Measurable Counterfactual Local Explanations for Any Classifier

Adam White; Artur d'Avila Garcez

REVIEW 3 major objections 5 minor 1 cited by

Measurable Counterfactual Local Explanations for Any Classifier

T0 review · 3 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read CLEAR generates local explanations that state the minimum changes to flip a prediction and measure their own fidelity, beating LIME by over 40% on average.

desk verdict CLEAR is a useful new local explanation method with a genuine counterfactual fidelity metric, but the headline fidelity gains from adding b-counterfactuals are circular and overstated; the non-circular version still beats LIME. read the letter →

arxiv 1908.03020 v2 pith:AE6R2DE5 submitted 2019-08-08 cs.AI cs.LG

Adam White , Artur d'Avila Garcez This is my paper · ORCID

classification cs.AIcs.LG

keywords counterfactualexplanationslocalmodel-agnosticfidelitymeasurementb-perturbationsregressiondecisionboundaryLIME

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

This paper argues that a satisfactory local explanation of a machine learning prediction has to do two things at once: answer counterfactual what-if questions about how the prediction would change if features were different, and state how accurately the explanation captures the classifier's actual behavior. To that end it introduces CLEAR, which first finds, by querying the classifier, the minimum change to each feature that would flip the prediction to another class, and then fits a local regression that goes through the original point. The regression is inverted to estimate those same minimum changes, and the gap between the estimated and actual changes defines a counterfactual fidelity error. Across five datasets, CLEAR's regressions achieve substantially higher fidelity than LIME's, averaging over 40% higher, because CLEAR uses a balanced neighborhood spanning the decision boundary and can include quadratic and interaction terms.

What carries the argument

The machinery has three parts. The b-perturbation is the unit of counterfactual truth: for each feature, a one-dimensional search moves away from x until the classifier's output crosses the decision boundary, yielding the minimal change required for a class flip. The balanced neighborhood dataset is constructed by sampling synthetic points and keeping about 200 distributed roughly evenly across probability bands between x and the classifier's decision boundary, so that the regression sees the region where counterfactual claims apply. Finally, the regression is constrained to pass through x, may include quadratic and interaction terms, and is inverted to estimate b-perturbations; the difference between those estimates and the queried values is the fidelity measure that lets CLEAR report when it does not know.

What would settle it

On a synthetic classifier with a known decision boundary, compute the true minimum feature changes by exhaustive grid search, run CLEAR with different values of the neighborhood margins b1 and b2, and compare its estimated changes with the true ones; if its fidelity changes materially across reasonable margin choices, especially when the regression's R-squared stays high, then the balanced-neighborhood assumption is not carrying the reported accuracy.

Watch

Extended reading notes

Core claim

The central claim is that explanation quality should be measured by counterfactual fidelity, not by how well a regression fits its training neighborhood or how often it reproduces the predicted class. The paper introduces the b-perturbation: the change in one feature's value, with all other features held fixed, that is just sufficient to move the classifier's output to a different class. CLEAR computes actual b-perturbations by one-dimensional searches that query the classifier, builds a local regression on a balanced neighborhood dataset, and estimates each b-perturbation by solving the regression equation at the decision boundary. The absolute difference between the estimated and actual b-perturbation is the fidelity error; a b-perturbation counts as faithful when that error falls below a user-chosen threshold. The paper reports that on its five case studies CLEAR's percentage fidelity is higher than LIME's in every configuration, with an average improvement of over 40%.

Load-bearing premise

The load-bearing assumption is that the balanced neighborhood of synthetic points, selected in probability bands between x and the decision boundary, is representative enough of the counterfactual region that regressions fit on it yield trustworthy estimates of the minimum changes that flip a prediction.

Editorial extensions

If this is right

Explanation reports can include a per-explanation fidelity error, so a user can see whether a local explanation is trustworthy rather than assume it is.
LIME-style local regressions can be badly misleading about how much a feature must change to alter a prediction, even when their adjusted R-squared and classification accuracy look high.
Neighborhood design matters: data balanced across probability bands up to and beyond the decision boundary supports counterfactual estimation much better than points weighted tightly around x.
Including the actual b-counterfactual points as weighted constraints in the regression improves fidelity on most of the tested datasets.
Fidelity to b-perturbations is a stricter test than reproducing predicted probabilities, so it can serve as a common metric for comparing local explainers.

Reading between the lines

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

The paper leaves implicit that the same fidelity measure could select among explanation configurations: choose the regression family, neighborhood size, or term set with the lowest average counterfactual fidelity error, turning the interpretability/fidelity trade-off into an empirical search.
Because the balanced-neighborhood selection depends on two unspecified margins, b1 and b2, a robustness test would sweep these margins and report how fidelity changes; the reported averages would be stronger if they hold across a range of margin choices.
The single-feature search for b-perturbations could be extended to joint counterfactuals involving two or more features by inverting the regression along a shortest path to the decision boundary, with fidelity measured against multi-feature searches; the paper lists this as future work.
The fidelity protocol suggests a general benchmark for any local explainer: generate a test set, compute actual b-perturbations by querying the classifier, and compare each explainer's estimates, making explanation accuracy comparable across methods.

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.

Desk Editor's Note

read the letter

CLEAR is a reasonable new entry in the local-explainability space: it adds counterfactual fidelity measurement to LIME-style regression, and on the non-circular comparison it beats LIME. But the paper's headline 'using b-counterfactuals' numbers are inflated by an in-sample evaluation loop. The stress-test note is right.

What's actually new: the paper defines b-perturbations, estimates them from a locally weighted regression, and reports a fidelity error against the actual classifier. That combination—measuring how well a local regression reproduces the classifier's own minimum-change counterfactuals—is genuinely useful and not in the cited LIME, SHAP, or Anchor work. The balanced-neighborhood selection and constrained regression through x are sensible choices. The GitHub code and honest discussion of cases where fidelity is poor (e.g., the Adult example with R² 0.97 but fidelity 59%) are to their credit.

Soft spots: Step 8 is the big one. Adding the actual b-counterfactual points to the regression training set with weight 10, then computing fidelity against those same points, means the regression is essentially interpolating its own training targets. That explains the sharp jumps from 57% to 77% (Pima) and 54% to 81% (Breast). The non-circular row (CLEAR not using b-counterfactuals) still beats LIME by a wide margin—80% vs 26% on Adult, for instance—so the method itself has value. But the specific claim that adding b-counterfactuals improves fidelity is not established by the reported protocol. A hold-out or cross-validation split is needed.

Minor issues: the margins b1 and b2 in Algorithm 2 are never specified; the step size in the one-dimensional search for actual b-perturbations is unclear; and configuration choices (kernel width for LIME, number of synthetic points, number of independent variables) appear to be tuned on the test set without a clear validation split. None of these are fatal, but they should be documented.

Overall: a solid idea with a flawed evaluation protocol. The paper is worth a serious referee—the fidelity metric and the balanced-neighborhood idea are contributions—but it needs revision before the quantitative claims can be trusted.

Recommendation: send to peer review, request an out-of-sample fidelity protocol and full parameter disclosure.

Referee Report

3 major / 5 minor

Summary. The paper proposes CLEAR, a local explanation method for arbitrary classifiers. CLEAR finds, for a prediction x, actual 'b-counterfactual' feature changes that flip the classifier's decision, builds a locally weighted regression (through x) on a balanced synthetic neighbourhood, and then uses that regression to estimate b-perturbations. Fidelity is defined as the proportion of estimated b-perturbations whose error against actual b-perturbations is below a threshold. The paper claims three contributions: counterfactual explanations with regression coefficients and interactions, a measurable fidelity definition, and the use of b-counterfactuals to improve regression fidelity. Experiments on five UCI datasets compare CLEAR with LIME and report substantially higher fidelity, with further gains when b-counterfactuals are added to the neighbourhood dataset.

Significance. The paper addresses an important gap in local explanation methods: the lack of a quantitative fidelity measure tied to counterfactual behaviour. If the evaluation were sound, the proposed metric and the idea of using boundary-counterfactual points to constrain local regressions would be a useful contribution. The non-circular comparison (CLEAR without b-counterfactuals versus LIME) is interesting and appears to favour CLEAR, and the authors provide prototype code for reproducibility. However, the headline quantitative claim that adding b-counterfactuals sharply improves fidelity is undermined by an in-sample evaluation loop, so the significance of the third contribution is not currently established.

major comments (3)

[3 (Algorithm 1) and Table 1] The headline row 'CLEAR using b-counterfactuals' is evaluated in-sample. Step 8 appends the actual b-counterfactual points w to the neighbourhood dataset N_tc, each with weight 10, and the regression r is then fit on this augmented dataset. The fidelity error e is computed by comparing the estimated b-perturbations w' (derived from r) with the same w. Because r has been fit to w, solving the regression for the boundary will largely reproduce these training targets; the reported increases (e.g., 57% to 77% on Pima, 54% to 81% on Breast) therefore largely reflect interpolation rather than explanatory generalization. The claim that adding b-counterfactuals improves fidelity (the paper's third contribution) is not established by this protocol. Please report a hold-out or cross-validated evaluation in which the actual b-counterfactual points used as evaluation targets are excluded from the regression fit, or use the w points only for evaluation and not for training. The non-circular row 'CLEAR not using b-counterfactuals' can remain and supports the weaker comparison with LIME.
[3 (Algorithm 2)] The balanced-neighbourhood construction depends on the margins b1 and b2 around the decision boundary, used to split the selected points into N1, N2, and N3. The paper never specifies the values of b1 and b2 for any dataset, nor does it provide a sensitivity analysis. Since Section 4 and Figure 3 argue that balanced neighbourhood selection is a key source of CLEAR's advantage, the absence of these parameter values makes the experiments unreproducible and leaves open whether the particular choice of margins drives the reported results. Please state the values used for each dataset and show how fidelity varies with b1 and b2.
[4 (Table 1 and LIME comparison)] The paper repeatedly states that CLEAR's fidelity is 'significantly' higher than LIME's, but no statistical significance tests are reported. Although the gaps in Table 1 are large (e.g., 20% vs 57% on Pima), a formal test (or at least confidence intervals with a stated test procedure) is needed to support the word 'significantly'. In addition, LIME is evaluated on a metric it was not designed to optimize; this may be defensible given the paper's argument from Woodward's theory of explanation, but the comparison should be framed explicitly as an evaluation of LIME on CLEAR's counterfactual-fidelity metric rather than as a measure of LIME's overall quality.

minor comments (5)

[Abstract and Introduction] The abstract states 'averaging over 45% higher in this paper's four case studies', while the introduction and full-text abstract say 'averaging over 40% higher in this paper's five case studies' and Table 1 lists five datasets; please harmonize the number of case studies and the reported average improvement.
[Section 3] The definitions are numbered as 'Definition 5.1' and 'Definition 5.3', but these are the first and third definitions in Section 3; the numbering should be corrected.
[Throughout] There are spelling errors in proper names: 'Woodard' should be 'Woodward', and 'Riberio' should be 'Ribeiro'.
[Algorithm 1] In the pseudocode, the inner loop 'for each feature f do w<- Find Counterfactuals(x,m)' does not pass f to the function, and it is not clear how w accumulates the per-feature b-counterfactuals; please make the pseudocode explicit.
[3 (Step 1) and 4] The one-dimensional search for actual b-perturbations uses 'regular amounts' but the step size is never specified, and the error threshold T is only mentioned as 0.25 in the text; please report these implementation details for reproducibility.

Circularity Check

1 steps flagged · score 6.0 of 10

Fidelity improvement from optional b-counterfactual training is evaluated in-sample, making that specific claim partially circular; the CLEAR-vs-LIME comparison is independent.

fitted input called prediction [Section 3, method step 8 and Algorithm 1; Table 1]
"8. CLEAR also provides the option of adding x’s b-counterfactuals, minf (x), to x’s neighbourhood dataset. The b-counterfactuals are weighted and act as soft constraints on CLEAR’s subsequent regression. ... Optional: Ntc← Ntc∪ w ... w′← Estimate Counterfactuals(r,x) ... e← Calculate Fidelity(w,w’,T)"

The actual b-counterfactuals w are first discovered by querying the classifier (step 1). Step 8 then inserts those same w into the neighbourhood dataset Ntc with a weight of 10, and the regression r is fit on this augmented data. The fidelity error e is computed by comparing the estimated b-perturbations w′ derived from r with the same w. Thus w functions simultaneously as training input and as evaluation target; small errors show in-sample interpolation rather than predictive generalization. Table 1's jumps (Iris 80% to 99.8%, Pima 57% to 77%, Breast 54% to 81%) therefore do not establish the claimed benefit of adding b-counterfactuals. A hold-out or cross-validated protocol would be needed. The non-circular 'CLEAR not using b-counterfactuals' row still independently beats LIME.

full rationale

The paper's central method and its comparison against LIME are not circular: CLEAR's balanced neighbourhood sampling, centered regression, and quadratic/interaction terms are evaluated on b-counterfactuals computed from the classifier, and these steps do not presuppose the conclusion. The main circularity is localized to the optional Step 8 evaluation: actual b-counterfactuals are added to the regression training set and then used as the fidelity targets, so the reported improvement from using b-counterfactuals is at least partly a fitted-value artifact. Because this is one of the paper's three advertised contributions and drives the headline Table 1 numbers, the paper is partially circular, but the core comparison with LIME retains independent content.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests on several chosen numbers (b1, b2, weighting 10, threshold T, neighborhood size, max features) and on domain assumptions about local regression expressiveness and the representative power of balanced synthetic neighborhoods. No new physical or mathematical entities are introduced.

free parameters (7)

b1, b2 decision-boundary margins = not specified
Algorithm 2 divides the neighborhood into three probability bands using margins b1 and b2; values are never stated in the paper, and the balanced selection procedure depends on them.
b-counterfactual weight = 10
In step 8, each b-counterfactual is added to the neighborhood with weight 10; this weighting sharply affects the fitted regression and the reported fidelity.
fidelity error threshold T = 0.25
Percentage fidelity is defined as the fraction of estimated b-perturbations with error below T; the paper uses T=0.25 without sensitivity analysis.
neighborhood size n = 200
A balanced neighborhood of 200 synthetic observations is used as the default size.
synthetic data count = 50,000
CLEAR generates 50,000 synthetic points by default before selecting the balanced neighborhood.
max independent variables = 14 (best configuration)
The best configuration limits regressions to 14 features; Table 2 shows fidelity depends strongly on this choice.
search step size for b-perturbations = regular increments, value not given
Step 1 searches for flips by moving away from x in regular increments; the step size is not specified, affecting the accuracy of the ground-truth b-perturbations.

assumptions (5)

domain assumption A local regression with second-degree and interaction terms is sufficiently expressive to approximate the classifier's input-output behavior between x and its decision boundary.
CLEAR's estimated b-perturbations are computed by solving this regression equation; if the true classifier is strongly non-linear, estimates will be biased. Invoked in Section 3, step 4.
domain assumption The balanced neighborhood dataset is representative of the region of feature space relevant for counterfactual explanation.
Algorithm 2 assumes that uniformly sampled synthetic points, filtered by probability bands, cover the region around x and the boundary; the margins b1 and b2 are not specified.
domain assumption Woodward's account of counterfactual explanation is the correct normative standard for evaluating local explanations.
The paper asserts that satisfactory explanations must be counterfactual and must include a generalization; this philosophical premise justifies the fidelity metric, but is not empirically tested.
domain assumption Class probability outputs of the classifier are reliable and can be treated as ground truth for explanation quality.
The method labels synthetic data with the model's probabilities; the fidelity metric measures agreement with this oracle, not with any true causal structure.
domain assumption Euclidean distance is the appropriate measure of locality in the standardized feature space.
Both LIME and CLEAR use Euclidean distance for neighborhood selection and kernels.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Measurable Counterfactual Local Explanations for Any Classifier." pith.science (2026). https://pith.science/paper/AE6R2DE5

@misc{pith2026190803020,
  author       = {Pith},
  title        = {Pith review of: Measurable Counterfactual Local Explanations for Any Classifier},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AE6R2DE5}},
  note         = {Machine review of arXiv:1908.03020}
}

read the original abstract

We propose a novel method for explaining the predictions of any classifier. In our approach, local explanations are expected to explain both the outcome of a prediction and how that prediction would change if 'things had been different'. Furthermore, we argue that satisfactory explanations cannot be dissociated from a notion and measure of fidelity, as advocated in the early days of neural networks' knowledge extraction. We introduce a definition of fidelity to the underlying classifier for local explanation models which is based on distances to a target decision boundary. A system called CLEAR: Counterfactual Local Explanations via Regression, is introduced and evaluated. CLEAR generates w-counterfactual explanations that state minimum changes necessary to flip a prediction's classification. CLEAR then builds local regression models, using the w-counterfactuals to measure and improve the fidelity of its regressions. By contrast, the popular LIME method, which also uses regression to generate local explanations, neither measures its own fidelity nor generates counterfactuals. CLEAR's regressions are found to have significantly higher fidelity than LIME's, averaging over 45% higher in this paper's four case studies.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

An Explainable Gaussian Process Auto-encoder for Tabular Data
cs.LG 2025-08 conditional novelty 6.0 of 10

A Gaussian-process autoencoder with a latent-space density estimator generates counterfactual examples for tabular data, with competitive or better scores on several evaluation metrics.

Reference graph

Works this paper leans on

22 extracted references · 18 canonical work pages · cited by 1 Pith paper

[1]

Tickle, ‘Survey and critique of techniques for extracting rules from trained artiﬁ- cial neural networks’,Knowledge-Based Systems, 8(6), 373–389, (December 1995)

Robert Andrews, Joachim Diederich, and Alan B. Tickle, ‘Survey and critique of techniques for extracting rules from trained artiﬁ- cial neural networks’,Knowledge-Based Systems, 8(6), 373–389, (December 1995)

work page 1995
[2]

Elaine Angelino, Nicholas Larus-Stone, Daniel Alabi, Margo Seltzer, and Cynthia Rudin, ‘Learning certiﬁably optimal rule lists for categorical data’, Journal of Machine Learning Re- search, 18, 234:1–234:78, (2017)

work page 2017
[3]

CVPR 2017, Honolulu, USA, (2017)

David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and An- tonio Torralba, ‘Network dissection: Quantifying interpretability of deep visual representations’, In Proc. CVPR 2017, Honolulu, USA, (2017)

work page 2017
[4]

Craven and Jude W

Mark W. Craven and Jude W. Shavlik, ‘Extracting tree-structured representations of trained networks’, inProc. NIPS 1995, pp. 24–

work page 1995
[5]

Zoubin Ghahramani, ‘Probabilistic machine learning and artiﬁ- cial intelligence’,Nature, 521(7553), 452–459, (2015)

work page 2015
[6]

Riccardo Guidotti, Anna Monreale, Franco Turini, Dino Pe- dreschi, and Fosca Giannotti, ‘A survey of methods for explain- ing black box models’,CoRR, abs/1802.01933, (2018)

arXiv 2018
[7]

Machine learning interpretability with H2O driverless AI, 2017

Patrick Hall, Navdeep Gill, Megan Kurka, and Wen Phan. Machine learning interpretability with H2O driverless AI, 2017. URL: http://docs.h2o.ai/ driverless-ai/latest- stable/docs/booklets/MLIBooklet.pdf

work page 2017
[8]

Irina Higgins, Loic Matthey, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner, ‘beta-vae: Learning basic visual concepts with a constrained variational framework.’, ICLR, 2(5), 6, (2017)

work page 2017

Show all 22 references

[9]

Nair, and Agus Sudjianto, ‘Lo- cally interpretable models and effects based on supervised parti- tioning (LIME-SUP)’,CoRR, abs/1806.00663, (2018)

Linwei Hu, Jie Chen, Vijayan N. Nair, and Agus Sudjianto, ‘Lo- cally interpretable models and effects based on supervised parti- tioning (LIME-SUP)’,CoRR, abs/1806.00663, (2018)

2018 arXiv
[10]

Henrik Jacobsson, ‘Rule extraction from recurrent neural net- works: A taxonomy and review’, Neural Computation , 17(6), 1223–1263, (June 2005)

2005
[11]

4765–4774, (2017)

Scott M Lundberg and Su-In Lee, ‘A uniﬁed approach to inter- preting model predictions’, in Advances in Neural Information Processing Systems, pp. 4765–4774, (2017)

2017
[12]

Tim Miller, ‘Contrastive explanation: A structural-model ap- proach’,arXiv preprint arXiv:1811.03163, (2018)

2018 arXiv
[13]

Tim Miller, ‘Explanation in artiﬁcial intelligence: Insights from the social sciences’,Artiﬁcial Intelligence, (2018)

2018
[14]

537–543, (2017)

Saumitra Mishra, Bob L Sturm, and Simon Dixon, ‘Local inter- pretable model-agnostic explanations for music content analy- sis.’, inISMIR, pp. 537–543, (2017)

2017
[15]

Conference on Fairness, Accountability, and Transparency , FAT’19, pp

Brent Mittelstadt, Chris Russell, and Sandra Wachter, ‘Explain- ing explanations in AI’, in In Proc. Conference on Fairness, Accountability, and Transparency , FAT’19, pp. 279–288, New York, USA, (2019)

2019
[16]

Judea Pearl, Causality: Models, Reasoning and Inference, Cam- bridge University Press, New York, NY , USA, 2nd edn., 2009

2009
[17]

ACM SIGKDD 2016, KDD ’16, pp

Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin, ‘Why should i trust you? explaining the predictions of any classiﬁer’, in Proc. ACM SIGKDD 2016, KDD ’16, pp. 1135–1144, New York, NY , USA, (2016). ACM

2016
[18]

Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin, ‘An- chors: High-precision model-agnostic explanations’, in Thirty- Second AAAI Conference on Artiﬁcial Intelligence, (2018)

2018
[19]

Tran and Artur S

Son N. Tran and Artur S. d’Avila Garcez, ‘Deep logic networks: Inserting and extracting knowledge from deep belief networks’, IEEE Transactions on Neural Networks and Learning Systems , 29(2), 246–258, (2018)

2018
[20]

Mittelstadt, and Chris Russell, ‘Coun- terfactual explanations without opening the black box: Auto- mated decisions and the GDPR’,CoRR, abs/1711.00399, (2017)

Sandra Wachter, Brent D. Mittelstadt, and Chris Russell, ‘Coun- terfactual explanations without opening the black box: Auto- mated decisions and the GDPR’,CoRR, abs/1711.00399, (2017)

2017 arXiv
[21]

Ororbia, II, Xinyu Xing, Xue Liu, and C

Qinglong Wang, Kaixuan Zhang, Alexander G. Ororbia, II, Xinyu Xing, Xue Liu, and C. Lee Giles, ‘An empirical evalua- tion of rule extraction from recurrent neural networks’, Neural Computation, 30(9), 2568–2591, (September 2018)

2018
[22]

Woodward, Making things happen: a theory of causal expla- nation, Oxford University Press, Oxford, England, 2003

J. Woodward, Making things happen: a theory of causal expla- nation, Oxford University Press, Oxford, England, 2003

2003

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Tickle, ‘Survey and critique of techniques for extracting rules from trained artiﬁ- cial neural networks’,Knowledge-Based Systems, 8(6), 373–389, (December 1995)

Robert Andrews, Joachim Diederich, and Alan B. Tickle, ‘Survey and critique of techniques for extracting rules from trained artiﬁ- cial neural networks’,Knowledge-Based Systems, 8(6), 373–389, (December 1995)

work page 1995

[2] [2]

Elaine Angelino, Nicholas Larus-Stone, Daniel Alabi, Margo Seltzer, and Cynthia Rudin, ‘Learning certiﬁably optimal rule lists for categorical data’, Journal of Machine Learning Re- search, 18, 234:1–234:78, (2017)

work page 2017

[3] [3]

CVPR 2017, Honolulu, USA, (2017)

David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and An- tonio Torralba, ‘Network dissection: Quantifying interpretability of deep visual representations’, In Proc. CVPR 2017, Honolulu, USA, (2017)

work page 2017

[4] [4]

Craven and Jude W

Mark W. Craven and Jude W. Shavlik, ‘Extracting tree-structured representations of trained networks’, inProc. NIPS 1995, pp. 24–

work page 1995

[5] [5]

Zoubin Ghahramani, ‘Probabilistic machine learning and artiﬁ- cial intelligence’,Nature, 521(7553), 452–459, (2015)

work page 2015

[6] [6]

Riccardo Guidotti, Anna Monreale, Franco Turini, Dino Pe- dreschi, and Fosca Giannotti, ‘A survey of methods for explain- ing black box models’,CoRR, abs/1802.01933, (2018)

arXiv 2018

[7] [7]

Machine learning interpretability with H2O driverless AI, 2017

Patrick Hall, Navdeep Gill, Megan Kurka, and Wen Phan. Machine learning interpretability with H2O driverless AI, 2017. URL: http://docs.h2o.ai/ driverless-ai/latest- stable/docs/booklets/MLIBooklet.pdf

work page 2017

[8] [8]

Irina Higgins, Loic Matthey, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner, ‘beta-vae: Learning basic visual concepts with a constrained variational framework.’, ICLR, 2(5), 6, (2017)

work page 2017