Interventional Causal Discovery in a Mixture of DAGs

Causal interactions among a group of variables are often modeled by a single causal graph. In some domains, however, these interactions are best described by multiple co-existing causal graphs, e.g., in dynamical systems or genomics. This paper addresses the hitherto unknown role of interventions in learning causal interactions among variables governed by a mixture of causal systems, each modeled by one directed acyclic graph (DAG). Causal discovery from mixtures is fundamentally more challenging than single-DAG causal discovery. Two major difficulties stem from (i)~an inherent uncertainty about the skeletons of the component DAGs that constitute the mixture and (ii)~possibly cyclic relationships across these component DAGs. This paper addresses these challenges and aims to identify edges that exist in at least one component DAG of the mixture, referred to as the true edges. First, it establishes matching necessary and sufficient conditions on the size of interventions required to identify the true edges. Next, guided by the necessity results, an adaptive algorithm is designed that learns all true edges using $O(n^2)$ interventions, where $n$ is the number of nodes. Remarkably, the size of the interventions is optimal if the underlying mixture model does not contain cycles across its components. More generally, the gap between the intervention size used by the algorithm and the optimal size is quantified. It is shown to be bounded by the cyclic complexity number of the mixture model, defined as the size of the minimal intervention that can break the cycles in the mixture, which is upper bounded by the number of cycles among the ancestors of a node.

Paper

References (53)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer Tiwj5/10 · confidence 2/52024-07-11

Summary

This paper deals with using interventions to learn the causal links in a mixture-of-DAGs model. They find the minimum number of intentions that are required to learn "true edges" where a true edge from a node X to a node Y indicates that at least in one mixture component, X is a parent of Y. They also present an algorithm that learns true edges via interventions. They show that in terms of the number of interventions, the proposed algorithm is close to optimal.

Strengths

1. Up to my knowledge this is the first time that the problem of using interventions to learn the mixture of causal DAGs is studied. 2. As far as I could follow, the presented theorems on the number of sufficient and necessary interventions as well as the proposed algorithm are correct.

Weaknesses

1. To present their theorems and algorithm, they require several definitions and this can be confusing in the first read. If they could provide a simple running example (i.e. a mixture of DAGs) by which they could illustrate/clarify all definitions, the paper would be easier to follow --in my opinion. 2. The provided experimental results are on synthetic data only. This, I think is the main weakness of this paper. If (as claimed in the abstract and intro) many causal systems are mixtures of DAGs, it would be helpful if the authors could present and analyze at least one such real-world system in the experiments section.

Questions

Is there any existing and open-source real-world dataset where the data generation mechanism is a mixture of DAGs? If yes, is there a reason that such a dataset is not utilized in this paper?

Rating

5

Confidence

2

Soundness

3

Presentation

3

Contribution

3

Limitations

N/A

Reviewer 2inT6/10 · confidence 4/52024-07-12

Summary

The paper studies the setting where data is generated from a mixture of DAGs and one wishes to recover the "true edges" (edges that exist in at least one of the underlying DAGs). Similar to the usual causal discovery setting, observational data alone is insufficient and interventions are required. The paper characterize the set of interventions needed to learn these true edges and provide an algorithm for doing so. Some small scale experiments and source code were given.

Strengths

The concept of mixture of DAGs provide an alternative framework to capture cyclic causal system and, to my understanding, is under-explored. This paper aims to fill some of these gaps.

Weaknesses

The paper strives to recover "true edges" (objective 1 on Line 195) but I don't think they can map these edges to the appropriate underlying DAG within within the mixture of DAGs. It is unclear why it is interesting to be able to know these edges without the context of the DAG they belong to. To me, this severely weakens the motivation/usefulness of this work. The experimental evaluation felt rather weak.

Questions

General questions: - [1] also studied the universal lower bound problem that you mentioned on Line 92, and [2] eventually provided an exact characterization for the number of interventions required to recover the DAG from the observational essential graph. You may want to consider adding these references in your revision. - Can you motivate why it is sufficient/useful enough to recover "true edges" (objective 1 on Line 195)? Your approach cannot identify which DAG each of the recovered edges belong to within the mixture of DAGs, right? - Is the number $K$ of mixture models given to the algorithm as input? The phrase "a priori knowledge of the number of mixture components" on Line 251-252 seem to suggest that $K$ is typically not given. Also, Algorithm 1 is independent of $K$. However, $K$ features prominently in the characterization (e.g. Theorem 2), so how does one perform the correct set of interventions without $K$ as input? - On Line 282, you mention "intervening on any set $I$ that contains $B(i)$ breaks all the cyclic relationships in $C(i)$". Why is this so? I thought some "edge cutting" (i.e. exactly one endpoint of the edge is in intervention set) is needed in order to distinguish edge directions? For example, intervening on the entire set of vertices will trivially include $B(i)$ but this does not provide any useful information, right? - In Line 290, do you want to perhaps write $(A \setminus (S_1(i) \cup \ldots \cup S_{n-1}(i)))$ since the algorithm is "removing the layer $S_u(i)$ from $A$ after each iteration", or at least reference Line 26 in Algorithm 1? - In equations (23) and (24) of Appendix C, shouldn't are you missing $\cup \{i\}$? I don't think you defined descendants and ancestors to include $i$ itself, or maybe I'm mistaken? - On Line 620, you wrote $S_1$. I don't see this defined anywhere (please correct me if I missed it). I don't think this is a typo of $S_1(i)$ here since $j \in S_1(i)$ and $j \not\in pa_m(j)$. This undefined notation also appears in the remaining of this proof and is inhibiting my ability to verify Theorem 3. - The mixture in Figure 2 assumes that $P_{G_1}(1) = P_{G_2}(1)$ since the $y$ node does not point to the $I$-mixture DAG (definition 5 and equation (2)), right? This assumption was not mentioned in Appendix D. Experiment questions: - The graph sizes in the experimental evaluation were extremely small. Is there a reason for this? In Appendix E (line 727), you say "only a marginal decrease in performance" when $K = \{2,3,4\}$ and $n \in \{5,\ldots,10\}$. I'm not sure if that is enough scale to draw such a conclusion... - By setting p = 2/n in the Erdos-Renyi G(n,p) graph generation, the graphs will likely be disconnected (random graph theory tells us we need about p ~= log n / n for the graph to be connected). Will this be a problem for your experiments if you scale to larger graphs? - Consider mentioning the edge weights of your linear Gaussian model in the main paper instead of the appendix. Line 348 claims that s = 1000 samples are sufficient for almost perfect precision but finite sample guarantees for "good recovery of the graph" scale non-trivially with the correlation strength, e.g. see the finite sample analysis for linear Gaussians in [3]. - What is the mixture skeleton learning algorithm? - In Appendix E, you mentioned that "the same edge weight is assigned to all realizations of a true edge across all component DAGs". Why is this reasonable? I thought the whole point of mixture models is that $p_\ell \neq p_{\ell'}$ in general? See Line 117. Possible typos: - Extra bold of "ixture" on Line 255? - Equation (57): Do you mean $C(5) = \emptyset$? References: [1] Porwal, Vibhor, Piyush Srivastava, and Gaurav Sinha. "Almost Optimal Universal Lower Bound for Learning Causal DAGs with Atomic Interventions." International Conference on Artificial Intelligence and Statistics. PMLR, 2022. [2] Choo, Davin, Kirankumar Shiragur, and Arnab Bhattacharyya. "Verification and search algorithms for causal DAGs." Advances in Neural Information Processing Systems 35 (2022): 12787-12799. [3] Kalisch, Markus, and Peter Bühlman. "Estimating high-dimensional directed acyclic graphs with the PC-algorithm." Journal of Machine Learning Research 8.3 (2007).

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Nil

Reviewer bGit7/10 · confidence 3/52024-07-13

Summary

This work studies an important problem in causal discovery for its relevance in the real world -- identifying the causal relationship when the underlying data-generating process comes from a mixture of different DAGs. They give the necessary and sufficient size of intervention set to identify the union of all the parents of a node across components. They also give an algorithm that requires $O(n^{2})$ interventions to identify all the directed edges of individual DAG. Finally, they also quantify the gap between the number of interventions used by the proposed algorithm and the optimal size in terms of cyclic complexity number.

Strengths

1. The paper is well-written and easy to follow. 2. Studying a mixture of DAG is underexplored but will help bring the application of causality closer to the real world. This paper furthers the line of work in this direction and is thus important. 3. This paper characterizes the necessary and sufficient conditions to identify the "true" edges, i.e., the edges that are actually present in at least one of the components of the mixture from the emergent edges. Also, they don't impose restricting assumptions on component DAGs like poset compatibility from previous work, thereby generalizing the results to a richer mixture family. 4. Also, to the best of my knowledge, this is the first work in causal discovery for a mixture of DAGs that allows for the use of interventional data, thereby improving identifiability.

Weaknesses

1. It is understandable that this paper doesn't show the sample complexity of identifying the true edge set, but maybe the empirical section could be more diverse with a larger number of nodes and components to get a sense of the proposed algorithm's statistical efficiency.

Questions

1. This is not directly relevant to the proposed algorithm and might be a direction for future work. But is it possible to partition the identified true edges into subsets for each of the individual components? Can we even identify the number of components in the mixture?

Rating

7

Confidence

3

Soundness

3

Presentation

4

Contribution

4

Limitations

Yes

Reviewer 5C9J7/10 · confidence 3/52024-07-15

Summary

In the case of a single DAG, conditional independence tests specify the skeleton (under faithfulness) and interventions are limited to orienting edges. In the case of data coming from a mixture of DAGs, it is possible for two variables to not be adjacent in any of the components but still be conditionally dependent for every conditioning set. This paper proves that the necessary and sufficient intervention size to identify the "true" edges of the mixture, i.e. node pairs where there is an edge in at least one component DAG is the size of the union of parents in all components + 1. It also designs an algorithm to do so.

Strengths

Causal discovery given data from a mixture of DAGs seems to be an important problem with applications in several domains. Existing work has focused on the observational data regime. But this is not sufficient in the mixture case because of the phenomenon of "emergent pairs". This is the first paper that studies this problem assuming presence of interventional data. Using interventional data to pinpoint DAGs among an equivalence class is also a relevant thread of interest. The paper is written well with a clear flow of ideas. I didn't check all the proofs but the ones that I didn't, appear sound. The analysis of the problem appears quite complete in a sense with the optimal intervention sizes characterized and the algorithm's gap to optimality also characterized.

Weaknesses

The results heavily depend on the interventional model where the modified distribution of the intervened variable does not change across component DAGs. I was wondering if there was any practical motivation for considering such an intervention model. If so, it would make sense to include it in the paper. I also could not find connections to existing results on intervention sizes for single DAG which should be a special case of thie mixture case. Is this because of the interventional model, again?

Questions

Already asked in the weakness section.

Rating

7

Confidence

3

Soundness

4

Presentation

3

Contribution

3

Limitations

Addressed implicitly in a conclusions section

Reviewer 2inT2024-08-09

Thank you for your detailed responses! I am very satisfied with them and intend to maintain my positive score. I have a non-technical follow-up question that I am curious to hear from the authors. **Recovering true edges**: I understand and can appreciate that in certain mixtures, just learning the direction of 1 true edge is sufficient to have a real-world impact. However, in the example of your response, the edges are (i) "consistent" and (ii) acyclic. What if there is a mixture such that (i) $1 \to 2$ in one DAG and $2 \to 1$ in another, and (ii) $1 \to 2$ and $2 \to 3$ in one DAG while $3 \to 1$ in another In such scenarios, how should one make use of the recovered edges in a meaningful manner?

Authorsrebuttal2024-08-09

We are glad to hear that our response addressed your questions! We comment on your follow-up examples as follows. - Having cyclic relationships in your examples fits the motivation of mixture models well. Consider your first example with $\mathcal{G}_1 : 1 \rightarrow 2$ and $\mathcal{G}_2 : 2 \rightarrow 1$. In this case, learning the true edges tells us about what we *should not do*, more than what to do. Suppose that an experimenter does not know which model they should adopt for the data; a DAG or a mixture of DAGs. If they elect a single DAG with a single edge, say $1 \rightarrow 2$ (possibly, the stronger among the two ground truth edges), there may be unintended consequences. For instance, if the mixture model is due to a feedback loop, then trying to control the level of $X_2$ via $X_1$ will result in an unintended increase in $X_1$. On the other hand, by taking the cautious route and considering a mixture mode, we identify the true edges in both directions and avoid making incorrect inferences. - The second example is a very interesting case. Note that, in Stage 1 of the algorithm, we learn the "mixture ancestors" as an intermediate step. Hence, we have the information $1 \in {\rm an}\_{\rm m}(3), 1 \notin {\rm pa}\_{\rm m}(3), 2 \in {\rm pa}\_{\rm m}(3), 3 \in {\rm pa}\_{\rm m}(1), 2 \notin {\rm an}\_{\rm m}(1), 3 \notin {\rm pa}\_{\rm m}(2), 1 \in {\rm pa}\_{\rm m}(2)$. Suppose that we know $K=2$. Then, we can see that the only possible mixture is $E_1 = \\{1 \rightarrow 2, \ 2 \rightarrow 3\\}$ and $E_2 = \\{3 \rightarrow 1\\}$, so we learn the individual DAGs in this case! (this is because $3 \rightarrow 1$ cannot be in the same DAG as either of the other two edges due to the known ancestral relationships). Without learning the true edges though, we cannot say whether we can learn the individual DAGs from the mixture model. - The second example also shows that our work can inspire future work for the use of interventions in a mixture of models, e.g., establishing graphical conditions for (partial) recovery of individual DAGs, leveraging the knowledge of $K$ when provided.

Reviewer 2inT2024-08-10

I see! Thanks for the responses.

Reviewer Tiwj2024-08-12

Thanks for the response

Reviewer bGit2024-08-12

I thank the author for their response. I will maintain my score.

Reviewer 5C9J2024-08-13

Response to rebuttal

Thanks for the response. I am happy to maintain my original score.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC