GLIME: General, Stable and Local LIME Explanation

As black-box machine learning models grow in complexity and find applications in high-stakes scenarios, it is imperative to provide explanations for their predictions. Although Local Interpretable Model-agnostic Explanations (LIME) [22] is a widely adpoted method for understanding model behaviors, it is unstable with respect to random seeds [35,24,3] and exhibits low local fidelity (i.e., how well the explanation approximates the model's local behaviors) [21,16]. Our study shows that this instability problem stems from small sample weights, leading to the dominance of regularization and slow convergence. Additionally, LIME's sampling neighborhood is non-local and biased towards the reference, resulting in poor local fidelity and sensitivity to reference choice. To tackle these challenges, we introduce GLIME, an enhanced framework extending LIME and unifying several prior methods. Within the GLIME framework, we derive an equivalent formulation of LIME that achieves significantly faster convergence and improved stability. By employing a local and unbiased sampling distribution, GLIME generates explanations with higher local fidelity compared to LIME. GLIME explanations are independent of reference choice. Moreover, GLIME offers users the flexibility to choose a sampling distribution based on their specific scenarios.

Paper

Similar papers

Peer review

Reviewer jFSL7/10 · confidence 4/52023-06-26

Summary

This work is embedded in the research on model-agnostic explanations, i.e., to provide the user an understanding on the outputs of otherwise black-box prediction methods without knowing about the model's internals. While LIME is a popular approach to solve this problem, prior work has demonstrated LIME to suffer from a strong dependence on the random seed, leading to instability and inconsistency. Since reliable, model-agnostic explanations will be a crucial tool for research and application alike to afford the use of otherwise black-box machine learning models, this paper is tackling an important issue considering LIME's popularity yet evident short-comings. GLIME is presented as a step towards more general, stable and consistent model-explanations. Due to the free choice of its sampling distribution and weights, it is shown how GLIME not only improves on LIME but generalizes over other methods.

Strengths

1.The method GLIME is presented very clearly. Not only are text and equations supporting the reader in understanding the method well, but its motivation as successor of LIME in terms of stability and local fidelity are easy to follow and well justified by both the presented related work as well as this paper's own evaluation. 2. The unification of various model-explanation methods not only gives the reader an overview of how these methods relate to each other but shows well how GLIME is not only succeeding at stability and local fidelity but also a more general framework then LIME.

Weaknesses

1. This work is strongly focused on comparing GLIME and its variants to LIME. While the relation of LIME and GLIME are made clear and well supported by the experiments, a more comprehensive overview on the field of explanation methods other than LIME could help the reader to better understand how GLIME fits into current research. Similarly, a discussion of GLIME's short-comings and an outline of Future Work would reinforce the contribution. Along the same line, a discussion on GLIME's quality as model-explainer and human-interpretability of the achieved results would greatly support the claims. 2. While the figures present the concepts and results of this paper quite well, they could benefit from some additional attention and polishing. For example, Fig 1a misses an explanation of the employed colormap. Fig. 1b shows GLIME-Gauss as blue dot in the legend but not the graphic itself. In Fig. 4a, the legend occupies important parts of the plot such that GLIME-Binomial and GLIME-Gauss curves are hard to see. 3. The use of inline-math can at times be overwhelming, e.g., in Theorem 4.2. While it is important to state all the relevant equations and relations, reverting to display- rather than inline-math for the key concepts might help the reader to better digest the underlying theory and assumptions.

Questions

1. What are GLIME's short-comings and what are plans to improve on the method in the future? 2. Has it been evaluated if there is a difference in how human-interpretable GLIME's and LIME's explanations are?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed 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

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

Overall I would say it's a technically sound paper, but for a model-explanation paper I am missing the human-side a bit. My understanding is that the improvements the work shows are only meaningful if the consistent/stable explanations are also still good explanations, which I think is not really discussed here.

Reviewer sX7G7/10 · confidence 3/52023-07-01

Summary

This paper proposes a new explanation method, GLIME, which provides more general, stable and local LIME explanations over ML models. Specifically, the authors demonstrate that small sample weights cause the instability of LIME, which results in dominance of regularization slow convergence and worse local fidelity.To address those issues, the authors proposed GLIME framework, which takes a slightly different form of LIME. Through rigorous theoretical analysis and some experiments, the authors can demonstrate that GLIME addressed the above issues and outperformed LIME.

Strengths

+ The authors addressed a very important problem, i.e., the well-known instability issue of LIME, and proposed an effective solution to address it. + The authors conducted a rigorous theoretical analysis to support their claims, which is very convincing. + The overall presentation is very clear and easy to follow.

Weaknesses

+ The experiments are only conducted on one dataset, i.e., ImageNet dataset. It would be better if the authors could show more results on more benchmark datasets + Some properties that are studied in theory for GLIME are not empirically verified. For example, in Section 4.1, the authors showed that their method can converge faster than LIME. Although they provide clear proof for it, the authors did not demonstrate it in experiments. So it would be better if some empirical experiments can cover this.

Questions

+ How about the experimental results of using GLIME on other datasets? + How about the empirical comparison between GLIME and LIME in terms of the convergence speed?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

Not applicable.

Reviewer DXPw7/10 · confidence 4/52023-07-05

Summary

In this paper, the authors present GLIME an approach for explainable ai that generalizes LIME. Here, the authors present a framework that encompasses different explainability methods as instantiations of different aspects such as loss function, sampling function, etc. The authors also present an analysis of problematic cases for LIME. More precisely, they show how the interaction of the weighting and regularization can cause instability in the explanations and how the samples generated in LIME might not be close to the original sampling space. The paper then presents different sampling procedures and show empirically how they converge and how stable the explanations are given different parameterizations.

Strengths

I find the paper insightful, in particular the aspect of the weights becoming all zeros in the standard case for low values of sigma. The paper is easy to read, technically sound and guides the reader through the concepts in a solid yet understandable way. The technical contributions are solid and overall provides a good foundation for further research. Overall the paper is original and I would say significant as it has the potential to become the standard replacement for LIME.

Weaknesses

The main concern I have is regarding the empirical section. In particular, you mention two main issues with LIME being the interaction of the weighting with regularization and sub-par sampling. However, it would seem like ALIME does not suffer from those two issues. It would be good to see a comparison of GLIME and ALIME in Fig. 4. There are some minor improvements I would suggest on the presentation. I would suggest you unify the color scheme in Fig. 4 and if possible present as many of the methods in both graphs. Is there a typo in the norm of the weighting function in line 171? shouldn't it be 2 and 2? The language on the sub-section in Feature attributions could be improved.

Questions

My main concern is regarding the contribution and the comparison to ALIME. Could you mention a bit more in terms of your contributions regarding that method? This does not diminish your theoretical contributions in terms of the convergence in comparison to LIME, however it can potentially make GLIME less appealing to the general user than ALIME. Regarding your GLIME-gauss, and I believe ALIME is similar, wouldn't the sampling space be too close to the original image we want to explain? It seems like simply a noise addition similar to difussion models.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed 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

3 good

Limitations

I don't consider the paper to have potential negative societal impact.

Reviewer NTaD7/10 · confidence 4/52023-07-06

Summary

The paper introduces GLIME as a solution to tackle the issues of instability and diminished local fidelity encountered in the original LIME method. To address the problem of instability, GLIME employs a novel sampling scheme that guaranteed to have a faster sampling rate. The diminished local fidelity problem is resolved by modifying sampling distribution so that nearby samples have higher probability to be sampled. Disclaimer: I only read the main text and do not check correctness of the proposed sample complexity argument.

Strengths

1. The problem they tackle with is specific and well-formulated. The proposed solution is simple and effective. 2. Their methods are supported by sample complexity analysis. This analysis not only provides mathematical evidence that the original LIME approach necessitates an exponentially large number of samples to achieve convergence, but also demonstrates that their proposed method requires only a polynomial number of samples, offering a significant improvement in efficiency.

Weaknesses

One weakness would be limited applicability of the proposed GLIME. The paper only demonstrates it can only be applied to the image domain. As other features from different domains, such as texts or categorical features, are not necessarily to be continuous, GLIME equipped with continuous distributions may not resolve the local fidelity issue.

Questions

In Section 4.3 of the paper, it is stated that the local fidelity problem arises due to the utilization of a high regularization weight. Could this issue be addressed by reducing the regularization weight or, in more extreme cases, completely eliminating the regularization.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed 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

3 good

Limitations

The potential societal impact is not stated in the main paper.

Reviewer sX7G2023-08-12

Thanks for the authors' efforts

I think the authors addressed my concerns. I also read other reviewers' comments and I think overall the responses look good to me. So I raised my score. But I would love to communicate with other reviewers if any other significant issues are raised.

Reviewer NTaD2023-08-12

Thanks for the author's reply

I believe the authors have taken into account the points I raised, so I raised my score to 7. Additionally, I've gone through feedback from other reviewers, and on the whole, I find the authors' responses satisfactory.

Reviewer DXPw2023-08-14

I'm satisfied with the rebuttal provided by the authors and the discussion. Therefore I'm raising my score.

Reviewer jFSL2023-08-14

Thank you for the response

I would like to thank the authors for their response. Although I still feel that this paper feels incomplete without the human angle, I will not fight for rejection due to this. Overall, other concerns have been satisfactorily addressed and thus I will raise my score to 5.

Authorsrebuttal2023-08-21

Human-interpretability experiment results

Thank you for your feedback and suggestions. We have incorporated human-interpretability experiments as you recommended. Our experimental design is as follows: * Our experiment consists of two parts: 1. Selecting images where the model's predictions are correct and presenting the original images along with explanations from LIME and GLIME to the participants. The participants are asked to rate the degree of matching between the explanations provided by the algorithms and their intuitive understanding on a scale of 1-5, where 1 indicates a significant mismatch and 5 indicates a strong match. 2. Selecting images where the model's predictions are incorrect and presenting the original images along with explanations from LIME and GLIME to the participants. The participants are asked to rate the degree of help provided by the explanations in understanding the model's behavior and identifying the reasons for the incorrect predictions on a scale of 1-5, where 1 indicates no help at all and 5 indicates significant help. * Each part consists of ten randomly selected images. * The participants in the experiment are college students from diverse backgrounds with no prior knowledge of machine learning. There are ten participants for each part, and the participants for the two parts are different. Here are the results of the experiment: * When participants were shown images where the model's predictions were correct, along with explanations from LIME and GLIME, they gave an average score of 2.96 to LIME and an average score of 3.37 to GLIME. Overall, GLIME had an average score 0.41 higher than LIME. For seven out of the ten images, GLIME had a higher average score than LIME. We performed a t-test on the scores of LIME and GLIME, resulting in a t-value of -1.36 and a p-value of 0.29. * When participants were shown images where the model's predictions were incorrect, along with explanations from LIME and GLIME, they gave an average score of 2.33 to LIME and an average score of 3.42 to GLIME. GLIME had an average score 1.09 higher than LIME for all ten images. We performed a t-test on the scores of LIME and GLIME, resulting in a t-value of -8.75 and a p-value of $1.08\times 10^{-5}$. The results of the second part of the experiment indicate that GLIME is more effective in helping people understand the model's behavior and debug the model. Although the results of the first part of the experiment are not statistically significant, they still reflect the relative advantage of GLIME over LIME to some extent. We will conduct more experiments to obtain statistically significant results and include results in the final version. We hope that our current experimental results address your concerns.

Reviewer jFSL2023-08-21

Thank you for all your hardwork

My concerns stand resolved and I am raising my rating to 7. I would like to thank the authors for their hardwork and good luck with the paper.

Authorsrebuttal2023-08-21

Thank you!

Thank you for your prompt response. We are glad that we were able to address your concerns. We appreciate the time and effort you have put into reviewing our paper. Wishing you a pleasant day!

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC