Glitches in Decision Tree Ensemble Models

Ashutosh Gupta; Kaushik Mallik; Krishna Shankaranarayanan; Namrita Varshney; Satyankar Chandra

REVIEW 3 major objections 2 minor 39 references

Glitches in Decision Tree Ensemble Models

T0 review · 3 major / 2 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 Output oscillations along one feature define a model 'glitch'; detecting one in tree ensembles is NP-complete even at depth 4, and MILP search still finds them in real models.

desk verdict The glitch definition is a real contribution, but the paper's NP-completeness proof rests on a false lemma—a single tree can already glitch—so the headline result isn't proven as written. read the letter →

arxiv 2507.14492 v1 pith:JBKWM4VK submitted 2025-07-19 cs.LG stat.ML

Satyankar Chandra , Ashutosh Gupta , Kaushik Mallik , Krishna Shankaranarayanan , Namrita Varshney This is my paper

classification cs.LGstat.ML MSC 68Q1768T0590C11

keywords glitchesdecisiontreeensemblesgradient-boostedtreesNP-completenessverificationmixed-integerlinearprogrammingrobustnessmonotonicity

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

Machine-learned models used in high-stakes decisions can behave inconsistently in ways that neither plain robustness nor monotonicity captures: as a single input feature rises through a tiny interval, the output can dip and then rise, or rise and then fall. The paper formalizes this pattern as a glitch, demonstrates it in gradient-boosted decision-tree models trained on standard datasets, and argues that finding one is a genuine model anomaly worth inspecting. It proves that deciding whether a tree ensemble contains a glitch above a given magnitude is NP-complete even when every tree has depth 4. It then shows that encoding the search as a mixed-integer linear program makes exact glitch detection feasible for many real-world ensembles, including large ones with up to a thousand trees. The broader point is that glitch-freedom is a sharper reliability criterion than robustness or monotonicity alone.

What carries the argument

The central object is the $\alpha$-glitch: an ordered triple $x^- \le_i x \le_i x^+$ that differs only in coordinate $i$, where the model output at $x$ is strictly lower than both endpoints (canyon) or strictly higher than both (hill), and $\min\{d(f(x),f(x^-)),d(f(x),f(x^+))\}/d(x^-,x^+) \ge \alpha$. For tree ensembles, whether such a triple exists is governed by the threshold predicates $v_i \le \eta$ along the chosen feature: the three inputs must be placed so that the active tree paths differ across those thresholds. The MILP machinery encodes this by making three synchronized copies of the tree-path activation variables, one for each input, and adds linear constraints for equal other coordinates, ordered center coordinate, the canyon/hill condition, and the magnitude constraint; the hardness proof instead builds pairs of clause trees whose outputs switch on a control feature, so the ensemble output is $m$ at both endpoints and $0$ in the middle exactly when the formula is satisfiable.

What would settle it

To test the main theorem, look for a polynomial-time exact solver for glitch detection in depth-4 ensembles, which would refute NP-completeness unless P=NP. To test the lemma, evaluate the paper's own definition on one tree with thresholds $0.5$ and $1.5$ on a single feature and leaf outputs $0,1,0$: inputs $0,1,2$ form an $\alpha$-glitch for $\alpha=1/2$, yet no pair of distinct trees exists, contradicting Proposition 3.1.

Watch

Extended reading notes

Core claim

The paper's central claim is that glitches—triples of nearby inputs along one feature where the output falls then rises (a canyon) or rises then falls (a hill)—are a distinct and useful failure mode for decision-tree ensembles, and their existence can be decided exactly but is computationally hard. On the formal side, the paper proves Theorem 3.2: the problems TE_GLITCH(α,i) and TE_GLITCH(α), which ask for a glitch of magnitude greater than α in a tree ensemble, are NP-complete, and the hardness already holds when every tree has depth 4. On the practical side, it encodes the verification and maximum-magnitude search as mixed-integer linear programs, and experiments on benchmark gradient-boosted models find glitches in almost every model considered, with the MILP route solving most instances and comfortably outperforming an SMT formulation.

Load-bearing premise

The load-bearing premise is that every glitch in a tree ensemble has to come from two different trees with cutoff values on the same feature that straddle the middle input, because the hardness proof needs to exploit that cross-tree gap; if a single tree can already create the oscillation with two of its own cutoffs, that premise fails and the proof as written stops working.

Editorial extensions

If this is right

Every monotone decision-maker is glitch-free, and every Lipschitz-continuous model with constant $L$ has glitch magnitudes at most $L/2$ (Proposition 2.2).
Exact global glitch detection in tree ensembles is NP-complete already for depth-4 trees, so no polynomial-time exact algorithm is expected under standard complexity assumptions.
A single MILP encoding covers all three problem variants—verification with a fixed feature, verification over all features, and search for the largest glitch—and solved all or most instances on the tested benchmarks within the time limit.
Glitches are rarer and more specific than robustness violations: many robustness violations in the case studies are anticipated by training data, whereas glitches additionally require the simultaneous monotonicity violation, making them a more precise anomaly signal.
The same encoding supports local glitch search by restricting the input-variable ranges to the region of interest.

Reading between the lines

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

Editorial caution: Proposition 3.1, a lemma the NP-hardness proof invokes, is contradicted by a single tree with two split points on one feature that satisfies the paper's own glitch definition; Theorem 3.2 may still be true, but the proof as written needs an additional argument.
Editorial inference: because exact detection is NP-complete, glitch-avoidance at training time will likely rely on heuristics or bounded-depth searches; a direct testable extension is to penalize adjacent split points that create canyon or hill output patterns during boosted-tree training.
Editorial inference: the same one-dimensional glitch definition transfers to neural networks by replacing tree paths with piecewise-linear activation regions; the paper reports promising preliminary results, and a concrete next step is a glitch-search MILP for ReLU networks.

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.

Desk Editor's Note

read the letter

Short version: this is a paper with a good idea and a broken proof. The formal definition of glitches—abrupt oscillation along one feature—is genuinely new and useful. It connects robustness and monotonicity in a way that matches the intuitive phenomenon, and the motivating examples are convincing. If the theory held up, it would give verification people a new target property and a practical search tool.

The MILP encoding is also a sensible adaptation of Kantchelian's robustness encoding. The experiments show glitches are widespread, and the scalability data (up to 1000 trees, depth 8) is plausible, though I note the paper doesn't release code or data, and the 'TI' entry in Table 2 looks like a typo for 'time'. Those are fixable.

The big problem is Proposition 3.1. The paper claims that any glitch in a tree ensemble requires two distinct trees whose thresholds straddle the middle point. That's false. A single tree with thresholds 0.5 and 1.5 on the same feature gives outputs 0, 1, 0 for inputs 0, 1, 2, which is a glitch for α=0.5. The proof's partition argument forgets that one tree can have multiple thresholds on the same feature along a path. This matters because the only-if direction of the NP-hardness proof (Theorem C.2) uses Prop 3.1 to conclude that a glitch must have r−, r, r+ in the three regions R1, R2, R3. Without that lemma, the inference isn't justified.

I don't think the paper is beyond rescue. The reduction itself may be salvageable: in the constructed ensemble, a direct cell argument could probably show that any glitch must straddle both thresholds, but that argument isn't written down. As submitted, the central NP-completeness result is unproven. That's a load-bearing flaw, not a minor gap.

Who is this for? Verification researchers working on tree ensembles will find the glitch definition worth thinking about, and the empirical observations are a useful prompt for tool-building. But anyone relying on the NP-completeness theorem will be misled. The paper deserves a serious referee because the concept is novel and the flaw is specific and possibly fixable, but it should not be accepted in this state.

My recommendation: send it to peer review, but with a clear signal that the proof of Theorem 3.2 must be corrected and the counterexample addressed. If the authors can fix Prop 3.1 or revise the proof, this could become a solid paper.

Referee Report

3 major / 2 minor

Summary. The paper proposes a formal notion of 'glitches' for AI decision-makers, defined as a monotone triple of inputs along one coordinate whose outputs drop-then-rise or rise-then-drop with a normalized jump of at least alpha. It motivates the notion through case studies, proves a Lipschitz bound on glitch magnitude, and studies glitch detection in sum-of-trees ensembles. The main theoretical contribution is Theorem 3.2, claiming NP-completeness of TE_GLITCH(alpha,i) and TE_GLITCH(alpha), already for depth-4 trees. The paper also presents MILP/SMT encodings and reports experiments with a tool called VIKRITI on benchmark GBDT models.

Significance. If correct, this would be a genuinely useful contribution: the glitch definition gives a quantitative way to localize simultaneous robustness and monotonicity failures, the Lipschitz bound is clean, and an NP-completeness result for tree ensembles would place the problem next to robustness verification. The experimental scope is broad and the MILP approach is sensible. However, the paper does not provide code or artifacts, and the two most load-bearing technical pieces--the proof of NP-completeness and the MILP encodings--are not correct as written; the central claims therefore need a major revision that supplies correct proofs and revalidates the experiments.

major comments (3)

[Prop. 3.1, Sec. 3.1 and App. B/C] Proposition 3.1 is false. A single decision tree with root v<=0.5 (leaf 0), false child v<=1.5 with leaves 1 and 0 gives f(0)=0, f(1)=1, f(2)=0. For x^-=0, x=1, x^+=2, Eq. (2) evaluates to min{|1-0|,|1-0|}/|2-0|=0.5 and Eq. (3) holds, so this is a 0.5-glitch, but there is no pair of distinct trees. The proof in Appendix B is invalid exactly at the step 'any two values among x_i^-, x_i, x_i^+ would fall in the same partition', since one tree can have two thresholds on the same feature. The only-if direction of Theorem C.2 uses Prop. 3.1 to infer the r-region pattern, and Theorem C.3 uses it to exclude glitches in other dimensions; both inferences are unsupported. The reduction itself appears salvageable by a direct cell argument on the three r-regions, but that proof is not in the manuscript.
[Sec. 3.2, Fig. 5(C)] The MILP for TE_GLITCH does not maximize glitch magnitude. In the displayed objective, alpha is a free variable and no constraint links it to the ratio in Eq. (2). For any fixed triple (W^-, W, W^+), the term -alpha|x_i^+ - x_i^-| is decreased by increasing alpha, so the optimum over alpha is attained at the minimal admissible value (0 if alpha>=0); the objective then reduces to min{|M(W^-)-M(W)|, |M(W)-M(W^+)|}, not to min-jump / distance. The correct encoding needs a constraint such as min{...} >= alpha|x_i^+ - x_i^-| with alpha maximized. As a result, the 'largest magnitude' column in Table 2 and the TE_GLITCH runs in RQ2/RQ3 do not correspond to the paper's stated definition.
[Sec. 3.2, predicate Delta] The proposed encoding of Eq. (3) is not equivalent to the glitch condition. Delta requires c^- and c^+ to have the same sign relative to zero and c to have the opposite sign, but Eq. (3) is an ordering condition. For example, a hill glitch with outputs (1,2,1) satisfies Eq. (3) but fails Delta because all three values are nonnegative; a canyon glitch with outputs (1,0,1) also fails Delta. Thus the MILP and SMT instances in Fig. 5 can miss valid glitches, and the 'no-glitch' results in Table 2 cannot be interpreted as certificates of glitch absence.

minor comments (2)

[Sec. 3.1, after Theorem 3.2] 'adding 4 - d dummy nodes' should read 'adding d - 4 dummy nodes' when the target depth d is larger than 4.
[Problem 1 and Fig. 5] Problem 1 asks for a glitch with magnitude 'larger than alpha', while Definition 1 and the MILP constraint use '>= alpha'; the strict versus non-strict convention should be made consistent.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the glitch definition, complexity proofs, and MILP encodings are self-contained, and the main proof gap is a correctness issue rather than a circularity.

full rationale

The chain of claims in this paper is not circular. Definition 1 introduces glitches from raw input-output triples, Proposition 2.2 derives a Lipschitz bound on glitch magnitude from the definition, and Theorem 3.2's NP-hardness is attempted via an explicit reduction from 3-CNF-SAT. The MILP encodings in Section 3.2 and Appendix D directly encode Definition 1's conditions (1)-(3) plus tree semantics; they do not fit any parameter and then reuse that fit as a prediction. The experimental sections compare glitch counts with robustness violations on external pretrained models and standard benchmarks, so the empirical claims are measurements against outside data rather than outputs forced by construction. The one citation involving a co-author, Ahmad et al. on sensitivity verification, appears only in related-work discussion and is not load-bearing for the paper's central claims. The skeptical counterexample to Proposition 3.1 shows a possible soundness gap in the NP-hardness proof: the single-tree example satisfies Definition 1 without the asserted pair of distinct trees. That is a correctness concern for the proof of Theorem 3.2, not a circularity, because the theorem is not assumed by the auxiliary proposition and no fitted input is renamed as a derived result. Accordingly, no specific circular step can be quoted, and the appropriate score is 0.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

No new physical entities, forces, or dimensions are introduced. 'Glitch' is a formal property, not a postulated entity with independent evidence.

free parameters (2)

glitch threshold alpha = 0.001
Chosen for all TE_GLITCH(alpha) experiments in Table 2; the number of detected glitches is sensitive to this choice.
robustness neighborhood epsilon per dataset = 0.004 to 0.278 by model
Used in Table 1 to compare glitches with robustness violations; no selection rule is given.

assumptions (4)

standard math Decision tree ensemble semantics: output is the sum of leaf values reached by each tree.
Section 3 defines JMK as sum over trees; the MILP encoding relies on this.
ad hoc to paper For a glitch in a tree ensemble, thresholds from two distinct trees must straddle the middle point (Prop. 3.1).
Used in App. C NP-hardness proof; the statement is false, so the theorem proof is unsound as written.
domain assumption Training data in the case studies are free of outliers, so nearby training labels are a reliable basis for calling a robustness violation anticipated or unanticipated.
Stated in Sec. 1 Case Study I; the qualitative comparison of robustness versus glitches depends on this.
domain assumption Abrupt output oscillations in a small neighborhood usually indicate model inconsistency.
Abstract and Sec. 5; the evidence is one breast-cancer glitch assessed by three oncologists, not a systematic study.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Glitches in Decision Tree Ensemble Models." pith.science (2026). https://pith.science/paper/JBKWM4VK

@misc{pith2026250714492,
  author       = {Pith},
  title        = {Pith review of: Glitches in Decision Tree Ensemble Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JBKWM4VK}},
  note         = {Machine review of arXiv:2507.14492}
}

read the original abstract

Many critical decision-making tasks are now delegated to machine-learned models, and it is imperative that their decisions are trustworthy and reliable, and their outputs are consistent across similar inputs. We identify a new source of unreliable behaviors-called glitches-which may significantly impair the reliability of AI models having steep decision boundaries. Roughly speaking, glitches are small neighborhoods in the input space where the model's output abruptly oscillates with respect to small changes in the input. We provide a formal definition of glitches, and use well-known models and datasets from the literature to demonstrate that they have widespread existence and argue they usually indicate potential model inconsistencies in the neighborhood of where they are found. We proceed to the algorithmic search of glitches for widely used gradient-boosted decision tree (GBDT) models. We prove that the problem of detecting glitches is NP-complete for tree ensembles, already for trees of depth 4. Our glitch-search algorithm for GBDT models uses an MILP encoding of the problem, and its effectiveness and computational feasibility are demonstrated on a set of widely used GBDT benchmarks taken from the literature.

Figures

Figures reproduced from arXiv: 2507.14492 by the authors.

**Figure 3.** Variation of computation time of VIKRITI for solving TE_GLITCH(α, i) on an average over different choices of feature i and for α = 0.001. Model ϵ Robustness #Glitch #A #U #I BCR 0.278 190 10 0 1 BCU 0.067 104 16 0 1 DR 0.036 389 2 1 23 DU 0.004 349 2 1 15 IJR 0.004 35 137 86 34 IJU 0.004 133 217 46 31 WSR 0.004 0 333 128 0 WSU 0.004 10 96 19 72 BMR 0.004 784 0 0 0 BMU 0.004 2341 11 0 0 [PITH_FULL_IMAGE:figures/full… view at source ↗

**Figure 4.** Illustration: reducing the toy 3-CNF-SAT instance [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗

Figures from the paper (2 more)

**Figure 5.** Figure 5: MILP encodings of the glitch-search problems. [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]

**Figure 6.** Figure 6: breast cancer glitch points 17 [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 28 canonical work pages

[1]

Sensitivity verification for decision tree ensembles

Arhaan Ahmad, Tanay Vineet Tayal, Ashutosh Gupta, and S Akshay. Sensitivity verification for decision tree ensembles. In The Thirteenth International Conference on Learning Representations

work page
[2]

Satisfiability modulo theories

Clark Barrett and Cesare Tinelli. Satisfiability modulo theories. Handbook of model checking, pages 305--343, 2018

work page 2018
[3]

Treant: training evasion-aware decision trees

Stefano Calzavara, Claudio Lucchese, Gabriele Tolomei, Seyum Assefa Abebe, and Salvatore Orlando. Treant: training evasion-aware decision trees. Data Min. Knowl. Discov., 34 0 (5): 0 1390--1420, 2020. doi:10.1007/S10618-020-00694-9. URL https://doi.org/10.1007/s10618-020-00694-9

work page doi:10.1007/s10618-020-00694-9 2020
[4]

Boning, and Cho - Jui Hsieh

Hongge Chen, Huan Zhang, Duane S. Boning, and Cho - Jui Hsieh. Robust decision trees against adversarial examples. In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors, Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA , volume 97 of Proceedings of Machine Learning Research, pag...

work page 2019
[5]

Boning, and Cho - Jui Hsieh

Hongge Chen, Huan Zhang, Si Si, Yang Li, Duane S. Boning, and Cho - Jui Hsieh. Robustness verification of tree-based models. In Hanna M. Wallach, Hugo Larochelle, Alina Beygelzimer, Florence d'Alch \' e - Buc, Emily B. Fox, and Roman Garnett, editors, Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing ...

work page 2019
[6]

Xgboost: A scalable tree boosting system

Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In Balaji Krishnapuram, Mohak Shah, Alexander J. Smola, Charu C. Aggarwal, Dou Shen, and Rajeev Rastogi, editors, Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, August 13-17, 2016 , pages 785--794. ACM , 2...

arXiv 2016
[7]

Learning security classifiers with verified global robustness properties

Yizheng Chen, Shiqi Wang, Yue Qin, Xiaojing Liao, Suman Jana, and David Wagner. Learning security classifiers with verified global robustness properties. In Proceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security, pages 477--494, 2021

work page 2021
[8]

Query-efficient hard-label black-box attack: An optimization-based approach

Minhao Cheng, Thong Le, Pin - Yu Chen, Huan Zhang, Jinfeng Yi, and Cho - Jui Hsieh. Query-efficient hard-label black-box attack: An optimization-based approach. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019 . OpenReview.net, 2019. URL https://openreview.net/forum?id=rJlk6iRqKX

work page 2019

Show all 39 references

[9]

Fair prediction with disparate impact: A study of bias in recidivism prediction instruments

Alexandra Chouldechova. Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. Big data, 5 0 (2): 0 153--163, 2017

2017
[10]

Robustbench: a standardized adversarial robustness benchmark

Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robustbench: a standardized adversarial robustness benchmark. arXiv preprint arXiv:2010.09670, 2020

2010 arXiv
[11]

Satisfiability modulo theories: introduction and applications

Leonardo De Moura and Nikolaj Bj rner. Satisfiability modulo theories: introduction and applications. Communications of the ACM, 54 0 (9): 0 69--77, 2011

2011
[12]

Z3: an efficient smt solver

Leonardo de Moura and Nikolaj Bjørner. Z3: an efficient smt solver. In 2008 Tools and Algorithms for Construction and Analysis of Systems, pages 337--340. Springer, Berlin, Heidelberg, March 2008. URL https://www.microsoft.com/en-us/research/publication/z3-an-efficient-smt-solver/

2008
[13]

Versatile verification of tree ensembles

Laurens Devos, Wannes Meert, and Jesse Davis. Versatile verification of tree ensembles. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event , volume 139 of Proceedings of Machi...

2021
[14]

Sampathkrishna Reddy, and G

Srichand Doki, Siddhartha Devella, Sumanth Tallam, Sai Sujeeth Reddy Gangannagari, P. Sampathkrishna Reddy, and G. Pradeep Reddy. Heart disease prediction using xgboost. In 2022 Third International Conference on Intelligent Computing Instrumentation and Control Technologies (I...

2022
[15]

Fairness through awareness

Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Richard Zemel. Fairness through awareness. In Proceedings of the 3rd innovations in theoretical computer science conference, pages 214--226, 2012

2012
[16]

Verifying robustness of gradient boosted models

Gil Einziger, Maayan Goldstein, Yaniv Sa'ar, and Itai Segall. Verifying robustness of gradient boosted models. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, ...

2019 doi
[17]

Breast Cancer Models

Fares Elmenshawii. Breast Cancer Models . https://www.kaggle.com/code/fareselmenshawii/breast-cancer-various-models, 2023. [Online; accessed 8-Jan-2025]

2023
[18]

Runaway feedback loops in predictive policing

Danielle Ensign, Sorelle A Friedler, Scott Neville, Carlos Scheidegger, and Suresh Venkatasubramanian. Runaway feedback loops in predictive policing. In Conference on fairness, accountability and transparency, pages 160--171. PMLR, 2018

2018
[19]

Additive logistic regression: a statistical view of boosting (with discussion and a rejoinder by the authors)

Jerome Friedman, Trevor Hastie, and Robert Tibshirani. Additive logistic regression: a statistical view of boosting (with discussion and a rejoinder by the authors). The annals of statistics, 28 0 (2): 0 337--407, 2000

2000
[20]

Comparative analysis of boosting algorithms over mnist handwritten digit dataset

Soumadittya Ghosh. Comparative analysis of boosting algorithms over mnist handwritten digit dataset. In Evolutionary Computing and Mobile Sustainable Networks: Proceedings of ICECMSN 2021, pages 985--995. Springer, 2022

2021
[21]

Gurobi Optimizer Reference Manual , 2024

Gurobi Optimization, LLC . Gurobi Optimizer Reference Manual , 2024. URL https://www.gurobi.com

2024
[22]

Cooper, Mohamed Siala, Emmanuel Hebrard, and Jo \ a o Marques - Silva

Alexey Ignatiev, Martin C. Cooper, Mohamed Siala, Emmanuel Hebrard, and Jo \ a o Marques - Silva. Towards formal fairness in machine learning. In Helmut Simonis, editor, Principles and Practice of Constraint Programming - 26th International Conference, CP 2020, Louvain-la-Neuv...

2020 doi
[23]

Alex Kantchelian, J. D. Tygar, and Anthony D. Joseph. Evasion and hardening of tree ensemble classifiers. In Maria - Florina Balcan and Kilian Q. Weinberger, editors, Proceedings of the 33nd International Conference on Machine Learning, ICML 2016, New York City, NY, USA, June ...

2016
[24]

Bankruptcy Prediction-XGB F1 0.84

Kuralamuthan Kathirvelan. Bankruptcy Prediction-XGB F1 0.84 . https://www.kaggle.com/code/kuralamuthan300/bankruptcy-prediction-xgb-f1-0-84, 2023. [Online; accessed 8-Jan-2025]

2023
[25]

Huang, Duligur Ibeling, Kyle Julian, Christopher Lazarus, Rachel Lim, Parth Shah, Shantanu Thakoor, Haoze Wu, Aleksandar Zelji \' c , David L

Guy Katz, Derek A. Huang, Duligur Ibeling, Kyle Julian, Christopher Lazarus, Rachel Lim, Parth Shah, Shantanu Thakoor, Haoze Wu, Aleksandar Zelji \' c , David L. Dill, Mykel J. Kochenderfer, and Clark Barrett. The marabou framework for verification and analysis of deep neural ...

2019
[26]

Globally-robust neural networks

Klas Leino, Zifan Wang, and Matt Fredrikson. Globally-robust neural networks. In International Conference on Machine Learning, pages 6212--6222. PMLR, 2021

2021
[27]

Delayed impact of fair machine learning

Lydia T Liu, Sarah Dean, Esther Rolf, Max Simchowitz, and Moritz Hardt. Delayed impact of fair machine learning. In International Conference on Machine Learning, pages 3150--3158. PMLR, 2018

2018
[28]

Global robustness evaluation of deep neural networks with provable guarantees for the hamming distance

Wenjie Ruan, Min Wu, Youcheng Sun, Xiaowei Huang, Daniel Kroening, and Marta Kwiatkowska. Global robustness evaluation of deep neural networks with provable guarantees for the hamming distance. IJCAI-19, 2019

2019
[29]

Higher income, larger loan? monotonicity testing of machine learning models

Arnab Sharma and Heike Wehrheim. Higher income, larger loan? monotonicity testing of machine learning models. In Proceedings of the 29th ACM SIGSOFT International Symposium on Software Testing and Analysis, pages 200--210, 2020

2020
[30]

Gagandeep Singh, Timon Gehr, Markus P \" u schel, and Martin T. Vechev. An abstract domain for certifying neural networks. Proc. ACM Program. Lang. , 3 0 ( POPL ): 0 41:1--41:30, 2019. doi:10.1145/3290354. URL https://doi.org/10.1145/3290354

2019 doi
[31]

Nuclear feature extraction for breast tumor diagnosis

Street, Nick, Wolberg, William, Mangasarian, and O. Nuclear feature extraction for breast tumor diagnosis. Proc. Soc. Photo-Opt. Inst. Eng., 1993, 01 1999. doi:10.1117/12.148698

1993 doi
[32]

Machine Failure

Yan Teixeira. Machine Failure . https://www.kaggle.com/code/yantxx/xgboost-binary-classifier-machine-failure , 2023 a . [Online; accessed 8-Jan-2025]

2023
[33]

Steel Plate Defect

Yan Teixeira. Steel Plate Defect . https://www.kaggle.com/code/aspillai/steel-plate-defect-binary-xgboost-0-89590 , 2023 b . [Online; accessed 8-Jan-2025]

2023
[34]

Zico Kolter

Shiqi Wang, Huan Zhang, Kaidi Xu, Xue Lin, Suman Jana, Cho-Jui Hsieh, and J. Zico Kolter. Beta-crown: Efficient bound propagation with per-neuron split constraints for complete and incomplete neural network robustness verification, 2021. URL https://arxiv.org/abs/2103.06624

2021 arXiv
[35]

H. P. Williams. Model building in mathematical programming. Wiley, Hoboken, N.J., 2013. ISBN 9781118443330 1118443330. URL https://ebookcentral.proquest.com/lib/uvic/detail.action?docID=1120846

2013
[36]

Understanding local robustness of deep neural networks under natural variations, 2021

Ziyuan Zhong, Yuchi Tian, and Baishakhi Ray. Understanding local robustness of deep neural networks under natural variations, 2021. URL https://arxiv.org/abs/2010.04821

2021 arXiv
[37]

Adversarial robustness of neural networks from the perspective of lipschitz calculus: A survey

Monty-Maximilian Z \"u hlke and Daniel Kudenko. Adversarial robustness of neural networks from the perspective of lipschitz calculus: A survey. ACM Computing Surveys, 2024

2024
[38]

, " * write output.state after.block = add.period write

ENTRY address author booktitle chapter doi edition editor eid howpublished institution journal key month note number organization pages publisher school series title type url volume year label INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION in...
[39]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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

[1] [1]

Sensitivity verification for decision tree ensembles

Arhaan Ahmad, Tanay Vineet Tayal, Ashutosh Gupta, and S Akshay. Sensitivity verification for decision tree ensembles. In The Thirteenth International Conference on Learning Representations

work page

[2] [2]

Satisfiability modulo theories

Clark Barrett and Cesare Tinelli. Satisfiability modulo theories. Handbook of model checking, pages 305--343, 2018

work page 2018

[3] [3]

Treant: training evasion-aware decision trees

Stefano Calzavara, Claudio Lucchese, Gabriele Tolomei, Seyum Assefa Abebe, and Salvatore Orlando. Treant: training evasion-aware decision trees. Data Min. Knowl. Discov., 34 0 (5): 0 1390--1420, 2020. doi:10.1007/S10618-020-00694-9. URL https://doi.org/10.1007/s10618-020-00694-9

work page doi:10.1007/s10618-020-00694-9 2020

[4] [4]

Boning, and Cho - Jui Hsieh

Hongge Chen, Huan Zhang, Duane S. Boning, and Cho - Jui Hsieh. Robust decision trees against adversarial examples. In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors, Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA , volume 97 of Proceedings of Machine Learning Research, pag...

work page 2019

[5] [5]

Boning, and Cho - Jui Hsieh

Hongge Chen, Huan Zhang, Si Si, Yang Li, Duane S. Boning, and Cho - Jui Hsieh. Robustness verification of tree-based models. In Hanna M. Wallach, Hugo Larochelle, Alina Beygelzimer, Florence d'Alch \' e - Buc, Emily B. Fox, and Roman Garnett, editors, Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing ...

work page 2019

[6] [6]

Xgboost: A scalable tree boosting system

Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In Balaji Krishnapuram, Mohak Shah, Alexander J. Smola, Charu C. Aggarwal, Dou Shen, and Rajeev Rastogi, editors, Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, August 13-17, 2016 , pages 785--794. ACM , 2...

arXiv 2016

[7] [7]

Learning security classifiers with verified global robustness properties

Yizheng Chen, Shiqi Wang, Yue Qin, Xiaojing Liao, Suman Jana, and David Wagner. Learning security classifiers with verified global robustness properties. In Proceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security, pages 477--494, 2021

work page 2021

[8] [8]

Query-efficient hard-label black-box attack: An optimization-based approach

Minhao Cheng, Thong Le, Pin - Yu Chen, Huan Zhang, Jinfeng Yi, and Cho - Jui Hsieh. Query-efficient hard-label black-box attack: An optimization-based approach. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019 . OpenReview.net, 2019. URL https://openreview.net/forum?id=rJlk6iRqKX

work page 2019