Pith. sign in

REVIEW 3 major objections 6 minor 300 references

When recourse recommendations move through a causal graph, stable models exist and beat naive training once intervention strength is not too large.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-31 05:29 UTC pith:MSLJBQWR

load-bearing objection Clean causal-performative packaging of recourse with solid math, but the title claim that causality specifically beats non-causal performative baselines is not isolated in the experiments. the 3 major comments →

arxiv 2607.28497 v1 pith:MSLJBQWR submitted 2026-07-30 cs.LG cs.CYcs.GT

The Role of Causality in Algorithmic Recourse

Srikanth Avasarala , Varun Gupta , Shahin Jabbari , Saber Salehkaleybar , Juba Ziani This is my paper
classification cs.LG cs.CYcs.GT
keywords algorithmic recourseperformative predictionstructural causal modelsstrategic classificationcausal interventionsstable predictorscredit scoring
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Standard algorithmic recourse tells people how to change their features so a fixed model flips their score, without asking whether those changes improve the true outcome or only game the predictor. Once people follow the advice, the data distribution shifts, the model becomes stale, and the advice itself can stop working after retraining. This paper treats that feedback loop as a performative prediction problem inside a structural causal model: recommended actions propagate through feature dependencies and also move the true label. Under quadratic action costs and ordinary strong-convexity and smoothness conditions on the loss, performatively stable linear predictors exist and can be reached by simple repeated risk minimization or repeated gradient descent whenever the intervention strength stays below a threshold set by the loss constants and a causal-cost sensitivity factor. Empirically, those stable (and nearby optimal) models cut performative risk relative to ordinary empirical risk minimization as strategic pressure grows, and the same causal machinery shows why action locations and induced feature changes can diverge.

Core claim

Causal recourse induces a non-convex performative risk even under convex losses, yet performatively stable linear models still exist and are efficiently computable by repeated risk minimization or repeated gradient descent whenever intervention strength κ is smaller than a threshold controlled by strong convexity, joint smoothness, and the causal-cost factor Λ built from the contribution matrices and the inverse cost matrix; those stable points stay near performative optima and substantially improve risk over naive ERM as κ grows.

What carries the argument

The causal contribution matrices Gx and Gy that map an additive intervention through the DAG into feature and label shifts, together with the closed-form best response a* = (κ/2) C^{-1} Gx w under quadratic costs; these define the distribution map whose sensitivity ε = (κ/2)Λ governs contraction of RRM and RGD.

Load-bearing premise

The learner must already know the true causal graph and how interventions change both features and the real outcome label; if that map is wrong, the stability guarantees do not describe actual behavior.

What would settle it

On the semi-synthetic or Taiwan credit setups, raise κ past the paper’s contraction threshold while keeping the stated cost and causal matrices fixed: if RRM or RGD still converges to a fixed point whose performative risk is no better than naive ERM, the central claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Deployed recourse systems that ignore causal structure will systematically incentivize gaming of non-causal features and force more frequent retraining.
  • Stable causal-recourse models reduce the gap between recommended actions and genuine qualification improvement once people adapt.
  • Intervention cost geometry and the causal graph play distinct roles: costs choose where people act, the graph chooses which features actually move.
  • Sufficiently strong interventions or strong causal amplification can destroy the contraction needed for simple iterative methods to reach stability.

Where Pith is reading between the lines

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

  • If the causal graph must be learned from the same strategic data stream, misspecification error would itself become performative and could push the system outside the contraction regime.
  • The same Gx–Gy sensitivity bound could be used as a design constraint: choose which features are actionable so that Λ stays small enough for stability at the desired κ.
  • Extending the analysis beyond linear scorers would require replacing the closed-form best response, but the performative framing itself would still apply.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper formalizes algorithmic recourse under strategic agent responses as a causal performative prediction problem. Agents take utility-maximizing interventions that propagate through a linear SCM via contribution matrices Gx and Gy, shifting both features and true labels; the learner then minimizes loss on the induced distribution. Under quadratic costs the best response is closed-form (Lemma 1), the resulting ERM objective can be non-convex (Lemma 2), and the induced map is ϵ-sensitive with ϵ = (κ/2)Λ(C,Gx,Gy) (Lemma 3). Convergence of RRM and RGD to performatively stable (and near-optimal) linear predictors is obtained by reducing to Perdomo et al. when κ is below a threshold involving loss constants and Λ. Experiments on a 7D semi-synthetic loan SCM and the Taiwan credit data show RRM/RGD convergence for moderate κ, large performative-risk gains of stable/near-optimal models over naive ERM as κ grows, and that cost geometry and causal propagation jointly shape actions versus induced Δx.

Significance. If the results hold, the paper cleanly connects causal recourse to performative prediction and supplies an explicit, checkable sensitivity parameter Λ that makes standard RRM/RGD guarantees apply. Lemma 1–3 and the reduction to Perdomo are technically sound and useful; the finite-sample algorithms and multi-profile cost experiments are reproducible in spirit. The main novelty is the causal response model inside the performative loop rather than a new optimization theory. The practical significance of the “role of causality” claim is currently limited by missing non-causal performative controls and by the assumption that the learner knows the true SCM (Gx, Gy), which the Discussion flags but which underpins both guarantees and experiments.

major comments (3)
  1. [Abstract; §1; Theorems 1–2; Lemma 3; Fig. 2–3] Abstract and §1 claim that “recourse policies that ignore causal structure can induce large, misaligned behavioral responses, whereas causal recourse leads to stable equilibria that reduce incentives for gaming.” Theorems 1–2 and Lemma 3 hold for arbitrary fixed Gx, Gy (including Gx = I). Empirically, Fig. 2 only compares causal-performative stable/Grid-search models to naive ERM, confounding (i) modeling distribution shift with (ii) using causal G. There is no performative but non-causal control (e.g., agents and learner with Gx = I, or learner using identity while agents use the true SCM). Fig. 3 illustrates propagation under a fixed causal G but does not show better performative risk, tighter stability, or less gaming than a non-causal performative equilibrium. Either add that ablation or substantially qualify the title-level causality claims to match what is proven and measured.
  2. [§2; Appendix B.1; §5 Discussion] The framework and algorithms assume the learner knows the true DAG, edge weights, Gx, and Gy used both to form a*(x,w) and to run RRM/RGD (§2, App. B.1). Stability and the map Dw are with respect to that assumed response; misspecified G yields guarantees that need not apply to real behavior. The Discussion notes joint learning of the graph as future work, but given that this is the load-bearing modeling assumption for the central claim, the paper should either provide misspecification experiments (wrong edges/weights) or state clearly that all guarantees and empirical gains are conditional on a correctly specified SCM available to the learner.
  3. [§4.3; Theorem 1; Corollary-style bounds in App. B.4] §4.3 reports that empirically estimated γ, β certify convergence only for κ ≲ 10^{-8}, while Fig. 1 shows stable behavior for κ up to O(1)–O(10). The text correctly calls the bound conservative, but this gap means Theorems 1–2 do not explain the observed regime. A tighter, instance-dependent sensitivity analysis or a clear statement that the theorems are only qualitative sufficient conditions would strengthen the link between theory and experiments.
minor comments (6)
  1. [§2; Algorithms 1–2] Notation: adjacency matrix is called A while the action set is also A and the induced feature map uses A := (κ/2)G_x^T C^{-1} G_x in Algorithms 1–2. Disambiguate (e.g., adjacency A_graph vs. shift matrix A_w).
  2. [Theorem 2] Theorem 2 item 2 refers to “Algorithm 3” but only Algorithms 1–2 are defined; almost certainly RGD (Algorithm 2).
  3. [Assumption 2] Assumption 2 writes C ∈ R^{n×n} while features are d-dimensional; should be d×d.
  4. [Appendix B.2 Lemma 3] Lemma 3 proof uses 1/(2λ) while the main text uses κ/2; align the cost scaling notation with Eq. (1).
  5. [§4.2–4.3] Fig. 1 caption and §4.2: RGD max iterations stated as 8000 in text but “1000 iterations” in the results paragraph; make consistent.
  6. [§1.1] Related work on improvement-focused causal recourse (König et al.) and performative validity of recourse is cited; a short explicit contrast table (fixed model vs. joint retraining) would help readers place the contribution.

Circularity Check

0 steps flagged

No significant circularity: performative stability and sensitivity are derived, not defined to equal the claimed gains.

full rationale

The paper’s load-bearing chain is: quadratic-cost best response (Lemma 1) → induced map Dw with explicit feature/label shifts via Gx, Gy → ε-sensitivity (Lemma 3) → application of external Perdomo et al. (2020) RRM/RGD convergence and stability–optimum proximity (Theorems 1–2). None of these steps equates the optimization target to the reported improvement by construction. Performative risk and stability are standard fixed-point / risk-on-induced-distribution notions; beating naive ERM (Figure 2) is an empirical comparison, not a tautology from fitted parameters. Gx is built from the assumed SCM and Gy by regressing scores on features—ordinary model specification used as input to the dynamics, not a quantity later “predicted” as if independent. Quadratic costs and strong convexity/smoothness are stated assumptions, not smuggled uniqueness theorems from overlapping authors. Citations to Perdomo are external and machine-checkable in the usual sense of published theorems with stated hypotheses; self-citations (e.g., Avasarala et al., Bechavod/Ziani) appear only as related work or “standard cost” context and are not load-bearing for the convergence claims. Gaps such as missing non-causal performative controls affect claim isolation, not circularity of the derivation.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 2 invented entities

Load-bearing structure is mostly imported: linear SCM/DAG, quadratic action costs, linear scoring rules, and Perdomo-style strong convexity/smoothness plus EMD sensitivity. The paper adds the Gx/Gy response embedding and the κΛ contraction threshold. Free knobs (κ, C profiles, ρ, η, heuristic optimum search) materially affect reported gaps. The largest external premise is knowing the causal graph used to define agent responses.

free parameters (4)
  • κ (intervention / performativity strength) = swept; e.g. values up to 5–15 in figures
    Scales inverse costs and thus action magnitude and distribution shift; convergence region and ERM-gap plots are parameterized by hand-chosen κ grids.
  • Cost matrix C diagonal profiles = Table 1 normalized unit-mean profiles
    Default/uniform/upstream/downstream relative costs are designer choices; immutable features set to ∞; finite entries jittered for error bars.
  • ℓ2 regularization ρ and RGD step η = η=5×10^-4; ρ implicit in γ bound
    Enter strong-convexity γ=2ρ and practical optimization; η=5e-4 fixed in experiments.
  • Estimated edge weights and Gy regression = dataset-specific regressions (Appendix C)
    Signed weights from regressing children on parents; Gy from regressing features on scores—fit to the evaluation distributions.
axioms (5)
  • domain assumption Feature causal structure is a known linear DAG SCM; interventions are additive and propagate as Δx=(I-A^T)^{-1}a = Gx^T a
    Section 2 and Appendix B.1; common in causal recourse but load-bearing for both agent response and learner updates.
  • domain assumption Action cost is quadratic a^T C a with C≻0, yielding closed-form a*=(κ/2)C^{-1}Gx w
    Assumption 2 and Lemma 1; standard in strategic classification but restrictive versus general recourse costs.
  • domain assumption Hypothesis class is linear scores w^T x; loss jointly convex, γ-strongly convex in w, β-jointly smooth
    Section 2 and Assumptions 1,3,4; needed for Perdomo-style contraction.
  • standard math Performative stability/optimality framework and RRM/RGD convergence theorems of Perdomo et al. (2020) apply once EMD sensitivity is bounded
    Theorems 1–2 explicitly corollary Perdomo Thm 3.5/3.10/4.3 plus Lemma 3.
  • ad hoc to paper Finite exponential α-moment condition on features under shifts x+Aw so ξ_{α,μ}<∞
    Technical tail condition for finite-sample guarantees in Theorem 1; Remark 1 offers a milder D0 moment surrogate.
invented entities (2)
  • Feature/label contribution matrices Gx, Gy in the performative recourse map independent evidence
    purpose: Encode cumulative causal effects of actions on features and true scores so Dw and best responses are explicit linear functions of w
    Derived from the linear SCM rather than postulated particles; still a paper-specific packaging that the whole analysis hangs on.
  • Sensitivity aggregate Λ(C,Gx,Gy)=||Gx||^2||C^{-1}||+||Gx||||C^{-1}||||Gy|| no independent evidence
    purpose: Single constant controlling ϵ-sensitivity and allowed κ for contraction
    Definitional composite of norms; useful but not an external physical entity.

pith-pipeline@v1.2.0-daily-grok45 · 21714 in / 3684 out tokens · 81597 ms · 2026-07-31T05:29:31.234746+00:00 · methodology

0 comments
read the original abstract

Algorithmic recourse aims to provide individuals with actionable changes to improve their predicted outcomes in high-stakes classification settings, such as loan and mortgage applications. However, most existing approaches focus only on flipping a model's prediction, without accounting for whether the recommended changes lead to genuine improvement in an individual's true qualifications or merely enable strategic gaming of the classifier. Consequently, deployed recourse policies can induce behavioral responses that degrade predictive accuracy and become ineffective after model retraining. In this work, we formalize this failure mode through a causal performative framework for recourse. We model how recourse actions propagate through a structural causal model, capturing interactions among features as well as their effect on the true label. These causal responses induce a non-convex optimization problem, even under standard convex losses. We characterize conditions under which performatively stable solutions exist and can be efficiently computed via simple iterative dynamics. Our analysis reveals that recourse policies that ignore causal structure can induce large, misaligned behavioral responses, whereas causal recourse leads to stable equilibria that reduce incentives for gaming. Experiments on both semi-synthetic and real credit datasets demonstrate that our approach consistently outperforms standard empirical risk minimization while reducing the need for repeated model retraining to accommodate distribution shifts caused by strategic agent behavior.

Figures

Figures reproduced from arXiv: 2607.28497 by Juba Ziani, Saber Salehkaleybar, Shahin Jabbari, Srikanth Avasarala, Varun Gupta.

Figure 1
Figure 1. Figure 1: Convergence of RRM and RGD for varying values of [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Ratio of the naive ERM performative objective to the corresponding performative objective as a function of [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Normalized RGD action magnitudes and induced feature changes on the semi-synthetic and Taiwan datasets [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The actions a and how they affect features and label changes using a linear relationship, i.e., ∆x(a) = GT x a and ∆y(a) = GT y a. We model recourse actions as additive interventions on the structural equations. Given an intervention vector a = (a1, . . . , ad) ⊤, the post-intervention features satisfy x ′ = A⊤x ′ + u + a. Thus, the intervention ai acts as an additive shift to the structural equation gener… view at source ↗
Figure 5
Figure 5. Figure 5: Causal graph for a) Semi-synthetic dataset, and b) Taiwan dataset [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Convergence behavior of RRM and RGD under randomized cost matrices for the semi-synthetic and Taiwan [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Ratio of the performative objective under the naive model to that under the stable (RGD) and Grid-search [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Relative model-weights’ distance ∥wRGD − wgrid∥2/∥wgrid∥2 as a function of the performativity parameter κ under four intervention cost profiles. Each panel reports results for both the semi-synthetic and Taiwan datasets. Error bars denote one standard deviation over randomized cost matrix realizations. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗

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

300 extracted references

  1. [1]

    42nd International Conference on Machine Learning , year =

    Kentaro Kanamori and Ken Kobayashi and Satoshi Hara and Takuya Takagi , title =. 42nd International Conference on Machine Learning , year =

  2. [2]

    CoRR , volume =

    Valia Efthymiou and Ekaterina Fedorova and Chara Podimata , title =. CoRR , volume =

  3. [3]

    38th International Conference on Machine Learning , pages =

    Zachary Izzo and Lexing Ying and James Zou , title =. 38th International Conference on Machine Learning , pages =

  4. [4]

    29th Annual Conference on Artificial Intelligence and Statistics , year =

    Antonio Gois and Sophia Gunluk and Nir Rosenfeld and Nidhi Hegde and Simon Lacoste-Julien and Dhanya Sridhar , title =. 29th Annual Conference on Artificial Intelligence and Statistics , year =

  5. [5]

    4th Symposium on Computer Science and Law , pages =

    Benjamin Laufer and Manish Raghavan and Solon Barocas , title =. 4th Symposium on Computer Science and Law , pages =

  6. [6]

    2026 , booktitle =

    Chris Hays and Benjamin Laufer and Solon Barocas and Manish Raghavan , title =. 2026 , booktitle =

    2026

  7. [7]

    Revisiting

    Ainhize Barrainkua and Giovanni De Toni and Jos. Revisiting. 40th

  8. [8]

    29th Annual Conference on Artificial Intelligence and Statistics , year =

    Eoin Kenny and Allan Anzagira and Tom Bewley and Freddy Lecue and Manuela Veloso , title =. 29th Annual Conference on Artificial Intelligence and Statistics , year =

  9. [9]

    Neill , title =

    Gordon Dai and Pavan Ravishankar and Rachel Yuan and Emily Black and Daniel B. Neill , title =. 5th

  10. [10]

    Advances in Neural Information Processing Systems 38 , year =

    Hsiang Hsu and Ivan Brugere and Shubham Sharma and Freddy L. Advances in Neural Information Processing Systems 38 , year =

  11. [11]

    CoRR , volume =

    Evzenie Coupkova and Mireille Boutin , title =. CoRR , volume =

  12. [12]

    Stephen Boyd and Lieven Vandenberghe , title =

  13. [13]

    24th International Conference on Autonomous Agents and Multiagent Systems , pages =

    Erfaun Noorani and Pasan Dissanayake and Faisal Hamman and Sanghamitra Dutta , title =. 24th International Conference on Autonomous Agents and Multiagent Systems , pages =

  14. [14]

    Statistical modeling: the two cultures , volume =

    Breiman, Leo , journal =. Statistical modeling: the two cultures , volume =

  15. [15]

    Theory of Computing , volume =

    Arora, Sanjeev and Hazan, Elad and Kale, Satyen , title =. Theory of Computing , volume =. 2012 , pages =

    2012

  16. [16]

    6th International Conference on Learning Representations , year =

    Constantinos Daskalakis and Andrew Ilyas and Vasilis Syrgkanis and Haoyang Zeng , title =. 6th International Conference on Learning Representations , year =

  17. [17]

    Advances in Neural Information Processing Systems 26 , pages =

    Alexander Rakhlin and Karthik Sridharan , title =. Advances in Neural Information Processing Systems 26 , pages =

  18. [18]

    Chris Russell , title =. 2nd

  19. [19]

    Advances in Neural Information Processing Systems 35 , year =

    Rui Xin and Chudi Zhong and Zhi Chen and Takuya Takagi and Margo Seltzer and Cynthia Rudin , title =. Advances in Neural Information Processing Systems 35 , year =

  20. [20]

    CoRR , volume =

    Lesia Semenova and Cynthia Rudin , title =. CoRR , volume =

  21. [21]

    Heller and Dan Moldovan and Ben Adlam and Babak Alipanahi and Alex Beutel and Christina Chen and Jonathan Deaton and Jacob Eisenstein and Matthew D

    Alexander D'Amour and Katherine A. Heller and Dan Moldovan and Ben Adlam and Babak Alipanahi and Alex Beutel and Christina Chen and Jonathan Deaton and Jacob Eisenstein and Matthew D. Hoffman and Farhad Hormozdiari and Neil Houlsby and Shaobo Hou and Ghassen Jerfel and Alan Karthikesalingam and Mario Lucic and Yi. Underspecification Presents Challenges fo...

  22. [22]

    Journal of Machine Learning Research , volume =

    Aaron Fisher and Cynthia Rudin and Francesca Dominici , title =. Journal of Machine Learning Research , volume =

  23. [23]

    Algorithmic Recourse: from Counterfactual Explanations to Interventions , booktitle =

    Amir. Algorithmic Recourse: from Counterfactual Explanations to Interventions , booktitle =

  24. [24]

    13th International Conference on Learning Representations , year =

    Ally Yalei Du and Dung Daniel Ngo and Zhiwei Steven Wu , title =. 13th International Conference on Learning Representations , year =

  25. [25]

    Dropout-Based

    Hsiang Hsu and Guihong Li and Shaohan Hu and Chun. Dropout-Based. 12th International Conference on Learning Representations , year =

  26. [26]

    Reconciling Predictive Multiplicity in Practice , booktitle =

    Tina Behzad and S. Reconciling Predictive Multiplicity in Practice , booktitle =

  27. [27]

    Krishna Acharya and Eshwar Ram Arunachaleswaran and Sampath Kannan and Aaron Roth and Juba Ziani , title =. 1st

  28. [28]

    34th International Conference on Machine Learning , pages =

    Joshua Achiam and David Held and Aviv Tamar and Pieter Abbeel , title =. 34th International Conference on Machine Learning , pages =

  29. [29]

    Advances in Neural Information Processing Systems 31 , pages =

    Julius Adebayo and Justin Gilmer and Michael Muelly and Ian Goodfellow and Moritz Hardt and Been Kim , title =. Advances in Neural Information Processing Systems 31 , pages =

  30. [30]

    A Reductions Approach to Fair Classification , booktitle =

    Alekh Agarwal and Alina Beygelzimer and Miroslav Dud. A Reductions Approach to Fair Classification , booktitle =

  31. [31]

    Fair Regression: Quantitative Definitions and Reduction-Based Algorithms , booktitle =

    Alekh Agarwal and Miroslav Dud. Fair Regression: Quantitative Definitions and Reduction-Based Algorithms , booktitle =

  32. [32]

    38th International Conference on Machine Learning , year =

    Sushant Agarwal and Shahin Jabbari and Chirag Agarwal and Sohini Upadhyay and Steven Wu and Himabindu Lakkaraju , title =. 38th International Conference on Machine Learning , year =

  33. [33]

    Saba Ahmadi and Hedyeh Beyhaghi and Avrim Blum and Keziah Naggita , title =. 22nd

  34. [34]

    Saba Ahmadi and Avrim Blum and Kunhe Yang , title =. 24th

  35. [35]

    Advances in Neural Information Processing Systems 38 , year =

    Saba Ahmadi and Kunhe Yang and Hanrui Zhang , title =. Advances in Neural Information Processing Systems 38 , year =

  36. [36]

    Model extraction from counterfactual explanations , journal =

    Ulrich A. Model extraction from counterfactual explanations , journal =

  37. [37]

    1999 , publisher =

    Altman, Eitan , title =. 1999 , publisher =

    1999

  38. [38]

    Machine Learning , volume =

    Angluin, Dana , title =. Machine Learning , volume =. 1988 , issn =

    1988

  39. [39]

    12th Innovations in Theoretical Computer Science Conference , pages =

    Eshwar Arunachaleswaran and Sampath Kannan and Aaron Roth and Juba Ziani , title =. 12th Innovations in Theoretical Computer Science Conference , pages =

  40. [40]

    Sepehr Assadi and Justin Hsu and Shahin Jabbari , title =. 3rd

  41. [41]

    35th International Conference on Machine Learning , pages =

    Anish Athalye and Nicholas Carlini and David Wagner , title =. 35th International Conference on Machine Learning , pages =

  42. [42]

    8th AAAI/ACM Conference on AI, Ethics, and Society , year =

    Srikanth Avasarala and Serena Wang and Juba Ziani , title =. 8th AAAI/ACM Conference on AI, Ethics, and Society , year =

  43. [43]

    Aparna Balagopalan and Haoran Zhang and Kimia Hamidieh and Thomas Hartvigsen and Frank Rudzicz and Marzyeh Ghassemi , title =. 5th

  44. [44]

    SIGecom Exchanges , volume =

    Eric Balkanksi and Vasilis Gkatzelis and Xizhi Tan , title =. SIGecom Exchanges , volume =

  45. [45]

    Fairness and Machine Learning: Limitations and Opportunities , author =

  46. [46]

    Solon Barocas and Andrew Selbst and Manish Raghavan , title =. 3rd

  47. [47]

    23rd International Conference on Artificial Intelligence and Statistics , pages =

    Elnaz Barshan and Marc. 23rd International Conference on Artificial Intelligence and Statistics , pages =

  48. [48]

    4th Annual Workshop on Computational Learning Theory , pages =

    Peter Bartlett and Robert Williamson , title =. 4th Annual Workshop on Computational Learning Theory , pages =

  49. [49]

    Advances in Neural Information Processing Systems 35 , year =

    Osbert Bastani and Varun Gupta and Christopher Jung and Georgy Noarov and Ramya Ramalingam and Aaron Roth , title =. Advances in Neural Information Processing Systems 35 , year =

  50. [50]

    Verifiable Reinforcement Learning via Policy Extraction , booktitle =

    Osbert Bastani and Yewen Pu and Armando Solar. Verifiable Reinforcement Learning via Policy Extraction , booktitle =

  51. [51]

    Understanding the role of individual units in a deep neural network , journal =

    David Bau and Jun. Understanding the role of individual units in a deep neural network , journal =

  52. [52]

    24th International Conference on Artificial Intelligence and Statistics , pages =

    Yahav Bechavod and Katrina Ligett and Zhiwei Steven Wu and Juba Ziani , title =. 24th International Conference on Artificial Intelligence and Statistics , pages =

  53. [53]

    39th International Conference on Machine Learning , pages =

    Yahav Bechavod and Chara Podimata and Zhiwei Steven Wu and Juba Ziani , title =. 39th International Conference on Machine Learning , pages =

  54. [54]

    Fairness in Algorithmic Recourse Through the Lens of Substantive Equality of Opportunity , journal =

    Andrew Bell and Jo. Fairness in Algorithmic Recourse Through the Lens of Substantive Equality of Opportunity , journal =

  55. [55]

    A theory of learning from different domains , journal =

    Shai Ben. A theory of learning from different domains , journal =

  56. [56]

    CoRR , volume =

    Richard Berk and Hoda Heidari and Shahin Jabbari and Matthew Joseph and Michael Kearns and Jamie Morgenstern and Seth Neel and Aaron Roth , title =. CoRR , volume =

  57. [57]

    Sociological Methods & Research , year =

    Richard Berk and Hoda Heidari and Shahin Jabbari and Michael Kearns and Aaron Roth , title =. Sociological Methods & Research , year =

  58. [58]

    Twitter Blog , month=

    Introducing Twitter’s first algorithmic bias bounty challenge , author=. Twitter Blog , month=. 2021 , url=

    2021

  59. [59]

    Emily Black and Manish Raghavan and Solon Barocas , title =. 5th

  60. [60]

    10th International Conference on Learning Representations , year =

    Emily Black and Zifan Wang and Matt Fredrikson , title =. 10th International Conference on Learning Representations , year =

  61. [61]

    41st International Conference on Machine Learning , year =

    Tom Bewley and Salim Amoukou and Saumitra Mishra and Daniele Magazzeni and Manuela Veloso , title =. 41st International Conference on Machine Learning , year =

  62. [62]

    1st Symposium on Foundations of Responsible Computing , pages =

    Mark Braverman and Sumegha Garg , title =. 1st Symposium on Foundations of Responsible Computing , pages =

  63. [63]

    Statistical modeling:

    Breiman, Leo , journal =. Statistical modeling:

  64. [64]

    Stackelberg games for adversarial prediction problems , booktitle =

    Michael Br. Stackelberg games for adversarial prediction problems , booktitle =

  65. [65]

    39th Annual AAAI Conference on Artificial Intelligence , pages =

    Andrei Buliga and Chiara Di Francescomarino and Chiara Ghidini and Marco Montali and Massimiliano Ronzani , title =. 39th Annual AAAI Conference on Artificial Intelligence , pages =

  66. [66]

    Data Min

    Toon Calders and Sicco Verwer , title =. Data Min. Knowl. Discov. , volume =

  67. [67]

    State Augmented Constrained Reinforcement Learning: Overcoming the Limitations of Learning With Rewards , journal =

    Miguel Calvo. State Augmented Constrained Reinforcement Learning: Overcoming the Limitations of Learning With Rewards , journal =

  68. [68]

    Advances in Neural Information Processing Systems 34 , pages =

    Elisa Celis and Anay Mehrotra and Nisheeth Vishnoi , title =. Advances in Neural Information Processing Systems 34 , pages =

  69. [69]

    28th International Conference on Artificial Intelligence and Statistics , volume =

    Yatong Chen and Andrew Estornell and Yevgeniy Vorobeychik and Yang Liu , title =. 28th International Conference on Artificial Intelligence and Statistics , volume =

  70. [70]

    Advances in Neural Information Processing Systems 33 , year =

    Yiling Chen and Yang Liu and Chara Podimata , title =. Advances in Neural Information Processing Systems 33 , year =

  71. [71]

    Yiling Chen and Chara Podimata and Ariel Procaccia and Nisarg Shah , title =. 19th

  72. [72]

    Performative Prediction with Bandit Feedback: Learning through Reparameterization , booktitle =

    Yatong Chen and Wei Tang and Chien. Performative Prediction with Bandit Feedback: Learning through Reparameterization , booktitle =

  73. [73]

    Yatong Chen and Jialu Wang and Yang Liu , title =. Trans. Mach. Learn. Res. , volume =

  74. [74]

    37th Annual Conference on Learning Theory , pages =

    On Finding Small Hyper-Gradients in Bilevel Optimization: Hardness Results and Improved Analysis , author =. 37th Annual Conference on Learning Theory , pages =

  75. [75]

    13th International Conference on Learning Representations , year =

    Seung Hyun Cheon and Anneke Wernerfelt and Sorelle Friedler and Berk Ustun , title =. 13th International Conference on Learning Representations , year =

  76. [76]

    American Economic Review , Volume =

    Chetty, Raj and Hendren, Nathaniel and Kline, Patrick and Saez, Emmanuel and Turner, Nicholas , Title =. American Economic Review , Volume =. 2014 , Pages =

    2014

  77. [77]

    Advances in Neural Information Processing Systems 30 , pages =

    Flavio Chierichetti and Ravi Kumar and Silvio Lattanzi and Sergei Vassilvitskii , title =. Advances in Neural Information Processing Systems 30 , pages =

  78. [78]

    Big Data , volume =

    Alexandra Chouldechova , title =. Big Data , volume =

  79. [79]

    Advances in Neural Information Processing Systems 38 , year =

    Lee Cohen and Saeed Sharifi. Advances in Neural Information Processing Systems 38 , year =

  80. [80]

    Algorithmic Decision Making and the Cost of Fairness , booktitle =

    Sam Corbett. Algorithmic Decision Making and the Cost of Fairness , booktitle =

Showing first 80 references.