Trade-off Between Efficiency and Consistency for Removal-based Explanations

In the current landscape of explanation methodologies, most predominant approaches, such as SHAP and LIME, employ removal-based techniques to evaluate the impact of individual features by simulating various scenarios with specific features omitted. Nonetheless, these methods primarily emphasize efficiency in the original context, often resulting in general inconsistencies. In this paper, we demonstrate that such inconsistency is an inherent aspect of these approaches by establishing the Impossible Trinity Theorem, which posits that interpretability, efficiency, and consistency cannot hold simultaneously. Recognizing that the attainment of an ideal explanation remains elusive, we propose the utilization of interpretation error as a metric to gauge inefficiencies and inconsistencies. To this end, we present two novel algorithms founded on the standard polynomial basis, aimed at minimizing interpretation error. Our empirical findings indicate that the proposed methods achieve a substantial reduction in interpretation error, up to 31.8 times lower when compared to alternative techniques. Code is available at https://github.com/trusty-ai/efficient-consistent-explanations.

Paper

Similar papers

Peer review

Reviewer JLEA5/10 · confidence 2/52023-07-02

Summary

This paper studies the consistency and efficiency of removal-based explanation methods. By establishing the impossible trinity theorem, the authors show that consistency and efficiency cannot be achieved simutaneously. They further propose the interpretation error as a metric to gauge the trade-off between consistency and efficiency. Then, two algorithms based on standard polynomial basis are developed to minimize the interpretation error for more consistent explanations, when relaxing the efficiency to truthfulness.

Strengths

1. Consistency and efficiency are important properties for evaluating explanation methods. 2. The finding of trade-off between efficiency and consistency is interesting and leads to some insights.

Weaknesses

1. My major concern is that this paper is quite hard to follow and the presentation needs significant improvement. The authors present a lot of definitions and claims but not sufficient explanations and motivations for them. 2. The definition for consistency and efficiency is not rigorous and seems to be different from that in previous literature. The authors should justify the correctness of definitions and give more explanations here.

Questions

1. At line 136, if the consistency should be conditioned on a local input region which shares the same ground truth label? 2. At line 139, how to match the definition of efficiency with that at line 194? 3. How to approximately calculate the equation 2 and 3 in experiments? 4. Apart from the proposed metrics, truthful gap and interpretation error, are there any existing metrics in current literature for consistency and efficiency? The existing ones should be compared with the proposed ones.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

Please see the weakness and questions.

Reviewer ngxu5/10 · confidence 5/52023-07-02

Summary

This paper addresses the problem of tradeoffs between interpretability, consistency, and efficiency of AI models. The authors present “the Impossible Trinity Theorem”, which posits that interpretability, efficiency and consistency cannot hold simultaneously. To quantify the tradeoff, the authors define the utilization of interpretation error as a metric to gauge inconsistencies and inefficiencies of AI models in order to minimize an interpretation error. With such definitions of terms and metrics, the authors achieve a reduction in interpretation error, up to 31.8 times lower when compared to alternative techniques.

Strengths

The strength of the paper lies in the effort to formulate a mathematical framework including efficiency, consistency, and interpretability of AI models. The framework separates linear and non-linear parts of input-output functions represented by trained AI models by stating “only linear functions are readable” and therefore interpretable.

Weaknesses

The weaknesses of the paper lie in (1) forcing reviewers to read Appendices because the main text does not contain sufficient results supporting the claims. This is unfair to other submissions. (2) obfuscating many simple concepts with too many theorems and definitions to formulate relationships. (3) missing discussion why 2nd order Harmonica results are worse than Shapley in Table 1 which implies that the 3rd order Harmonica is needed. BTW, Standard deviation would help to understand statistical significance of the values when comparing the results.

Questions

In reference to “Note that all of these methods are applied to the same image segmentation using the SLIC superpixels [61]. “ it is not clear why the simple linear iterative clustering algorithm applied to 224 x 224 images split into 16 superpixels is used as ground truth for ImageNet rather ground truth segmentation obtained by manual segmentation.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

Overall, the framework has introduced too much complexity for a problem that can be described very clearly. For example, Theorem 1(Impossible Trinity for Interpretation) could just state that the input-output relationship in a trained AI model is either accurate and non-linear or inaccurate and linear. The focus should be rather on Fourier spectrum and the order of Harmonica algorithm (i.e., LASSO regression over the coefficients on the polynomial basis) needed to meet a given epsilon error.

Reviewer PY9W6/10 · confidence 4/52023-07-04

Summary

This paper explores the interplay between two properties of removal based explanations: efficiency and consistency. It established that these two properties cannot coexist when the set of predictive models (requiring explanation), and the set of interpreting models (doing the explaining) are disjoint. It proposes a new property: “truthfulness”, which is a relaxation of efficiency, and measures the information-capture within a functional sub-space (a chosen subset of the interpretable models). It proposes several algorithms for learning truthful interpreting models, grounded in methods from Boolean functional analysis, which allow for different trade-offs between consistency and efficiency. It compares these algorithms to existing methods by looking at the interpretation error on a few practical NLP and Vision tasks.

Strengths

The authors demonstrate a solid understanding of existing removal-based (and other) explainability methods. The paper surveys a large number of these related works. The paper explains a tension in post-hoc explanations: efficiency vs. consistency in a concise and easy to understand way. The front half of the paper (sections 1 to 4.1), where most of the concepts are presented, is clearly written and easy to understand. The paper draws insightful connections between producing interpreting models and work in Boolean functional analysis. Between the theoretical discussion, the methods introduced, and the subsequent experiments, the paper makes numerous useful contributions.

Weaknesses

The paper defines consistency for explanations (Definition 3.2). It is unclear whether the authors are proposing this particular definition, or if it has been previously defined in this way. This should be clarified. Furthermore, the term consistency has been used differently in the explainability and ML literature previously, albeit more casually. It would be helpful to contrast this definition to other ways the term has been used, and to relate/contrast it to other previously defined properties of explanations, such as fidelity. The motivation for consistency could be improved. The extent to which consistency is desirable, and the size of the neighborhood in which it would be desired, both seem to be highly dependent on how the explanation is going to be used. As it stands, there is a single (simple) motivating example in the main text. Since the primary reason one would use the methods proposed in the paper is to improve consistency, it seems reasonable to further motivate this property. I question the choice of the words “truthful” and “truthfulness” for Definition 3.4 .When T_v = 0, g is a sort of best approximation to f within a space of interpretable models. The property is interesting, but it does not intuitively align with “truthfulness” for me. With the number of overloaded terms that already exist in the explainability space (many of which invoke their own “layman” concepts), I caution against introducing another, especially a term as loaded as “truth”. I strongly encourage the use of a more verbose and precise term. It seems unfair to call Theorem 1 the “Impossible Trinity for Interpretability” when the set of interpretable models is separated from the set of predictive models in order to get the result. Interpretable models are a very important class of predictive models. If the authors wish to name the result, I would suggest something along the lines of “Impossible Trinity for Post-Hoc Explanation”. There is something of a conceptual gap in the flow of the paper. Section 3, and all definitions consider an arbitrary input space X. While Section 4 onwards deals with Boolean functions. What is missing is any real discussion of the process of binarizing real-valued (or other) inputs and how that affects or relates to consistency. From what I understand, ensuring the consistency of an explanation in a neighborhood of removed features (the binarized input space) is different from ensuring the consistency of that explanation in some ball around the original real-valued input. This is simply because moving to another region of the original input space (for example by changing a word/token in an input sentence or by perturbing the input image) will require a different binarization, because what it means to include/exclude one (or more) of the features is different. This seems like it should be better explored and discussed. The paper discusses “amounts” of inconsistency, but only defines what it means to be consistent, not how to measure deviation from consistency. This seems like a considerable omission, especially since Algorithm 3 aims to reduce inconsistency. The second half of the paper (section 4.2 onwards) struggles with clarity. The paper should include concrete examples of output explanations. Especially for Harmonica-anchor and the constrained variant. As it stands, I do not understand how one is meant to take a set of coefficients for some polynomial bases and produce a heatmap over an image or textual input. I can guess, but this should be made clear. The algorithm boxes include some notation that is not explained elsewhere in the main text. For instance, C is introduced in the appendix. Algorithm 1 should define input and outputs. For clarity I would also consider using different symbols when referring to Boolean models vs. models having arbitrary inputs, instead of f(x) for both. Finally, the paper relies heavily on the Appendix, and in that sense the main text does not truly seem self-contained. For instance, in the opening paragraph of the introduction the authors ask the reader to refer to a figure in the Appendix (Figure 3). I suggest shuffling some content around. For instance, the definition of the Shapley value (which is well-known in the field) and even Figure 1 (which isn’t really discussed), would both seem like more appropriate Appendix content than Figure 3, which helps to motivate consistency. The current introduction could also be compressed somewhat. Some typos: Line 72: “the algorithm only retain” Line 131 “We call A model g is interpretable”

Questions

The paragraph starting on line 218 quantifies inconsistency, but does not explicitly explain how. What is an “amount” of inconsistency? It seems like what is being reported is the number of different explanation functions required in the neighbourhood. Is Figure 2 correctly labeled? Should the subspace radius for Imagenet not be 1-16 and 1-\inf be for IMDB the Should the C in Harmonica-anchor be indexed by i? In Table 1 and Section 6.4, it is not clear to me why we would expect (or even want) any of the other methods (other than Harm. 3rd) to have a small “truthful gap” in C^3. The other methods do not aim to explain with 3rd order polynomials. Why should we assess them in this way? It would seem to me that Harmonica’s “truthful gap” should be a function of T and \lambda, is this explored somewhere?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

The paper would benefit from an explicit discussion of its limitations.

Reviewer XVSh6/10 · confidence 4/52023-07-11

Summary

The paper analyses the trade-off between efficiency and consistency in the context of removal-based post-hoc explanations. The authors demonstrate that having an interpretability method that is both efficient and consistent is impossible. Further, on the problem of finding consistent explanations, the authors propose to relax the efficiency desiderata and instead propose a new desiderata called truthfulness (efficiency within a subspace). The authors argue that this maps to the problem of learning Boolean functions that can be tackled by using Fourier analysis.

Strengths

1. The paper is well-written. The motivation, definitions and key contributions are clearly explained. 2. The core idea of leveraging Fourier analysis in the context of removal-based explanations is interesting. 3. The experiments and bibliography sections are extensive.

Weaknesses

1. I do not understand why is consistency of explanations a more important desiderata than efficiency. Instead, one may argue that efficiency is critical since it acts as a proxy to the problem of lack of ground-truth. Specifically, one of the main advantages of using SHAP is that supports some mathematically properties which come handy in the absence of ground-truth. Since, due to impossibility, we can't achieve both, I would be keen to hear when is consistency more important than efficiency. Further, here we talk about consistency in Boolean sense which I guess is different in a non-Boolean sense (similar inputs should have similar explanations) as there this desiderata is referred to as robustness. 2. Using Fourier analysis for Boolean functions is a well-explored topic. Hence, although interesting, I see limited novelty in applying this to post-hoc removal-based methods. 3. Some experimental details seem to be missing. For example, all removal based methods use a baseline and the explanations are quite sensitive to the choice of the baseline [1]. I wonder what is baseline for SHAP/IG/LIME etc. and if the authors analysed the effect of choosing a different baseline. [1] https://distill.pub/2020/attribution-baselines/

Questions

1. In experiments mentioned in section 6.1, how did the authors choose the feature values to compute ground-truth? 2. What version of the SHAP algorithm was used? Further, as LIME is very sensitive to the chosen hyperparameters[2], I wonder what are the hyperparameters for the LIME algorithm. [2] https://arxiv.org/pdf/2005.07788.pdf

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors do not mention any limitations of their work.

Reviewer k1pA6/10 · confidence 4/52023-07-12

Summary

This paper proposes using interpretation error as a metric to gauge inconsistencies and inefficiencies. The authors proposed three algorithms founded on the standard polynomial basis to minimize the interpretation errors. Authors perform experiments using both image and text classification to demonstrate the superior performance in terms of lowest interpretation error to the baseline methods.

Strengths

A major issue for the removal-based explanation is that the change of input distribution with the removal of inputs. This makes the performance evaluation problematic. The trade-off between efficiency and consistency for removal-based explanations is well known. The novelty of this paper is to propose using a single metric, i.e., interpretation error, for evaluating the quality of model interpretation, followed by two algorithms for finding the best interpretation models assuming the interpretable model is polynomial. Authors also conducted comprehensive experiments to evaluate both interpretation error and truthful gap of Harmonica and other baseline algorithms.

Weaknesses

To establish your approach is superior, comparison in terms of prediction accuracy and time complexity across multiple DNN architectures would be helpful. Indeed, the trade-off between prediction performance and interpretability can also be important. It would be nice to see your approach achieves lower interpretation errors without compromising prediction performance.

Questions

Besides efficiency and consistency, there are other factors that would influence the evaluation of model interpretability. Can you also show the prediction accuracy together with interpretability error for your experiments? What about computational complexity?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

3 good

Contribution

4 excellent

Limitations

Not really a limitation but rather an assumption: using higher order polynomial to approximate the nonlinearity of DNN.

Reviewer JLEA2023-08-12

Response to Rebuttal

I thank authors for the rebuttal. I have carefully checked the response and additional experiment. Since the authors promise to improve the presentation quality and show the comparable performance of removal-based algorithms on existing metrics, I am willing to change my score.

Reviewer ngxu2023-08-14

I read the authors' replies.

I do not have any other comments.

Reviewer PY9W2023-08-15

I have read the authors response, the attached pdf, and the other reviews and associated rebuttals. I am glad that my feedback was appreciated and that the authors intend to incorporate many of the suggestions.

Reviewer XVSh2023-08-16

I thank the authors for providing clarifications and sharing more results. I suggest the authors to clarify in the paper how consistency relates with robustness in the context of feature attributions [1][2][3]. In the light of the new results, I am happy to raise my score. [1] https://arxiv.org/pdf/2111.00358.pdf [2] https://arxiv.org/abs/1711.00867 [3] https://arxiv.org/pdf/1906.07983.pdf

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC