Pith. sign in

REVIEW 3 major objections 4 minor 47 references

Explanation Design in Strategic Learning: Sufficient Explanations that Induce Non-harmful Responses

T0 review · 3 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read Under conditional homogeneity of agents' responses, any non-harmful response that any explanation method can induce can also be induced by an action recommendation-based explanation, so decision makers lose nothing by restricting to…

desk verdict The necessary-condition theorem is solid; the sufficiency result is really a statement about how strong Assumption 3.5 is, not a general dominance result for ARexes. read the letter →

arxiv 2502.04058 v2 pith:2QZEAKLW submitted 2025-02-06 cs.AI

Kiet Q. H. Vo , Siu Lun Chau , Masahiro Kato , Yixin Wang , Krikamol Muandet This is my paper · ORCID
classification cs.AI
keywords explanationdesignstrategiclearningactionrecommendation-basedexplanationscounterfactualno-harmresponsesinformationperformativepredictionalgorithmicrecourse
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

This paper asks how a decision maker who will not or cannot disclose its full predictive model should explain decisions to strategic agents who may change their features in response. The authors establish a necessary condition for surrogate-model explanations to avoid harming agents: the surrogate must not overstate the gain an agent would get from moving to a lower-scoring profile. They then define action recommendation-based explanations (ARexes), explanations of the form "change your covariates to x and your predicted outcome will be y," and prove that under conditional homogeneity of agents' responses, ARexes are sufficient: any non-harmful response that any explanation method can induce can also be induced by some ARex. If this sufficiency result is right, decision makers can restrict attention to ARexes when optimizing their own objectives under a no-harm constraint, without losing anything. The paper also gives a joint optimization procedure for the predictive model and ARex policy, with synthetic and German credit experiments showing predictive gains while agents' utility is preserved.

What carries the argument

The central object is the ARex explanation policy, a mapping $\sigma: X \times G \to X \times Y$ that, for a base covariate and a predictive model, outputs a recommended covariate profile $\vec{x}_t$ and the predicted outcome $\hat{\vec{y}}_t = g(\vec{x}_t)$. The agent's reaction model compares utilities and adopts the recommendation if $u_t(g, \vec{x}_t) \ge u_t(g, \ddot{x}_t)$, otherwise stays at the base covariate. The conditional homogeneity assumption (Assumption 3.5) ensures that agents sharing the same observed covariates and the same explanation respond identically, which lets one ARex work for an entire group and drives the proof of Theorem 3.6. For surrogate explanations, the key mechanism is the inequality in Theorem 3.2, which prevents the surrogate from exaggerating the gain of moving to a lower-scoring profile.

What would settle it

One concrete check is to take agents with identical observable covariates, give them the same explanation, and vary only aspects the decision maker cannot observe, such as the cost of changing a feature: under conditional homogeneity these agents must respond identically, so observing a substantial share of agents in the same covariate cell responding differently, or some declining the same ARex recommendation, would falsify Assumption 3.5 and remove the sufficiency guarantee of Theorem 3.6.

Watch

Extended reading notes

Core claim

The central claim is a sufficiency result for a class of explanations the paper calls action recommendation-based explanations (ARexes). Theorem 3.6 states: for any group of agents with the same base covariate and conditionally homogeneous responses, and for any explanation $e$ from any method that induces a response $x^\bullet$ that does not reduce any agent's utility, there is an ARex explanation $(x^\bullet, g(x^\bullet))$ that induces the same response. Because an ARex explicitly tells the agent which covariate profile to adopt and what prediction it will yield, the agent can compare it with the status quo and can never be led to a response outside their non-harmful set (Remark 3.4). The paper also proves a necessary condition for surrogate-model explanations: a surrogate $f_t$ must satisfy $f_t(\ddot{x}_t) - f_t(x) \le g(\ddot{x}_t) - g(x)$ on the set of lower-scoring profiles $X^{g\downarrow}_t$, otherwise there exists a cost function under which the agent's best response lowers true utility.

Load-bearing premise

The load-bearing premise is conditional homogeneity (Assumption 3.5): agents who share the same observed covariates and receive the same explanation must give identical responses. This requires the decision maker's observed features to capture all response-relevant heterogeneity; if unobserved costs or preferences differ within a covariate cell, the proof that a single ARex works for the whole group no longer goes through.

Editorial extensions

If this is right

  • When the no-harm constraint is in force, any explanation method outside the ARex class is weakly dominated by some ARex: regardless of the decision maker's objective, an optimal ARex policy achieves at least as much.
  • Counterfactual explanations are a special case of ARexes, so they inherit the no-harm guarantee, but because ARexes need not minimize feature changes, the decision maker can steer agents toward regions where the predictive model is accurate.
  • Any surrogate explanation that overstates the true gain from a lower-scoring profile violates the necessary condition of Theorem 3.2, and for some cost function some agent will be harmed.
  • Jointly optimizing the predictive model and the ARex policy via repeated risk minimization yields lower strategic prediction error than fixing a counterfactual policy, while maintaining high compliance in simulated agents.
  • The sufficiency result is stated independently of the decision maker's objective, so the same ARex framework supports error minimization, welfare maximization, or other goals without re-deriving the safety guarantee.

Reading between the lines

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

  • An implication the paper leaves implicit is that the practical burden shifts from writing interpretable models to collecting enough covariates about each agent to make responses conditionally homogeneous; otherwise the sufficiency guarantee is not available.
  • A testable extension not pursued in the paper is to replace the deterministic response rule in Equation (3) with a stochastic compliance model estimated from user studies and check whether a population-level ARex still achieves the no-harm rate the deterministic theory predicts.
  • Connected to neighbouring work on algorithmic recourse, the paper's no-harm framing sidesteps the causal assumptions usually required for recourse, so ARex-style recommendations could serve as a weaker, safer fallback when causal knowledge is unavailable.
  • The necessary condition for surrogates suggests a practical diagnostic: before deploying any feature-attribution or local-model explanation, compute whether the surrogate's implied gain exceeds the true model's gain on plausible agent moves; if it does, the explanation is unsafe for at least some cost structure.
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

3 major / 4 minor

Summary. The paper studies how a decision maker (DM) should explain a predictive model to strategic agents who may modify their covariates before resubmission. It defines non-harmful responses (Definition 3.1), derives a necessary condition for surrogate-model explanations to avoid misleading agents (Theorem 3.2), and proposes action recommendation-based explanations (ARexes), where the DM recommends a specific covariate update and discloses its predicted outcome. The central theoretical result (Theorem 3.6) claims that, under a conditional homogeneity assumption (Assumption 3.5), ARexes are sufficient: any response induced by an arbitrary explanation method and lying in the agents' non-harmful set can also be induced by an ARex. The paper also presents two empirical studies, one synthetic and one on the German credit dataset, in which the DM jointly optimizes the predictive model and the ARex policy via repeated risk minimization.

Significance. If the sufficiency result were established under reasonably general conditions, it would provide a useful information-design perspective on explainable ML: DMs could restrict attention to action recommendations when they care about avoiding harm to agents, regardless of their own objective. The paper is clearly written, makes its modeling assumptions explicit, and provides proofs in the appendix. The framework also gives a clean distinction between explanations that merely describe a model and explanations that directly guide agent actions. However, the main sufficiency claim rests on a very strong homogeneity assumption that, as the paper itself notes in Section 3.2, effectively requires the DM to have 'just enough' information to make unobserved heterogeneity irrelevant. Read literally, this assumption removes most of the strategic-heterogeneity structure introduced in Section 2, and it is violated in the paper's own synthetic experiment (Appendix C.2). The contribution is therefore best seen as a formal characterization under a restrictive condition rather than a broadly applicable design principle.

major comments (3)
  1. [Section 3.2, Assumption 3.5] Assumption 3.5 is much stronger than the surrounding discussion suggests. Since the explanation policy in Definition 2.1 maps (x,g) to e, agents with the same observed covariate x automatically receive the same explanation under any fixed method. The assumption therefore requires the response function psi(x,e,z,c) to be constant in the unobservable (z,c) for every explanation e, i.e., the DM's observed covariates fully determine each agent's reaction. This contradicts the modeling choice in Section 2, where C is explicitly allowed to be a random function correlated with X and Z and psi depends on z_t and c_t. Moreover, the synthetic setup in Appendix C.2 generates z_t and alpha_t that vary within covariate cells, so Assumption 3.5 is violated in the very experiments used to illustrate ARexes. The paper should either weaken the assumption, provide a concrete way to enforce it in the experiments, or clearly state that Theorem 3.6 applies only to settings with no unobserved heterogeneity in reactions.
  2. [Appendix B.4, proof of Theorem 3.6] The first step of the proof states that conditional homogeneity implies all T' agents have the same non-harmful set nu_t = nu, but this inference is not immediate and the proof currently skips the key argument. To justify it, one must consider, for each x in X, the ARex explanation (x,g(x)). If x lies in nu_t but not nu_{t'}, then agent t adopts the recommendation while agent t' rejects it, producing different responses to the same explanation, which violates Assumption 3.5. Hence nu_t = nu for all t. This missing argument should be written out explicitly, because it is load-bearing for the sufficiency claim.
  3. [Section 4.2, Table 1] The empirical claims in Table 1 and Figure 2 are based on single runs with no error bars, standard deviations, or multiple seeds. As reported, the differences between Joint-Opt and the fixed counterfactual baselines cannot be distinguished from optimization noise or random initialization. The paper should report means and variances over several random seeds, and ideally a paired test, before claiming that joint optimization of g and sigma 'achieves the lowest test error' or 'highest test score' in a robust sense.
minor comments (4)
  1. [Abstract and Section 1] There are several typographical issues, including 'su fficient', 'e ffective', and 'counterfactual explanations [42], by design, can prevent such misalignment' in the contributions list, where the intended meaning would be clearer as 'counterfactual explanations, by design, prevent such misalignment'.
  2. [Section 2, Definition 2.1] The definition of an explanation method does not explicitly state that the response function psi is part of the model, yet Assumption 3.5 and Theorem 3.6 rely on responses across agents. The paper would benefit from stating psi as an explicit component of the interaction model, rather than only in prose.
  3. [Section 4.1] The sentence 'ARexes cannot be misled, even if the recommended actions are generated arbitrarily' is imprecise: Remark 3.4 guarantees that the agent cannot be made worse off relative to the base covariate, but the recommendation itself is not the agent's response. Rephrasing to 'ARexes cannot induce a harmful response' would be more accurate.
  4. [Appendix C.2] The statement that lower-level settings such as learning rates and model sizes 'do not affect our main conclusions' is an assertion, not a demonstrated claim. Since the main empirical conclusion is comparative, at least the key hyperparameters of the baselines and the joint method should be reported explicitly.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: Theorem 3.6 is a formal consequence of the paper's stated agent-response model and conditional homogeneity assumption, not a fitted or self-citational prediction.

full rationale

We walked the derivation chain. Theorem 3.2 and Corollary 3.3 are proven by contrapositive from the surrogate-utility model and Definition 3.1; no fitted parameter enters and the necessary condition is derived, not assumed. Theorem 3.6 has a definitional flavor: Equation (3) defines the ARex agent's response as adopting the recommended action exactly when it is non-harmful, so Remark 3.4 is true by design and the existence of an ARex that induces a given common non-harmful response x• is shown by recommending x• (Appendix B.4). This is a construction, not a circular derivation: the proof does not assume the theorem's conclusion, and the common-ν premise is obtained from Assumption 3.5 rather than from the desired result. Assumption 3.5 is strong and genuinely load-bearing - it is essentially equivalent, under Equation (3), to requiring agents with the same ẋ to share the same ν set - but a strong assumption is a robustness/validity concern, not circularity. The paper's self-citations ([6], [7], [40]) appear only as examples or related work and do not support the main theorems; no uniqueness claim is imported from the authors' prior work. The empirical sections use heuristically chosen simulation parameters (e.g., the 0.01 cost scaling, CTGAN-generated data) and a learned compliance model, but these are presented as simulations and do not generate the theoretical 'predictions'; they are not fitted inputs renamed as predictions. Overall, the formal claims are derived from explicitly stated assumptions and hold independently of the experimental setup.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central theory rests on standard strategic learning assumptions plus the strong conditional homogeneity assumption. No new physical entities are introduced. The experimental free parameters are heuristic choices that affect the demonstrations but not the main theorems.

free parameters (3)
  • cost scaling factor = 0.01
    In the German credit experiment, the cost function uses c(x_t,x)=0.01 times a normalized L1 distance. The paper states this value is heuristically chosen to balance agents who can and cannot change features, and it directly affects compliance rates and the optimization outcome.
  • CE regularizer lambda = 0.1, 1.0, 4.0
    Baseline counterfactual explanations use lambda to trade off prediction gain against feature distance. These values are chosen by the authors and only affect baselines, not the theoretical claim.
  • sampler pi variance = N(x*, 4)
    In the synthetic Joint-Opt, random ARexes are sampled from a normal distribution with variance 4 to train the compliance predictor. This choice influences the learned policy but is not part of the theory.
assumptions (6)
  • domain assumption Agents maximize an additive utility u_t(g,x) = -g(x) - c_t(x_t,x) (Eq. 1).
    Standard strategic learning assumption, used to define non-harmful responses.
  • domain assumption Cost function satisfies c_t(x_t,x_t)=0 and c_t>0 otherwise (Def. 2.2).
    Ensures staying has zero cost and any change is costly.
  • domain assumption ARex agents choose only between staying at x_t and following the recommended x (Eq. 3).
    Critical for Remark 3.4; if agents infer more from the recommendation and choose other points, no-harm is not guaranteed.
  • domain assumption Conditional homogeneity of responses (Assumption 3.5).
    Load-bearing for Theorem 3.6; it asserts that unobserved heterogeneity does not affect responses once covariate and explanation are fixed.
  • domain assumption Surrogate agents best-respond to the surrogate utility u_t(f_t,.) (Section 3.1).
    Needed for Theorem 3.2; alternative behavioral models would require re-derivation.
  • domain assumption Potential outcome h is deterministic and DM learns g to approximate h.
    Standard setup; the empirical objective uses squared loss against h.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Explanation Design in Strategic Learning: Sufficient Explanations that Induce Non-harmful Responses." pith.science (2026). https://pith.science/paper/2QZEAKLW

@misc{pith2026250204058,
  author       = {Pith},
  title        = {Pith review of: Explanation Design in Strategic Learning: Sufficient Explanations that Induce Non-harmful Responses},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2QZEAKLW}},
  note         = {Machine review of arXiv:2502.04058}
}
read the original abstract

We study explanation design in algorithmic decision making with strategic agents, individuals who may modify their inputs in response to explanations of a decision maker's (DM's) predictive model. As the demand for transparent algorithmic systems continues to grow, most prior work assumes full model disclosure as the default solution. In practice, however, DMs such as financial institutions typically disclose only partial model information via explanations. Such partial disclosure can lead agents to misinterpret the model and take actions that unknowingly harm their utility. A key open question is how DMs can communicate explanations in a way that avoids harming strategic agents, while still supporting their own decision-making goals, e.g., minimising predictive error. In this work, we analyse well-known explanation methods, and establish a necessary condition to prevent explanations from misleading agents into self-harming actions. Moreover, with a conditional homogeneity assumption, we prove that action recommendation-based explanations (ARexes) are sufficient for non-harmful responses, mirroring the revelation principle in information design. To demonstrate how ARexes can be operationalised in practice, we propose a simple learning procedure that jointly optimises the predictive model and explanation policy. Experiments on synthetic and real-world tasks show that ARexes allow the DM to optimise their model's predictive performance while preserving agents' utility, offering a more refined strategy for safe and effective partial model disclosure.

Figures

Figures reproduced from arXiv: 2502.04058 by the authors.

Figure 1
Figure 1. (left) With full access to the DM’s model, the agent with a covariate [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. (a) Taylor-ex mislead agents into reducing their utility while ARexes do not. The box plot shows the change in agents’ utility after best responding. (b) & (c): Joint-Opt has the lowest training-loss curves (nMSE and BCE) against the three baselines, showing that jointly optimising both g and σ is more beneficial to the DM. distribution of agents’ responses PX. 4 This objective reflects a typical goal in insurance p… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 35 canonical work pages

  1. [1]

    The strategic perceptron

    Saba Ahmadi, Hedyeh Beyhaghi, Avrim Blum, and Keziah Naggita. The strategic perceptron. In Proceedings of the 22nd ACM Conference on Economics and Computation, pages 6–25, 2021

    work page 2021

  2. [2]

    Fundamental bounds on online strategic classifi- cation

    Saba Ahmadi, Avrim Blum, and Kunhe Yang. Fundamental bounds on online strategic classifi- cation. In Proceedings of the 24th ACM Conference on Economics and Computation, pages 22–58, 2023

    work page 2023

  3. [3]

    Information discrepancy in strategic learning

    Yahav Bechavod, Chara Podimata, Steven Wu, and Juba Ziani. Information discrepancy in strategic learning. In International Conference on Machine Learning, pages 1691–1715. PMLR, 2022

    work page 2022

  4. [4]

    Information design: A unified perspective

    Dirk Bergemann and Stephen Morris. Information design: A unified perspective. Journal of Economic Literature, 57(1):44–95, 2019

    work page 2019

  5. [5]

    Static prediction games for adversar- ial learning problems

    Michael Brückner, Christian Kanzow, and Tobias Scheffer. Static prediction games for adversar- ial learning problems. The Journal of Machine Learning Research, 13(1):2617–2654, 2012

    work page 2012

  6. [6]

    Rkhs-shap: Shapley values for kernel methods

    Siu Lun Chau, Robert Hu, Javier Gonzalez, and Dino Sejdinovic. Rkhs-shap: Shapley values for kernel methods. Advances in neural information processing systems, 35:13050–13063, 2022

    work page 2022

  7. [7]

    Explaining the uncertain: Stochastic shapley values for gaussian process models

    Siu Lun Chau, Krikamol Muandet, and Dino Sejdinovic. Explaining the uncertain: Stochastic shapley values for gaussian process models. Advances in Neural Information Processing Systems, 36:50769–50795, 2023

    work page 2023

  8. [8]

    Interpretable machine learning: A guide for making black box models explainable

    Molnar Christoph. Interpretable machine learning: A guide for making black box models explainable. Leanpub, 2020

    2020

Show all 47 references
  1. [9]

    Sequential strategic screening

    Lee Cohen, Saeed Sharifi-Malvajerdi, Kevin Stangl, Ali Vakilian, and Juba Ziani. Sequential strategic screening. In International Conference on Machine Learning, pages 6279–6295. PMLR, 2023

    2023

  2. [10]

    Bayesian strategic classification

    Lee Cohen, Saeed Sharifi-Malvajerdi, Kevin Stangl, Ali Vakilian, and Juba Ziani. Bayesian strategic classification. arXiv preprint arXiv:2402.08758, 2024

    2024 arXiv

  3. [11]

    Council regulation (EU) no 679/2016, 2016

    Council of European Union. Council regulation (EU) no 679/2016, 2016. https://eur-lex.europa.eu/legal-content/EN/TXT/PDF/?uri=CELEX:32016R0679

    2016

  4. [12]

    Strategic classification from revealed preferences

    Jinshuo Dong, Aaron Roth, Zachary Schutzman, Bo Waggoner, and Zhiwei Steven Wu. Strategic classification from revealed preferences. InProceedings of the 2018 ACM Conference on Economics and Computation, pages 55–70, 2018

    2018

  5. [13]

    Strategic classification in the dark

    Ganesh Ghalme, Vineet Nair, Itay Eilat, Inbal Talgam-Cohen, and Nir Rosenfeld. Strategic classification in the dark. In International Conference on Machine Learning, pages 3672–3681. PMLR, 2021. 11 Sufficient Explanations that Induce Non-harmful Responses A Preprint

    2021

  6. [14]

    right to explanation

    Bryce Goodman and Seth Flaxman. European union regulations on algorithmic decision- making and a “right to explanation”. AI magazine, 38(3):50–57, 2017

    2017

  7. [15]

    Calibrated stackelberg games: Learning optimal commitments against calibrated agents

    Nika Haghtalab, Chara Podimata, and Kunhe Yang. Calibrated stackelberg games: Learning optimal commitments against calibrated agents. Advances in Neural Information Processing Systems, 36, 2024

    2024

  8. [16]

    Strategic classifi- cation

    Moritz Hardt, Nimrod Megiddo, Christos Papadimitriou, and Mary Wootters. Strategic classifi- cation. In Proceedings of the 2016 ACM conference on innovations in theoretical computer science, pages 111–122, 2016

    2016

  9. [17]

    Stateful strategic regression

    Keegan Harris, Hoda Heidari, and Steven Z Wu. Stateful strategic regression. Advances in Neural Information Processing Systems, 34:28728–28741, 2021

    2021

  10. [18]

    Bayesian persuasion for algorithmic recourse.Advances in Neural Information Processing Systems, 35:11131–11144, 2022

    Keegan Harris, Valerie Chen, Joon Kim, Ameet Talwalkar, Hoda Heidari, and Steven Z Wu. Bayesian persuasion for algorithmic recourse.Advances in Neural Information Processing Systems, 35:11131–11144, 2022

    2022

  11. [19]

    Strategic instrumental variable regression: Recovering causal relationships from strategic responses

    Keegan Harris, Dung Daniel T Ngo, Logan Stapleton, Hoda Heidari, and Steven Wu. Strategic instrumental variable regression: Recovering causal relationships from strategic responses. In International Conference on Machine Learning, pages 8502–8522. PMLR, 2022

    2022

  12. [20]

    Statlog (German Credit Data)

    Hans Hofmann. Statlog (German Credit Data). UCI Machine Learning Repository, 1994. DOI: https://doi.org/10.24432/C5NC77

    1994 doi

  13. [21]

    The disparate effects of strategic manipulation

    Lily Hu, Nicole Immorlica, and Jennifer Wortman Vaughan. The disparate effects of strategic manipulation. CoRR, abs/1808.08646, 2018. URL http://arxiv.org/abs/1808.08646

    2018 arXiv

  14. [22]

    Alternative microfoundations for strategic classification

    Meena Jagadeesan, Celestine Mendler-Dünner, and Moritz Hardt. Alternative microfoundations for strategic classification. In International Conference on Machine Learning, pages 4687–4697. PMLR, 2021

    2021

  15. [23]

    Bayesian persuasion

    Emir Kamenica and Matthew Gentzkow. Bayesian persuasion. American Economic Review, 101 (6):2590–2615, 2011

    2011

  16. [24]

    Algorithmic recourse: from counterfactual explanations to interventions

    Amir-Hossein Karimi, Bernhard Schölkopf, and Isabel Valera. Algorithmic recourse: from counterfactual explanations to interventions. In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency, pages 353–362, 2021

    2021

  17. [25]

    A survey of algorithmic recourse: contrastive explanations and consequential recommendations

    Amir-Hossein Karimi, Gilles Barthe, Bernhard Schölkopf, and Isabel Valera. A survey of algorithmic recourse: contrastive explanations and consequential recommendations. ACM Computing Surveys, 55(5):1–29, 2022

    2022

  18. [26]

    How do classifiers induce agents to invest effort strategi- cally? ACM Transactions on Economics and Computation (TEAC), 8(4):1–23, 2020

    Jon Kleinberg and Manish Raghavan. How do classifiers induce agents to invest effort strategi- cally? ACM Transactions on Economics and Computation (TEAC), 8(4):1–23, 2020

    2020

  19. [27]

    Strategic classification made practical

    Sagi Levanon and Nir Rosenfeld. Strategic classification made practical. In International Conference on Machine Learning, pages 6243–6253. PMLR, 2021

    2021

  20. [28]

    A unified approach to interpreting model predictions

    Scott Lundberg. A unified approach to interpreting model predictions. arXiv preprint arXiv:1705.07874, 2017

    2017 arXiv

  21. [29]

    Strategic classification is causal modeling in disguise

    John Miller, Smitha Milli, and Moritz Hardt. Strategic classification is causal modeling in disguise. In International Conference on Machine Learning, pages 6917–6926. PMLR, 2020

    2020

  22. [30]

    The social cost of strategic classification

    Smitha Milli, John Miller, Anca D Dragan, and Moritz Hardt. The social cost of strategic classification. In Proceedings of the Conference on Fairness, Accountability, and Transparency , pages 230–239, 2019

    2019

  23. [31]

    Interpretable machine learning

    Christoph Molnar. Interpretable machine learning. Lulu. com, 2020

    2020

  24. [32]

    Performative prediction

    Juan Perdomo, Tijana Zrnic, Celestine Mendler-Dünner, and Moritz Hardt. Performative prediction. In International Conference on Machine Learning, pages 7599–7609. PMLR, 2020

    2020

  25. [33]

    why should i trust you?

    Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. " why should i trust you?" explaining the predictions of any classifier. InProceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135–1144, 2016. 12 Sufficient Explanati...

    2016

  26. [34]

    From predictions to decisions: Using lookahead regularization

    Nir Rosenfeld, Anna Hilgard, Sai Srivatsa Ravindranath, and David C Parkes. From predictions to decisions: Using lookahead regularization. Advances in Neural Information Processing Systems, 33:4115–4126, 2020

    2020

  27. [35]

    Meaningful information and the right to explanation

    Andrew D Selbst and Julia Powles. Meaningful information and the right to explanation. International Data Privacy Law, 7(4):233–242, 2017

    2017

  28. [36]

    Strategic classification under unknown person- alized manipulation

    Han Shao, Avrim Blum, and Omar Montasser. Strategic classification under unknown person- alized manipulation. Advances in Neural Information Processing Systems, 36, 2024

    2024

  29. [37]

    Causal strategic linear regression

    Yonadav Shavit, Benjamin Edelman, and Brian Axelrod. Causal strategic linear regression. In International Conference on Machine Learning, pages 8676–8686. PMLR, 2020

    2020

  30. [38]

    Pac-learning for strategic classifica- tion

    Ravi Sundaram, Anil Vullikanti, Haifeng Xu, and Fan Yao. Pac-learning for strategic classifica- tion. Journal of Machine Learning Research, 24(192):1–38, 2023

    2023

  31. [39]

    Decisions, counterfactual explanations and strategic behavior

    Stratis Tsirtsis and Manuel Gomez Rodriguez. Decisions, counterfactual explanations and strategic behavior. Advances in Neural Information Processing Systems, 33:16749–16760, 2020

    2020

  32. [40]

    Causal strategic learning with competitive selection

    Kiet QH Vo, Muneeb Aadil, Siu Lun Chau, and Krikamol Muandet. Causal strategic learning with competitive selection. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 38, pages 15411–15419, 2024

    2024

  33. [41]

    Why a right to explanation of auto- mated decision-making does not exist in the general data protection regulation

    Sandra Wachter, Brent Mittelstadt, and Luciano Floridi. Why a right to explanation of auto- mated decision-making does not exist in the general data protection regulation. International data privacy law, 7(2):76–99, 2017

    2017

  34. [42]

    Counterfactual explanations without opening the black box: Automated decisions and the gdpr

    Sandra Wachter, Brent Mittelstadt, and Chris Russell. Counterfactual explanations without opening the black box: Automated decisions and the gdpr. Harv. JL & Tech., 31:841, 2017

    2017

  35. [43]

    Non-linear welfare-aware strategic learning

    Tian Xie and Xueru Zhang. Non-linear welfare-aware strategic learning. arXiv preprint arXiv:2405.01810, 2024

    2024 arXiv

  36. [44]

    Learning under imitative strategic behavior with unforeseeable outcomes

    Tian Xie, Zhiqun Zuo, Mohammad Mahdi Khalili, and Xueru Zhang. Learning under imitative strategic behavior with unforeseeable outcomes. arXiv preprint arXiv:2405.01797, 2024

    2024 arXiv

  37. [45]

    Modeling tabular data using conditional gan

    Lei Xu, Maria Skoularidou, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni. Modeling tabular data using conditional gan. In Advances in Neural Information Processing Systems, 2019. 13 Sufficient Explanations that Induce Non-harmful Responses A Preprint A Additional illustrat...

    2019

  38. [46]

    Computational details

    For simplicity, we generate the AR-based explanations randomly as follows: ⃗Xt∼N (0,0.4), ˆ⃗Yt :=g(⃗Xt). Computational details. This experiment took less than 5 seconds to run on a standard MacBook Pro with an M2 chip and 16GB of RAM. 18 Sufficient Explanations that Induce Non...

  39. [47]

    Then, we interact with the next 106 agents to collect another data setD2 ={ ¨xt,⃗ xt,wt, ∆⃗ gt}t∈[106]

    We use a datasetD1 ={xt,yt}t∈[5000] for this step. Then, we interact with the next 106 agents to collect another data setD2 ={ ¨xt,⃗ xt,wt, ∆⃗ gt}t∈[106]. The collected dataset will later be used to train the compliance predictorξ. To do this, we construct a samplerπ to genera...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.