Pith. sign in

REVIEW 2 major objections 3 minor 108 references

Statistical Inference for Responsiveness Verification

T0 review · 2 major / 3 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper introduces a formal, black-box procedure to estimate and test how much a model's predictions respond to realistic interventions on an individual's features, with controlled false-claim rates.

desk verdict The paper's statistical framework is sound and its empirical comparisons are useful, but Algorithm 1 as printed does not sample from the claimed distribution, so the central guarantees don't apply to its output as written. read the letter →

arxiv 2507.02169 v1 pith:TOBTS6WN submitted 2025-07-02 cs.LG

Seung Hyun Cheon , Meredith Stewart , Bogdan Kulynych , Tsui-Wei Weng , Berk Ustun This is my paper · ORCID
classification cs.LG
keywords responsivenessverificationreachablesetsinterventionsanddownstreameffectsbinomialconfidenceintervalsblack-boxmodelauditingpreclusiongamingcounterfactualinvariance
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

The paper introduces a statistical validation procedure for a quantity it calls responsiveness: the proportion of feasible interventions at a given point that move a model's prediction into a target set, such as a denied loan applicant reaching approval or a bot account reaching 'human'. The central claim is that responsiveness can be estimated and tested for any black-box model by sampling reachable points and counting successes, which makes the estimate binomially distributed and gives exact confidence intervals and type-I error bounds. This matters because responsiveness failures such as preclusion, gaming, and monotonicity violations are safety problems in lending, hiring, organ allocation, and content moderation, and the procedure exposes them with only query access. The paper demonstrates the machinery on recidivism prediction, the UK liver-transplant benefit score, and bot detection, where it finds fixed predictions, monotonicity violations, and a convex-relaxation audit that under-reports gaming.

What carries the argument

The load-bearing object is the reachable set $X^{\mathrm{reach}}_A(x)$, defined through interventions $a \in A(x)$ and downstream effects $r \sim P_{x,a}$ as $x' = x + a + r$, and the load-bearing statistical tool is the exact Clopper-Pearson binomial confidence interval applied to the count $\hat S_n$ of sampled reachable points whose predictions land in $\hat Y^{\mathrm{reach}}_x$. Algorithm 1 generates these points by sampling interventions uniformly from $A(x)$, enforcing feature-level and joint constraints such as integrality, monotonicity, encoding, and directional-linkage constraints with a mixed-integer feasibility check, and sampling downstream effects; this uniform sampling is what makes $\hat S_n$ binomial and the finite-sample guarantees valid regardless of model type.

What would settle it

Run the procedure on a finite discrete reachable set where true responsiveness can be enumerated, as the paper does on its credit-data validation: if the empirical coverage of the Clopper-Pearson interval across many independent runs is materially below $1-\alpha$, or if the observed false-claim rate exceeds $\alpha$, then the i.i.d. uniform-sampling premise behind $\hat S_n \sim \mathrm{Bin}(n, \rho(x))$ is broken.

Watch

Extended reading notes

Core claim

The discovery is that responsiveness verification reduces to binomial inference. Given $n$ i.i.d. reachable points sampled uniformly over the intervention set, the number of points whose prediction falls in the target set is $\mathrm{Bin}(n, \rho(x))$, so the Clopper-Pearson interval of Proposition 3 covers the true responsiveness with probability at least $1-\alpha$, and the one-sided test of Proposition 5 rejects the claim 'responsiveness at least $\varepsilon$' with false-rejection probability at most $\alpha$. The paper further gives a necessary sample-size bound, $n > \log \alpha / \log(1-\varepsilon)$, for the test ever to reject, and an algorithm that samples a non-convex reachable set over discrete and continuous features by sampling interventions, checking feasibility through a mixed-integer program, and drawing downstream effects from deterministic, random, or causal models. Because the estimator uses only black-box queries, the guarantees hold for any model class.

Load-bearing premise

The practitioner must be able to specify the set of interventions a real decision subject could actually perform and how each intervention changes downstream features; if that specification is unfaithful, the statistically valid verdicts are not meaningful.

Editorial extensions

If this is right

  • Any model with black-box query access can be audited for preclusion, gaming, and monotonicity with a declared confidence level, because the estimator and tests do not depend on model architecture.
  • Practitioners can choose sample size before running the audit: $N^{\min}(\alpha, L)$ controls interval width, and $N^{\min}(\alpha, \beta, \varepsilon, \Delta)$ controls the power to detect unresponsiveness.
  • Uniform sampling can replace exhaustive enumeration on large discrete reachable sets, cutting storage and compute while keeping estimation error and false-claim rates bounded; the validation study reports about $4\%$ absolute error with $n=30$.
  • Every failed test comes with concrete reachable points that falsify the model-level claim, which supports debugging, regression testing, and refining the intervention specification.
  • Validating responsiveness with a convex relaxation of the intervention set can certify a model as robust when the exact intervention model shows that more than half of bot predictions are responsive, so exact reachable sets matter for safety verdicts.

Reading between the lines

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

  • If the specification burden can be met, the same binomial machinery extends naturally to any target predicate, including regression intervals, multiclass sets, or arbitrary safety conditions, so responsiveness could serve as a general validation primitive across model types.
  • The minimal-response-model idea suggests a robustness property worth formalizing: if preclusion is detected under a conservative intervention set, responsiveness under any superset is at least as large, so conservative failures imply failures under realistic interventions; a proof would strengthen the auditing interpretation.
  • The paper notes its sampler is uniform and not optimized for finding counterexamples near the decision boundary; pairing the same binomial confidence machinery with importance sampling or boundary-aware proposals would make failure discovery more efficient, at the cost of weighting the counts.
  • A natural stress test for deployment would be to re-run the audit under several intervention models, narrow, neutral, and broad, and report verdicts that are stable across the family; this would convert the specification assumption from a caveat into a quantified sensitivity analysis.
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

2 major / 3 minor

Summary. The paper introduces a statistical framework for auditing the responsiveness of black-box model predictions to feature interventions. Responsiveness is defined as the probability that a point sampled from a user-specified reachable distribution P^reach_x, obtained by sampling an intervention uniformly from A(x) and then drawing downstream effects from P_{x,a}, falls into a target prediction set. The authors propose estimating this quantity from i.i.d. samples via the Clopper-Pearson interval (Proposition 3) and a one-sided hypothesis test for whether responsiveness exceeds a threshold (Proposition 5), with sample-size rules. They present Algorithm 1 for generating the reachable samples, validate the estimation and testing guarantees against full enumeration on the german dataset, and demonstrate the procedure on recidivism prediction, liver transplant scoring, and content moderation.

Significance. The statistical core of the paper is sound and self-contained: given genuinely i.i.d. samples from P^reach_x, the Clopper-Pearson interval and the one-sided test provide exactly the stated coverage and type-I error bounds, and the validation study against full enumeration on the german dataset is a genuine external check. If the sampling algorithm actually produced such samples, the framework would be a practical, model-agnostic tool for preclusion, gaming, and monotonicity verification. The main weaknesses are internal inconsistencies between the printed Algorithm 1 and the distributional assumption, and a use case that violates the paper's own sample-size condition. Both are fixable without changing the statistical theory.

major comments (2)
  1. [§3.2 (Algorithm 1) and Propositions 3/5] Algorithm 1 as printed does not sample from P^reach_x. In Line 4, CheckFeasibility(x, a, C_S) is applied to the intervention vector a before downstream effects r are sampled in Line 5, and Line 6 adds x+a+r without any subsequent feasibility check. For constraints that couple an intervention to downstream features—such as the Directional Linkage equalities in Eqs. (3)–(4) or the Missing Values row of Table 1—this check is wrong: a valid intervention that requires a compensating downstream change is rejected (e.g., setting no_posts=TRUE forces num_posts=0, but with a_num_posts=0 the constraint x_j·x_k=0 cannot be verified on a alone), and, conversely, when no coupling is present, the missing final check can accept x+a+r outside the reachable set. Consequently, the accepted samples are not distributed as a∼Uniform[A(x)] and r∼P_{x,a}; the resulting S_n is not necessarily Bin(n, ρ(x)), and the guarantees of Propositions 3 and 5 do not apply to the printed algorithm's output. Appendix A.1 states that the implementation instead checks feasibility on the final intervention a+r, which would fix the distributional claim, but this is not reflected in the main-text Algorithm 1 or its stated guarantees. Please revise Algorithm 1 to include the final feasibility check and state explicitly what distribution the accepted samples follow (e.g., conditional on x+a+r satisfying the constraints).
  2. [§4.1 (recidivism)] The text states 'We construct reachable sets with 20 samples per point, satisfying Remark 7.' This is numerically false. For ε=0.1 and α=0.05, Remark 7 requires n > log(0.05)/log(0.9) ≈ 28.4, so n=20 is below the necessary threshold; the formal test can never reject H0 for any point. As a result, the claim that 15,986 of 18,614 recidivism predictions are 'fixed' cannot be supported by the testing procedure described in Section 3. If the authors instead mean an estimation-based criterion (e.g., ρ̂<0.01), they should say so explicitly and separate that analysis from the formal hypothesis test, which is invalid at this sample size.
minor comments (3)
  1. [Figure 2 caption] The caption defines Specificity as P(Claim Responsive | ρ ≥ ε) and calls it 'analogous to statistical power: 1−β', while main text Section 3.2 calls the same quantity 'the probability of correctly identifying responsiveness (Specificity)' and claims it is at least 1−α. These are inconsistent: P(Claim Responsive | ρ≥ε) is 1 minus the type-I error, i.e., specificity (≥1−α), not power. Please correct the caption or the main-text terminology so that Specificity and Recall are labeled consistently with their formal definitions.
  2. [Appendix B.1.2] The numbered constraint list contains apparent duplicates: items 25–27 repeat items 13–15 (both describe Directional Linkage from prior_arrests≥2, prior_arrests=1, and prior_arrests≥5 to prior_arrests_for_felony). This makes the stated count of 27 constraints inconsistent with the distinct constraints listed. Please reconcile the numbering or remove the duplicates.
  3. [Throughout] The paper refers to a provided Python library and to the enumeration procedure from Kothari et al. [32], but it does not specify the software version or reproducibility commands; adding a brief 'Reproducibility' statement or an artifact appendix would help readers verify the validation study in Appendix A.2.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: statistical derivation is self-contained and validated against external enumeration.

full rationale

The paper's load-bearing results, Proposition 3 and Proposition 5, are standard exact binomial inference (Clopper-Pearson intervals and one-sided tests). Proposition 3 states a confidence interval for rho(x) under the explicit assumption that the n reachable points are i.i.d. from P^reach_x; Proposition 5 bounds type-I error under the same binomial assumption. These results do not define rho in terms of the estimator, fit a parameter to the quantity being predicted, or invoke a self-citation to force the conclusion. The estimation is Monte Carlo evaluation of a Bernoulli probability, which is a legitimate, non-circular derivation. The validation study in Appendix A.2 compares sampled estimates to ground-truth responsiveness obtained by full enumeration of reachable sets on the german dataset; the enumeration procedure from Kothari et al. [32] is an external benchmark rather than an input fitted into the sampling estimate. Self-citations to [9], [32], and [35] are contextual building blocks and are not load-bearing in the statistical argument. A separate correctness concern, not a circularity, is that Algorithm 1 as printed checks feasibility on the intervention vector a before sampling downstream effects r and does not re-check x+a+r, while Appendix A.1 states the implementation checks feasibility on a+r; this is an internal consistency issue about whether the sampler produces the intended distribution, not an equivalence-by-construction between the paper's inputs and its conclusions.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central estimation procedure is self-contained and introduces no fitted parameters. The case-study SCM contains hand-chosen constants, and the method depends on the domain assumption that intervention models are correctly specified.

free parameters (1)
  • Transplant simulation SCM parameters = Sigma in Eq. (11), constants in Eqs. (12)-(15)
    Hand-chosen to approximately match the simulated cohort statistics in Attia et al. [3]; they define the synthetic cohort used in the Section 4.2 case study, not the central estimation procedure.
assumptions (3)
  • standard math Clopper-Pearson exact binomial confidence intervals are valid for i.i.d. Bernoulli samples.
    Propositions 3 and 5 rely on S_n ~ Bin(n, rho(x)) and the exact method [11].
  • domain assumption Features are semantically meaningful and a practitioner can specify a plausible intervention set A(x) and downstream effect distribution P_{x,a}.
    Required by Definition 2 and throughout Section 2; the paper acknowledges that misspecification biases responsiveness.
  • domain assumption The user-specified intervention model captures real-world actionability well enough that 'minimal response models' are indisputable.
    Section 2 'Discussion' proposes minimal response models to avoid blatant misspecification, but the validity of any responsiveness claim depends on this specification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Statistical Inference for Responsiveness Verification." pith.science (2026). https://pith.science/paper/TOBTS6WN

@misc{pith2026250702169,
  author       = {Pith},
  title        = {Pith review of: Statistical Inference for Responsiveness Verification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TOBTS6WN}},
  note         = {Machine review of arXiv:2507.02169}
}
read the original abstract

Many safety failures in machine learning arise when models are used to assign predictions to people (often in settings like lending, hiring, or content moderation) without accounting for how individuals can change their inputs. In this work, we introduce a formal validation procedure for the responsiveness of predictions with respect to interventions on their features. Our procedure frames responsiveness as a type of sensitivity analysis in which practitioners control a set of changes by specifying constraints over interventions and distributions over downstream effects. We describe how to estimate responsiveness for the predictions of any model and any dataset using only black-box access, and how to use these estimates to support tasks such as falsification and failure probability estimation. We develop algorithms that construct these estimates by generating a uniform sample of reachable points, and demonstrate how they can promote safety in real-world applications such as recidivism prediction, organ transplant prioritization, and content moderation.

Figures

Figures reproduced from arXiv: 2507.02169 by the authors.

Figure 1
Figure 1. Responsiveness verification with reachable sets. Left: Given an instance x, we generate a uniform sample of reachable points Xreach A (x), i.e., feature vectors that can be reached following an intervention on x. Here, x1 ∈ [10000, 60000] and x2 ∈ {0, . . . , 5} are monotonically increasing features. Middle: We use Xreach A (x) to determine the vulnerability to gaming in a bot detection task; accounts that are flagg… view at source ↗
Figure 2
Figure 2. Convergence of responsiveness estimates and test metrics for a lending model built from the german dataset [13]. We compute the true responsiveness of all instances in the dataset by enumeration, build sampled reachable sets to estimate and test responsiveness (ε = 0.1, α = 0.05). Left: Absolute Estimation Error (|ρˆn − ρ|). Middle: Specificity (P(Claim Responsive | ρ ≥ ε), analogous to statistical power: 1−β) and R… view at source ↗
Figure 3
Figure 3. Distribution of unresponsive predictions in demographic groups. Left: Train sample. Middle: Test sample. Right: CDF of responsiveness proportion by demographic group 4.2 Testing Counterfactual Invariance in Organ Transplant Prioritization Predictive statistical models are routinely used in allocation of organ transplants [23]. Recently, they have attracted scrutiny both from the public and the academic circles becau… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Average proportion of predictions that violate monotonicity across a cohort of simulated non-cancer patients, across two intervention sets in which counterfactual simulated patients are assigned cancer with size of either < 2cm (“small”) or < 5cm (“large”). Error bars …
Figure 5
Figure 5. Figure 5: CDF of points by responsiveness percentage 10 [PITH_FULL_IMAGE:figures/full_fig_p027_5.png]
Figure 6
Figure 6. Figure 6: Test AUC of the best model that has less than 10% "Bot" predictions that have higher responsiveness than ε = 0.01, 0.05, 0.1, 0.2 for each procedure. Model does not change for Inspection since features are immutable. 12 [PITH_FULL_IMAGE:figures/full_fig_p029_6.png]
Figure 7
Figure 7. Figure 7: Pairwise relationships of the four liver parameter distributions according to our probabilistic model. These 1 statistics are similar to those obtained by Attia et al. [3]. where ε ∼ N (0, Σ) represents correlated noise. The counterfactual endogenous variables x (1) ar…
Figure 8
Figure 8. Figure 8: Average predictions of the TBS model and its components (need model on the left, utility model in the middle, combined on the right) over the reachable sets in the simulated cohorts. We can see that only for the middle-age group the average predicted survival w/o trans…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

108 extracted references · 72 canonical work pages

  1. [1]

    Fake news, disinformation and misinformation in social media: a review.Social Network Analysis and Mining, 13(1):30, 2023

    Esma Aïmeur, Sabrine Amri, and Gilles Brassard. Fake news, disinformation and misinformation in social media: a review.Social Network Analysis and Mining, 13(1):30, 2023

    2023

  2. [2]

    Effect of recipient age on prioritisation for liver transplantation in the uk: a population-based modelling study.The Lancet Healthy Longevity, 5 (5):e346–e355, 2024

    Anthony Attia, Jamie Webb, Katherine Connor, Chris JC Johnston, Michael Williams, Tim Gordon- Walker, Ian A Rowe, Ewen M Harrison, and Ben M Stutchfield. Effect of recipient age on prioritisation for liver transplantation in the uk: a population-based modelling study.The Lancet Healthy Longevity, 5 (5):e346–e355, 2024

    2024

  3. [3]

    Implausible algorithm output in uk liver transplantation allocation scheme: importance of transparency.The Lancet, 401(10380):911–912, 2023

    Antony Attia, Ian A Rowe, Ewen M Harrison, Tim Gordon-Walker, and Ben M Stutchfield. Implausible algorithm output in uk liver transplantation allocation scheme: importance of transparency.The Lancet, 401(10380):911–912, 2023

    2023

  4. [4]

    Kentucky pretrial risk assessment instrument validation

    James Austin, Roger Ocker, and Avi Bhati. Kentucky pretrial risk assessment instrument validation. Bureau of Justice Statistics, 2010

    2010

  5. [5]

    Monotonicity Maintenance in Information-Theoretic Machine Learning Algorithms

    Arie Ben-David. Monotonicity Maintenance in Information-Theoretic Machine Learning Algorithms. Machine Learning, 19(1):29–43, 1995

    1995

  6. [6]

    Help wanted: An examination of hiring algorithms, equity, and bias

    Miranda Bogen and Aaron Rieke. Help wanted: An examination of hiring algorithms, equity, and bias. Upturn, December, 7, 2018

    2018

  7. [7]

    Young people wait four times longer for liver transplants, 2023

    Cathering Burns and Vicki Loader. Young people wait four times longer for liver transplants, 2023

    2023

  8. [8]

    Model transferability with responsive decision subjects

    Yatong Chen, Zeyu Tang, Kun Zhang, and Yang Liu. Model transferability with responsive decision subjects. In International Conference on Machine Learning, pages 4921–4952. PMLR, 2023

    2023

Show all 108 references
  1. [9]

    Feature responsiveness scores: Model-agnostic explanations for recourse

    Seung Hyun Cheon, Anneke Wernerfelt, Sorelle Friedler, and Berk Ustun. Feature responsiveness scores: Model-agnostic explanations for recourse. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  2. [10]

    Pre-existing conditions and medical underwriting in the individual insurance market prior to the aca.Menlo Park, CA, 2016: 1–11, 2016

    Gary Claxton, Cynthia Cox, Anthony Damico, Larry Levitt, and Karen Pollitz. Pre-existing conditions and medical underwriting in the individual insurance market prior to the aca.Menlo Park, CA, 2016: 1–11, 2016

    2016

  3. [11]

    The use of confidence or fiducial limits illustrated in the case of the binomial

    Charles J Clopper and Egon S Pearson. The use of confidence or fiducial limits illustrated in the case of the binomial. Biometrika, 26(4):404–413, 1934

    1934

  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. ACM, 2018

    2018

  5. [13]

    UCI machine learning repository, 2017

    Dheeru Dua and Casey Graff. UCI machine learning repository, 2017. URLhttp://archive.ics.uci. edu/ml

    2017

  6. [14]

    Sacrificing accuracy for transparency in recidivism risk assessment: The impact of classification method on predictive performance.Corrections, 1(3):155–176, 2016

    Grant Duwe and KiDeuk Kim. Sacrificing accuracy for transparency in recidivism risk assessment: The impact of classification method on predictive performance.Corrections, 1(3):155–176, 2016

    2016

  7. [15]

    Constructing recidivism risk.Emory LJ, 67:59, 2017

    Jessica M Eaglin. Constructing recidivism risk.Emory LJ, 67:59, 2017

    2017

  8. [16]

    Incentivizing recourse through auditing in strategic classification

    Andrew Estornell, Yatong Chen, Sanmay Das, Yang Liu, and Yevgeniy Vorobeychik. Incentivizing recourse through auditing in strategic classification. In Edith Elkind, editor,Proceedings of the Thirty- Second International Joint Conference on Artificial Intelligence, IJCAI-23, pa...

    2023 doi

  9. [17]

    Regularization paths for generalized linear models via coordinate descent.Journal of Statistical Software, 33(1):1–22, 2010

    Jerome Friedman, Trevor Hastie, and Robert Tibshirani. Regularization paths for generalized linear models via coordinate descent.Journal of Statistical Software, 33(1):1–22, 2010

    2010

  10. [18]

    Strategic classification in the dark

    Ganesh Ghalme, Vineet Nair, Itay Eilat, Inbal Talgam-Cohen, and Nir Rosenfeld. Strategic classification in the dark. InInternational Conference on Machine Learning, pages 3672–3681. PMLR, 2021

    2021

  11. [19]

    Classification of twitter accounts into automated agents and human users

    Zafar Gilani, Ekaterina Kochmar, and Jon Crowcroft. Classification of twitter accounts into automated agents and human users. InProceedings of the 2017 IEEE/ACM international conference on advances in social networks analysis and mining 2017, pages 489–496, 2017

    2017

  12. [20]

    Content moderation, ai, and the question of scale

    Tarleton Gillespie. Content moderation, ai, and the question of scale. Big Data & Society, 7(2): 2053951720943234, 2020

    2020

  13. [21]

    Explaining and harnessing adversarial examples

    Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. ICLR, 2015

    2015

  14. [22]

    Algorithmic content moderation: Technical and political challenges in the automation of platform governance.Big Data & Society, 7(1):2053951719897945, 2020

    Robert Gorwa, Reuben Binns, and Christian Katzenbach. Algorithmic content moderation: Technical and political challenges in the automation of platform governance.Big Data & Society, 7(1):2053951719897945, 2020

    2020

  15. [23]

    The promise of machine learning applications in solid organ transplantation.NPJ digital medicine, 5(1):89, 2022

    Neta Gotlieb, Amirhossein Azhie, Divya Sharma, Ashley Spann, Nan-Ji Suo, Jason Tran, Ani Orchanian- Cheff, Bo Wang, Anna Goldenberg, Michael Chassé, et al. The promise of machine learning applications in solid organ transplantation.NPJ digital medicine, 5(1):89, 2022

    2022

  16. [24]

    Monotonic calibrated interpolated look-up tables

    Maya Gupta, Andrew Cotter, Jan Pfeifer, Konstantin Voevodski, Kevin Canini, Alexander Mangylov, Wojciech Moczydlowski, and Alexander Van Esbroeck. Monotonic calibrated interpolated look-up tables. Journal of Machine Learning Research, 17(109):1–47, 2016

    2016

  17. [25]

    Strategic classification

    Moritz Hardt, Nimrod Megiddo, Christos Papadimitriou, and Mary Wootters. Strategic classification. In Proceedings of the 2016 ACM Conference on Innovations in Theoretical Computer Science, pages 111–122. ACM, 2016

    2016

  18. [26]

    Prior record and recidivism risk.American Journal of Criminal Justice, 44:353–375, 2019

    Rhys Hester. Prior record and recidivism risk.American Journal of Criminal Justice, 44:353–375, 2019

    2019

  19. [27]

    Credit scoring in the era of big data.Yale JL & Tech., 18:148, 2016

    Mikella Hurley and Julius Adebayo. Credit scoring in the era of big data.Yale JL & Tech., 18:148, 2016

    2016

  20. [28]

    The model for end-stage liver disease (meld).Hepatology, 45(3): 797–805, 2007

    Patrick S Kamath and W Ray Kim. The model for end-stage liver disease (meld).Hepatology, 45(3): 797–805, 2007

    2007

  21. [29]

    Adversarial robustness for tabular data through cost and utility awareness

    Klim Kireev, Bogdan Kulynych, and Carmela Troncoso. Adversarial robustness for tabular data through cost and utility awareness. InNetwork and Distributed System Security (NDSS) Symposium, 2023

    2023

  22. [30]

    Kochenderfer, Sydney M

    Mykel J. Kochenderfer, Sydney M. Katz, Anthony L. Corso, and Robert J. Moss. Algorithms for validation. https://algorithmsbook.com/validation/files/val.pdf, 2025. GitHub repository PDF, accessed May 16, 2025

    2025

  23. [31]

    Understanding black-box predictions via influence functions

    Pang Wei Koh and Percy Liang. Understanding black-box predictions via influence functions. InICML, 2017

    2017

  24. [32]

    Prediction without preclusion: Recourse verification with reachable sets

    Avni Kothari, Bogdan Kulynych, Tsui-Wei Weng, and Berk Ustun. Prediction without preclusion: Recourse verification with reachable sets. InThe Twelfth International Conference on Learning Repre- sentations, 2024

    2024

  25. [33]

    Counterfactual fairness.Advances in neural information processing systems, 30, 2017

    Matt J Kusner, Joshua Loftus, Chris Russell, and Ricardo Silva. Counterfactual fairness.Advances in neural information processing systems, 30, 2017

    2017

  26. [34]

    The creation and validation of the ohio risk assessment system (oras).Fed

    Edward J Latessa, Richard Lemke, Matthew Makarios, and Paula Smith. The creation and validation of the ohio risk assessment system (oras).Fed. Probation, 74:16, 2010

    2010

  27. [35]

    Understanding fixed predictions via confined regions, 2025

    Connor Lawless, Tsui-Wei Weng, Berk Ustun, and Madeleine Udell. Understanding fixed predictions via confined regions, 2025. URLhttps://arxiv.org/abs/2502.16380. 14 Statistical Inference for Responsiveness Verification

    2025 arXiv

  28. [36]

    Strategic classification made practical

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

    2021

  29. [37]

    A human-is-the-loop approach for semi-automated content moderation

    Daniel Link, Bernd Hellingrath, and Jie Ling. A human-is-the-loop approach for semi-automated content moderation. In ISCRAM, 2016

    2016

  30. [38]

    Towards deep learning models resistant to adversarial attacks.ICLR, 2018

    Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks.ICLR, 2018

    2018

  31. [39]

    Predictive multiplicity in classification

    Charles Marx, Flavio Calmon, and Berk Ustun. Predictive multiplicity in classification. InProceedings of Machine Learning and Systems 2020, pages 9215–9224. 2020

    2020

  32. [40]

    But are you sure? an uncertainty-aware perspective on explainable ai

    Charles Marx, Youngsuk Park, Hilaf Hasson, Yuyang Wang, Stefano Ermon, and Luke Huan. But are you sure? an uncertainty-aware perspective on explainable ai. InInternational Conference on Artificial Intelligence and Statistics, pages 7375–7391. PMLR, 2023

    2023

  33. [41]

    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

  34. [42]

    Algorithms are deciding who gets organ transplants

    Madhumita Murgia. Algorithms are deciding who gets organ transplants. are their decisions fair?, 2023

    2023

  35. [43]

    Unsafe at Any Speed: The Designed-in Dangers of the American Automobile

    Ralph Nader. Unsafe at Any Speed: The Designed-in Dangers of the American Automobile. Pocket Books, 1966

    1966

  36. [44]

    Regretful decisions under label noise

    Sujay Nagaraj, Yang Liu, Flavio P Calmon, and Berk Ustun. Regretful decisions under label noise. arXiv preprint arXiv:2504.09330, 2025

    2025 arXiv

  37. [45]

    Distributionally robust parametric maximum likelihood estimation, 2020

    Viet Anh Nguyen, Xuhui Zhang, Jose Blanchet, and Angelos Georghiou. Distributionally robust parametric maximum likelihood estimation, 2020. URLhttps://arxiv.org/abs/2010.05321

    2020 arXiv

  38. [46]

    Causality: Models, Reasoning and Inference

    Judea Pearl. Causality: Models, Reasoning and Inference. Cambridge University Press, USA, 2nd edition, 2009. ISBN 052189560X

    2009

  39. [47]

    Sentence Risk Assessment Instrument, April 2017

    Pennsylvania Bulletin. Sentence Risk Assessment Instrument, April 2017

    2017

  40. [48]

    Mitigating bias in algorithmic hiring: Evaluating claims and practices

    Manish Raghavan, Solon Barocas, Jon Kleinberg, and Karen Levy. Mitigating bias in algorithmic hiring: Evaluating claims and practices. InProceedings of the 2020 conference on fairness, accountability, and transparency, pages 469–481, 2020

    2020

  41. [49]

    Automated trouble: The role of algorithmic selection in harms on social media platforms.Media and Communication, 9(4):222–233, 2021

    Florian Saurwein and Charlotte Spencer-Smith. Automated trouble: The role of algorithmic selection in harms on social media platforms.Media and Communication, 9(4):222–233, 2021

    2021

  42. [50]

    Constrained adaptive attack: Effective adversarial attack against deep neural networks for tabular data.arXiv preprint arXiv:2406.00775, 2024

    Thibault Simonetto, Salah Ghamizi, and Maxime Cordy. Constrained adaptive attack: Effective adversarial attack against deep neural networks for tabular data.arXiv preprint arXiv:2406.00775, 2024

    2024 arXiv

  43. [51]

    Saver: A toolbox for sampling-based, probabilistic verification of neural networks

    Vignesh Sivaramakrishnan, Krishna C Kalagarla, Rosalyn Devonport, Joshua Pilipovsky, Panagiotis Tsiotras, and Meeko Oishi. Saver: A toolbox for sampling-based, probabilistic verification of neural networks. arXiv preprint arXiv:2412.02940, 2024

    2024 arXiv

  44. [52]

    Counterfactual explanations for arbitrary regression models.arXiv preprint arXiv:2106.15212, 2021

    Thomas Spooner, Danial Dervovic, Jason Long, Jon Shepard, Jiahao Chen, and Daniele Magazzeni. Counterfactual explanations for arbitrary regression models.arXiv preprint arXiv:2106.15212, 2021

    2021 arXiv

  45. [53]

    Ai and the risk of consumer harm

    Staff in the Office of Technology and the Division of Advertising Practices. Ai and the risk of consumer harm. https://www.ftc.gov/policy/advocacy-research/tech-at-ftc/2025/01/ ai-risk-consumer-harm, January 3 2025

    2025

  46. [54]

    Interpretable predictions of tree-based ensembles via actionable feature tweaking

    Gabriele Tolomei, Fabrizio Silvestri, Andrew Haines, and Mounia Lalmas. Interpretable predictions of tree-based ensembles via actionable feature tweaking. InProceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining, pages 465–474, 2017. 15

    2017

  47. [55]

    Office Of Justice Programs

    United States Department Of Justice. Office Of Justice Programs. Bureau Of Justice Statistics. Recidi- vism of prisoners released in 1994, 2002. URLhttps://www.icpsr.umich.edu/web/NACJD/studies/3355/ versions/V8

    1994

  48. [56]

    Actionable recourse in linear classification

    Berk Ustun, Alexander Spangher, and Yang Liu. Actionable recourse in linear classification. pages 10–19, 2019. doi: 10.1145/3287560.3287566

    2019

  49. [57]

    Counterfactual invariance to spurious correlations: Why and how to pass stress tests.arXiv preprint arXiv:2106.00545, 2021

    Victor Veitch, Alexander D’Amour, Steve Yadlowsky, and Jacob Eisenstein. Counterfactual invariance to spurious correlations: Why and how to pass stress tests.arXiv preprint arXiv:2106.00545, 2021

    2021 arXiv

  50. [58]

    Proven: Verifying robustness of neural networks with a probabilistic approach

    Tsui-Wei Weng, Pin-Yu Chen, Lam Nguyen, Mark Squillante, Akhilan Boopathy, Ivan Oseledets, and Luca Daniel. Proven: Verifying robustness of neural networks with a probabilistic approach. InICML, 2019

    2019

  51. [59]

    John Wiley & Sons, 2020

    Laurence A Wolsey.Integer programming. John Wiley & Sons, 2020

    2020

  52. [60]

    Interpretable classification models for recidivism prediction

    Jiaming Zeng, Berk Ustun, and Cynthia Rudin. Interpretable classification models for recidivism prediction. Journal of the Royal Statistical Society: Series A (Statistics in Society), 180(3):689–722, 2017

    2017

  53. [61]

    source variable

    Fei Zhang, Patrick PK Chan, Battista Biggio, Daniel S Yeung, and Fabio Roli. Adversarial feature selection against evasion attacks.IEEE transactions on cybernetics, 46(3):766–777, 2015. 16 A Supplementary Material for Section 3 1 A.1 Description of Routines in Algorithm 1 . . ...

    2015

  54. [62]

    [0, 0, 0, 0]: account is less than 180 days old

  55. [63]

    [1, 0, 0, 0]: account is older than 180 days but less than 365 days old

  56. [64]

    [1, 1, 0, 0]: account is older than 365 days but less than 730 days old

  57. [65]

    [1, 1, 1, 0]: account is older than 730 days but less than 1825 days old

  58. [66]

    reliability

    [1, 1, 1, 1]: account is more than 1825 days old Given an ordered set of feasible values V, like above, we also define a reachability matrix E ∈ {0, 1}|V |×|V |, where the(i, j)-th entry ofE is 1 when we can reach from thei-th element ofV to its jth element and 0 otherwise. No...

    1994

  59. [67]

    YearsAtResidence and 20 YearsEmployed≥1, which form a partition. • Thermometer Encoding constraints enforce conceptual requirements in this dataset - 1) requir- ing CheckingAcct≥0=True to be reachable only ifCheckingAcct_exists is also True, and 2) requiring SavingsAcct≥100=Tr...

  60. [68]

    Each unit change in YearsAtResidence leads to a unit change inAge

    DirectionalLinkage: Actions onYearsAtResidence will induce actions on [‘Age’]. Each unit change in YearsAtResidence leads to a unit change inAge

  61. [69]

    Each unit change in YearsEmployed≥1 leads to a unit change inAge

    DirectionalLinkage: Actions onYearsEmployed≥1 will induce actions on [‘Age’]. Each unit change in YearsEmployed≥1 leads to a unit change inAge

  62. [70]

    Actions can only turn on higher-level dummies that are off, whereCheckingAcctexists is the lowest-level dummy andCheckingAcct≥0 is the highest-level-dummy

    ThermometerEncoding: Actions on [CheckingAcctexists, CheckingAcct≥0] must preserve thermometer encoding of CheckingAcct., which can only increase. Actions can only turn on higher-level dummies that are off, whereCheckingAcctexists is the lowest-level dummy andCheckingAcct≥0 is...

  63. [71]

    Recidivism of Prisoners Released in 1994

    ThermometerEncoding: Actions on [SavingsAcctexists, SavingsAcct≥100] must preserve thermometer encoding of SavingsAcct., which can only increase. Actions can only turn on higher-level dummies that are off, whereSavingsAcctexists is the lowest-level dummy andSavingsAcct≥100 is ...

  64. [72]

    Each unit change in priorarrests≥2 leads to a unit change intimeserved≤1year

    DirectionalLinkage: Actions onpriorarrests≥2 will induce actions on [timeserved≤1year]. Each unit change in priorarrests≥2 leads to a unit change intimeserved≤1year

  65. [73]

    Each unit change in priorarrests=1 leads to a unit change intimeserved≤1year

    DirectionalLinkage: Actions onpriorarrests=1 will induce actions ontimeserved≤1year. Each unit change in priorarrests=1 leads to a unit change intimeserved≤1year

  66. [74]

    Each unit change in priorarrests≥5 leads to a unit change intimeserved≤1year

    DirectionalLinkage: Actions onpriorarrests≥5 will induce actions ontimeserved≤1year. Each unit change in priorarrests≥5 leads to a unit change intimeserved≤1year

  67. [75]

    Each unit change in priorarrests≥2 leads to a unit change intimeservedg1year

    DirectionalLinkage: Actions on priorarrests≥2 will induce actions ontimeservedg1year. Each unit change in priorarrests≥2 leads to a unit change intimeservedg1year

  68. [76]

    Each unit change in priorarrests=1 leads to a unit change intimeservedg1year 7

    DirectionalLinkage: Actions on priorarrests=1 will induce actions ontimeservedg1year. Each unit change in priorarrests=1 leads to a unit change intimeservedg1year 7

  69. [77]

    Each unit change in priorarrests≥5 leads to a unit change intimeservedg1year

    DirectionalLinkage: Actions on priorarrests≥5 will induce actions ontimeservedg1year. Each unit change in priorarrests≥5 leads to a unit change intimeservedg1year

  70. [78]

    Each unit change in priorarrests≥2 leads to a unit change intimeservedg2years

    DirectionalLinkage: Actions onpriorarrests≥2 will induce actions ontimeservedg2years. Each unit change in priorarrests≥2 leads to a unit change intimeservedg2years

  71. [79]

    Each unit change in priorarrests=1 leads to a unit change intimeservedg2years

    DirectionalLinkage: Actions onpriorarrests=1 will induce actions ontimeservedg2years. Each unit change in priorarrests=1 leads to a unit change intimeservedg2years

  72. [80]

    Each unit change in priorarrests≥5 leads to a unit change intimeservedg2years

    DirectionalLinkage: Actions onpriorarrests≥5 will induce actions ontimeservedg2years. Each unit change in priorarrests≥5 leads to a unit change intimeservedg2years

  73. [81]

    Each unit change in priorarrests≥2 leads to a unit change intimeservedg5years

    DirectionalLinkage: Actions onpriorarrests≥2 will induce actions ontimeservedg5years. Each unit change in priorarrests≥2 leads to a unit change intimeservedg5years

  74. [82]

    Each unit change in priorarrests=1 leads to a unit change intimeservedg5years

    DirectionalLinkage: Actions onpriorarrests=1 will induce actions ontimeservedg5years. Each unit change in priorarrests=1 leads to a unit change intimeservedg5years

  75. [83]

    Each unit change in priorarrests≥5 leads to a unit change intimeservedg5years

    DirectionalLinkage: Actions onpriorarrests≥5 will induce actions ontimeservedg5years. Each unit change in priorarrests≥5 leads to a unit change intimeservedg5years

  76. [84]

    Each unit change inpriorarrests≥2 leads to a unit change inpriorarrestsforfelony

    DirectionalLinkage: Actions on priorarrests≥2 will induce actions onpriorarrestsforfelony. Each unit change inpriorarrests≥2 leads to a unit change inpriorarrestsforfelony

  77. [85]

    Each unit change inpriorarrests=1 leads to a unit change inpriorarrestsforfelony

    DirectionalLinkage: Actions on priorarrests=1 will induce actions onpriorarrestsforfelony. Each unit change inpriorarrests=1 leads to a unit change inpriorarrestsforfelony

  78. [86]

    Each unit change inpriorarrests≥5 leads to a unit change inpriorarrestsforfelony

    DirectionalLinkage: Actions on priorarrests≥5 will induce actions onpriorarrestsforfelony. Each unit change inpriorarrests≥5 leads to a unit change inpriorarrestsforfelony

  79. [87]

    Each unit change inpriorarrests≥2 leads to a unit change inpriorarrestsformisdemeanor

    DirectionalLinkage: Actions onpriorarrests≥2will induce actions onpriorarrestsformisdemeanor. Each unit change inpriorarrests≥2 leads to a unit change inpriorarrestsformisdemeanor

  80. [88]

    Each unit change inpriorarrests=1 leads to a unit change inpriorarrestsformisdemeanor

    DirectionalLinkage: Actions onpriorarrests=1 will induce actions onpriorarrestsformisdemeanor. Each unit change inpriorarrests=1 leads to a unit change inpriorarrestsformisdemeanor

  81. [89]

    Each unit change inpriorarrests≥5 leads to a unit change inpriorarrestsformisdemeanor

    DirectionalLinkage: Actions onpriorarrests≥5will induce actions onpriorarrestsformisdemeanor. Each unit change inpriorarrests≥5 leads to a unit change inpriorarrestsformisdemeanor

  82. [90]

    Each unit change inpriorarrests≥2 leads to a unit change inpriorarrestsforgeneralviolence

    DirectionalLinkage: Actions onpriorarrests≥2 will induce actions onpriorarrestsforgeneralviolence. Each unit change inpriorarrests≥2 leads to a unit change inpriorarrestsforgeneralviolence

  83. [91]

    Each unit change inpriorarrests=1 leads to a unit change inpriorarrestsforgeneralviolence

    DirectionalLinkage: Actions onpriorarrests=1 will induce actions onpriorarrestsforgeneralviolence. Each unit change inpriorarrests=1 leads to a unit change inpriorarrestsforgeneralviolence

  84. [92]

    Each unit change inpriorarrests≥5 leads to a unit change inpriorarrestsforgeneralviolence

    DirectionalLinkage: Actions onpriorarrests≥5 will induce actions onpriorarrestsforgeneralviolence. Each unit change inpriorarrests≥5 leads to a unit change inpriorarrestsforgeneralviolence

  85. [93]

    Each unit change in priorarrests≥2 leads to a unit change inanypriorprborfine

    DirectionalLinkage: Actions onpriorarrests≥2 will induce actions onanypriorprborfine. Each unit change in priorarrests≥2 leads to a unit change inanypriorprborfine

  86. [94]

    Each unit change in priorarrests=1 leads to a unit change inanypriorprborfine

    DirectionalLinkage: Actions onpriorarrests=1 will induce actions onanypriorprborfine. Each unit change in priorarrests=1 leads to a unit change inanypriorprborfine

  87. [95]

    Each unit change in priorarrests≥5 leads to a unit change inanypriorprborfine 8 Statistical Inference for Responsiveness Verification

    DirectionalLinkage: Actions onpriorarrests≥5 will induce actions onanypriorprborfine. Each unit change in priorarrests≥5 leads to a unit change inanypriorprborfine 8 Statistical Inference for Responsiveness Verification

  88. [96]

    Each unit change inpriorarrests≥2 leads to a unit change inpriorarrestsforfelony

    DirectionalLinkage: Actions onpriorarrests≥2 will induce actions on [’priorarrestsforfelony’]. Each unit change inpriorarrests≥2 leads to a unit change inpriorarrestsforfelony

  89. [97]

    Each unit change inpriorarrests=1 leads to a unit change inpriorarrestsforfelony

    DirectionalLinkage: Actions onpriorarrests=1 will induce actions on [’priorarrestsforfelony’]. Each unit change inpriorarrests=1 leads to a unit change inpriorarrestsforfelony

  90. [98]

    Each unit change inpriorarrests≥5 leads to a unit change inpriorarrestsforfelony

    DirectionalLinkage: Actions onpriorarrests≥5 will induce actions on [’priorarrestsforfelony’]. Each unit change inpriorarrests≥5 leads to a unit change inpriorarrestsforfelony

  91. [99]

    ReachabilityConstraint: The values of [priorarrests≥2, priorarrests=1, nopriorarrests, priorarrests≥5] must belong to one of 4 values with custom reachability conditions

  92. [100]

    ReachabilityConstraint: The values of [drugabuser, drugtreatment] must belong to one of 4 values with custom reachability conditions

  93. [101]

    ReachabilityConstraint: The values of [alcoholabuser, alcoholtreatment] must belong to one of 4 values with custom reachability conditions. B.1.3 Additional Results This table includes additional model training and performance statistics.p is the percent of positive points, n ...

    1927

  94. [102]

    IfThenConstraint: If notweets = 0.0, thennumtweets > 1.0

  95. [103]

    IfThenConstraint: If noretweets = 0.0, thennumretweets > 1.0

  96. [104]

    IfThenConstraint: If noreplies = 0.0, thennumreplies > 1.0

  97. [105]

    Bot” predictions with certified responsiveness≥ ε = 0.05. % Responsiveshow % of

    DirectionalLinkage: Actions onnumtweets will induce to actions on [‘urlscount’]. Each unit change in numtweets leads to at least 1.00-unit change inurlscount B.2.3 Additional Results Model Pool % Resp. (Perceived) % Resp. (True) AUC Procedure Description # Models # Cert. Robus...

  98. [106]

    https://github.com/SurgicalInformatics/transplantbenefit/ 13 • Xbili represents bilirubin levels (clipped to[15, 200]) • XNa represents sodium levels (clipped to[125, 145]) • XINR represents international normalized ratio (clipped to[0.9, 2.4]) • Xcreat represents creatinine l...

  99. [107]

    IfThenConstraint: If rtumour_number_tbs ∈ {‘2’,‘3+’}, then rdisease_primary_tbs = 1 (cancer)

  100. [108]

    Concretely, to generate counterfactual patients with cancer, we define two intervention sets for small and large tumours, following Attia et al.[3]

    IfThenConstraint: If rdisease_primary_tbs = 1, thenrmax_tumour_size_tbs > 0. Concretely, to generate counterfactual patients with cancer, we define two intervention sets for small and large tumours, following Attia et al.[3]. In the small intervention set, we consider interven...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.