acceptodds
Under review as a conference paper at ICLR 2027

The Role of Local Background Knowledge in Maximally Oriented Partially Directed Acyclic Graphs

Abstract

In observational causal inference, background knowledge about specific edge orientations can refine a Markov equivalence class (MEC), with the resulting restricted class represented by a maximally oriented partially directed acyclic graph (MPDAG). Focusing on the causal effect of a treatment on an outcome , this paper addresses the question of how and when local background knowledge about edges incident to improves the identifiability of this effect. To answer this, we establish a suite of local graphical criteria. First, we derive a necessary and sufficient condition for the validity of proposed orientations of edges incident to . Second, we provide local orientation rules that determine whether an adjacent undirected edge of can be directed by the given background knowledge, without enumerating all DAGs in the MEC. Finally, via a notion of possible mediators, we give a necessary and sufficient, polynomial-time criterion for determining when background knowledge improves identifiability, validated on benchmark networks.

Primary area:
causal reasoning
Keywords:
Background knowledge, Maximally partially directed acyclic graph, Causal effects

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_fyi8OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper considers the problem of identifying interventional distributions in graphical models when the underlying DAG is not fully known. One such setting is when the DAG is known only up to its Markov Equivalence Class (MEC), which is a natural setting since using purely observational data, a graphical model can only be distinguished up to its MEC. It is well known that such an MEC can be described in terms of a partially directed mixed graph (sometimes called CPDAG) in which the directed edges are those for which every DAG in the MEC has the same direction.

The paper considers the more interesting setting where in addition to the Markov Equivalence Class, background knowledge of the the directions of some of the edges left undirected in the description of the MEC is available. Such extra information can, for example, from domain knowledge.

The paper builds upon earlier work of Perkovic and Guo and Perkovic who consider the problem of which interventions remain identifiable, in the MEC and MEC with background knowledge settings respectively. Here, identifiability means that the intervention distribution has the same value in all the DAGs in the MEC (or the MEC restricted by the available background knowledge). This paper asks a more fine-grained question: when does the number of possibles values of the intervention distribution decrease when background knowledge is added.

The results involve use of combinatorial properties of Markov equivalence (in the form of the classical Meek's rules), to reduce the problem (at least in part) to the work of Guo of Perkovic (e.g. in the proof of one of the main theorems, Theorem 4.3, in the appendix). Towards this, the paper also isolates some important combinatorial properties of MPDAGs (the version of CPDAGs with background knowledge) in Section 3.

Critical strengths
  1. The problem studied is quite natural, and the paper gives a satisfying answer.

  2. The paper also gives some numerical examples of how the reduction in the number of possible interventional distributions turns out in practice in some standard benchmark models.

Critical weaknesses
  1. The main results (Algorithm 1, Theorems 4.3/4.4) only consider a very special class of background knowledge: we are trying to identity the effect of node X on Y, and the background knowledge edges are all incident on X. What happens if background knowledge edges elsewhere in the graph are revealed?

  2. some parts of the paper may benefit from adding further discussion of related work/notions. These are described in the Questions section.

Questions for authors
  • Section 3. The results in this section are reminiscent of similar ideas used in various previous works on the combinatorial properties of CPDAGs; in particular results such as Wienöbst, M., Bannach, M., & Liskiewicz, M., "Polynomial-Time Algorithms for Counting and Sampling Markov Equivalent DAGs." AAAI 2021. https://doi.org/10.1609/aaai.v35i13.17448, and follow up works extending it to the setting of background knowledge.

    One reason to believe why such a connection may exist is that there is a short proof of Corollary 3.3 of the paper under review that follows easily from the above previous results, bypassing all the machinery in paper under review. To see this, we start with a characterization of Andersson et al. (1997) which says that a CPDAG is a mixed graph in which, in particular, the undirected components are chordal. It also follows from their characterization that (see, e.g., the Wienöbst et al. paper cited above) the DAGs in the MEC represented by the CPDAG are all obtained by directing each chordal component independently according to a perfect elimination ordering of the component. In the setting of Corollary 3.3, S1∪{X}S_1 \cup \{X\} is a clique in a chordal component containing XX and S1S_1, and hence standard algorithms for constructing perfect elimination orderings of chordal graphs (such as maximum cardinality search or lexicographic breadth first search) imply that there is a perfect elimination ordering of the component containing S1∪XS_1 \cup {X} that starts with S1S_1, followed by XX. In the DAG corresponding to this ordering, we thus have S1→XS_1 \rightarrow X, and hence the corollary follows.

    I could not immediately find a similar simple proof of Theorem 3.2 of the paper. However, I think it would further strengthen the paper to include a discussion of these connections.

  • Comparison to Perkovic 2020. l.304. The identifiability criterion of Perkovic (2020) does not seem to be a direct corollary of Theorem 4.3 as stated here since in Theorem 4.3, X and Y are singleton sets while the criterion in Perkovic (2020) is stated for the more general case where X and Y can be sets of nodes (I believe this is Theorem 3.6 in Perkovic (2020)). See also the first comment in the "Weaknesses" section.

  • Complexity of the algorithm. In the final discussion section, the paper highlights how the combinatorial results yield finally a polynomial time algorithm. But if only a polynomial time algorithm is needed, wouldn't just completing what the paper calls the "Meek cascade" by brute force (known to be in polynomial time) and then applying the criterion of Theorems 4.3/4.4 (existence of "possible mediator", a notion also present in the cited works of Perkovic and Guo and Perkovic) also work in polynomial time, without any of the combinatorial optimizations coming from the results in Section 3? (Of course, these results are also important for the proof of Theorem 4.3.)

Minor typos etc.

  • l. 234. The notion of a "semi-directed" cycle has been used undefined.
  • l. 310. "consistent to" -> "consistent with"
  • l. 789. "an DAG" -> "a DAG"
  • l. 798-805. "contradicts with" -> "contradicts"
  • l. 807. "only orient" -> "only orients"
  • l. 817-818: "connection" -> "concatenation"
  • Proof of Lemma A.5, l. 993. In cases 2, 3 and 4, shouldn't one only explicitly exclude the case that one of the directed edges in the "triggering" configuration was already directed in G∗\mathcal{G}^{*}? The proof I think is correct: this cannot be true for both the directed edges, and picking either works. But the wording of the proof seems to ignore this possibility.
3 comments
AuthorsOpenReview ↗

Reply to Reviewer fyi8

We sincerely thank the reviewer for the careful reading, the positive assessment, and the insightful and constructive comments. We address each point as follows.

W1: background knowledge revealed elsewhere in the graph.

This is a good question. Background knowledge outside the BB-component [4] of XX cannot affect the locality of XX: [1, Theorem 10] shows that the local structure around XX is unchanged, so our criterion is unaffected. For knowledge inside the BB-component of XX, one can run Meek's rules within that component to derive the knowledge incident on XX and then apply Algorithm 1; since the BB-component of XX is typically small, this remains efficient.

Our focus on knowledge incident on XX follows the local identification literature: the parent set of XX is a standard choice of back-door set, and IDA and its variants identify the possible causal effects by enumerating the parent sets of XX arising in the DAGs of the MEC. Knowledge incident on XX is precisely what changes those parent sets directly.

Q1: relation to Wienöbst et al. (2021).

We thank the reviewer for this observation, which is correct for Corollary 3.3: [2] gives a much shorter argument than the one in the paper, and we have added this argument, with the citation, immediately before Corollary 3.3.

It seems difficult to extend to Theorem 3.2. In a CPDAG no directed edge joins two siblings of XX, so the second condition of Theorem 3.2 cannot fail and a perfect elimination ordering of the undirected component settles validity. In an MPDAG that condition is exactly what rules out a directed cycle through XX. Indeed, this is one of the largest differences between MPDAGs and CPDAGs.

Q2: comparison with Perkovic (2020).

We agree and thank the reviewer for the correction. The criterion of [3] is stated for node sets X,Y\mathbf{X}, \mathbf{Y}, whereas Theorem 4.3 handles only singletons. We have removed the claim that [3] is a special case of ours; the set-valued case is left for future work.

Q3: is the Section 3 machinery necessary for a polynomial-time algorithm?

We agree that the alternative route you describe is polynomial. The submitted version answered this question with practical differences only: deciding by local orientations at XX requires fewer applications of Meek's rules, and the procedure never materialises the “Meek cascade”. The revised version replaces that answer with a stronger one.

Theory. We added Theorem 4.10, which shows that PM(X,Y,G∗)≠PM(X,Y,G′)\mathrm{PM}(X,Y,\mathcal{G}^{*})\neq\mathrm{PM}(X,Y,\mathcal{G}') if and only if PM(X,Y,G∗)∩⋃i=15Si≠∅\mathrm{PM}(X,Y,\mathcal{G}^{*})\cap\bigcup\nolimits_{i=1}^{5}\mathbf{S}_i\neq\varnothing. The sets S1,…,S5\mathbf{S}_1,\dots,\mathbf{S}_5 of Section 3 are therefore not merely an implementation device: they are exactly the local quantities that determine whether the PM\mathrm{PM} set changes, whereas two global closures tell us only that it changed, not why.

Algorithm. Algorithm 1 was rewritten corresponding to this theorem. Whether PM\mathrm{PM} changes is now decided by performing a single set intersection after the initial PM\mathrm{PM} is found. As a result, Algorithm 3 is no longer needed.

Experiments. The empirical evaluation was re-run on this rewritten algorithm, so every reported number corresponds to the procedure described in the revised paper.

A detailed complexity comparison with MIDA is given in Appendix C.4.

Minor typos. We thank the reviewer for these corrections. All of them have been corrected in the revised version.

On Lemma A.5 (cases 2–4): we agree with the reviewer's reading of the proof, and we thank the reviewer for pointing out the gap in the wording. The proof should explicitly exclude the case where one of the two directed edges in the triggering configuration was already directed in G∗\mathcal{G}^{*} — the two cannot both have been present, and either may be chosen. We have added this exclusion to the wording; the argument itself is unchanged.

We hope that our clarifications have adequately addressed your concerns, and would greatly appreciate it if you would consider revisiting your initial rating.

References

  • [1] Zheng, Q., Liu, Y., & He, Y. (2026). Local Causal Discovery With Background Knowledge. IEEE Transactions on Pattern Analysis and Machine Intelligence, 48(7), 7969–7980.

  • [2] Wienöbst, M., Bannach, M., & Liskiewicz, M. (2021). Polynomial-Time Algorithms for Counting and Sampling Markov Equivalent DAGs. Proceedings of the AAAI Conference on Artificial Intelligence, 35(13).

  • [3] Perkovic, E. (2020). Identifying causal effects in maximally oriented partially directed acyclic graphs. In Conference on Uncertainty in Artificial Intelligence (UAI), pp. 530–539.

  • [4] Fang, Z., Zhao, R., Liu, Y., & He, Y. (2025). On the Representation of Pairwise Causal Background Knowledge and Its Applications in Causal Inference. Journal of Machine Learning Research, 26(229), 1–73.

Thanks for the response. I maintain my original positive rating for the paper.

AuthorsOpenReview ↗

Thank you very much for your positive feedback and support!

Thank you very much for your positive feedback and support! We sincerely appreciate your constructive suggestions, which have greatly helped us improve the paper.

Reviewer_WrCuOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how local background knowledge regarding edge orientations can enhance the identifiability of causal effects in MPDAGs. The authors provide necessary and sufficient conditions for validating such background knowledge and for determining the orientations of additional edges. By introducing the concept of possible mediators, the authors characterize when background knowledge can narrow down the set of possible causal effects and propose a polynomial-time decision algorithm. Experiments conducted on benchmark networks validate these theoretical results.

Critical strengths
  • Determining whether background knowledge can enhance the identifiability of a causal effect is a significant question, especially in cases where the causal effect remains unidentifiable after incorporating such knowledge.
  • Theoretical results provide useful graphical criteria for validating background knowledge and locally determining the orientations of additional edges, without the need to enumerate all directed acyclic graphs (DAGs) within a Markov equivalence class.
Critical weaknesses

The proposed decision-making process is not entirely local. Although Theorem 3.4 utilizes local information surrounding XX to determine edge orientations, Algorithms 2 and 3 require searching the graph for paths from potential mediating variables to YY. Consequently, decisions regarding "improvements in identifiability" may depend on information lying outside the local neighborhood of XX. The authors should further clarify in what sense this overall decision-making process can be considered "local".

Questions for authors
  1. Does the claimed necessary and sufficient condition for identifiability improvement hold for a fixed observational distribution, or only in the sense that improvement is possible for some compatible distribution?
  2. How does the computational efficiency of the proposed algorithm compare with MIDA when the number of siblings of the treatment variable increases?
4 comments
AuthorsOpenReview ↗

Reply to Reviewer WrCu

We sincerely thank the reviewer for the careful reading, the positive assessment, and the insightful and constructive comments. We address each point as follows.

W1: in what sense is the decision process local?

This is a good and insightful question, and the submitted version was not accurate on this point. Obtaining PM(X,Y,G∗)\mathrm{PM}(X,Y,\mathcal{G}^{*}) for given XX and YY does require searching the graph for paths, and Algorithms 2 and 3 did exactly that, so the procedure as a whole was not local in the graph-theoretic sense.

The revised version separates the two steps. We first proved Theorem 4.10: PM(X,Y,G∗)≠PM(X,Y,G′)\mathrm{PM}(X,Y,\mathcal{G}^{*})\neq\mathrm{PM}(X,Y,\mathcal{G}') if and only if there exists Z∈PM(X,Y,G∗)Z\in\mathrm{PM}(X,Y,\mathcal{G}^{*}) such that X→ZX\to Z or X←ZX\leftarrow Z in G′\mathcal{G}'; equivalently, if and only if PM(X,Y,G∗)∩⋃i=15Si≠∅\mathrm{PM}(X,Y,\mathcal{G}^*)\cap\bigcup\nolimits_{i=1}^{5}\mathbf{S}_i\neq\varnothing. The sets S1,…,S5\mathbf{S}_1,\dots,\mathbf{S}_5 are computed by Theorem 3.4 and depend only on the adjacency relations among XX, its siblings, and S1∪S2\mathbf{S}_1\cup\mathbf{S}_2, that is, on local information around XX.

Algorithm 1 was rewritten accordingly. Computing PM(X,Y,G∗)\mathrm{PM}(X,Y,\mathcal{G}^*) on the original MPDAG (line 4) still requires searching paths, and we keep that step. But deciding whether identifiability changes (lines 6–10) is now local: it is a single set intersection. In particular, one no longer has to run Meek's rules to obtain the MPDAG after incorporating the background knowledge and then search paths again for the new PM\mathrm{PM} set; Algorithm 3, which performed exactly that second global pass, has been removed. The step that is repeated for every candidate piece of background knowledge is therefore local, even though the one-off computation of the initial PM\mathrm{PM} set is not. We also re-ran all experiments on the rewritten procedure, so the reported accuracies refer to the algorithm described here.

We have revised the statements in the introduction (Section 1) and in the conclusion: the procedure is now described as a polynomial-time decision procedure that avoids MEC enumeration and global Meek propagation, and the phrase “purely local” has been removed.

Q1: does the necessary and sufficient condition hold for a fixed observational distribution, or only for some compatible distribution?

As stated in Algorithm 1 (l.381–382), the guarantee is for some compatible distribution.

To see why, consider the simplest case of two nodes, XX and YY. Because the effect is not identifiable, the MEC contains both X→YX \to Y and Y→XY \to X, and in the latter the average causal effect of XX on YY is zero. In most cases, background knowledge X→YX \to Y improves our estimate of the causal effect, because we then know that the effect cannot be zero. But there exist distributions in which the average causal effect of XX on YY is in fact zero — for instance, half of the individuals have an individual causal effect of +1+1 and half have −1-1. Therefore we can guarantee that background knowledge improves the effect estimate only for some compatible distributions, not for every compatible distribution.

Q2: computational efficiency compared with MIDA as the number of siblings grows.

The two procedures answer different questions. MIDA enumerates the possible distinct causal effects; since the number of possible distinct causal effects can grow exponentially in the number of siblings of XX, MIDA cannot run in polynomial time in the worst case. Applied to our question, MIDA would also have to be run twice — once on G∗\mathcal{G}^{*} and once on the MPDAG obtained after incorporating B\mathcal{B} — and the two effect sets compared. Each run costs O(2M(G)⋅poly(m))O(2^{M(G)}\cdot\mathrm{poly}(m)), where M(G)M(G) counts the undirected edges at XX on possibly causal paths to YY, so both are exponential in the number of siblings of XX. Our procedure instead decides whether a given piece of background knowledge changes the set of possible causal effects without enumerating it. Specifically, our procedure finds whether the PM\mathrm{PM} set changes, which yields an O(n4)O(n^{4}) procedure rather than an exponential one. MIDA does answer a richer question, since it returns the effect values themselves; our decision problem does not require them, which is what allows it to be settled without enumeration.

We hope that our clarifications have adequately addressed your concerns, and would greatly appreciate it if you would consider revisiting your initial rating.

AuthorsOpenReview ↗

Thank you very much!

Thank you again for your constructive feedback. We have addressed your concerns in our rebuttal and updated the manuscript with additional clarifications and experimental results.

As the discussion period is approaching its end, we would greatly appreciate it if you could review our responses and let us know whether they address your concerns.

Thank you for your time and consideration.

Thanks for the response. My concerns have been mostly addressed and I would like to keep my rating.

AuthorsOpenReview ↗

Thank you very much for your positive feedback and support!

Thank you very much for your positive feedback and support! We sincerely appreciate your constructive suggestions, which have greatly helped us improve the paper.

Reviewer_cWUSOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper provides a polynomial-time algorithm for deciding whether adding background knowledge to a known MPDAG improves identifiability of the casual effect for a specified par of nodes (X,Y). Supporting theory, including a local test for whether the added background knowledge is consistent and a notion of "possible mediator", are also developed. All the claims are validated with proofs and an application to the ALARM dataset.

Critical strengths
  • The theoretical claims and proofs look solid, and the examples and application are helpful for understanding. I believe the paper is provides a good solution to the questions it poses.
Critical weaknesses
  • It's not clear to me that these results are especially helpful or relevant for the causality community. I can't imagine when a practitioner would realistically benefit from the algorithm, and I don't see how these results could be put to use in answering other pressing theoretical questions in the field.
Questions for authors

Main questions:

  1. Do you have any realistic use cases for the algorithm? Or speculation about how the results here connect to other important theoretical questions in the field?
  2. If one has background knowledge, why not use all of it to begin with? The paper suggests to me that one would first learn an MPDAG using only some background knowledge and then make use of this algorithm with some additional background knowledge.

Two important clarification questions:

  1. On page 3, starting around line 152, you're not just discussing an MEC but a restricted MEC, right? Namely, [G∗][G^*]?
  2. Do any of the results assume linearity? I saw it mentioned in the proofs, but I though that was just an example and not actually needed for the general result. The only mention of linearity I saw in the main text is the the second-to-last sentence "Extending the possible-mediator framework to...non-linear models...would broaden its applicabality", which I find puzzling.

Not actually questions, but hopefully helpful for the authors and future readers:

  • Parents, children, and related things are operators, so it's conventional to typeset them using \operatorname instead of bare math text
  • Is there a difference between the treatment XX in line 16 and the exposure XX on line 40?
  • Line 107: I think d-connecting is more conventional, since its describing a path and not a pair of sets of nodes
  • In (1), pa(Xi∣G)pa(X_i\mid \mathcal{G}) isn't quite the right object, is it? You need the realizations of these RVs?
  • Maybe the paragraph at line 170 should come earlier in the intro, since it's the first reference for things like MPDAGs and things already discussed on the preceding 3 pages
  • Line 342: why the change in notation from G∗∗G^{\ast\ast} to G′G'?
  • Figure 3(a): the DAG is not complete (no missing edges)
  • I guess it should be "IDA" on line 502
  • Figure 4(a) is missing the actual edge that R1 adds
6 comments
AuthorsOpenReview ↗

Relpy to Reviewer cWUS

We thank the reviewer for the careful reading. We understand the principal concern to be the practical relevance of the results, and we agree that the paper does not make the intended use case sufficiently clear. We address each point as follows.

W1 & Q1: practical value, realistic use cases, and connections to other theoretical questions.

Determining whether background knowledge can enhance the identifiability of a causal effect is an important question, especially when the effect remains unidentifiable after that knowledge is incorporated, as Reviewer WrCu also notes. A typical case is the effect of attending college (XX) on employment (YY). Structure learning returns an MPDAG in which the edge between college and skills training (ZZ) may remain undirected, since the data do not say whether college leads to training or the reverse. This matters: if college precedes training, training may carry part of the effect and enter the possible-mediator set; if training precedes college, it is a parent of XX and cannot. Randomising college attendance is infeasible, but assigning a training programme is not, and such an experiment orients that edge. Algorithm 1 decides, before the experiment is paid for, whether that orientation improves identifiability at all. The same test applies to every undirected edge at XX, so the candidate experiments can be screened before any of them is run. The gain need not be point identification: shrinking the set of possible causal effects is already a weaker but practically important form of improvement.

The question also connects to a gap in the identifiability literature: existing results characterise whether a fixed effect is identifiable in a given MPDAG, but not whether additional knowledge improves it, which naively requires enumerating the whole MEC. Theorem 4.10 settles it from the local sets alone. This composes with local causal discovery, where [1] learns the local structure around XX and infers causal relationships from it, while our criterion decides which further knowledge is worth acquiring; the same line of work includes an application to fair machine learning [2]. We have added both points to the related work (Section 2.3).

Q2: why not use all background knowledge at once?

  • Background knowledge is usually acquired incrementally, and any individual item may contradict the current MPDAG. Theorem 3.2 provides a necessary and sufficient condition for validating each item, so an invalid item can be identified.

  • Expert knowledge may be expensive, slow, or difficult to observe. As the example above illustrates, the only experiments available may be on neighbours of XX, and each has to be paid for. This trade-off between the cost of an experiment and the value of the knowledge it yields is the subject of active learning [3], where the goal is precisely to choose which intervention to perform next.

C1: is the MEC under discussion the restricted MEC [G∗][\mathcal{G}^{*}]?

Yes. The MEC under discussion is the restricted MEC, i.e. the set of Markov equivalent DAGs that are consistent with the MPDAG. We have made this explicit in Section 2.2.

C2: do the results assume linearity?

No result in the paper assumes linearity. Linear-Gaussian models appear only in the proof of Theorem 4.4, where they serve to construct an explicit witness distribution. Theorems 3.2, 3.4, 4.3 and 4.9 and Algorithms 1–2 are purely graphical results; the numbers in Table 1 come from the discrete CPTs of the ASIA network.

The concluding paragraph of the submitted version listed non-linear models among possible extensions, which was easy to misread as an assumption of the paper. We have removed that paragraph. The criterion and the possible-effect set are both defined graphically, and only computing the numerical values of the effect set requires a parametric model.

Minor points. We thank the reviewer for these careful observations. We have adopted all of them and incorporated the corresponding corrections in the revised version.

We hope that our clarifications have adequately addressed your concerns, and would greatly appreciate it if you would consider revisiting your initial rating.

References

  • [1] Zheng, Q., Liu, Y., & He, Y. (2026). Local Causal Discovery With Background Knowledge. IEEE Transactions on Pattern Analysis and Machine Intelligence, 48(7), 7969–7980.

  • [2] Zuo A, Wei S, Gong M. (2024). Interventional fairness on partially known causal graphs: A constrained optimization approach. International Conference on Learning Representations. 2024: 45791-45825.

  • [3] Hauser A, Bühlmann P. (2012). Two optimal strategies for active learning of causal models from interventions. Proceedings of the 6th European Workshop on Probabilistic Graphical Models. 2012: 123-130.

AuthorsOpenReview ↗

Thank you very much!

Thank you again for your constructive feedback. We have addressed your concerns in our rebuttal and updated the manuscript with additional clarifications and experimental results.

As the discussion period is approaching its end, we would greatly appreciate it if you could review our responses and let us know whether they address your concerns.

Thank you for your time and consideration.

Thanks for the rebuttal. It answers my clarification questions, but I have remaining concerns about motivation and use cases of the results:

  1. The rebuttal example about college vs training is about interventions/experiments. I don't see how designing an experiment to force a particular edge orientation can act as "background" knowledge for an existing observational dataset.
  2. If the motivation is about helping decide which interventions to do and how much the experiments help, then the work needs to engage with the substantial literature on that topic. The rebuttal (but not the updated draft) mentions one paper, but the updated draft should do this more thoroughly, including for example

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Revisions

We first summarise the main changes in the revised version, in which the substantive changes are marked in red.

  • A new theorem and a rewritten algorithm. We added Theorem 4.10, which shows that $\mathrm{PM}(X,Y,\mathcal{G}^{*}) $$ \neq \mathrm{PM}(X,Y,\mathcal{G}')$ if and only if some possible mediator ceases to be a sibling of XX in G′\mathcal{G}', equivalently if and only if PM(X,Y,G∗)∩⋃i=15Si≠∅\mathrm{PM}(X,Y,\mathcal{G}^{*})\cap\bigcup\nolimits_{i=1}^{5}\mathbf{S}_i\neq\varnothing. Algorithm 3 in the submitted version is no longer needed, and Algorithm 1 was rewritten accordingly: the decision is now a single set intersection rather than one auxiliary-graph search per possible mediator, and the auxiliary-graph construction is used only once, in Algorithm 2, to compute the initial PM\mathrm{PM} set.

  • Clarifications of definitions and claims. The MEC under discussion is now explicitly the restricted MEC (Section 2.2), and the connection to polynomial-time counting and sampling of Markov equivalent DAGs is stated before Corollary 3.3. The phrase “purely local” has been replaced by a precise description of what is local.

  • Corrections to the text. The concluding paragraph that mentioned non-linear models has been removed, and the typos and minor errors pointed out by the reviewers have been corrected.

  • Re-run of the experiments. All experiments reported in the paper have been re-run on the revised algorithm.

Dear Reviewers,

The author–reviewer discussion period is nearing its end, and the authors have responded to your reviews. Please engage with them as soon as possible, indicating whether their responses address your concerns and raising any remaining questions while there is still time for clarification.

Many thanks, AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.