 3–44

    work page 2021

  8. [8]

    Umang Bhatt, Javier Antorán, Yunfeng Zhang, Q Vera Liao, Prasanna Sattigeri, Riccardo Fogliato, Gabrielle Melançon, Ranganath Krishnan, Jason Stanley, Omesh Tickoo, et al. 2021. Uncertainty as a form of transparency: Measuring, communicating, and using uncertainty. In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society. 401–413

    work page 2021

  9. [9]

    Simon Caton and Christian Haas. 2024. Fairness in machine learning: A survey. Comput. Surveys 56, 7 (2024), 1–38

    work page 2024

  10. [10]

    Gaebler, Hamed Nilforoshan, Ravi Shroff, and Sharad Goel

    Sam Corbett-Davies, Johann D. Gaebler, Hamed Nilforoshan, Ravi Shroff, and Sharad Goel. 2023. The measure and mismeasure of fairness. J. Mach. Learn. Res. 24, 1, Article 312 (Jan. 2023), 117 pages

    work page 2023

  11. [11]

    Amanda Coston, Ashesh Rambachan, and Alexandra Chouldechova. 2021. Characterizing Fairness Over the Set of Good Models Under Selective Labels. In Proceedings of the 38th International Conference on Machine Learning. 2144–2155

    work page 2021

  12. [12]

    Elliot Creager, David Madras, Toniann Pitassi, and Richard Zemel. 2020. Causal Modeling for Fairness In Dynamical Systems. In Proceedings of the 37th International Conference on Machine Learning. 2185–2195

    work page 2020

  13. [13]

    Sculley, and Yoni Halpern

    Alexander D’Amour, Hansa Srinivasan, James Atwood, Pallavi Baljekar, D. Sculley, and Yoni Halpern. 2020. Fairness is Not Static: Deeper Understanding of Long Term Fairness via Simulation Studies. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency. 525–534

    work page 2020

  14. [14]

    Maria De-Arteaga, Artur Dubrawski, and Alexandra Chouldechova. 2018. Learning under selective labels in the presence of expert consistency. arXiv preprint arXiv:1807.00905 (2018)

    work page arXiv 2018

  15. [15]

    Giovanni De Toni, Stefano Teso, Bruno Lepri, and Andrea Passerini. 2025. Time can invalidate algorithmic recourse. In Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency. 89–107

    work page 2025

  16. [16]

    Parkes, and Goran Radanovic

    Christos Dimitrakakis, Yang Liu, David C. Parkes, and Goran Radanovic. 2019. Bayesian Fairness. Proceedings of the AAAI Conference on Artificial Intelligence 33, 01 (2019), 509–516

    work page 2019

  17. [17]

    Michele Donini, Luca Oneto, Shai Ben-David, John Shawe-Taylor, and Massimiliano Pontil. 2018. Empirical Risk Minimization under Fairness Constraints. In Advances in Neural Information Processing Systems 32. 2796–2806

    work page 2018

  18. [18]

    Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Richard Zemel. 2012. Fairness through awareness. In Proceedings of the 3rd Innovations in Theoretical Computer Science Conference. 214–226. doi:10.1145/2090236.2090255

    work page doi:10.1145/2090236.2090255 2012

  19. [19]

    Kfir Eliaz and Ariel Rubinstein. 2014. On the fairness of random procedures. Economics Letters 123, 2 (2014), 168–170

    work page 2014

  20. [20]

    Dorothy F Garrison-Wade and Chance W Lewis. 2004. Affirmative Action: History and Analysis. Journal of College Admission 184 (2004), 23–26

    work page 2004

  21. [21]

    Talia B Gillis. 2022. The input fallacy. Minnesota Law Review, forthcoming (2022)

    work page 2022

  22. [22]

    Usman Gohar, Zeyu Tang, Jialu Wang, Kun Zhang, Peter Spirtes, Yang Liu, and Lu Cheng. 2025. Long-Term Fairness Inquiries and Pursuits in Machine Learning: A Survey of Notions, Methods, and Challenges. Transactions on Machine Learning Research (2025)

    work page 2025

  23. [23]

    Moritz Hardt, Eric Price, Eric Price, and Nati Srebro. 2016. Equality of Opportunity in Supervised Learning. In Advances in Neural Information Processing Systems 30, Vol. 29. 3315–3323

    work page 2016

  24. [24]

    Kenneth Holstein, Jennifer Wortman Vaughan, Hal Daumé III, Miro Dudik, and Hanna Wallach. 2019. Improving fairness in machine learning systems: What do industry practitioners need?. In Proceedings of the 2019 CHI conference on human factors in computing Fairness under uncertainty in sequential decisions FAccT ’26, June 25–28, 2026, Montreal, QC, Canada sy...

    work page 2019

  25. [25]

    Yaowei Hu and Lu Zhang. 2022. Achieving Long-Term Fairness in Sequential Decision Making. Proceedings of the AAAI Conference on Artificial Intelligence 36 (2022), 9549–9557

    work page 2022

  26. [26]

    Wen Huang, Lu Zhang, and Xintao Wu. 2021. Achieving Counterfactual Fairness for Causal Bandit. InNeurIPS Workshopon Algorithmic Fairness through the Lens of Causality and Robustness, Vol. 34

    work page 2021

  27. [27]

    Eyke Hüllermeier and Willem Waegeman. 2021. Aleatoric and epistemic uncertainty in machine learning: An introduction to concepts and methods. Machine Learning 110, 3 (2021), 457–506

    work page 2021

  28. [28]

    J. D. Hunter. 2007. Matplotlib: A 2D graphics environment. Computing in Science & Engineering 9, 3 (2007), 90–95. doi:10.1109/MCSE. 2007.55

    work page doi:10.1109/mcse 2007

  29. [29]

    Shahin Jabbari, Matthew Joseph, Michael Kearns, Jamie Morgenstern, and Aaron Roth. 2017. Fairness in Reinforcement Learning. In Proceedings of the 34th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 70). 1617– 1626

    work page 2017

  30. [30]

    Matthew Joseph, Michael Kearns, Jamie H Morgenstern, and Aaron Roth. 2016. Fairness in Learning: Classic and Contextual Bandits. In Advances in Neural Information Processing Systems, Vol. 29

    work page 2016

  31. [31]

    Kirthevasan Kandasamy, Jeff Schneider, and Barnabás Póczos. 2015. High dimensional Bayesian optimisation and bandits via additive models. In International conference on machine learning. PMLR, 295–304

    work page 2015

  32. [32]

    Heena Khanna, Manik Mehra, and Jordao Fortunato Diogo. 2025. Unveiling the Impact: Societal Implications of Reinforcement Learning Algorithms. In Reinforcement Learning: Foundations and Applications. Bentham Science Publishers, 77–95

    work page 2025

  33. [33]

    Niki Kilbertus, Manuel Gomez Rodriguez, Bernhard Schölkopf, Krikamol Muandet, and Isabel Valera. 2020. Fair Decisions Despite Imperfect Predictions. In Proceedings of the 23rd International Conference on Artificial Intelligence and Statistics (Proceedings of Machine Learning Research, Vol. 108). 277–287

    work page 2020

  34. [34]

    Jon Kleinberg, Sendhil Mullainathan, and Manish Raghavan. 2017. Inherent Trade-Offs in the Fair Determination of Risk Scores. In 8th Innovations in Theoretical Computer Science Conference, Vol. 67. 43.1–43.23

    work page 2017

  35. [35]

    Selim Kuzucu, Jiaee Cheong, Hatice Gunes, and Sinan Kalkan. 2024. Uncertainty as a Fairness Measure. Journal of Artificial Intelligence Research 81 (2024), 307–335

    work page 2024

  36. [36]

    Himabindu Lakkaraju, Jon Kleinberg, Jure Leskovec, Jens Ludwig, and Sendhil Mullainathan. 2017. The Selective Labels Problem: Evaluating Algorithmic Predictions in the Presence of Unobservables. InProceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 275–284

    work page 2017

  37. [37]

    Michelle Seng Ah Lee, Luciano Floridi, and Jatinder Singh. 2021. Formalising trade-offs beyond algorithmic fairness: lessons from ethical philosophy and welfare economics. AI and Ethics 1, 4 (2021), 529–544

    work page 2021

  38. [38]

    Michelle Seng Ah Lee and Jat Singh. 2021. The landscape and gaps in open source fairness toolkits. In Proceedings of the 2021 CHI conference on human factors in computing systems. 1–13

    work page 2021

  39. [39]

    Michelle Seng Ah Lee and Jatinder Singh. 2021. Risk Identification Questionnaire for Detecting Unintended Bias in the Machine Learning Development Lifecycle. In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society. 704–714

    work page 2021

  40. [40]

    Liu, Sarah Dean, Esther Rolf, Max Simchowitz, and Moritz Hardt

    Lydia T. Liu, Sarah Dean, Esther Rolf, Max Simchowitz, and Moritz Hardt. 2019. Delayed Impact of Fair Machine Learning. InProceedings of the 28th International Joint Conference on Artificial Intelligence. 6196–6200

    work page 2019

  41. [41]

    David Loshin. 2010. The practitioner’sguide to data quality improvement. Elsevier

    work page 2010

  42. [42]

    Razieh Nabi, Daniel Malinsky, and Ilya Shpitser. 2019. Learning Optimal Fair Policies. InProceedings of the 36th International Conference on Machine Learning, Kamalika Chaudhuri and Ruslan Salakhutdinov (Eds.). 4674–4682

    work page 2019

  43. [43]

    Judea Pearl. 2009. Causality. Cambridge University Press, New York

    work page 2009

  44. [44]

    Bhagyashree Puranik, Upamanyu Madhow, and Ramtin Pedarsani. 2022. A Dynamic Decision-Making Framework Promoting Long-Term Fairness. In Proceedings of the 2022 AAAI/ACM Conference on AI, Ethics, and Society. 547–556

    work page 2022

  45. [45]

    Miriam Rateike, Isabel Valera, and Patrick Forré. 2024. Designing Long-term Group Fair Policies in Dynamical Systems. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency. 20–50

    work page 2024

  46. [46]

    Anka Reuel and Devin Ma. 2024. Fairness in reinforcement learning: A survey. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, Vol. 7. 1218–1230

    work page 2024

  47. [47]

    Yaniv Romano, Rina Foygel Barber, Chiara Sabatti, and Emmanuel Candès. 2020. With malice toward none: Assessing uncertainty via equalized coverage. Harvard Data Science Review 2, 2 (2020), 4

    work page 2020

  48. [48]

    Sutton and Andrew G

    Richard S. Sutton and Andrew G. Barto. 2018. Reinforcement Learning: An Introduction (second ed.). The MIT Press, Cambridge, MA

    work page 2018

  49. [49]

    Rui Tuo and Wenjia Wang. 2022. Uncertainty quantification for Bayesian optimization. In International Conference on Artificial Intelligence and Statistics. PMLR, 2862–2884

    work page 2022

  50. [50]

    Sandra Wachter, Brent Mittelstadt, and Chris Russell. 2021. Why fairness cannot be automated: Bridging the gap between EU non- discrimination law and AI. Computer Law & Security Review 41 (2021), 105567

    work page 2021

  51. [51]

    Piette, Joshua R

    Jitao Wang, Chengchun Shi, John D. Piette, Joshua R. Loftus, Donglin Zeng, and Zhenke Wu. 2025. Counterfactually Fair Reinforcement Learning via Sequential Data Preprocessing. arXiv preprint arXiv:2501.06366 (2025). FAccT ’26, June 25–28, 2026, Montreal, QC, Canada Lee, Padh, Watson, Kilbertus, and Singh

    work page arXiv 2025

  52. [52]

    Min Wen, Osbert Bastani, and Ufuk Topcu. 2021. Algorithms for Fairness in Sequential Decision Making. In Proceedings of The 24th International Conference on Artificial Intelligence and Statistics, Vol. 130. 1144–1152

    work page 2021

  53. [53]

    Jess Whittlestone, Kai Arulkumaran, and Matthew Crosby. 2021. The societal implications of deep reinforcement learning. Journal of Artificial Intelligence Research 70 (2021), 1003–1030

    work page 2021

  54. [54]

    Tongxin Yin, Reilly Raab, Mingyan Liu, and Yang Liu. 2023. Long-Term Fairness with Unknown Dynamics. In Advances in Neural Information Processing Systems, Vol. 36. 55110–55139

    work page 2023

  55. [55]

    Mert Yuksekgonul, Linjun Zhang, James Y Zou, and Carlos Guestrin. 2024. Beyond confidence: Reliable models should also consider atypicality. Advances in Neural Information Processing Systems 36 (2024)

    work page 2024

  56. [56]

    Muhammad Bilal Zafar, Isabel Valera, Manuel Gomez-Rodriguez, and Krishna P. Gummadi. 2019. Fairness Constraints: A Flexible Approach for Fair Classification. J. Mach. Learn. Res. 20, 75 (2019), 1–42

    work page 2019

  57. [57]

    Junzhe Zhang and Elias Bareinboim. 2018. Fairness in Decision-Making — The Causal Explanation Formula. Proceedings of the AAAI Conference on Artificial Intelligence 32, 1 (2018)

    work page 2018

  58. [58]

    Xueru Zhang and Mingyan Liu. 2021. Fairness in Learning-Based Sequential Decision Algorithms: A Survey. 525–555

    work page 2021

  59. [59]

    Zeyu Zhou, Tianci Liu, Ruqi Bai, Jing Gao, Murat Kocaoglu, and David I Inouye. 2024. Counterfactual fairness by combining factual and counterfactual predictions. Advances in Neural Information Processing Systems 37 (2024), 47876–47907. A Synthetic data The underlying phenomenon takes the following form: 𝑌=𝑓(𝑋) +𝜀 𝑋 obs =𝑔(𝑋) 𝑌 obs =ℎ(𝑌), where 𝑋 denotes t...

    work page 2024