Pith. sign in

REVIEW 3 major objections 5 minor 30 references

Leo Breiman, the Rashomon Effect, and the Occam Dilemma

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a large Rashomon effect—many equally good models—and an accuracy-simplicity tradeoff cannot coexist; the effect makes the tradeoff vanish.

desk verdict A clear, well-written perspective, but the nullification claim rests on an unquantified 'large enough' inherited from prior work; the stress-test counterexample is a fair hit. read the letter →

arxiv 2507.03884 v1 pith:MOHPVFNL submitted 2025-07-05 stat.ML cs.LG

Cynthia Rudin This is my paper · ORCID
classification stat.MLcs.LG
keywords RashomoneffectOccamdilemmaaccuracy-simplicitytradeoffinterpretablemachinelearningsettheorynoisytabulardatasparsedecisiontreesTwoCultures
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 revisits Leo Breiman's 2001 'Two Cultures' essay and challenges one of its central warnings, the Occam Dilemma: that accurate predictive models must be complicated and hard to interpret. It argues that a large Rashomon effect—many different models with nearly equal accuracy—is incompatible with an accuracy-simplicity tradeoff. If many models fit the data almost equally well, then, under the Rashomon Set Theory the paper relies on, some of those models will be simple as well as accurate. The author concludes that for noisy tabular data, interpretable models such as sparse additive models and optimized decision trees can match black-box performance, and that causality can be sought without fitting a single explanatory model.

What carries the argument

The central object is the Rashomon set, the collection of all models whose predictive loss is within a chosen tolerance of the best loss on a dataset. The argument uses the geometry of function balls in that set: if the set is large enough to contain whole neighborhoods of functions, and if simpler function classes approximate complex ones well, then at least one delta-close simpler model sits inside each neighborhood. This converts noise-driven variance in the loss into an abundance of equally good models, and the abundance forces simpler models into the set. The size of the set can be measured by packing numbers, the number of separated balls of models it contains.

What would settle it

For a synthetic problem where the true complexity is known, increase the label noise and measure the smallest model that still reaches near-best accuracy. The paper's logic says that size should not grow with noise; if it does, the central claim fails.

Watch

Extended reading notes

Core claim

On the paper's own terms: the Rashomon effect does not coexist with the Occam Dilemma. A dataset with sufficiently noisy outcomes admits a large Rashomon set, the collection of models whose loss is within a small tolerance of the best achievable loss. Within that set, function classes are often nested and approximating, so a ball of complex models contains at least one model from a simpler class. Hence a large Rashomon set contains simpler, accurate models, and the supposed tradeoff between accuracy and simplicity disappears. Breiman's contrary observation is attributed to his narrow notion of simplicity: linear models and unoptimized CART trees, both of which are stunted compared to modern sparse generalized additive models and optimized sparse decision trees. The paper maintains Breiman's goals—accuracy, simplicity, causal insight—but argues they can be reached by model-free causal methods and interpretable prediction models instead of a single explanatory data model.

Load-bearing premise

The argument depends on the assumption that whenever there is enough noise to create many equally good models, at least one accurate model in that crowd is simple enough to be interpretable; the paper gives no quantitative test for when that must happen.

Editorial extensions

If this is right

  • On noisy tabular problems, practitioners should expect sparse GAMs, optimized decision trees, and risk scorecards to match more complex black-box models, so interpretable models can be used in high-stakes decisions without an accuracy penalty.
  • The accuracy-simplicity tradeoff that motivated black-box model use is not a general law; it appears only when simplicity is defined so narrowly that the intermediate class of expressive but interpretable models is excluded.
  • Non-noisy problems, such as image classification, behave differently: the Rashomon effect is weaker, so the arguments that erase the tradeoff for noisy data do not transfer unchanged to deterministic-label settings.
  • Causal insight does not require a single simple model: variable-importance distributions over many good models and case-based matching methods can give interpretable causal information without claiming to model the data-generation process.

Reading between the lines

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

  • If the claim is right, benchmark competitions on noisy tabular data that report tiny accuracy differences among methods are largely measuring noise within the Rashomon set; the more consequential objective is to find the simplest member of the set.
  • This suggests a practical design rule: on any noisy tabular dataset, first enumerate or approximate the Rashomon set, then choose the simplest model inside it, rather than optimize a single regularized objective.
  • A testable extension would be to measure the minimal complexity of an epsilon-optimal model as noise increases on controlled datasets; the paper's logic predicts that complexity should stay flat while the Rashomon set grows.
  • The noisy/non-noisy split implies that claims about the Occam Dilemma should be stratified by noise level; a tradeoff may reappear in near-deterministic problems even if it vanishes in noisy ones.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. This perspective paper revisits Leo Breiman's "Two Cultures" dichotomy and argues that Breiman's accuracy/simplicity tradeoff, the Occam Dilemma, is generally false. The paper's central claim is that the Rashomon Effect—the existence of many approximately equally good models—does not coexist with an accuracy/simplicity tradeoff; rather, a sufficiently large Rashomon set forces the appearance of simpler models that are also accurate. The paper attributes Breiman's contrary observation to his narrow notion of simplicity (linear models and unoptimized CART trees), presents Bruce Hoadley's 2001 credit-scoring experience as early empirical evidence, and argues that causal questions can be addressed without fitting a single interpretable data model, citing model-free methods such as Rashomon Importance Distributions and Almost Matching Exactly.

Significance. If the central claim were rigorously established, it would substantially reframe a foundational debate in statistics and machine learning, strengthening the case for interpretable modeling in noisy tabular settings and converting the Occam Dilemma from a presumed law into a symptom of restricted model classes and insufficient computation. The paper is valuable as a synthesis: it crisply identifies the logical tension in Breiman's position, brings forward Hoadley's striking field evidence, and connects the discussion to concrete modern tools such as optimized sparse decision trees, sparse GAMs, and risk scores. However, the paper's primary theoretical assertion is not derived here; it is cited to the author's earlier work, and the conditions under which the assertion holds are only stated qualitatively. The significance is therefore conditional on the cited theorem and on the breadth of the empirical claim being better quantified.

major comments (3)
  1. [Section 2.2] The inference from "a large Rashomon Effect" to "there exist simpler accurate models" is not established in this manuscript. The text states that "if the ball is large enough, the simplest models in the ball may be polynomials of a smaller degree," but "large enough" is never quantified, and the later condition that "functions from a simpler class can represent those from the more complex class sufficiently well" is also left qualitative. This is the load-bearing step of the paper's main claim. A concrete counterexample shows why quantification matters: take target f(x)=x^10, complex class of degree-10 polynomials, and simple class of degree-1 polynomials. For moderate noise, the set of degree-10 models within one standard error can be large (many near-optimal coefficient vectors), yet the nearest degree-1 polynomial is far outside the error tolerance, so a large Rashomon set and a tradeoff coexist. The paper's escape clause is exactly the condition that fails in this example, but no argument is given that the clause holds broadly. If the theorem in [SRP22] contains precise conditions, those conditions should be stated here; otherwise the central claim is a non-sequitur.
  2. [Sections 2.2-2.3] The paper's key explanatory mechanism—that noise in the world causes increased variance, which forces simpler models and a larger Rashomon set—is presented narratively rather than formally. Neither "sufficient noise" nor "sufficiently well" is defined in terms of loss functions, model classes, or dataset properties. As a result, the central assertion is not falsifiable in its current form. To make the claim testable, the paper should specify the hypothesis class, the loss, the error tolerance defining the Rashomon set, and a quantitative condition under which the distance between the complex and simple classes is smaller than the Rashomon set's radius. Without such a specification, the statement that the Rashomon Effect causes the Occam Dilemma to vanish cannot be distinguished from a tautology about a specially chosen notion of "large enough."
  3. [Section 2.3] The empirical generalization "For most tabular data problems with noise in the outcomes, very sparse models perform similarly to complicated black box models" is supported in the manuscript mainly by a series of studies from the author's group and by the single historical case of Hoadley. While these are suggestive, the breadth of the claim exceeds the evidence shown here. The paper would be strengthened by a small systematic comparison table listing datasets, model classes, and accuracy differences, or by explicitly framing the statement as a research hypothesis rather than an established fact. This matters because the paper's conclusion about the nonexistence of the tradeoff in practice depends on this generalization.
minor comments (5)
  1. [Abstract and Section 1] The abstract and Section 1 state that the Rashomon Effect "causes" the accuracy/simplicity tradeoff to vanish, while Section 2.2 concludes only under two qualitative conditions (sufficient noise and sufficient representability). The wording should be aligned so that the earlier statements match the conditional nature of the later argument.
  2. [Figure 1 caption] The phrase "its 2 δ-packing number" should be typeset as "its 2δ-packing number," and the symbol δ should be defined in the caption or nearby text.
  3. [Section 5] In the sentence "If we want to estimate the effect that a chosen variablev has on the outcome," there is a missing space: "variable v."
  4. [Section 3.2] The quotation from Breiman contains "more accurate predictors in classification and regression then neural nets"; if this is a direct quote the spelling should be checked against the original, and if it is the author's own sentence, "then" should be "than."
  5. [Section 2.3] The phrase "They haven't made a dent on a multitude of noisy problems – because they can't" is rhetorically strong but overstated. A more precise phrasing, such as "they have not been shown to improve accuracy over simpler methods in many noisy problems," would better match the evidence presented.

Circularity Check

2 steps flagged · score 7.0 of 10

The central no-tradeoff claim is a conditional whose stated sufficient condition already contains the conclusion, and the underlying mechanism is imported from the author's own prior work.

  1. self definitional [Section 2.2, paragraph beginning "Based on this logic"]
    "Based on this logic, as long as there is sufficient noise in the world to produce a Rashomon Effect, and that functions from a simpler class can represent those from the more complex class sufficiently well, the Rashomon Effect does not simultaneously exist with the accuracy/simplicity tradeoff."

    The target claim is that the Rashomon Effect makes the accuracy/simplicity tradeoff vanish, i.e., that accurate simple models exist alongside accurate complex ones. The stated sufficient condition already contains this target: "functions from a simpler class can represent those from the more complex class sufficiently well" is essentially the existence of accurate simple approximations. No quantitative bridge connects the size of the Rashomon set to the representability condition; if representability holds, the conclusion follows independently of the Rashomon effect, and if it fails, no amount of noise or model multiplicity yields the conclusion. The argument is therefore an implication whose hypothesis is the conclusion.

  2. self citation load bearing [Section 2.2, first paragraph]
    "I will follow the "Rashomon Set Theory" argument of [SRP22] to show that a large Rashomon Effect (the existence of many good models) gives rise to simpler models that are also accurate. That is, the Rashomon Effect itself suppresses the accuracy/simplicity tradeoff."

    The central mechanism of the paper—large Rashomon set implies simpler accurate models—is not derived in this manuscript. It is imported from [SRP22], co-authored by the present author, and the following noise-to-simplicity chain is delegated to additional self-authored citations [SCPR23; BCS+24; RZS+24]. The paper gives no statement of the theorem's assumptions and no reproduction of its proof, so the load-bearing step rests entirely on the author's own prior work. This is a self-citation chain rather than an independent derivation presented to the reader.

full rationale

The paper's main theoretical assertion, that the Rashomon Effect nullifies the Occam Dilemma, is not independently established here. In Section 2.2, the proof is explicitly delegated to the author's own Rashomon Set Theory paper, and the formal sufficient condition stated at the end of the section—"functions from a simpler class can represent those from the more complex class sufficiently well"—is effectively the no-tradeoff conclusion itself. This makes the central derivation circular by construction: the claimed implication assumes the target property as one of its antecedents. The unquantified phrase "if the ball is large enough" is a separate correctness gap, but the deeper circularity is that "large enough" would have to be defined as "large enough to contain a simpler model," which is just the conclusion restated. Some external grounding exists, notably Hoadley's independent credit-scoring observations and the modern empirical examples, which prevent the paper from being entirely vacuous. The presence of that external evidence is why the score is not higher. However, the paper's formal argument for the central claim reduces to a self-citation chain plus an assumption equivalent to the conclusion, so the circularity is substantial.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The central claim relies on domain assumptions about noise, approximation, and a historical interpretation of Breiman's usage. No numeric parameters are fitted in this paper; the conceptual threshold conditions are treated as qualitative axioms.

assumptions (4)
  • domain assumption Noisy real-world tabular problems have enough noise to produce a large Rashomon Effect (many approximately equally good models).
    Section 2.2 assumes sufficient noise to produce a Rashomon Effect and Section 2.3 asserts most relevant problems are noisy, but no quantitative or empirical threshold is given.
  • standard math Simple function classes can approximate complex ones sufficiently well.
    Used in Section 2.2 and Figure 1 to argue that balls of complex models contain simpler models; relies on standard approximation theory without a formal statement.
  • ad hoc to paper Breiman's notion of simplicity was limited to linear models and unoptimized CART trees.
    Section 3 makes this historical interpretation to reconcile Breiman's observed tradeoff with the author's claim; it is an interpretive assumption, not established by a direct quote.
  • domain assumption The practical settings of interest are noisy rather than non-noisy deterministic problems.
    Section 2.3 and the conclusion restrict the nullification claim to noisy problems; this exclusion is load-bearing because the author concedes the tradeoff may reappear for deterministic problems like computer vision.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Leo Breiman, the Rashomon Effect, and the Occam Dilemma." pith.science (2026). https://pith.science/paper/MOHPVFNL

@misc{pith2026250703884,
  author       = {Pith},
  title        = {Pith review of: Leo Breiman, the Rashomon Effect, and the Occam Dilemma},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MOHPVFNL}},
  note         = {Machine review of arXiv:2507.03884}
}
read the original abstract

In the famous Two Cultures paper, Leo Breiman provided a visionary perspective on the cultures of ''data models'' (modeling with consideration of data generation) versus ''algorithmic models'' (vanilla machine learning models). I provide a modern perspective on these approaches. One of Breiman's key arguments against data models is the ''Rashomon Effect,'' which is the existence of many different-but-equally-good models. The Rashomon Effect implies that data modelers would not be able to determine which model generated the data. Conversely, one of his core advantages in favor of data models is simplicity, as he claimed there exists an ''Occam Dilemma,'' i.e., an accuracy-simplicity tradeoff. After 25 years of powerful computers, it has become clear that this claim is not generally true, in that algorithmic models do not need to be complex to be accurate; however, there are nuances that help explain Breiman's logic, specifically, that by ''simple,'' he appears to consider only linear models or unoptimized decision trees. Interestingly, the Rashomon Effect is a key tool in proving the nullification of the Occam Dilemma. To his credit though, Breiman did not have the benefit of modern computers, with which my observations are much easier to make. Breiman's goal for interpretability was somewhat intertwined with causality: simpler models can help reveal which variables have a causal relationship with the outcome. However, I argue that causality can be investigated without the use of single models, whether or not they are simple. Interpretability is useful in its own right, and I think Breiman knew that too. Technically, my modern perspective does not belong to either of Breiman's Two Cultures, but shares the goals of both of them - causality, simplicity, accuracy - and shows that these goals can be accomplished in other ways, without the limitations Breiman was concerned about.

Figures

Figures reproduced from arXiv: 2507.03884 by the authors.

Figure 1
Figure 1. Illustration reproduced from [SRP22; RZS [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 29 canonical work pages

  1. [1]

    Using noise to infer aspects of simplicity without learning

    Zachery Boner, Harry Chen, Lesia Semenova, Ronald Parr, and Cynthia Rudin. Using noise to infer aspects of simplicity without learning. In Neural Information Processing Systems , 2024

    work page 2024

  2. [2]

    Breiman, J

    L. Breiman, J. Friedman, C. J. Stone, and R. A. Olshen. Classification and Regression Trees . CRC press, 1984

    work page 1984

  3. [3]

    Statistical modeling: The two cultures (with comments and a rejoinder by the author)

    Leo Breiman. Statistical modeling: The two cultures (with comments and a rejoinder by the author). Statistical Science , 16(3):199--231, 2001

    work page 2001

  4. [4]

    A holistic approach to interpretability in financial lending: Models, visualizations, and summary-explanations

    Chaofan Chen, Kangcheng Lin, Cynthia Rudin, Yaron Shaposhnik, Sijia Wang, and Tong Wang. A holistic approach to interpretability in financial lending: Models, visualizations, and summary-explanations. Decision Support Systems , 152:113647, 2022

    work page 2022

  5. [5]

    The rashomon importance distribution: Getting RID of unstable, single model-based variable importance

    Jon Donnelly, Srikar Katta, Cynthia Rudin, and Edward P Browne. The rashomon importance distribution: Getting RID of unstable, single model-based variable importance. In Neural Information Processing Systems ( NeurIPS ) , 2023

    work page 2023

  6. [6]

    Interpretable almost-exact matching for causal inference

    Awa Dieng, Yameng Liu, Sudeepa Roy, Cynthia Rudin, and Alexander Volfovsky. Interpretable almost-exact matching for causal inference. In Proceedings of Artificial Intelligence and Statistics (AISTATS) , pages 2445--2453, 2019

    2019

  7. [7]

    Variable importance matching for causal inference

    Quinn Lanners, Harsh Parikh, Alexander Volfovsky, Cynthia Rudin, and David Page. Variable importance matching for causal inference. In Uncertainty in Artificial Intelligence ( UAI ) , 2023

    work page 2023

  8. [8]

    Generalized and scalable optimal sparse decision trees

    Jimmy Lin, Chudi Zhong, Diane Hu, Cynthia Rudin, and Margo Seltzer. Generalized and scalable optimal sparse decision trees. In Proceedings of the International Conference on Machine Learning (ICML) , pages 6150--6160, 2020

    work page 2020

Show all 30 references
  1. [9]

    Fasterrisk: Fast and accurate interpretable risk scores

    Jiachang Liu, Chudi Zhong, Boxuan Li, Margo Seltzer, and Cynthia Rudin. Fasterrisk: Fast and accurate interpretable risk scores. In Neural Information Processing Systems ( NeurIPS ) , 2022

    2022

  2. [10]

    Fast sparse classification for generalized linear and additive models

    Jiachang Liu, Chudi Zhong, Margo Seltzer, and Cynthia Rudin. Fast sparse classification for generalized linear and additive models. In Proceedings of Artificial Intelligence and Statistics (AISTATS) , 2022

    2022

  3. [11]

    McCullagh and John A

    P. McCullagh and John A. Nelder. Generalized Linear Models, Second Edition (Chapman & Hall/CRC Monographs on Statistics & Applied Probability) . Chapman and Hall/CRC, 2 edition, 1989

    1989

  4. [12]

    Fast sparse decision tree optimization via reference ensembles

    Hayden McTavish, Chudi Zhong, Reto Achermann, Ilias Karimalis, Jacques Chen, Cynthia Rudin, and Margo Seltzer. Fast sparse decision tree optimization via reference ensembles. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 36, pages 9604--9613, 2022

    2022

  5. [13]

    MALTS : Matching after learning to stretch

    Harsh Parikh, Cynthia Rudin, and Alexander Volfovsky. MALTS : Matching after learning to stretch. Journal of Machine Learning Research , 23(240):1--42, 2022

    2022

  6. [14]

    Interpretable machine learning: Fundamental principles and 10 grand challenges

    Cynthia Rudin, Chaofan Chen, Zhi Chen, Haiyang Huang, Lesia Semenova, and Chudi Zhong. Interpretable machine learning: Fundamental principles and 10 grand challenges . Statistics Surveys , 16(none):1 -- 85, 2022

    2022

  7. [15]

    Analytics for power grid distribution reliability in new york city

    Cynthia Rudin, S eyda Ertekin, Rebecca Passonneau, Axinia Radeva, Ashish Tomar, Boyi Xie, Stanley Lewis, Mark Riddle, Debbie Pangsrivinij, and Tyler McCormick. Analytics for power grid distribution reliability in new york city. Interfaces , 44(4):364--–383, 2014

    2014

  8. [16]

    Why are we using black box models in AI when we don't need to? A lesson from an explainable AI competition

    Cynthia Rudin and Joanna Radin. Why are we using black box models in AI when we don't need to? A lesson from an explainable AI competition. Harvard Data Science Review , 1(2), 2019

    2019

  9. [17]

    Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead

    Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence , 1:206--215, May 2019

    2019

  10. [18]

    Anderson, Albert Boulanger, Ansaf Salleb-Aouissi, Maggie Chow, Haimonti Dutta, Philip Gross, Bert Huang, Steve Ierome, Delfina Isaac, Arthur Kressner, Rebecca J

    Cynthia Rudin, David Waltz, Roger N. Anderson, Albert Boulanger, Ansaf Salleb-Aouissi, Maggie Chow, Haimonti Dutta, Philip Gross, Bert Huang, Steve Ierome, Delfina Isaac, Arthur Kressner, Rebecca J. Passonneau, Axinia Radeva, and Leon Wu. Machine learning for the N ew Y ork C ...

    2012

  11. [19]

    Amazing things come from having many good models

    Cynthia Rudin, Chudi Zhong, Lesia Semenova, Margo Seltzer, Ronald Parr, Jiachang Liu, Srikar Katta, Jon Donnelly, Harry Chen, and Zachery Boner. Amazing things come from having many good models. In Proceedings of the International Conference on Machine Learning (ICML) , 2024

    2024

  12. [20]

    Evaluating pre-trial programs using machine learning matching algorithms

    Travis Seale-Carlisle, Saksham Jain, Courtney Lee, Caroline Levenson, Swathi Ramprasad, Brandon Garrett, Sudeepa Roy, Cynthia Rudin, and Alexander Volfovsky. Evaluating pre-trial programs using machine learning matching algorithms. In Proceedings of AAAI , 2024

    2024

  13. [21]

    A path to simpler models starts with noise

    Lesia Semenova, Harry Chen, Ronald Parr, and Cynthia Rudin. A path to simpler models starts with noise. In Neural Information Processing Systems ( NeurIPS ) , 2023

    2023

  14. [22]

    Libon, Rodney Swenson, Catherine C

    William Souillard-Mandar, Randall Davis, Cynthia Rudin, Rhoda Au, David J. Libon, Rodney Swenson, Catherine C. Price, Melissa Lamar, and Dana L. Penney. Learning classification models of cognitive conditions from subtle behaviors in the digital clock drawing test. Machine Lear...

    2016

  15. [23]

    On the existence of simpler machine learning models

    Lesia Semenova, Cynthia Rudin, and Ronald Parr. On the existence of simpler machine learning models. In ACM Conference on Fairness, Accountability, and Transparency ( ACM FAccT ) , 2022

    2022

  16. [24]

    Learning optimized risk scores

    Berk Ustun and Cynthia Rudin. Learning optimized risk scores. Journal of Machine Learning Research , 20(150):1--75, 2019

    2019

  17. [25]

    In Pursuit of Interpretable, Fair and Accurate Machine Learning for Criminal Recidivism Prediction

    Caroline Wang, Bin Han, Bhrij Patel, and Cynthia Rudin. In Pursuit of Interpretable, Fair and Accurate Machine Learning for Criminal Recidivism Prediction . Journal of Quantitative Criminology , pages 1--63, 2022

    2022

  18. [26]

    Usaid Awan, Yameng Liu, Sudeepa Roy, Cynthia Rudin, and Alexander Volfovsky

    Tianyu Wang, Marco Morucci, M. Usaid Awan, Yameng Liu, Sudeepa Roy, Cynthia Rudin, and Alexander Volfovsky. Flame: A fast large-scale almost matching exactly approach to causal inference. Journal of Machine Learning Research , 22(31):1--41, 2021

    2021

  19. [27]

    Exploring the whole R ashomon set of sparse decision trees

    Rui Xin, Chudi Zhong, Zhi Chen, Takuya Takagi, Margo Seltzer, and Cynthia Rudin. Exploring the whole R ashomon set of sparse decision trees. In Neural Information Processing Systems ( NeurIPS ) , volume 35, pages 14071--14084, 2022

    2022

  20. [28]

    Exploring and interacting with the set of good sparse generalized additive models

    Chudi Zhong, Zhi Chen, Jiachang Liu, Margo Seltzer, and Cynthia Rudin. Exploring and interacting with the set of good sparse generalized additive models. In Neural Information Processing Systems ( NeurIPS ) , 2023

    2023

  21. [29]

    Fast and interpretable mortality risk scores for critical care patients

    Chloe Qinyu Zhu, Muhang Tian, Lesia Semenova, Jiachang Liu, Jack Xu, Joseph Scarpa, and Cynthia Rudin. Fast and interpretable mortality risk scores for critical care patients. arXiv preprint arXiv:2311.13015 , 2023

    2023 arXiv

  22. [30]

    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

Pith tools

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