CoChat

An Empirical Comparison of Interpretable Models to Post-Hoc Explanations

Open access

The DOI matches its Crossref record; 22 of the 22 references it cites were found in a registry or an index; 1 of the 5 sources its claims cite were checked for support.

Abstract

Recently, some effort went into explaining intransparent and black-box models, such as deep neural networks or random forests. So-called model-agnostic methods typically approximate the prediction of the intransparent black-box model with an interpretable model, without considering any specifics of the black-box model itself. It is a valid question whether direct learning of interpretable white-box models should not be preferred over post-hoc approximations of intransparent and black-box models. In this paper, we report the results of an empirical study, which compares post-hoc explanations and interpretable models on several datasets for rule-based and feature-based interpretable models. The results seem to underline that often directly learned interpretable models approximate the black-box models at least as well as their post-hoc surrogates, even though the former do not have direct access to the black-box model.

Overview

CoChat Summary

TL;DR

This paper empirically compares directly learned interpretable models against post-hoc surrogate explanations for rule-based and feature-based settings. It finds that directly learned interpretable models often approximate black-box models at least as well as post-hoc surrogates in accuracy and fidelity, even without access to the black-box model.

Key findings

  • For rule-based models on the adult dataset, JRIP achieved accuracy 0.839, fidelity 0.957, and 28 rules, while GLOCALX’s best fidelity setting (α = 10) achieved accuracy 0.762, fidelity 0.942, and 10 rules. Across datasets, JRIP had slightly higher accuracy and fidelity than GLOCALX: adult 0.839/0.957 vs. 0.762/0.942; German 0.712/0.809 vs. 0.700/0.764; NHANES I 0.823/0.838 vs. 0.788/0.819; compas-two-year 0.782/0.836 vs. 0.756/0.804.
  • For feature-based models, the paper states GA2MS had higher accuracy and fidelity than MAPLE on all datasets; Table 4 shows, for example, Adult GA2MS 0.851/0.901 vs. MAPLE 0.723/0.734, and NHANES I GA2MS 0.826/0.867 vs. MAPLE 0.745/0.788. One table entry is an accuracy exception to the text’s “all datasets” claim: credit card fraud MAPLE accuracy 0.947 vs. GA2MS 0.939, though GA2MS fidelity is higher (0.989 vs. 0.955).
  • The paper reports Rashomon-effect differences: MAPLE and GA2MS give different feature importances on adult, e.g., age 0.187 vs. 0.665, marital status 0.002 vs. 0.629, capital gain 0.640 vs. 0.551, and relationship −0.050 vs. 0.487.
  • GLOCALX generally generated shorter rules with fewer conditions than JRIP, but JRIP obtained higher accuracy while generating more complex rules.

Methods

  • Binary classification experiments used datasets including adult, compas-two-year, German, NHANES I, credit card fraud, Bank, and Heart Disease, mostly from the UCI repository.
  • Rule-based comparison paired GLOCALX (post-hoc, using random forests as black-box and LORE for local explanations) against JRIP (direct rule learner). Data were split into 60% for black-box training, 20% for GLOCALX training, and 20% for validation; JRIP used the first 80% as training.
  • Feature-based comparison paired MAPLE (post-hoc) against GA2MS (interpretable GAM). Evaluation used accuracy, fidelity, and number of rules for rule-based models; feature-based fidelity used normalized feature-importance weights, a linear scoring rule, ROC-based threshold tuning, and AUC.

Limitations

  • The paper does not state a dedicated limitations section or explicit limitations.
  • Evident from the text: experiments are limited to binary classification datasets, and the rule-based setup uses one black-box model (random forests) and one local explainer (LORE) for GLOCALX.
  • Evident from the text: the feature-based fidelity comparison uses a proxy linear model with normalized weights and threshold tuning rather than native MAPLE/GA2MS prediction outputs, and no statistical tests are reported.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3390/ai4020023

Cites

22 of 22 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

1 Valid · 0 Not valid · 4 Not checked · 16 Own results and interpretations

2 more claims

Create a free account to see all 21 claims and every source verdict.

See all claims — free account
  • Probably not a substantial claim. 7
  • It may state two separate claims as one. 11

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

2 of 2

Citations

Cites

Show all 22 references

Research works printed but not linked: 6

  1. Krishnapuram, B., Shah, M., Smola, A.J., Aggarwal, C.C., Shen, D., and Rastogi, R. (2016). Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), ACM.
  2. Guyon, I., von Luxburg, U., Bengio, S., Wallach, H.M., Fergus, R., Vishwanathan, S.V.N., and Garnett, R. (2017). Advances in Neural Information Processing Systems 30, Proceedings of the NIPS 2017, Long Beach, CA, USA, 4–9 December 2017, Curran Associates Inc.
  3. Samek, W., Montavon, G., Vedaldi, A., Hansen, L.K., and Müller, K. (2019). Explainable AI: Interpreting, Explaining and Visualizing Deep Learning, Springer. · DOI · Paper page
  4. Prieditis, A., and Russell, S. (1995). Machine Learning Proceedings 1995, Proceedings of the Twelfth International Conference on Machine Learning, Tahoe City, CA, USA, 9–12 July 1995, Morgan Kaufmann.
  5. Bengio, S., Wallach, H.M., Larochelle, H., Grauman, K., Cesa-Bianchi, N., and Garnett, R. (2018). Advances in Neural Information Processing Systems 31, Proceedings of the NeurIPS 2018, Montreal, QC, Canada, 3–8 December 2018, Curran Associates Inc.
  6. Knobbe, A.J. (2008). From Local Patterns to Global Models: Proceedings of the ECML/PKDD-08 Workshop (LeGo-08), University and State Library Darmstadt.

Cited by

See all 17

Metadata

Authors
Mahya, P., & Fürnkranz, J.
Affiliations
Johannes Kepler University of Linz (AT)
Corresponding author
Mahya, Parisa
Venue
AI
Publisher
MDPI AG
Published
2023-05-19
Volume
4
Issue
2
Pages
426-436
ISSN
2673-2688 (ISSN-L)
OpenAlex
W4377115400
Cited by
17
References
14
Citations by year
2026: 5 · 2025: 9 · 2024: 2 · 2023: 1
FWCI
1.21
Type
journal article
Language
en
Open access
Locations
  • AI (publishedVersion, cc-by)
  • University Library Linz repository (Johannes Kepler Universitat Linz) (submittedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
  • Physical Sciences › Computer Science › Software › Software Testing and Debugging Techniques
  • Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning and Data Classification
Keywords
interpretable models, post-hoc explanation, model-agnostic explanation, black-box models, white-box models, rule-based models, feature-based models, explainable machine learning, model approximation

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-10-02, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-02 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: An Empirical Comparison of Interpretable Models to Post-Hoc Explanations. https://app.cochat.ai/papers/10.3390/ai4020023
The paper
Mahya, P., & Fürnkranz, J. (2023). An Empirical Comparison of Interpretable Models to Post-Hoc Explanations. AI. https://doi.org/10.3390/ai4020023

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.