Invariant Learning via Probability of Sufficient and Necessary Causes

Out-of-distribution (OOD) generalization is indispensable for learning models in the wild, where testing distribution typically unknown and different from the training. Recent methods derived from causality have shown great potential in achieving OOD generalization. However, existing methods mainly focus on the invariance property of causes, while largely overlooking the property of \textit{sufficiency} and \textit{necessity} conditions. Namely, a necessary but insufficient cause (feature) is invariant to distribution shift, yet it may not have required accuracy. By contrast, a sufficient yet unnecessary cause (feature) tends to fit specific data well but may have a risk of adapting to a new domain. To capture the information of sufficient and necessary causes, we employ a classical concept, the probability of sufficiency and necessary causes (PNS), which indicates the probability of whether one is the necessary and sufficient cause. To associate PNS with OOD generalization, we propose PNS risk and formulate an algorithm to learn representation with a high PNS value. We theoretically analyze and prove the generalizability of the PNS risk. Experiments on both synthetic and real-world benchmarks demonstrate the effectiveness of the proposed method. The details of the implementation can be found at the GitHub repository: https://github.com/ymy4323460/CaSN.

Paper

Similar papers

Peer review

Reviewer CQCE7/10 · confidence 1/52023-07-01

Summary

This paper proposes the new concept of 'essential causal information' to achieve OOD generalization and proposes the probability of necessarity and sufficiency (PNS) risk to estimate the representation if satisfies the new concept. Then, it proposes a new algorithm to achieve loss PNS risk on the test domain.

Strengths

(1) Introducing the PNS risk for out-of-distribution generalization is novel and interesting. The proof for it is rigorous. (2) The method is general and works well on various datasets. (3) The writing of this paper is good.

Weaknesses

(1) The fox also has long whiskers. The example in section 2.3 does not hold on.

Questions

Why you don't consider recent methods (2021~2023) for Domain Adaptation? I am not an expert in this domain but I think there have many strong and new baselines for out-of-distribution generalization.

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

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

3 good

Contribution

3 good

Limitations

I don't see any potential negative social impact of their work.

Reviewer pPMP7/10 · confidence 4/52023-07-04

Summary

This paper presents a novel approach to enhance out-of-distribution (OOD) generalization by identifying both sufficient and necessary conditions, thereby extending the existing literature that primarily focuses on causal invariance. The proposed method employs the probability of sufficient and necessary causes (PNS) to demonstrate its effectiveness in improving OOD generalization, through theoretical analysis and experimental results,.

Strengths

1. The proposed method is very intuitive and novel, as far as I know. It focuses on sufficient and necessary causes to explore the invariance. This method successfully combines existing ideas in invariant causality with PNS, and has good experimental results. 2. The proposed delta-PNS is reasonable to me. If the semantics have overlap, it is impossible to make good prediction. 3. The experiment part is extensive which covers a variety of common datasets, including synthetic data, an ablation study, and datasets for both OOD generalization and domain adaptation tasks.

Weaknesses

1. It is difficult to understand the definition of PNS in Definition 2.1, Could you provide more details? 2. Typo: missing ) in line 222. 3. Do you have discussed more general causal structure? ------- After reading the author's response, my concerns have been well addressed, so I will keep my score as accept.

Questions

Please answer the questions in Weakness.

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

3 good

Presentation

3 good

Contribution

3 good

Limitations

N/A

Reviewer qVZ25/10 · confidence 3/52023-07-05

Summary

The paper proposes a new formulation of the out-of-distribution generalization task via finding necessary and sufficient features, similar to invariant feature learning but with some added constraints. The paper uses the definition of necessary and sufficient from Pearl (2008) and proposes a related risk measure. Then, a connection between source and test risks is given along with assumptions needed for the bounds to hold. Finally, the paper proposes a practical adversarial algorithm to minimize the proposed objective and presents empirical results on DomainBed and a toy simulated dataset.

Strengths

- The paper proposes an interesting new direction to find necessary and sufficient features and proposes a risk based on this objective. - The paper theoretically connects the target risk and source risk based on the theoretic objective. - The paper provides reasonable results on PACS and VLCS and a toy simulated dataset to validate the theory.

Weaknesses

- PNS needs to be explained and discussed more. This is a very difficult concept to grasp theoretically. Also, Definition 2.1 seems incomplete. What are $c$ and $\bar{c}$? How can just two realizations be used to determine necessity and sufficiency? Is it some expectation over $c$ and $\bar{c}$? Even theoretically, how would I come up with $c$ and $\bar{c}$? After looking at (Pearl 2009), it seems that it only deals with binary variables which would make more sense. Is this an extension of (Pearl 2009)? If so, this needs to explained more. - What is the distribution of the $\bar{C}$ variable (even in theory)? This again is important missing detail. This is not explained except that there is another auxiliary distribution. It seems to be a completely independent random variable but it is unclear what distribution it is or should be in theory (or in practice). - It is not explained why the monotonicity term can be "absorbed" into the monotonicity measurement (Line 144). Also, why does this make sense as a monotonicity term? Why would this encourage monotonicity? Again, part of the confusion is that it is completely unclear what the distribution of $\bar{C}$ should be in theory. - For equation 6, is $\mathcal{T}(x,y) = P_t(x,y)$? If so, then this is the expected ratio to the $k$-th power? But this isn't a divergence because if $\mathcal{T} = \mathcal{S}$ this "divergence" would be 1. Thus, it does not even satisfy the most basic property of a distribution divergence. I also could not find the definition of $\beta$ divergence quickly in (Gong et al. 2016). Could you reference the specific definition? - Again, assumption 4.1 is unclear. And perhaps this could aid in answering the questions above. But it seems that the distributions of $C$ and $\bar{C}$ need to be different. Yet, again, I am not sure why as the paper does not explain $\bar{C}$ well. - Theorem 4.3 contains typos and seems to be incomplete. The statement after the equation is not a complete sentence and does not match the previous sentence. - The paper lacks a direct comparison with related works like IRM. How is the proposed objective or algorithm fundamentally different or similar to IRM or related methods? A direct comparison of the final equations would aid in understanding how this work fits in with related work. - The paper lacks discussion on the limitations of the practical optimization algorithm especially since this is a new adversarial optimization algorithm. See limitations comments. **Minor:** - The examples of necessary and sufficient features could be clarified. Perhaps just a 2 x 2 table of with necessary/ not necessary and sufficient/not sufficient. And then, just fill in each box with a feature. This might be simpler than having an example that takes careful analysis. Also perhaps showing with arrows, it might be easier to show. Like whiskers <-> cat, cat feet -> cat, pointy ears <- cat... - The figure is in the completely wrong place. Figures should always be placed after they are referenced and not before. **Review summary** Overall, I think the paper could be very good. The perspective is fresh and introduces some new concepts. However, the lack of conceptual and notational clarity and other weaknesses hinder it from being a strong paper.

Questions

- Do you limit the number of hyperparameter choices to a fixed number for every method? I believe DomainBed (or maybe WILDS) fixes the number of hyperparameter choices allowed for each method. This means that methods with more hyperparameters are harder to fit and thus may perform worse on average. - What is PA in Figure 2? I'm guessing that is a typo and should be SN? - Why not call OOD Generalization the more standard term of "domain generalization"?

Rating

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

3 good

Limitations

- The paper does not discuss the practical limitations or computational complexity of the method. The paper does not acknowledge that adversarial learning is inherently a very difficult problem both theoretically and practically. Were any instabilities encountered during training? How were these handled? It seems there were two encoders trained but this would seem to increase training instability. What about the constraint in Eq 8? How is that enforced in practice? Again, this seems to be a very difficult constraint to maintain. Furthermore, the KL term likely requires the distributions to be Gaussians to allow closed for computation, or did you use something else? - The paper acknowledges some of the limitations with respect to theory. More discussion on the assumptions and which are reasonably likely to hold in practice or not would be helpful. Also, how bad will the theory break if the assumptions do not hold?

Reviewer jeo88/10 · confidence 4/52023-07-06

Summary

A mostly theoretical paper that formulates the necessity and sufficiency of causal variables in order to optimize models to rely on necessary & sufficient variables. The paper provides rigorous theory connecting the empirical loss when models are optimized using the Probability of Necessity and Sufficiency Loss (PNS). Finally, there are some empirical results verifying the effectiveness of this approach.

Strengths

1) Rigorous theoretical guarantees for OOD generalization when optimizing with the introduced PNS loss on source domain S. Builds upon prior well-established theory. 2) Clear and well-explained intuition of using necessary and sufficient causal variables to prevent performance degradation when generalizing to unseen domains

Weaknesses

1) Rigorous empirical evaluation in the presence of more realistic datasets could help this paper be even more convincing. Evaluating on the recent more challenging dataset SpuCo for spurious correlations where existing methods fail can be a good starting point. Alternatively, considering other datasets from the WILDS package would be another useful direction.

Questions

N/A

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

4 excellent

Presentation

3 good

Contribution

4 excellent

Limitations

N/A

Reviewer aG697/10 · confidence 4/52023-07-28

Summary

This paper addresses the crucial issue of out-of-distribution (OOD) generalization in machine learning models, where the testing data distribution differs from the training data. To tackle this limitation, the authors introduce the concept of the Probability of Sufficiency and Necessary Causes (PNS), representing the likelihood of a feature being both necessary and sufficient. They propose a novel optimization approach, PNS risk, to enhance representations with high PNS values, effectively capturing the information of sufficient and necessary causes. The paper provides theoretical analysis and experimental results on synthetic and real-world datasets.

Strengths

* The authors propose a novel PNS risk and build the theoretical connections between PNS risk and generalization errors under some assumptions. * An approximate algorithm is derived to optimize the proposed risk. * Experimental results validate the effectiveness of PNS risk.

Weaknesses

I have the following comments: * The difference between the proposed PNS risk objective and other invariant learning methods: Although the PNS risk objective is newly incorporated in invariant learning, I would like to know the key difference from existing metrics/objectives. Since the objective is also to minimize the upper bound of PNS. I am confused about this. * The exact meaning of the derived bounds: In Equation (7), the authors provide an upper bound of the generalization error on the target distribution. I have some concerns about the RHS: when $k\rightarrow \infty$, $\beta_k$ will be extremely large, and it seems that the proposed method could not address the covariate shift problem. Therefore, I double the real effect of this bound. * Approximation of algorithms: In algorithms, conditional distributions are parameterized by some parameters, could you analyze the approximation error of such parameterizations? It seems that this is a very theoretical work, but there is a gap between theories and algorithms. Also, the authors do not mention how to calculate $\mathbb E_{c\sim P(C|X)}$, does this require multiple implementations $C$ for one given input $X$? Also, in Equation (8), why the minimization is taken over $\phi, w$? * Mismatch between algorithmic design and experiments: I think this algorithm & theories are not designed for covariate shifts, which could be seen from the remaining $\beta_k) term in Theorem 3.2. But for image classification tasks, the shift patterns are mainly covariate shifts, since one image is not likely to be given different labels in different domains. Also, for image datasets, the supports of different domains are likely to be disjoint, then why the proposed method could achieve good performance in these settings, which are not well-supported by theoretical analysis.

Questions

Please refer to weaknesses.

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

3 good

Presentation

2 fair

Contribution

3 good

Limitations

Please refer to weaknesses.

Authorsrebuttal2023-08-17

Rebuttal Feedback

Dear Reviewer aG69 As the discussion period ends soon, we just wanted to check if the response clarified your questions. Thanks again for your constructive feedback. Best, Authors

Reviewer aG692023-08-17

The authors have well addressed my concerns, and I hope these discussions could be included in the final version. I have updated my score to 7. Sorry for the delay!

Authorsrebuttal2023-08-21

Thank you for taking the time to review our rebuttal and adjusting the score. We genuinely appreciate your constructive feedback and will continue to make improvements based on your suggestions!

Reviewer jeo82023-08-10

Authors have addressed concerns (to some degree) regarding lack of larger-scale experiments

These are promising empirical results that I think further strengthen the paper and should definitely be included in the revision. I stand by my original assessment of the paper and recommend for acceptance.

Authorsrebuttal2023-08-21

We are genuinely grateful for the diligence you've shown in assessing our rebuttal and providing further comments. We will continuously improve the quality of the paper based on your constructive suggestions. Thank you!

Reviewer qVZ22023-08-12

Thank you for your response. I appreciated the added examples and further clarity. I would encourage the authors to push more for this in the final version if accepted. These are new concepts that need to be explained well and can be confusing. Both precise and clear should be the goals. Given both the general response and specific response, I have raised my score.

Authorsrebuttal2023-08-21

Thank you for taking the time to consider our rebuttal and providing further feedback! We deeply appreciate your constructive comments and we will continue to make improvements based on your suggestions!

Reviewer pPMP2023-08-21

Thank you for the detailed response. My concerns have been well addressed, so I will keep my score as accept.

Authorsrebuttal2023-08-21

We sincerely appreciate your efforts in reviewing our rebuttal and offering feedback. We will continuously improve the quality of the paper. Thank you!

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC