Distributionally Robust Skeleton Learning of Discrete Bayesian Networks

We consider the problem of learning the exact skeleton of general discrete Bayesian networks from potentially corrupted data. Building on distributionally robust optimization and a regression approach, we propose to optimize the most adverse risk over a family of distributions within bounded Wasserstein distance or KL divergence to the empirical distribution. The worst-case risk accounts for the effect of outliers. The proposed approach applies for general categorical random variables without assuming faithfulness, an ordinal relationship or a specific form of conditional distribution. We present efficient algorithms and show the proposed methods are closely related to the standard regularized regression approach. Under mild assumptions, we derive non-asymptotic guarantees for successful structure learning with logarithmic sample complexities for bounded-degree graphs. Numerical study on synthetic and real datasets validates the effectiveness of our method. Code is available at https://github.com/DanielLeee/drslbn.

Paper

Similar papers

Peer review

Reviewer Np2S8/10 · confidence 3/52023-06-25

Summary

The authors give novel algorithms to learn the structure of a Bayes net from noisy data.

Strengths

Originality: Work seems original. Quality: Quality is high, writing-wise and content-wise; see questions. Clarity: Writing is clear; see questions. Significance: Topic is significant.

Weaknesses

No significant issues detected; see questions.

Questions

Line 21: What do you mean by "ordinal?" Line 69 to Line 70: Unclear. Line 127: Can you please elaborate on faithfulness? Line 135: Please define "unweighted effects." Can you please explain Equation (2)? Can you please justify Assumption 2? Can you please explain Equation (5)? Proof of Theorem 9: I do not understand the union bound argument. Line 304: Why is it always possible to find such a solution? Line 313 to Line 314: Unclear. Figure 1: Can you please add some short *further* explanation in the caption? Experiments look OK!

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and 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

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

Yes.

Reviewer onGN5/10 · confidence 1/52023-06-30

Summary

The authors propose methods for learning Bayesian network skeletons based on discrete data when the data is corrupted by arbitrary noise. The authors show guarantees on discovering the correct skeleton in terms of the required sample size. Some empirical experiments are also presented.

Strengths

+ rigorous mathematical derivations

Weaknesses

- very dense mathematical derivations - limited and inconclusive empirical results

Questions

To be honest, the level of mathematical detail in the paper is too much for me. It's not that I wouldn't be able to refresh my measure theory and calculus but I just can't afford to spend a week on a review. This is of course primarily my problem, but I have reviewed and published at NeurIPS before, and based on my experience, this paper is mathematically too dense for at least 90% of the NeurIPS community. But perhaps this is fine in case the 10% find it relevant and useful -- I just couldn't possibly tell. In Sec. 4, explain what exactly you mean my the F1 score: the F1 score is used for binary classification, so I'm guessing you're perhaps looking at presence/absence of each of the edges in the DAG. Please clarify further which methods are proposed in this paper (Wass + KL?) and which are benchmarks (Reg, MMPC, GRASP?) so that it's easier to quickly grasp. Am I right in that the punchline is Figs 1.b and 1.c? If so, please make it easier for the reader to see this or otherwise they'll miss it. I take it the results indicate that the proposed methods are otherwise about as good as prior work except that they're better when the failure probability is close to 0.5 (Figs. 1.b and 1.c)? If so, doesn't this require an extremely high rate of failures, to the extent that the data is mostly random noise? detailed comments: - p. 1 "DAG structure of a Bayesian network is usually unknown": citation needed (or maybe just say "often") - p. 2 "contains as subgraphs the skeleton": I'm confused by the plural/singular tenses here: if there's only one skeleton, how is it contained as subgraphs - p. 2: "(fixed-parameter) tractable": what does this mean? - p. 2: "observational data is commonly contaminated [...]": again, citation needed - p. 2: "may produce false edges": may it also miss edges? - p. 2: "do not assume any specific form of conditional distributions" vs "set of distributions characterized by certain a priori properties": aren't these statements contradictory? please elaborate - p. 2: "small amount of samples (high-dimensional) and potential perturbations": the word 'high-dimensional' is grammatically speaking placed in an odd spot; also, does "small amount " refer only to samples or also to perturbations -- please clarify the sentence

Rating

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

Confidence

1: Your assessment is an educated guess. The submission is not in your area or the submission was difficult to understand. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

ok

Reviewer YEGV4/10 · confidence 3/52023-07-06

Summary

In this paper, the authors revisit the problem of structure learning of Bayesian networks. Here, we are interested to learn the underlying DAG structure that encodes the conditional independence properties of a Bayes net from random samples. A number of algorithms have been proposed to solve this problem over the years which can be broadly classified into two groups: score-based and constraint-based. Despite a lot of research activity over the last several decades the general problem is not yet known to be solvable in polynomial time and several approaches are shown to be NP-hard. The authors of this paper start with an approach first proposed by Bank and Honorio in 2020: to reduce this discrete structure learning problem to a continuous regression problem based on encoding schemes and surrogate parameters. The idea is that once cast as a regression problem, techniques from structure learning in continuous models such as regression can be employed. The main contribution of this paper is a distributionally robust optimization (DRO) method that builds upon the aforesaid framework to handle noisy data. In this proposed method, the best structure is searched among a ball of certain radius around the empirical distribution that minimizes the regression cost. Specifically, balls according to the Wasserstein distance and KL divergence are considered. The authors give two theorems for each of the above two distance functions under certain assumptions. They further verify the effectiveness of their methods by experimenting with benchmark real and synthetic datasets.

Strengths

The problem considered in extremely important from a theoretical and practical standpoint. The approach followed by the authors to reduce it to a regression problem as in the previous work could be promising for certain datasets.

Weaknesses

A major weakness of the paper is that a number of assumptions are made in order to derive the results, which looks ad-hoc and too restrictive in my opinion. See page 4 of the paper for these assumptions. See also line number 271. I could not help but notice several vague statements throughout the paper, a few of which are listed below. Line number 224-233 the NP-hardness argument is very vague. Looks like you are using a brute-force approach, if so, the algorithms wont scale beyond 15 nodes. Line number 240-241: the citation of Chen and Paschalidis [2018] here is for the first time and sudden. What question are they solving? How is it related to your problem? Unfortunately, I am not convinced from the experimental results that the proposed algorithms are overall performing better than the existing methods. Many of the entries in table 1 show that other methods are performing superior to the proposed algorithms. Finally, I found the paper very hard to follow, maybe it reads more complicated than it actually is. The author may try to see whether writing it in a simpler manner with small examples makes the paper more readable or not.

Questions

None

Rating

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

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

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

None

Reviewer AT8n7/10 · confidence 2/52023-07-06

Summary

The paper presents a method for learning the structure of discrete Bayesian networks from potentially corrupted data. It utilizes distributionally robust optimization and regression, optimizing the worst-case risk over a set of distributions. The approach works with general categorical variables without assuming specific conditions and provides efficient algorithms. Non-asymptotic guarantees for successful structure learning are derived, and numerical experiments validate its effectiveness.

Strengths

Originality: The paper addresses the challenge of missing data in distributionally robust skeleton learning. This aspect adds a unique and valuable contribution to the field. Quality: The quality of writing in the paper is commendable. The ideas and concepts are presented effectively, and the paper demonstrates a strong grasp of the subject matter. Clarity: The paper's clarity is notable, as it is organized in a logical manner where each section builds upon the previous ones. The flow of information is well-structured, making it easy for readers to follow the paper's content. Significance: The task of building discrete Bayesian networks from potentially corrupted data holds significant importance, as it has wide-ranging applications in various fields. By addressing the challenge of distributionally robust skeleton learning, the paper contributes to the advancement of techniques that can handle data uncertainties and enable more accurate modeling and inference in real-world scenarios. The significance of this research lies in its potential to enhance decision-making processes, predictive modeling, and knowledge discovery across diverse domains.

Weaknesses

- The author's assertion that they had limited space to relate their work to the existing literature is understandable. However, it is crucial to provide a clear positioning of their own work within the broader research landscape. Emphasizing the novelty and distinguishing features of their approach will strengthen the paper's contribution and demonstrate its unique value. - The paper lacks clarity on whether cross-validation or hyperparameter tuning was performed during the experiments. It would be beneficial to address this issue explicitly to ensure the robustness of the results. Additionally, to further improve the reliability of the findings, it is recommended to conduct a rigorous statistical test, such as the corrected Student's t-test proposed by Nadeau and Bengio [1], using 15 hold-out folds. Incorporating such statistical analyses will enhance the credibility and validity of the experimental results. [1] https://link.springer.com/article/10.1023/A:1024068626366

Questions

l 105: How is T defined in A_{ST}?

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

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

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

- The proposed method accounts for the effect of outliers by optimizing the worst-case risk over a family of distributions within bounded Wasserstein distance or KL divergence. However, the extent to which the method can effectively handle outliers and their impact on the learned network structure should be further explored and evaluated. - The proposed approach is designed for general categorical random variables without assuming faithfulness, an ordinal relationship, or a specific form of conditional distribution. While this makes the approach flexible, it may also limit its applicability to certain domains or types of data. Investigating the generalizability of the approach to a broader range of variable types or distributions could be a potential limitation.

Reviewer Hyho8/10 · confidence 4/52023-07-31

Summary

The paper considers the problem of learning the exact skeleton of general discrete Bayesian networks from potentially corrupted data. The estimator optimizes the worst case risk over a family of distributions within bounded Wasserstein distance and KL divergence to the empirical distribution. Under mild conditions, the paper presents non-asymptotic guarantees for successful structure learning with logarithmic sample complexities for bounded degree graphs. Numerical studies on synthetic and real datasets validates the performance of the estimator.

Questions

1) Is it possible to extend the DRO framework to learn the structure of continuous Bayesian networks from corrupted samples. A high level discussion regarding this would shed more light on the limitations of the present contributions. 2) Since no assumptions on the underlying true distribution is made, it is only possible to recover the structure upto a Markov equivalence class. However the proposed approach overcomes this limitation and recovers the true skeleton with high probability. A discussion as to why DRO framework achieves this will be highly appreciated.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and 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

3 good

Presentation

4 excellent

Contribution

3 good

Reviewer AT8n2023-08-14

Thank you for taking the time to read my comments. The authors have taken all my suggestions into account and provided valuable feedback on my questions. I think that by clarifying the limitations of the method and improving the related work section, the paper will be of more value to the community. I increase my vote from 6 to 7.

Reviewer Hyho2023-08-14

The authors have addressed my concerns and I am quite happy with the response.

Reviewer YEGV2023-08-18

Response

I have read the rebuttal by the authors. I am still not convinced about the contribution of the paper and am leaning towards rejection.

Authorsrebuttal2023-08-20

More specific

Dear Reviewer YEGV, Thanks for responding to our rebuttal. Could you please be more specific about which answer to your questions you are not convinced of? If you are not convinced about any of them, just let us know and we will reply with a more detailed rebuttal that responds to all your concerns. Wish you all the best in NeurIPS 2023 both as a reviewer and author. Authors

Reviewer YEGV2023-08-20

response

My main concern includes the assumptions of the paper which I believe is an inherent component of the paper.

Reviewer onGN2023-08-21

Changing my score borderline reject => borderline accept

I thank the authors for detailed rebuttals to my own review and those of the other reviewers. I believe that the numerous clarification that the authors promise to make in the revised version will improve the paper enough to bump my score up to borderline accept. The reason why I'm not increasing my score higher is that I'm still not entirely happy with a) the modest empirical results compared to prior methods, and b) the dense mathematical presentation that is making the paper unnecessarily(?) hard to follow, but perhaps this paper will inspire follow up work that can further improve empirical performance and simplify the assumptions and the theoretical results.

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC