Pith. sign in

REVIEW 5 major objections 4 minor 77 references

Simulating Eutopia: Revisiting Long-term Fairness with Outcomes, Performativity, and Dynamics

T0 review · 5 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read In a simulated credit market, lenders who model how applicants react to their decisions and optimize long-run group wealth achieve both higher profit and near-equal wealth growth, showing that fairness need not cost efficiency.

desk verdict A genuinely useful performative-fairness testbed whose headline claim about performative learning is not yet supported: PePG is handed the simulator's true Hawkes parameters, so the RL-vs-PERL comparison is an oracle-model learner against a model-free baseline, not a clean test. read the letter →

arxiv 2607.19389 v1 pith:VSZLKXUL submitted 2026-07-02 cs.CY cs.AIcs.LGstat.ML

Vedant Palit , Udvas Das , Brahim Driss , Debabrota Basu This is my paper · ORCID
classification cs.CYcs.AIcs.LGstat.ML
keywords long-termfairnessperformativereinforcementlearningcreditlendingwealthdynamicsalgorithmicsequentialdecisionmakingsimulationoutcome
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Most work on long-term fairness treats the population as fixed and measures bias at the moment of a single prediction. The paper argues this misses what matters in credit lending: approving or denying a loan changes borrowers' wealth, which changes who applies next, and this feedback loop determines long-run equity. To study that loop, the authors build Eutopia, a performative lending simulator, together with a data generator that supplies realistic applicant populations, and compare passive and performative reinforcement learners across a suite of fairness utilities. Their central results are (a) learning under performative dynamics outperforms learning that ignores them, and (b) fairness objectives evaluated on social outcomes — in particular a Fairness Lagrangian on group mean wealth — deliver better profit, welfare, and near-parity wealth growth. They also find that no utility wins on every metric, so the choice of fairness objective is a policy decision rather than a purely technical one.

What carries the argument

The central object is a performative Markov Decision Process (PeMDP) for credit lending: the transition and reward functions depend on the deployed policy because loan approvals alter borrower wealth, which in turn drives future application rates through a self-exciting arrival process whose base rate is an increasing linear function of group mean wealth and an individual self-selection term λ=0.5. The authors implement this in the Eutopia simulator with a continual (never-reset) environment and a data generator that builds the initial population and fixed financial types. Two learning mechanisms carry the comparison: a standard policy-gradient lender (RL) that ignores the feedback, and a pe

What would settle it

Run the same RL-versus-PERL comparison in Eutopia with performativity switched off (make application arrivals independent of group mean wealth and history, λ=0) while keeping everything else fixed; if PERL no longer beats RL on wealth gap, social welfare, or profit, the benefit is an artifact of the chosen feedback loop. More convincingly, estimate the application-rate response and wealth-gain parameters from a real lender's portfolio and re-run; if the outcome-fairness advantage vanishes under those parameters, the practical prescription fails.

Watch

Extended reading notes

Core claim

Using Eutopia, a simulator in which application arrivals follow a self-exciting process whose base rate depends on group mean wealth, and in which approved loans that are repaid add 20% of the loan to borrower wealth, the authors train a passive policy-gradient lender (RL) and a performative policy-gradient lender (PERL) under four utility families crossed with three fairness perspectives (decision-maker profit parity, outcome wealth parity, and a two-sided blend). Their strongest numerical claim is that the performative agent trained with an outcome-based Fairness Lagrangian—maximizing total group wealth minus a penalty on the inter-group gap—achieves the highest cumulative profit (174.408

Load-bearing premise

The paper's conclusions depend on Eutopia's hand-set behavioral response—application rates driven by an increasing linear function of group mean wealth, self-selection strength λ=0.5, and a 20% wealth gain per repaid loan—being representative of how real borrowers actually respond to lending policy; if the real performative response is weaker, stronger, or shaped differently, the claimed benefits of performative modeling and outcome-based rewards could disappear.

Editorial extensions

If this is right

  • If the simulator reflects real dynamics, fairness evaluations that ignore performative feedback understate long-run inequity; a lender that optimizes only short-term predictions can amplify wealth gaps.
  • Outcome-based fairness objectives like the Fairness Lagrangian can raise long-run profit and social welfare simultaneously, so equity need not come at the expense of profit.
  • The choice of fairness objective interacts with the learning algorithm: PERL favors outcome fairness, RL favors two-sided fairness, so the two should be chosen jointly rather than independently.
  • Decision-maker fairness that equalizes per-group profit can still allow large wealth divergence and exclusion of the disadvantaged group, so profit-parity constraints are not a substitute for outcome fairness.
  • No single utility dominates on all metrics (profit, wealth gap, welfare, inequality ratio), meaning regulatory and policy choices must weigh which dimension to prioritize.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's main practical message—design rewards around downstream group wealth rather than per-decision metrics—should generalize beyond lending to hiring, insurance, or any ADM whose decisions change the pool of future candidates; the mechanism is the feedback loop, not the credit domain.
  • A testable extension would vary the self-selection strength λ and the wealth-gain parameter κ across a grid to see whether PERL's superiority over RL holds when applicant response is weaker or returns to wealth are smaller.
  • Because loans concentrate among repeat recipients under every configuration, the paper implies that aggregate group-level metrics are insufficient as regulatory targets; reach-rate or Lorenz-type within-group measures should be monitored.
  • A natural next step is to re-run the same RL versus PERL comparison on real loan-portfolio data with estimated performativity parameters, which would test whether the conclusions transfer beyond the simulator.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 4 minor

Summary. The paper formalizes credit lending as a performative MDP, builds Eutopia (a no-reset Hawkes-process loan simulator) and LOANLY (an Adult-based performative data generator), and compares a model-free policy gradient (RL) with a performative policy gradient (PERL) under three fairness perspectives and four reward functions. The two headline claims are that performative learning improves long-term efficiency/equity and that outcome-based fairness utilities improve efficiency, equity, and inclusivity. The headline results (Tables 2–3) show, for example, PERL with outcome-FL reaching cumulative profit 174.408±74.473 and social welfare 0.029±0.010.

Significance. The contributions are potentially useful: a public simulator with continual performative feedback, the LOANLY pipeline, heuristic sanity checks, and seed-variance reporting. The paper is honest that the simulator is not meant to reproduce the real world. However, the central comparison supporting claim (a) is not controlled: PERL is given the exact Hawkes parameters of the environment while RL is not, and the implementations differ in additional ways. As a result, the paper's main empirical conclusion is currently not established. The issues are addressable by ablations and a cleaner experimental protocol.

major comments (5)
  1. [§4.2; Appendix D.3.2, Table 6] The RL-vs-PERL comparison is confounded by oracle access. PePG's gradient includes a transition term with 'Hawkes excitation (matched to env)' and 'Hawkes decay (matched to env)' (α_R=α_B=0.3, β_R=β_B=2.0). The PG baseline is model-free and does not receive these parameters. Therefore the profit/welfare advantage in Tables 2–3 under outcome fairness (e.g., 174.408 vs 138.420) may reflect knowledge of the data-generating process, not the benefit of modeling performativity. In addition, PePG uses a replay buffer and disables entropy bonus, while PG does not. A controlled comparison should include (i) PePG with estimated or misspecified Hawkes/wealth parameters, (ii) PG augmented with the same oracle transition model, and (iii) ablations matching hyperparameters. Without these, claim (a) is not supported.
  2. [§2, Algorithm 1, Table 5] The formalization adopts the average-reward criterion ('V̄_π = lim_{T→∞} ...'), but the implementation is REINFORCE with discount γ=0.99 and per-episode return normalization. In a continuing task with no reset, per-episode discounted returns do not approximate the average-reward objective, and the algorithm is not optimizing the stated criterion. The paper should either implement an average-reward estimator (e.g., differential return) or revise the formalization to discounted return.
  3. [§3.3 vs Appendix E.1] The inequality ratio is defined inconsistently. Section 3.3 defines ρ(t) = (μ_M(t)-μ_M(0))/(μ_F(t)-μ_F(0)), while Appendix E.1 states 'ρ(t) = μ_M/μ_F'. These differ unless initial wealths are equal, and the ratio-of-gains version is undefined for the Always-Reject policy (0/0). Since ρ≈1 is used as evidence of parity in the main text and tables, the definition and implementation must be clarified and applied uniformly.
  4. [§3.1, Appendix D.2, Algorithm 1] The default mechanism is inconsistent. Section 3.1 and Appendix D.2 say default outcomes are pre-drawn per individual and fixed across episodes, but Algorithm 1 (line 28) draws δ_i ∼ Bernoulli(p_def) at each approval. Fixed versus redrawn defaults change the latent risk structure and can alter policy comparisons. The code and algorithm should match.
  5. [Tables 2–3, Appendix B, Appendix F] The headline comparisons lack statistical support. Standard deviations overlap widely (e.g., profit 174.408±74.473 vs 138.420±86.722; welfare 0.029±0.010 vs 0.023±0.013), and no significance tests, paired differences, or effect sizes are reported. The seed count is also inconsistent: Appendix B says 5 seeds, Appendix F says 10. Please report per-seed paired analysis and resolve the seed-count discrepancy.
minor comments (4)
  1. [Algorithm 1, line 16] 'λ_g(t) ← toward f(μ_g) + ...' contains the stray word 'toward'; presumably it should be 'f(μ_g)'.
  2. [Section 6] The paper correctly disclaims real-world mimicry, but the abstract's wording 'lead to better long-term efficiency and equity' should be qualified as 'in Eutopia' to avoid over-generalization.
  3. [Appendix B] The reactivity model (Eqs. 1–3) is not connected to the main results; consider moving it to a separate technical note or explicitly linking it to the PerPG update.
  4. [Section 3.2] LOANLY is described as a 'performative data generator', but the paper only validates the simulated environment, not the generator's realism. Please clarify that this is a synthetic benchmark generator.

Circularity Check

2 steps flagged · score 7.0 of 10

PePG is handed Eutopia's true Hawkes parameters ('matched to env'), so the RL-vs-PERL comparison conflates performative modeling with oracle model access; outcome-fairness rewards are also the evaluation metrics.

  1. fitted input called prediction [Appendix D.3.2, Table 6; Section 3.1 (Hawkes process)]
    "αR = αB 0.3 Hawkes excitation (matched to env); βR = βB 2.0 Hawkes decay (matched to env) ... λg(t) = f(µg) + αg Σ e^{-βg(t-tj)} ... Parameters are set symmetrically across groups (αg=0.3, βg=2.0) ... The gradient estimate augments the standard policy gradient with ... a transition gradient capturing how the policy shapes future Hawkes intensities and group mean wealth."

    PePG's transition-gradient term is evaluated with the simulator's true Hawkes constants, explicitly labeled 'matched to env', while the RL baseline is model-free and never receives them. PERL is therefore a model-based agent with oracle access to the exact performative dynamics, not merely an agent that 'accounts for performativity'. The headline comparison (Tables 2–3) thus measures the value of being given the data-generating process, not the value of performative modeling per se; the claimed advantage of 'learning with performative dynamics' is built into the setup by construction.

  2. self definitional [Section 3.4 / Algorithm 3 / Section 3.3]
    "Outcome fairness asks whether the long-run wealth distribution is equitable across groups, captured by µg^t ... Social Welfare: rt ← µR_t + µB_t ... Social Welfare interpreted as the long-term welfare ... R_g(t) := lim (µg(t)−µg(0))/t."

    Under outcome fairness the reward is exactly the group mean wealth statistics used to define the evaluation metrics (wealth gap, social welfare, inequality ratio); Algorithm 3 sets rt = µR+µB (or min/penalized versions). Hence the observation that outcome-based FL/RMM improve equity is guaranteed by the reward definition — the agent is trained on the same quantities on which it is scored. Claim (b)'s equity component is thus self-definitional; only the profit comparison is an independent empirical finding.

full rationale

The formal PeMDP formulation and the Eutopia/LOANLY simulator construction are self-contained and not circular: the environment dynamics are explicitly specified and the heuristic sanity checks are sensible. The circularity is concentrated in the empirical support for the headline claims. First, claim (a) rests on comparing PERL to RL, but PePG is parameterized with the simulator's true Hawkes excitation and decay constants, explicitly 'matched to env', and its transition-gradient term uses those constants to model performative shift. This gives PERL oracle access to the data-generating process while RL is model-free; the reported PERL advantage is therefore forced by construction and does not demonstrate a general benefit of modeling performativity. Second, claim (b)'s equity component is partly self-definitional: outcome-fairness rewards are defined directly as group mean wealth, the same quantities used for the equity metrics, so improving those metrics is a direct consequence of the reward choice. The profit and efficiency comparisons retain some independent content, and the paper honestly reports that no single objective dominates, but the two central 'predictions' in the abstract are substantially built into the experimental design. A modest self-citation to the authors' own PePG work (Basu et al., 2025) is present but not the primary source of circularity. Score 7 reflects that the main claims are partially reduced to the inputs rather than fully independent discoveries.

Assumptions & free parameters 11 free parameters · 7 assumptions · 0 invented entities

The central claim rests on a chain of hand-set simulation parameters (Hawkes excitation/decay, self-selection strength, wealth mapping, default regimes) and domain assumptions about proxy data and loan wealth effects. No new physical entities are introduced; the simulator and data generator are software artifacts.

free parameters (11)
  • Initial wealth mapping w_i = 10+190*sc_i = 10, 190
    Maps Adult-derived credit score to wealth in $K; no empirical asset calibration; sets initial group inequality that all fairness metrics measure.
  • Performativity strength λ = 0.5
    Controls how strongly perceived approval probability changes application propensity; hand-chosen; central to the performative feedback being tested.
  • Hawkes excitation α_g = 0.3
    Self-exciting arrival strength, set symmetrically across groups; also handed to PePG as ground truth (Table 6), confounding the RL/PERL comparison.
  • Hawkes decay β_g = 2.0
    Sets excitation window ≈3.5 time units; chosen arbitrarily; also supplied to PePG.
  • Base application rate f(μ)=max(0.5, 2+0.05μ) = max(0.5, 2+0.05μ)
    Encodes Matthew effect: wealthier groups apply more; coefficients hand-picked and drive self-reinforcing inequality.
  • Loan gain fraction κ_i/l_i = 0.20 (τ_inv=0.35, τ_interest=0.15)
    Assumes loans are wealth-generating on average; without it the profit-welfare tension and wealth divergence mechanism disappear.
  • Default rate regimes = [0.02,0.15] pre-train; [0.14,0.16] eval
    Chosen to mimic subprime conditions; induces distribution shift between training and evaluation.
  • Ground-truth label threshold τ = 0.5
    Converts income score into repayment label; fairness reference hidden from agent.
  • Population size per group N_g = 3000
    Chosen for stable statistics and compute budget; affects variance of all metrics.
  • Loan amount distribution = N(30,10^2) clipped to [10,100]
    Hand-set financial type; scales profit and wealth units.
  • Wealth score logistic regression = coefficients fit on Adult data
    Learned from UCI Adult income; provides external data grounding but also encodes the gender/wealth disparity used as the fairness target.
assumptions (7)
  • domain assumption UCI Adult income and gender are a valid proxy for real credit-market wealth disparities
    LOANLY initializes wealth from Adult income predictions; the fairness conclusions are grounded in this proxy (Section 3.2, D.1).
  • domain assumption Approved loans increase borrower wealth on average (τ_inv > τ_interest)
    Wealth transition κ=0.20l on repayment makes lending wealth-generating; central to the simulated equity-profit dynamics (Section 3.1, D.2).
  • domain assumption Application propensity follows the specified logistic-Hawkes response
    p_apply = clip(λ_g Δt/N_g · (1+λσ(hθ)),0,1) models rational self-selection; this is the performative mechanism the paper tests (Section 3.1).
  • domain assumption Hawkes base rate tied to group mean wealth captures the Matthew effect
    f(μ)=max(0.5,2+0.05μ) couples lending to future demand; coefficients chosen by hand (Section 3.1).
  • standard math PePG convergence guarantees hold (Basu et al. 2025)
    The performative policy gradient's near-optimality is imported from the authors' prior work; assumptions not restated (Appendix D.3.2).
  • domain assumption Selective-label censoring does not invalidate the policy gradient estimates
    Denied applicants' repayment outcomes are unobserved (Section 2), but the training reward uses only approved applicants with no correction for censoring; the paper does not analyze this bias.
  • domain assumption Average-reward theory is compatible with the discounted PG implementation
    Section 2 motivates average-reward, but Tables 5-6 use γ=0.99 discounting; no bridge is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Simulating Eutopia: Revisiting Long-term Fairness with Outcomes, Performativity, and Dynamics." pith.science (2026). https://pith.science/paper/VSZLKXUL

@misc{pith2026260719389,
  author       = {Pith},
  title        = {Pith review of: Simulating Eutopia: Revisiting Long-term Fairness with Outcomes, Performativity, and Dynamics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VSZLKXUL}},
  note         = {Machine review of arXiv:2607.19389}
}
read the original abstract

As AI-driven Decision Makers (ADMs) influence our socioeconomic reality, their roles in both enhancing efficiency and amplifying the social biases have drawn attention. In this paper, we revisit the nuances of long-term `fairness' achievable by an ADM, specifically in the context of a credit lending induced wealth process. The literature on long-term fairness mostly (a) considers passive environments, i.e. the outcome of a predictor does not change the population's behaviour, and (b) measures bias in terms of disparity in instantaneous predictions rather than the downstream equity. These are not true for modern ADMs, like credit lenders. To address these caveats, we first formalise the wealth dynamics induced by a loan approving ADM interacting with a multi-demographic population as a performative Markov Decision Process with ADM level and social outcome level reward functions. Then, we mitigate the absence of such a performative test-bed by developing Eutopia: a lending-process simulator enabled with a novel performative data generator to learn long-term fair strategies. Finally, we test performative and classical RL algorithms with different fairness-aware and utilitarian utilities. Experimental results show that (a) learning with performative dynamics lead to better long-term efficiency and equity, and (b) learning with well-designed fairness-aware utility evaluated on social outcomes induces better efficiency, equity, and inclusivity.

Figures

Figures reproduced from arXiv: 2607.19389 by the authors.

Figure 1
Figure 1. Performative feedback loop of ADM [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Learning vs. Heuristics. Learned agents adapt to the performative feedback loop and maintain equitable wealth trajectories over time, while heuristic policies accumulate profit immedi￾ately but produce growing wealth disparities they cannot correct. Refer [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Lorenz curves under DM fairness (PERL). Red: male group, Blue: female group. Curves shown at episodes 100, 400, 800, and 1000 (light to dark). Dashed line is perfect equality. 0 200 400 600 800 1000 Episode 0.05 0.10 0.15 0.20 0.25 Reach rate (unique recipients / N) (a) RMM 0 200 400 600 800 1000 Episode 0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.45 Reach rate (unique recipients / N) (b) FL 0 200 400 600 800 1000 Episode … view at source ↗
Figures from the paper (15 more)
Figure 5
Figure 5. Figure 5: Reach rate under DM’s (a-c) and outcome fairness (d-f) (PERL). Red: male group, blue: female group. Shaded regions show standard deviation across seeds. 5 Measuring Actual Outcome of Bias Matters Metrics such as wealth gap and inequality ratio measure fairness between …
Figure 6
Figure 6. Figure 6: Episodic reactivity: ∥bt∥ over T = 1000 deploy episodes, one panel per reward family. Each panel overlays the available fairness objectives. Mean ± std across seeds. 0 200 400 600 800 1000 Episode (T) 1 0 1 2 3 4 s t / s 1e 3 (a) Utilitarian Profit (UP) 0 200 400 600 8…
Figure 7
Figure 7. Figure 7: Cumulative reactivity:Running sum P s≤t ∥bs∥ of reactivity over T = 1000 deploy episodes, one panel per reward family. Each panel overlays the available fairness objectives. Mean ± std across seeds. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: Evaluating Eutopia with Heuristic Algorithms Reverse Rich Becomes Richer produces the mirror pattern, with ρ falling sharply below one before recovering slightly. The near-symmetry between the two trajectories is informative: it confirms that the Hawkes process paramet…
Figure 9
Figure 9. Figure 9: Learning vs. Heuristics. Expanded results from [PITH_FULL_IMAGE:figures/full_fig_p025_9.png]
Figure 10
Figure 10. Figure 10: PeRL: Outcome Fairness. Long-term performance of PERL across Fairness Objectives under outcome fairness, tracked over T = 1000 episodes. Mean ± std across seeds. any RL configuration, and the best wealth gap (29.559 ± 66.819), though the high variance on the wealth ga…
Figure 11
Figure 11. Figure 11: RL: Outcome Fairness. Long-term performance of RL across Fairness Objective under outcome fairness, tracked over T = 1000 episodes. Mean ± std across seeds. gap metric are now measuring related quantities, the dynamics here are more directly interpretable than under D…
Figure 12
Figure 12. Figure 12: PeRL: DM’s Fairness. Long-term performance of PERL across Fairness Objectives under DM’s fairness, tracked over T = 1000 episodes. Mean ± std across seeds. that the performative gradient interacts most productively with the Lagrangian reward structure and provides the…
Figure 13
Figure 13. Figure 13: RL: DM’s Fairness. Long-term performance of RL across Fairness Objectives under DM’s fairness, tracked over T = 1000 episodes. Mean ± std across seeds. Inequality ratio. Under RL (Figure 13b), all fairness objectives begin near ρ ≈ 1.25–1.3. RMM spikes sharply in the …
Figure 14
Figure 14. Figure 14: PeRL: Two-Sided Fairness. Long-term performance of PERL across fairness objectives under Two-Sided fairness, tracked over T = 1000 episodes. Mean ± std across seeds. reward becomes welfare-enhancing when the agent models the performative consequences of its decisions,…
Figure 15
Figure 15. Figure 15: RL: Two-Sided Fairness. Long-term performance of RL across fairness objectives under Two-Sided fairness, tracked over T = 1000 episodes. Mean ± std across seeds. ping below one, indicating that some seeds over-correct and temporarily favour the initially weaker group.…
Figure 16
Figure 16. Figure 16: PERL: Radar Plots ity measured as proximity to ρ = 1, i.e. |ρ − 1| inverted so that a larger value indicates a more equal outcome. The axes are normalised to a common visual scale to enable shape comparison across constraints. A larger and more symmetric polygon indic…
Figure 17
Figure 17. Figure 17: Lorenz curves under outcome fairness (PERL). Red: male group, Blue: female group. Curves shown at episodes 100, 400, 800, and 1000 (light to dark). Dashed line is perfect equality. 0 500 1000 1500 2000 2500 3000 Individual index 0 50000 100000 150000 200000 250000 300…
Figure 18
Figure 18. Figure 18: Cumulative loan distributions under DM fairness (PERL). Each bar shows total loans received by an individual over the full deployment horizon, sorted by wealth index. Top row: male group. Bottom row: female group. Outcome fairness. The inter-group separation is genera…
Figure 19
Figure 19. Figure 19: Cumulative loan distributions under outcome fairness (PERL). Each bar shows total loans received by an individual over the full deployment horizon, sorted by wealth index. Top row: male group. Bottom row: female group. A universal pattern holds across all reach rate p…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

77 extracted references · 5 linked inside Pith

  1. [1]

    Frontiers in Artificial Intelligence , volume=

    On consequentialism and fairness , author=. Frontiers in Artificial Intelligence , volume=. 2020 , publisher=

    2020

  2. [2]

    1998 , publisher=

    Reinforcement learning: An introduction , author=. 1998 , publisher=

    1998

  3. [3]

    Nature Communications , year=

    Connecting algorithmic fairness and fair outcomes in a sociotechnical simulation case study of AI-assisted healthcare , author=. Nature Communications , year=

  4. [4]

    Proceedings of the 2018 World Wide Web Conference , pages=

    A short-term intervention for long-term fairness in the labor market , author=. Proceedings of the 2018 World Wide Web Conference , pages=

    2018

  5. [5]

    Advances in Neural Information Processing Systems , volume=

    Long-term fairness with unknown dynamics , author=. Advances in Neural Information Processing Systems , volume=

  6. [6]

    arXiv preprint arXiv:2210.06732 , year=

    Equal improvability: A new fairness notion considering the long-term impact , author=. arXiv preprint arXiv:2210.06732 , year=

    arXiv

  7. [7]

    Reproducibility Study: Equal Improvability: A New Fairness Notion Considering the Long-Term Impact , author=

  8. [8]

    Designing Long-term Group Fair Policies in Dynamical Systems , year =

    Rateike, Miriam and Valera, Isabel and Forr\'. Designing Long-term Group Fair Policies in Dynamical Systems , year =. doi:10.1145/3630106.3658538 , booktitle =

    arXiv

Show all 77 references
  1. [9]

    2023 , isbn =

    Estornell, Andrew and Das, Sanmay and Liu, Yang and Vorobeychik, Yevgeniy , title =. 2023 , isbn =. doi:10.1145/3593013.3594006 , booktitle =

    2023

  2. [10]

    2024 , isbn =

    Zhang, Aurora and Hosoi, Anette , title =. 2024 , isbn =. doi:10.1145/3630106.3658952 , booktitle =

    2024

  3. [11]

    2025 , eprint=

    Performativity and Prospective Fairness , author=. 2025 , eprint=

    2025

  4. [12]

    Performative Debias with Fair-exposure Optimization Driven by Strategic Agents in Recommender Systems , url=

    Xiang, Zhichen and Zhao, Hongke and Zhao, Chuang and He, Ming and Fan, Jianping , year=. Performative Debias with Fair-exposure Optimization Driven by Strategic Agents in Recommender Systems , url=. doi:10.1145/3637528.3671786 , booktitle=

  5. [13]

    Proceedings of the 3rd innovations in theoretical computer science conference , pages=

    Fairness through awareness , author=. Proceedings of the 3rd innovations in theoretical computer science conference , pages=

  6. [14]

    Advances in neural information processing systems , volume=

    Equality of opportunity in supervised learning , author=. Advances in neural information processing systems , volume=

  7. [15]

    Big data , volume=

    Fair prediction with disparate impact: A study of bias in recidivism prediction instruments , author=. Big data , volume=. 2017 , publisher=

    2017

  8. [16]

    arXiv preprint arXiv:1609.05807 , year=

    Inherent trade-offs in the fair determination of risk scores , author=. arXiv preprint arXiv:1609.05807 , year=

    arXiv

  9. [17]

    Proceedings of the 23rd acm sigkdd international conference on knowledge discovery and data mining , pages=

    Algorithmic decision making and the cost of fairness , author=. Proceedings of the 23rd acm sigkdd international conference on knowledge discovery and data mining , pages=

  10. [18]

    ACM Computing Surveys (CSUR) , volume=

    A review on fairness in machine learning , author=. ACM Computing Surveys (CSUR) , volume=. 2022 , publisher=

    2022

  11. [19]

    ACM computing surveys (CSUR) , volume=

    A survey on bias and fairness in machine learning , author=. ACM computing surveys (CSUR) , volume=. 2021 , publisher=

    2021

  12. [20]

    Recommender systems handbook , volume=

    Fairness and machine learning , author=. Recommender systems handbook , volume=. 2020 , publisher=

    2020

  13. [21]

    Advances in neural information processing systems , volume=

    Fairness in learning: Classic and contextual bandits , author=. Advances in neural information processing systems , volume=

  14. [22]

    Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency , pages=

    Fairness through robustness: Investigating robustness disparity in deep learning , author=. Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency , pages=

    2021

  15. [23]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Achieving long-term fairness in sequential decision making , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  16. [24]

    Proceedings of the 41st International Conference on Machine Learning , articleno =

    Xu, Yuancheng and Deng, Chenghao and Sun, Yanchao and Zheng, Ruijie and Wang, Xiyao and Zhao, Jieyu and Huang, Furong , title =. Proceedings of the 41st International Conference on Machine Learning , articleno =. 2024 , publisher =

    2024

  17. [25]

    International conference on machine learning , pages=

    Fairness in reinforcement learning , author=. International conference on machine learning , pages=. 2017 , organization=

    2017

  18. [26]

    arXiv preprint arXiv:2211.03994 , year=

    Reinforcement learning with stepwise fairness constraints , author=. arXiv preprint arXiv:2211.03994 , year=

    arXiv

  19. [27]

    The Twelfth International Conference on Learning Representations , year=

    Achieving fairness in multi-agent MDP using reinforcement learning , author=. The Twelfth International Conference on Learning Representations , year=

  20. [28]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Online restless multi-armed bandits with long-term fairness constraints , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  21. [29]

    International Conference on Machine Learning , pages=

    Performative prediction , author=. International Conference on Machine Learning , pages=. 2020 , organization=

    2020

  22. [30]

    International Conference on Machine Learning , pages=

    Strategic classification is causal modeling in disguise , author=. International Conference on Machine Learning , pages=. 2020 , organization=

    2020

  23. [31]

    2023 , eprint=

    Performative Reinforcement Learning , author=. 2023 , eprint=

    2023

  24. [32]

    2024 , eprint=

    Performative Reinforcement Learning in Gradually Shifting Environments , author=. 2024 , eprint=

    2024

  25. [33]

    2025 , eprint=

    Performative Policy Gradient: Optimality in Performative Reinforcement Learning , author=. 2025 , eprint=

    2025

  26. [34]

    International Conference on Machine Learning , pages=

    How to learn when data reacts to your model: performative gradient descent , author=. International Conference on Machine Learning , pages=. 2021 , organization=

    2021

  27. [35]

    International Conference on Artificial Intelligence and Statistics , pages=

    How to learn when data gradually reacts to your model , author=. International Conference on Artificial Intelligence and Statistics , pages=. 2022 , organization=

    2022

  28. [36]

    International Conference on Machine Learning , pages=

    Outside the echo chamber: Optimizing the performative risk , author=. International Conference on Machine Learning , pages=. 2021 , organization=

    2021

  29. [37]

    International Conference on Artificial Intelligence and Statistics , pages=

    State dependent performative prediction with stochastic approximation , author=. International Conference on Artificial Intelligence and Statistics , pages=. 2022 , organization=

    2022

  30. [38]

    Journal of Machine Learning Research , volume=

    Multiplayer performative prediction: Learning in decision-dependent games , author=. Journal of Machine Learning Research , volume=

  31. [39]

    Proceedings of the 24th ACM Conference on Economics and Computation , pages=

    Multi-agent performative prediction: From global stability and optimality to chaos , author=. Proceedings of the 24th ACM Conference on Economics and Computation , pages=

  32. [40]

    arXiv preprint arXiv:2408.05146 , year=

    Performative Prediction on Games and Mechanism Design , author=. arXiv preprint arXiv:2408.05146 , year=

    arXiv

  33. [41]

    Advances in Neural Information Processing Systems , volume=

    Stochastic optimization for performative prediction , author=. Advances in Neural Information Processing Systems , volume=

  34. [42]

    International Conference on Machine Learning , pages=

    Delayed impact of fair machine learning , author=. International Conference on Machine Learning , pages=. 2018 , organization=

    2018

  35. [43]

    Proceedings of the Conference on Fairness, Accountability, and Transparency , pages=

    From fair decision making to social equality , author=. Proceedings of the Conference on Fairness, Accountability, and Transparency , pages=

  36. [44]

    arXiv preprint arXiv:2503.00632 , year=

    Policy Design in Long-Run Welfare Dynamics , author=. arXiv preprint arXiv:2503.00632 , year=

  37. [45]

    Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency , pages=

    Fairness feedback loops: training on synthetic data amplifies bias , author=. Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency , pages=

    2024

  38. [46]

    , title =

    Hawkes, Alan G. , title =. Biometrika , volume =. 1971 , month =. doi:10.1093/biomet/58.1.83 , url =

    1971 doi

  39. [47]

    Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence,

    Delayed Impact of Fair Machine Learning , author =. Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence,. 2019 , month =. doi:10.24963/ijcai.2019/862 , url =

    2019 doi

  40. [48]

    A Theory of Justice: Original Edition , urldate =

    John Rawls , publisher =. A Theory of Justice: Original Edition , urldate =

  41. [49]

    2023 , publisher=

    Fairness and machine learning: Limitations and opportunities , author=. 2023 , publisher=

    2023

  42. [50]

    Journal of Machine Learning Research , volume=

    The measure and mismeasure of fairness , author=. Journal of Machine Learning Research , volume=

  43. [51]

    The Journal of Finance , volume=

    Predictably unequal? The effects of machine learning on credit markets , author=. The Journal of Finance , volume=. 2022 , publisher=

    2022

  44. [52]

    Proceedings of the 2020 conference on fairness, accountability, and transparency , pages=

    Mitigating bias in algorithmic hiring: Evaluating claims and practices , author=. Proceedings of the 2020 conference on fairness, accountability, and transparency , pages=

    2020

  45. [53]

    Science , volume=

    Dissecting racial bias in an algorithm used to manage the health of populations , author=. Science , volume=. 2019 , publisher=

    2019

  46. [54]

    Ethics of data and analytics , pages=

    Machine bias , author=. Ethics of data and analytics , pages=. 2022 , publisher=

    2022

  47. [55]

    Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency , pages=

    Fairness is not static: deeper understanding of long term fairness via simulation studies , author=. Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency , pages=

    2020

  48. [56]

    International Conference on Machine Learning , pages=

    Fairness without demographics in repeated loss minimization , author=. International Conference on Machine Learning , pages=. 2018 , organization=

    2018

  49. [57]

    1996 , howpublished =

    Becker, Barry and Kohavi, Ronny , title =. 1996 , howpublished =

    1996

  50. [58]

    2017 , publisher=

    The color of law: A forgotten history of how our government segregated America , author=. 2017 , publisher=

    2017

  51. [59]

    American economic review , volume=

    Consumer spending during unemployment: Positive and normative implications , author=. American economic review , volume=. 2019 , publisher=

    2019

  52. [60]

    international conference on machine learning , pages=

    Recommendations as treatments: Debiasing learning and evaluation , author=. international conference on machine learning , pages=. 2016 , organization=

    2016

  53. [61]

    Proceedings of the tenth ACM international conference on web search and data mining , pages=

    Unbiased learning-to-rank with biased feedback , author=. Proceedings of the tenth ACM international conference on web search and data mining , pages=

  54. [62]

    Journal of Financial Economics , volume=

    Consumer-lending discrimination in the FinTech era , author=. Journal of Financial Economics , volume=. 2022 , publisher=

    2022

  55. [63]

    Advances in neural information processing systems , volume=

    A fair classifier using kernel density estimation , author=. Advances in neural information processing systems , volume=

  56. [64]

    Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society , pages=

    Degenerate feedback loops in recommender systems , author=. Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society , pages=

    2019

  57. [65]

    International Conference on Machine Learning , pages=

    Performative reinforcement learning , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  58. [66]

    arXiv preprint arXiv:2402.09838 , year=

    Performative reinforcement learning in gradually shifting environments , author=. arXiv preprint arXiv:2402.09838 , year=

    arXiv

  59. [67]

    International Conference on Machine Learning , pages=

    Residual unfairness in fair machine learning from prejudiced data , author=. International Conference on Machine Learning , pages=. 2018 , organization=

    2018

  60. [68]

    Advances in Neural Information Processing Systems , volume=

    Performative power , author=. Advances in Neural Information Processing Systems , volume=

  61. [69]

    Advances in neural information processing systems , volume=

    Retiring adult: New datasets for fair machine learning , author=. Advances in neural information processing systems , volume=

  62. [70]

    , title=

    Williams, Ronald J. , title=. Machine Learning , year=. doi:10.1007/BF00992696 , url=

    doi

  63. [71]

    Hawkes , journal =

    Alan G. Hawkes , journal =. Spectra of Some Self-Exciting and Mutually Exciting Point Processes , urldate =

  64. [72]

    The Review of Financial Studies , volume =

    Demyanyk, Yuliya and Van Hemert, Otto , title =. The Review of Financial Studies , volume =. 2011 , month =. doi:10.1093/rfs/hhp033 , url =

    2011 doi

  65. [73]

    1955 , journal=

    Cardinal Welfare, Individualistic Ethics, and Interpersonal Comparisons of Utility , author=. 1955 , journal=

    1955

  66. [74]

    1789 , journal=

    An Introduction to the Principles of Morals and Legislation , author=. 1789 , journal=

  67. [75]

    Proceedings of the 2nd ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization , pages=

    On meritocracy in optimal set selection , author=. Proceedings of the 2nd ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization , pages=

  68. [76]

    nature , volume=

    Human-level control through deep reinforcement learning , author=. nature , volume=. 2015 , publisher=

    2015

  69. [77]

    The reward and communication systems of science are considered

    The Matthew effect in science. The reward and communication systems of science are considered. , author=. 1968 , journal=

    1968

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.