Causal de Finetti: On the Identification of Invariant Causal Structure in Exchangeable Data

Constraint-based causal discovery methods leverage conditional independence tests to infer causal relationships in a wide variety of applications. Just as the majority of machine learning methods, existing work focuses on studying $\textit{independent and identically distributed}$ data. However, it is known that even with infinite i.i.d.$\ $ data, constraint-based methods can only identify causal structures up to broad Markov equivalence classes, posing a fundamental limitation for causal discovery. In this work, we observe that exchangeable data contains richer conditional independence structure than i.i.d.$\ $ data, and show how the richer structure can be leveraged for causal discovery. We first present causal de Finetti theorems, which state that exchangeable distributions with certain non-trivial conditional independences can always be represented as $\textit{independent causal mechanism (ICM)}$ generative processes. We then present our main identifiability theorem, which shows that given data from an ICM generative process, its unique causal structure can be identified through performing conditional independence tests. We finally develop a causal discovery algorithm and demonstrate its applicability to inferring causal relationships from multi-environment data. Our code and models are publicly available at: https://github.com/syguo96/Causal-de-Finetti

Paper

Similar papers

Peer review

Reviewer 3rag7/10 · confidence 2/52023-07-05

Summary

This paper is about causal discovery with exchangeable data, thus relaxing the traditional iid assumption. This allows for causal structure identification by conditional independence tests. The result has application to multi-environment data.

Strengths

Despite my low confidence, I think the paper delivers strong results for causal analysis. Both the Causal De Finetti theorems and the identifiability result (Theorem 5) appear not trivial and of some impact on people working in the field.

Weaknesses

- The identifiability result seems to hold for Markovian models only. - The experiments are very promising, but only a very limited setup. - The paper is very technical and dense, thus more suitable for a journal version. Yet, it can be worth presenting such a condensed version at the Neurips conference.

Questions

I have found it surprising that a relaxation of the iid assumption leads to higher identifiability. Is it possible to better explain this point? Figure 1a could be clearer to me. Is it possible to make it more expressive (or drop it)? Is there any room to extend the result to non-Markovian 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

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

3 good

Presentation

3 good

Contribution

3 good

Limitations

No specific issues related to that.

Reviewer ikix6/10 · confidence 4/52023-07-06

Summary

Causal analogs to de Finetti's theorem are proven, showing that if in an exchangeable distribution certain conditional independences hold, the distribution can be seen as being generated by a DAG with a latent variable corresponding to each node, determining that node's causal mechanism. It is further shown that knowing these conditional independences allows unique identification of the DAG. A causal discovery algorithm is presented leveraging these results, and evaluated in a synthetic data setting where data come from very many different environments.

Strengths

* The causal de Finetti's theorems are philosophically interesting, similar to how the original de Finetti's theorem can play a role in the justification of Bayesian inference. The resulting graphical models, expressing disentangled causal mechanisms in terms of latent variables (Figure 1b, right), are very insightful and deserve to be commonly known in the community. * The resulting conditional independences, which are testable if data are available from multiple environments, are a very useful ingredient for causal discovery algorithms in such settings.

Weaknesses

* The related work section only considers causal discovery, not the causal de Finetti theorem itself. Such references should also be listed, because the paper is claiming a contribution in this area. For instance, [Dawid 2021] also uses exchangeability to do causal inference. * Also about the related work section: the sentence "These algorithms on non-i.i.d. grouped data all demonstrate success, though it is unclear why grouped data enable causal structure identification." - I disagree strongly with this, there is a very good understanding of why multiple environments (possibly including interventional data) help with causal discovery, in the papers you list as well as in the causal inference literature as a whole. * Implications of the causal de Finetti theorem are listed without sufficiently arguments, and I believe overstated. See question about line 139 below. * Here is a counterexample to Lemma 2: $X_i \rightarrow X_a \rightarrow X_b \leftarrow X_c \leftarrow X_d \rightarrow X_j$ with $X_b \rightarrow X_j$. Then $X_d \in S_n$, so only $X_b$ is conditioned on, but this opens the listed path. (Separate from this, in line 294 explaining the lemma, I think "non-directed" should be "open": also directed paths other than the 1-arrow path should be blocked.) * Experiment (some things are unclear to me now, see questions below): * other methods are not really fit for this scenario * 1 x-y pair per environment (?): this disconnects the experiment from the theory **References:** [Dawid 2004]: Probability, Causality and the Empirical World: A Bayes-de Finetti-Popper-Borel Synthesis, Statistical Science , Feb., 2004, Vol. 19, No. 1 (Feb., 2004), pp. 44-57 [Dawid 2021]: Decision-theoretic foundations for statistical causality, Journal of Causal Inference 2021; 9: 39-77

Questions

* Line 139: "one can separately manipulate each latent variable controlling different mechanisms" - Does this follow from equation 5? How? ([Dawid 2004] warns that de Finetti's theorem only establishes that the latent variable exists in our minds, not in the real world.) * In the experiment, I assume $\tilde{N}^e$ is a 2-element vector? What about $N^e$? And how many x-y pairs are sampled per environment? Suggestions to improve the language (not relevant for my assessment of the paper): * the spelling of "i.i.d." is inconsistent (sometimes with spaces, sometimes with the final . missing) * line 92 & 595: "infinite exchangeable" -> "infinitely exchangeable" * the final sentence of section 2.2 doesn't parse ("due to"&"that underlies"; "involving observations are i.i.d.") * below definition 2: "does not hold for all" -> "does not hold for any" * line 140 & 141: "supporting mechanisms" -> "supporting that mechanisms" (2x) * line 147: "one implicitly make" -> "makes". Similar in appendix A. * Theorem 3 & 4: I'd replace "The sequence is" by "If", add "the sequence is" to point 1, and remove "if" from point 2 * line 201: "decide" -> "deciding" * line 220: "process" -> "processes" (also elsewhere) * Definition 5: "Given $P$ is" -> "Let $P$ be"; "Given $G$ be" -> "Let $G$ be"; "a ADMG" -> "an ADMG" (this also appears in Def 6); "read-off" -> "read off" * Definition 7: There is only one mapping fitting this definition, so define it straightaway instead of defining when something "is an ICM operator". Line 242 has a double "other" * line 280: $<$ -> $\leq$ * line 287: "topological ordering" is not the right concept here, as that is a total order * line 629: "exchangeable" -> "exchangeability" * line 679: "$l < k + 1$ -> $l \leq k$

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

Yes (except as discussed above)

Reviewer idLT7/10 · confidence 4/52023-07-06

Summary

The authors describe how the combination of assumptions about exchangeability and specific statistical tests can enhance the causal implications that can be derived from observational data.

Strengths

The utility and implications of exchangeability for causal inference are well described by the authors. Under the assumptions that they outline (the availability of multiple environments), the additional causal implications that can be derived are clearly very useful. The authors help readers by following essentially all complex mathematical statements with an informal statement of the result in relatively plain language.

Weaknesses

A 2019 paper by Jensen et al. [1] on exchangeability and structure learning in causal graphical models makes a very similar point as this paper regarding the value of multi-environment data and exchangeability. That paper describes how to use knowledge of environments (what that paper calls “parent objects”) to provide additional constraints on graph structure by conditioning on environment (and the latent variables implied by that environment). The proofs in that paper rely on exchangeability and contrast exchangeability with conditional independence. That paper differs somewhat from this paper in goals and focus, but the current paper would be improved by clearly describing the contributions of that paper and delineating the unique contributions of the present paper. The authors cite the Independent Causal Mechanism principle without reference to the long history of the concept, stretching back almost a century. Pearl’s work [2] (in which it is sometimes called “modularity”) cites an earlier review by Aldrich [3] (in which it is called “autonomy”), which cites very early work by Haavelmo and others as far back as the 1930s. This history is worth at least noting. The difference between “environments” and random variables is not clearly explained until Section 4, and this distinction is critical to understanding how (for example) schools differ from students in the second example of Section 3. Each school corresponds to many students. The plate notation introduced in Section 4 would help readers if introduced earlier. *References* 1. Jensen, D., Burroni, J., & Rattigan, M. (2020, August). Object conditioning for causal inference. In *Uncertainty in Artificial Intelligence* (pp. 1072-1082). 2. Pearl, J. (2009). *Causality*. Cambridge University Press. 3. Aldrich, J. (1989). Autonomy. *Oxford Economic Papers*, *41*(1), 15-34.

Questions

(none)

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 authors do not sufficiently emphasize the special structure of the data (specifically, multiple environments) that is needed to apply the methods that they advocate. Judging from the results shown in Figure 3, the benefit to multivariate structure learning only becomes substantial at truly large numbers of environments. The need for multiple environments (and large numbers of them) should be emphasized more clearly earlier in the paper.

Reviewer TMn58/10 · confidence 4/52023-07-07

Summary

This paper examines a stronger notion of exchangeability which provides invariant causal structure, i.e., is capable of serving as the basis for causal reasoning. The main contribution is a theorem which states that for pairs of RV with certain conditional independence properties it is always possible to represent them as a mixture of i.i.d. sequences with identical Markov factorization. This idea is further extended to the multivariate case. The authors then connect this notion of causal exchangeability to causal identification and causal discovery. Experimental results are provided for bivariate orientation and causal discovery from multiple datasets which show favorable performance when taking causal exchangeability into account.

Strengths

I found this paper to be a simple, intuitive, and thorough piece of work that provides an interesting new lens on causal inference. The notion of defining what constitutes exchangeable sequences with invariant causal structures is very interesting, and the authors do a very nice job of motivating the problem, providing thorough and well written theorems, and contextualizing them within the context of application. This is a clear description of causal mechanisms from a purely statistical view and an interesting addition to the literature.

Weaknesses

The largest weakness I can see here is the lack of comprehensive empirical evaluations. The evaluations provided are fairly simplistic, limited in scope, and serve more as a demonstration rather than an evaluation finding relative strength and weakness of the proposed method.

Questions

Did you run any other causal structure learning algorithms such as FCI, GES or even optimization based approaches such as NOTEARS? Given the poor performance of PC I'm wondering if this is a weakness shared across all algorithms.

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

4 excellent

Contribution

4 excellent

Limitations

Yes.

Reviewer 3rag2023-08-16

Keeping the same (positive) recommendation

I thank the authors for their comments, which helped me understand some points better. As far as I can understand, the (serious) problem with the proof of Lemma 2 raised by Reviewer ikix has been fixed, and this had no impact on the empirical validation of the result. In this situation, I am happy to confirm my positive (but low confidence) opinion about the paper.

Authorsrebuttal2023-08-21

We thank the reviewer for responding and their positive feedback on the paper.

Reviewer ikix2023-08-17

I think this paper has the potential for high impact, but with this comes great responsibility to be precise and complete about positioning w.r.t. related work. I am happy that the authors promise to improve the paper in this regard. In the case of Lemma 2, they provide a concrete fix, and I agree that the new version of the lemma is correct; I am glad this had no impact on the experimental results. I am raising my score to 6. Not having seen the updated version, I am hesitant about raising it further.

Authorsrebuttal2023-08-17

We thank the reviewer for going the extra mile in finding an error and checking the corrected version. We thank the reviewer also for acknowledging the changes we made and for raising the score. We will take particular care regarding the related work section for the final version, and hope the updated version will not disappoint. In case you feel comfortable revealing your identity after the end of the reviewing process, we would be happy to add your name to the acknowledgement.

Reviewer idLT2023-08-20

I appreciate the authors' careful reading of my review and their thoughtful response. The changes that you outline will improve the paper. One side note: After reading Hoover (2008), I'm not convinced that it is a particularly good reference regarding autonomy/modularity. Hoover certainly mentions the Lucas Critique (a version of the modularity issue), but most of the entry is on other topics. Aldrich (1989) is much more on-point, as are some specific portions of Pearl (2009) (though the discussion of modularity is oddly dispersed throughout that book). I'll increase my rating.

Authorsrebuttal2023-08-21

We thank the reviewer for acknowledging our changes and raising the score. We thank the reviewer for pointing the appropriate references for ICM and will include the more relevant reference, e.g., Pearl (2009), Aldrich (1989) in the updated version.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC