pith. sign in

arxiv: 1907.02509 · v1 · pith:HOE3P6CDnew · submitted 2019-07-04 · 💻 cs.LG · cs.AI· cs.LO

On Validating, Repairing and Refining Heuristic ML Explanations

Pith reviewed 2026-05-25 09:06 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.LO
keywords machine learning explanationsheuristic explanationsboosted treesrigorous explanationsinstance spaceexplanation validationexplanation repair
0
0 comments X

The pith

Heuristic explanations for boosted-tree predictions are inadequate for most instances when checked over the full instance space.

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

The paper extends rigorous explanation methods to boosted trees and uses them to test common heuristic approaches on multiple datasets. It finds that heuristic explanations fail to remain valid for the vast majority of instances once the entire space of possible inputs is taken into account. This result matters because heuristic methods are the default choice for explaining non-interpretable models, yet their outputs can be unreliable without exhaustive verification. The work also outlines steps for validating, repairing, and refining those heuristic outputs using the rigorous methods as reference.

Core claim

When rigorous methods are applied as ground truth, state-of-the-art heuristic explanations for boosted trees prove inadequate on most datasets and for the vast majority of instances once validity is required across the complete instance space.

What carries the argument

Rigorous explanation computation for boosted trees, serving as an exact reference to measure whether heuristic explanations hold everywhere.

If this is right

  • Heuristic explanations require post-processing or repair to restore validity over the full instance space.
  • Rigorous methods can be used to detect and correct failures in heuristic outputs for boosted trees.
  • Explanation quality must be assessed against the complete input space rather than local approximations alone.
  • Many current heuristic tools produce outputs that cannot be trusted without additional validation.

Where Pith is reading between the lines

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

  • Practical explanation pipelines may need to combine heuristic speed with periodic rigorous checks rather than relying on heuristics alone.
  • The gap between heuristic and rigorous outputs could motivate new hybrid algorithms that start heuristic and finish with targeted rigorous repair.
  • If the inadequacy pattern holds beyond the tested datasets, explanation reliability becomes a property of the method rather than the model.

Load-bearing premise

The rigorous methods correctly identify every valid and invalid explanation over the entire instance space.

What would settle it

A dataset and boosted-tree model where a standard heuristic explanation remains valid for every possible input instance, contrary to the reported failure rates on most datasets.

Figures

Figures reproduced from arXiv: 1907.02509 by Alexey Ignatiev, Joao Marques-Silva, Nina Narodytska.

Figure 1
Figure 1. Figure 1: Example of a simplistic model trained by XGBoost. The model targets the well-known [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
read the original abstract

Recent years have witnessed a fast-growing interest in computing explanations for Machine Learning (ML) models predictions. For non-interpretable ML models, the most commonly used approaches for computing explanations are heuristic in nature. In contrast, recent work proposed rigorous approaches for computing explanations, which hold for a given ML model and prediction over the entire instance space. This paper extends earlier work to the case of boosted trees and assesses the quality of explanations obtained with state-of-the-art heuristic approaches. On most of the datasets considered, and for the vast majority of instances, the explanations obtained with heuristic approaches are shown to be inadequate when the entire instance space is (implicitly) considered.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The manuscript extends prior rigorous explanation techniques to boosted trees and compares them against state-of-the-art heuristic methods. It reports that, on most of the datasets examined, heuristic explanations are inadequate for the vast majority of instances when validity is assessed over the entire instance space.

Significance. If the rigorous procedures are sound, the empirical comparison would demonstrate concrete limitations of heuristic explanations for tree ensembles and motivate the use of validation/repair steps. The extension of rigorous methods to boosted trees is a useful technical step for the XAI literature.

major comments (1)
  1. [Abstract and methodology description of rigorous methods] The headline claim (heuristics inadequate on most datasets for the vast majority of instances) treats the rigorous explanation procedures as a complete and correct oracle over the full instance space. No independent soundness argument, enumeration check, or verification experiment for these rigorous procedures on boosted trees is supplied, which is load-bearing for the measured inadequacy rates.
minor comments (2)
  1. [Abstract] The abstract supplies no quantitative details on the number of datasets, models, or exact inadequacy percentages, making it difficult to assess the strength of the empirical claim.
  2. [Introduction] Notation for the instance space and the precise definition of 'inadequate' should be introduced earlier and used consistently.

Simulated Author's Rebuttal

1 responses · 0 unresolved

Thank you for the careful review and for highlighting the importance of the soundness of the rigorous procedures. We address the major comment below and will revise the manuscript to strengthen the justification.

read point-by-point responses
  1. Referee: The headline claim (heuristics inadequate on most datasets for the vast majority of instances) treats the rigorous explanation procedures as a complete and correct oracle over the full instance space. No independent soundness argument, enumeration check, or verification experiment for these rigorous procedures on boosted trees is supplied, which is load-bearing for the measured inadequacy rates.

    Authors: We agree that the soundness of the rigorous methods is foundational to the empirical claims. These procedures extend the constraint-based exact explanation methods from prior work on single decision trees, where formal soundness was established. For boosted trees the extension adapts the encoding to the additive model output while preserving exact equivalence to the model's decision over the full instance space; the validity guarantees therefore carry over by construction. We acknowledge that the manuscript did not include an explicit soundness subsection or small-scale verification experiment for the boosted-tree case. In revision we will add a dedicated subsection detailing the soundness argument (including how base-case properties are preserved) and a limited verification experiment on synthetic data where full enumeration is feasible. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical comparison to external rigorous methods

full rationale

The paper performs an empirical assessment of heuristic explanations against rigorous methods extended to boosted trees. No equations, derivations, fitted parameters, or self-referential definitions appear in the abstract or described claims. The central result (inadequacy of heuristics) is obtained by direct comparison rather than any reduction to inputs by construction. Prior work on rigorous methods is cited but functions as independent support for the ground truth, with no load-bearing self-citation chain or ansatz smuggling exhibited. The derivation chain is self-contained as a validation study.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no methods, parameters, or assumptions are detailed enough to populate the ledger.

pith-pipeline@v0.9.0 · 5647 in / 1032 out tokens · 34060 ms · 2026-05-25T09:06:41.114986+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

52 extracted references · 52 canonical work pages

  1. [1]

    In: NeurIPS

    Adebayo, J., Gilmer, J., Muelly, M., Goodfellow, I.J., Hardt, M., Kim, B.: Sanity checks for saliency maps. In: NeurIPS. pp. 9525–9536 (2018)

  2. [2]

    In: NeurIPS

    Alvarez-Melis, D., Jaakkola, T.S.: Towards robust interpretability with self-explaining neural networks. In: NeurIPS. pp. 7786–7795 (2018)

  3. [3]

    Angelino, E., Larus-Stone, N., Alabi, D., Seltzer, M., Rudin, C.: Learning certifiably optimal rule lists. In: KDD. pp. 35–44 (2017)

  4. [4]

    Journal of Machine Learning Research11, 1803– 1831 (2010)

    Baehrens, D., Schroeter, T., Harmeling, S., Kawanabe, M., Hansen, K., M ¨uller, K.: How to explain individual classification decisions. Journal of Machine Learning Research11, 1803– 1831 (2010)

  5. [5]

    In: Handbook of Model Checking., pp

    Barrett, C., Tinelli, C.: Satisfiability modulo theories. In: Handbook of Model Checking., pp. 305–343 (2018)

  6. [6]

    In: Biere, A., Heule, M., van Maaren, H., Walsh, T

    Barrett, C.W., Sebastiani, R., Seshia, S.A., Tinelli, C.: Satisfiability modulo theories. In: Biere, A., Heule, M., van Maaren, H., Walsh, T. (eds.) Handbook of Satisfiability, Frontiers in Artificial Intelligence and Applications, vol. 185, pp. 825–885. IOS Press (2009)

  7. [7]

    Boutilier, C., Brafman, R.I., Domshlak, C., Hoos, H.H., Poole, D.: Cp-nets: A tool for repre- senting and reasoning with conditional ceteris paribus preference statements. J. Artif. Intell. Res. 21, 135–191 (2004)

  8. [8]

    In: SODA

    Chandrasekaran, K., Karp, R.M., Moreno-Centeno, E., Vempala, S.: Algorithms for implicit hitting set problems. In: SODA. pp. 614–629 (2011)

  9. [9]

    Chen, T., Guestrin, C.: XGBoost: A scalable tree boosting system. In: KDD. pp. 785–794. ACM (2016)

  10. [10]

    In: NIPS

    Craven, M., Shavlik, J.W.: Extracting tree-structured representations of trained networks. In: NIPS. pp. 24–30 (1995)

  11. [11]

    DARPA: DARPA explainable Artificial Intelligence (XAI) program (2016)

  12. [12]

    In: CA V

    Dillig, I., Dillig, T., McMillan, K.L., Aiken, A.: Minimum satisfying assignments for SMT. In: CA V . pp. 394–409 (2012)

  13. [13]

    Eiter, T., Gottlob, G.: The complexity of logic-based abduction. J. ACM 42(1), 3–42 (1995)

  14. [14]

    EU Data Protection Regulation: Regulation (EU) 2016/679 of the European Parliament and of the Council (2016)

  15. [15]

    https://blog.fastforwardlabs.com/2017/ 03/09/fairml-auditing-black-box-predictive-models.html (2016)

    Auditing black-box predictive models. https://blog.fastforwardlabs.com/2017/ 03/09/fairml-auditing-black-box-predictive-models.html (2016)

  16. [16]

    Feldman, M., Friedler, S.A., Moeller, J., Scheidegger, C., Venkatasubramanian, S.: Certify- ing and removing disparate impact. In: KDD. pp. 259–268. ACM (2015)

  17. [17]

    Constraints 23(3), 296–309 (2018)

    Fischetti, M., Jo, J.: Deep neural networks and mixed integer linear optimization. Constraints 23(3), 296–309 (2018)

  18. [18]

    Friedler, S., Scheidegger, C., Venkatasubramanian, S.: On algorithmic fairness, discrimina- tion and disparate impact (2015)

  19. [19]

    In: CExAIIA (2017)

    Frosst, N., Hinton, G.E.: Distilling a neural network into a soft decision tree. In: CExAIIA (2017)

  20. [20]

    Dover (2003)

    Gallier, J.H.: Logic For Computer Science. Dover (2003)

  21. [21]

    AI Magazine 38(3), 50–57 (2017)

    Goodman, B., Flaxman, S.R.: European Union regulations on algorithmic decision-making and a ”right to explanation”. AI Magazine 38(3), 50–57 (2017)

  22. [22]

    In: CA V

    Huang, X., Kwiatkowska, M., Wang, S., Wu, M.: Safety verification of deep neural networks. In: CA V . pp. 3–29 (2017)

  23. [23]

    In: ECAI

    Ignatiev, A., Morgado, A., Marques-Silva, J.: Propositional abduction with implicit hitting sets. In: ECAI. pp. 1327–1335 (2016)

  24. [24]

    In: AAAI (2019) On Validating, Repairing and Refining Heuristic ML Explanations 17

    Ignatiev, A., Narodytska, N., Marques-Silva, J.: Abduction-based explanations for machine learning models. In: AAAI (2019) On Validating, Repairing and Refining Heuristic ML Explanations 17

  25. [25]

    In: IJCAR

    Ignatiev, A., Pereira, F., Narodytska, N., Marques-Silva, J.: A SAT-based approach to learn explainable decision sets. In: IJCAR. pp. 627–645 (2018)

  26. [26]

    In: CA V

    Katz, G., Barrett, C.W., Dill, D.L., Julian, K., Kochenderfer, M.J.: Reluplex: An efficient SMT solver for verifying deep neural networks. In: CA V . pp. 97–117 (2017)

  27. [27]

    In: NIPS

    Kim, B., Shah, J.A., Doshi-Velez, F.: Mind the gap: A generative approach to interpretable feature selection and extraction. In: NIPS. pp. 2260–2268 (2015)

  28. [28]

    Kohavi, R.: Scaling up the accuracy of naive-bayes classifiers: A decision-tree hybrid. In: KDD. pp. 202–207 (1996)

  29. [29]

    Lakkaraju, H., Bach, S.H., Leskovec, J.: Interpretable decision sets: A joint framework for description and prediction. In: KDD. pp. 1675–1684 (2016)

  30. [30]

    In: AAAI

    Li, O., Liu, H., Chen, C., Rudin, C.: Deep learning for case-based reasoning through proto- types: A neural network that explains its predictions. In: AAAI. pp. 3530–3537 (2018)

  31. [31]

    In: NIPS

    Lundberg, S.M., Lee, S.: A unified approach to interpreting model predictions. In: NIPS. pp. 4768–4777 (2017)

  32. [32]

    Marques-Silva, J., Previti, A.: On computing preferred muses and mcses. In: SAT. pp. 58–74 (2014)

  33. [33]

    In: FAIR

    Marquis, P.: Extending abduction from propositional to first-order logic. In: FAIR. pp. 141– 155 (1991)

  34. [34]

    In: Handbook of Defeasible Reasoning and Uncertainty Management Systems, pp

    Marquis, P.: Consequence finding algorithms. In: Handbook of Defeasible Reasoning and Uncertainty Management Systems, pp. 41–145 (2000)

  35. [35]

    Digital Signal Processing 73, 1–15 (2018)

    Montavon, G., Samek, W., M ¨uller, K.: Methods for interpreting and understanding deep neural networks. Digital Signal Processing 73, 1–15 (2018)

  36. [36]

    In: TACAS

    de Moura, L.M., Bjørner, N.: Z3: an efficient SMT solver. In: TACAS. pp. 337–340 (2008)

  37. [37]

    In: IJCAI

    Narodytska, N., Ignatiev, A., Pereira, F., Marques-Silva, J.: Learning optimal decision trees with SAT. In: IJCAI. pp. 1362–1368 (2018)

  38. [38]

    Ribeiro, M.T., Singh, S., Guestrin, C.: ”Why should I trust you?”: Explaining the predictions of any classifier. In: KDD. pp. 1135–1144 (2016)

  39. [39]

    In: AAAI (2018)

    Ribeiro, M.T., Singh, S., Guestrin, C.: Anchors: High-precision model-agnostic explana- tions. In: AAAI (2018)

  40. [40]

    AI Commun

    Rosa, E.D., Giunchiglia, E.: Combining approaches for solving satisfiability problems with qualitative preferences. AI Commun. 26(4), 395–408 (2013)

  41. [41]

    In: AAAI

    Ross, A.S., Doshi-Velez, F.: Improving the adversarial robustness and interpretability of deep neural networks by regularizing their input gradients. In: AAAI. pp. 1660–1669 (2018)

  42. [42]

    In: IJCAI

    Ross, A.S., Hughes, M.C., Doshi-Velez, F.: Right for the right reasons: Training differen- tiable models by constraining their explanations. In: IJCAI. pp. 2662–2670 (2017)

  43. [43]

    In: IJCAI

    Ruan, W., Huang, X., Kwiatkowska, M.: Reachability analysis of deep neural networks with provable guarantees. In: IJCAI. pp. 2651–2659 (2018)

  44. [44]

    Saikko, P., Wallner, J.P., J ¨arvisalo, M.: Implicit hitting set algorithms for reasoning beyond NP. In: KR. pp. 104–113 (2016)

  45. [45]

    Inter- University Consortium for Political and Social Research (1988)

    Schmidt, P., Witte, A.D.: Predicting recidivism in north carolina, 1978 and 1980. Inter- University Consortium for Political and Social Research (1988)

  46. [46]

    In: IJCAI

    Shih, A., Choi, A., Darwiche, A.: A symbolic approach to explaining bayesian network clas- sifiers. In: IJCAI. pp. 5103–5111 (2018)

  47. [47]

    Journal of Machine Learning Research 11, 1–18 (2010)

    Strumbelj, E., Kononenko, I.: An efficient explanation of individual classifications using game theory. Journal of Machine Learning Research 11, 1–18 (2010)

  48. [48]

    Data Knowl

    Strumbelj, E., Kononenko, I., Robnik-Sikonja, M.: Explaining instance classifications with interactions of subsets of feature values. Data Knowl. Eng. 68(10), 886–904 (2009)

  49. [49]

    In: ASRU

    Tan, S., Sim, K.C., Gales, M.J.F.: Improving the interpretability of deep neural networks with stimulated learning. In: ASRU. pp. 617–623. IEEE (2015) 18 Ignatiev et al

  50. [50]

    IEEE Trans

    Umans, C., Villa, T., Sangiovanni-Vincentelli, A.L.: Complexity of two-level logic mini- mization. IEEE Trans. on CAD of Integrated Circuits and Systems25(7), 1230–1246 (2006)

  51. [51]

    In: ICDM

    Wang, T., Rudin, C., Doshi-Velez, F., Liu, Y ., Klampfl, E., MacNeille, P.: Bayesian rule sets for interpretable classification. In: ICDM. pp. 1269–1274 (2016)

  52. [52]

    In: AAAI

    Wu, M., Hughes, M.C., Parbhoo, S., Zazzi, M., Roth, V ., Doshi-Velez, F.: Beyond sparsity: Tree regularization of deep models for interpretability. In: AAAI. pp. 1670–1678 (2018)