Causal Discovery from Subsampled Time Series with Proxy Variables

Inferring causal structures from time series data is the central interest of many scientific inquiries. A major barrier to such inference is the problem of subsampling, i.e., the frequency of measurement is much lower than that of causal influence. To overcome this problem, numerous methods have been proposed, yet either was limited to the linear case or failed to achieve identifiability. In this paper, we propose a constraint-based algorithm that can identify the entire causal structure from subsampled time series, without any parametric constraint. Our observation is that the challenge of subsampling arises mainly from hidden variables at the unobserved time steps. Meanwhile, every hidden variable has an observed proxy, which is essentially itself at some observable time in the future, benefiting from the temporal structure. Based on these, we can leverage the proxies to remove the bias induced by the hidden variables and hence achieve identifiability. Following this intuition, we propose a proxy-based causal discovery algorithm. Our algorithm is nonparametric and can achieve full causal identification. Theoretical advantages are reflected in synthetic and real-world experiments.

Paper

Similar papers

Peer review

Reviewer 3AR77/10 · confidence 2/52023-06-20

Summary

The authors suggest a method of causal discovery for multivariate time series under the regime when the data is being sampled at constant skips in the time dimension. Under mild assumptions, they prove that their method works asymptotically.

Strengths

The paper's contribution is clear, the methods are interesting and novel, and the subject (causal inference in multivariate time series) is of interest to many. Other than some places I will mention, the explanations are clear.

Weaknesses

I would like to see a little more discussion on previous work. I see citations about work that inspired the current one, for example [17, 19, 25] on using descendants of an unobserved variable to differentiate direct causation from hidden mediation, but I do not see citations for competing methods for causality under the regime of sampling on the time dimension. It would be good to be explicit about what the current state of the art is in that direction, and how the current work compares. You mention them in Section 5 (SVAR-FCI, NG-EM, Dynotears, PC-GCE) but I'd like to see a bigger discussion on the differences. Specifically, it would be great if we had a basic example in mind where "naive" approaches clearly fail, and make us understand how the proposed method avoids the issue intuitively.

Questions

1. I feel like Definition 2.3 can be phrased more simply. The way you wrote it makes it look circular. 2. I am a little confused by the figures. In Figure 2.c., I don't understand why there wouldn't be an arrow from A(t_3) to M(t_5). Or perhaps the faded arrows count? Why is there a difference between faded and non-faded arrows? 3. In Figure 3, where are some vertices faded and some not? Is there a difference? It looks like it is meant that faded means unobserved, but I don't think that's what you meant. 4. Put Remark 4.2 outside of the algorithm. 5. Line 224: random is misspelled.

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

None.

Reviewer Y9xG6/10 · confidence 4/52023-06-29

Summary

# Summary In this paper, the authors address the problem of inferring causal structures from subsampled time series data, where the frequency of measurement is much lower than that of causal influence. This presents challenges in identifying the causal structure, as hidden variables at unobserved time steps can induce bias. Existing methods that tackle this problem are limited to linear cases or fail to achieve identifiability. The main contribution of this paper is a constraint-based algorithm that can identify the entire causal structure from subsampled time series without any parametric constraints. The authors propose a proxy-based causal discovery algorithm that leverages the temporal structure of time series data to remove the bias induced by hidden variables. The algorithm is nonparametric and can achieve full causal identification. Specifically, the author leverages the proxy variables to test the edge directions of the summary DAG from the uniquely identified MAG. The authors demonstrate the theoretical advantages of their method and provide experiments on both synthetic and real-world data, showcasing improved performance over existing methods.

Strengths

# Originality The paper presents a novel approach to causal discovery in subsampled time series data by proposing a constraint-based algorithm that leverages proxy variables to handle the challenges posed by hidden variables at unobserved time steps. The originality of the method lies in its ability to identify the entire causal structure without any parametric constraints, setting it apart from existing methods that are limited to linear cases or fail to achieve full identifiability. The authors draw inspiration from the recent progress in proximal causal discovery and adapt it to the subsampled time series setting. # Quality The quality of the paper is high, as it presents a well-formulated methodology with solid theoretical foundations. The authors provide rigorous proofs for their proposed algorithm's identifiability properties, ensuring that the algorithm is grounded in strong theoretical underpinnings. Although I have some questions regarding its generality and required assumptions, I will elaborate about them later. As for the experiments, the author conduct 1 synthetic and 1 real-world experiment, which demonstrates its effectiveness against the baselines. # Clarity The paper is well-written and clear in its presentation. The methodogy is clear, thanks to the intuitive explanation provided by the authors. They also provide necessary background information on causal discovery and related literature. # Significance The problem that this paper aims to talks is significant in the field of tmeproal causal discovery. Although I don't think this paper fully addresses it, it made a reasonable progress towards the final destination.

Weaknesses

# Weakness Although the paper provide a solid progress towards the final destination, here are some limitations I am a bit worried about: ## Assumptions are too strong: One assumption is that only 1 step window is considered in this paper. However, is this true in practice? Is it possible that some factors has long-standing effect (higher-order Markovian) to the target variable? Can your method handle such scenario? For summary DAG, does it has to be a DAG? Since from general full-time DAG, the summary graph may not be a DAG and can contain cycles. Does your method handle such cases? For example, A(1) -> B(2) and B(1) -> A(2), and the summary graph is A<->B. What are the consequence if some variables are not self-caused? That means the proxy variables may not exists. In that case, full identifiablity cannot be establish, right? You may want to add something like to what extend of identifiability your method can establish without such assumption. ## Comparison with baselines The author includes some of the well-known baselines. Surprisingly, Dynotears performes reasonably ok considering that it is a linear model (in its original form). However, after Dynotears, several state-of-the-art baselines are proposed, like Rhino [1]. Can you compare your method with this nonlinear model? [1]Gong, Wenbo, et al. "Rhino: Deep Causal Temporal Relationship Learning With History-dependent Noise." arXiv preprint arXiv:2210.14706 (2022). Code at https://github.com/microsoft/causica/tree/v0.0.0 ## Scalability How expensive it is to run such algorithm, since you need to test every pair. For real-world experiment, you scale it to 90 variables, how long does it need to run it? If your method can be scaled to higher-order SVAR, how does it scales with the window size?

Questions

1. In figure 4 proxy model, do you mean $M(t_2)$?

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

3 good

Contribution

3 good

Limitations

The author did discuss its limitations but they only discuss the limitations of conditional independence test. Howver, in the weakness section, I have raised several improvements the author can consider. I suggest including these discussions in the limitations as well.

Reviewer Y9xG2023-08-16

Thanks for the authors' effort on addressing my concerns. They managed to addressed most of my concerns. But still, longer dependencies (longer than window 1) can still be present in practice, and this is a strong assumption in my personal opinions. I will keep my current score.

Reviewer 7zAn5/10 · confidence 5/52023-07-04

Summary

This paper proposed a non-parametric constraint-based algorithm that can identify the entire causal structure from subsampled data, which leverages the proxy variable to adjust the bias induced by the hidden variable.

Strengths

- Concise and clear theoretical derivation. Introduce the proposed method by discussing the connections between different graphs step by step.

Weaknesses

- The paper is rather incremental as the core of the method is based on [19] to use proxy variables to detect and eliminate the confounding effect brought by the subsampling. - Moreover, the method proposed by [19] seems to require more assumptions but is not disclosed in this paper. And it is unclear whether those specific assumptions can be satisfied in this work. - In addition, this work assumes the self-exciting property of the time-series which is not necessarily held and it is interesting to see how to find the proxy variables and what would be the result in this case. - It could be better to provide more than 5 vertex numbers in the experiment. - What would be the result if the causal graph becomes denser?

Questions

See the weakness above.

Rating

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

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

Yes.

Reviewer XP2k5/10 · confidence 3/52023-07-04

Summary

In this paper, the author(s) propose a new technique to learn the summary graph of time-series data. As a motivation for their work, the author(s) discuss the interesting application of learning causal pathways in Alzheimer’s disease. The time series model studied in this work is quite general, and suitable for many applications. The author(s) provide a simple algorithmic approach to learn the summary graph. This algorithm is essentially based on verifying d-separation, by testing conditional independence. The author(s) conclude with an experimental comparison, showing that their method achieves superior performance than a baseline, using various synthetic datasets and a real-world medical application.

Strengths

This paper studies an very important and difficult problem. The application of learning causal pathways in Alzheimer’s disease is very relevant. The paper is also well-written and easy to follow.

Weaknesses

My main concern pertains the faithfulness assumption. Faithfulness it is useful for getting identifiability results, and it allows to recover the causal structure by testing conditional independence. However, in many scenarios there is no practical reason to assume faithfulness. The use of faithfulness significantly limits the novelty of their work. If I understand correctly, the implementation of their algorithm essentially uses conditional independence to recover the causal structure. Testing conditional independence is practically problematic. Hence, I also have doubts on the scalability of their method.

Questions

Can you please describe exactly step 1 in your algorithm? How do you test conditional independence?

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

3 good

Presentation

3 good

Contribution

2 fair

Limitations

Yes, the author address limitations in Section 6.

Reviewer GTBD6/10 · confidence 3/52023-07-15

Summary

This paper studies the problem of subsampled time series in causal discovery, in which the unobserved time steps may lead to the existence of latent confounders. To this end, this paper proposes a constraint-based algorithm by leveraging proxy variables to remove the bias induced by latent confounders. The experimental results verify the effectiveness of the proposed algorithm.

Strengths

1. This paper addresses the problem of subsampled time series in causal discovery, which is important but challenging. 2. The paper is well-structured and written. 3. The experimental results show that the proposed method outperforms several representative baselines.

Weaknesses

1. In Theorem 2.8, extra assumptions are required for testing conditional independent relations in related literature, but they are not discussed in this paper. 2. How to search the proxy variable of target hidden variables. If an invalid proxy variable is selected, what is the output of the proposed algorithm? 3. In simulation data, the dimension of the random graph is only five. Can you show the performance of the proposed method in a larger-scale network?

Questions

Refer to Weaknesses

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, 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

2 fair

Presentation

3 good

Contribution

3 good

Limitations

Refer to Weaknesses

Reviewer XP2k2023-08-15

I would like to thank the reviewers for answering my questions. Given that the author(s) addressed my concerns in detail, and given the overall discussion and positive scores from the other reviewers, I will raise my score accordingly.

Reviewer 3AR72023-08-17

Thanks, I will keep the rating.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC