Causal Discovery in Semi-Stationary Time Series

Discovering causal relations from observational time series without making the stationary assumption is a significant challenge. In practice, this challenge is common in many areas, such as retail sales, transportation systems, and medical science. Here, we consider this problem for a class of non-stationary time series. The structural causal model (SCM) of this type of time series, called the semi-stationary time series, exhibits that a finite number of different causal mechanisms occur sequentially and periodically across time. This model holds considerable practical utility because it can represent periodicity, including common occurrences such as seasonality and diurnal variation. We propose a constraint-based, non-parametric algorithm for discovering causal relations in this setting. The resulting algorithm, PCMCI$_Ω$, can capture the alternating and recurring changes in the causal mechanisms and then identify the underlying causal graph with conditional independence (CI) tests. We show that this algorithm is sound in identifying causal relations on discrete time series. We validate the algorithm with extensive experiments on continuous and discrete simulated data. We also apply our algorithm to a real-world climate dataset.

Paper

References (39)

Scroll for more · 27 remaining

Similar papers

Peer review

Reviewer ztgZ6/10 · confidence 4/52023-06-25

Summary

The problem of causal discovery is tackled for time series data in those cases where stationarity of multivariate time series data can not be assumed. A problem formulation based on structural causal model is given by considering the case of semi-stationary time series. Under this assumptin, the paper considers the case where a finite number of different causal mechanisms happen in a sequential manner and periodically across time. A constraint based, and non parametric, structural learning algorithm is designed and developed by leveraging on the PCMCI, described in the specialized literature. Such an algorithm is extended and adapted to cope with the considered semi-stationary setting. The causal discovery problem is tackled by standard conditional independence tests. The paper proves that the proposd algorithm is sound when asked to discover and identify causal relationships between variables of discrete time series. Numerical experiments, both from synthetic and real world data, are reported and described to comment on the performance of the proposed approach. The main contributions of the paper are in my humble opinion the following: - designed and developing a new causal discovery algorithm for semi-stationary time series where the causal graph changes periodically. - validation of the proposed approach through syntethic data

Strengths

- the tackled problem which is of both theoretical and practical interest - the theoretical framework of the considered problem

Weaknesses

- the proposed algorithm wisely exploits the PCMCI algorithm while not being original, it looks a little as an incremental result - numerical experiments seem to compare the proposed algorithm to algortihms which are not conceived to tackle the considered problem - numerical experiments on real world data are disappointing. Indeed, the paper itself recognizes that not having knowledge about the considered domain the results obtained by the proposed algorithm can not be validated nor commented on. Therefore, I ask myself which is the value of presenting such experiments? - computational complexity is not mentioned at all - the paper structure and organization makes it not trivial to follow it - I found many typos

Questions

- which is the coputational complexity of the algorithm you proposed? - how does the proposed algrithm scales with respect to the number of variables of the multivariate time series? - could you please explain why the following works had not been taken into account? maybe they are slighlty out of the scope but still linked to the problem you tackled (Learning Continuous Time Bayesian Networks in Non-stationary Domains), (On-homogeneous dynamic bayesian networks with bayesian regularization for inferring gene regulatory networks with gradually time-varying structure) (Learning non-stationary dynamic bayesian networks) - line 104: could you please explain the meaning of setting P(V) different from 0? - line 105: the notation used for the pair of variables is not properly coherent, indeed Xj is a univariate timeseries contained in V, which has not been defined. Furthermore, what do you mean by variables here? two time serires? - line 115: what do you mean by finite maximum lag? - would you be that kind to comment explicitly and not only formally on the assumptions made about the time when nonstationarity occurs, i.e., the casual graph changes its structure? synchronous? asynchronous for each time series? could you discusse more on this aspect? what about parameters non stationarity? I guess you should consider this is excluded to happen, or not? - line 156: the definition of n is a little confusing. Indeed, n belongs to a set that is defined as a function of n again, would you be that kind to help me understand? - formula (14), the expernal p is the same probability ad the p inside brackets? - line 243, I read "continuous time series in four steps", but I found only three of them - would you also consider the structrual hamming distance to measure the performance of the proposed algorithm

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

- computational complexity of the proposed is not mentioned at all - comparison has not been performed with algorithm natively designed for semi-stationary time series data. Maybe no such algorithms exist? - results on real world data are of no help because no ground thruth is available and no domain expert comments are given to corroborate it

Reviewer ZpAB6/10 · confidence 3/52023-06-26

Summary

The paper describes a constraint-based causal discovery method for semi-stationary SCMs, where the causal structure is periodically changing over time. This type of structure is relevant in real-life scenarios where seasonal or diurnal variation is present for example. The problem is rigorously formalized and and algorithm is give to reconstruct the SCM. As a first step the algorithm identifies a superset of parents using PCMCI naively, then searches for the correct periodicity for all variables by minimizing parent set size. The author provides theoretical guarantees that the reconstructed causal graph is indeed the generating graph in the infinite sample size limit, if all assumptions hold.

Strengths

The paper describes a novel extension of PCMCI for semi-stationary SCMs, and gives theoretical guarantees. The probrem formulation is sufficiently rigorous, and the algorithm is described in sufficient details. This makes the works reproducible. The method relaxes a quite frequent assumption of causal discovery methods in the temporal domain: stationarity. Even if this new relaxed assumption also likely violated by many real life dataset, it is clearly an improvement in cases when stationarity is clearly a wrong assumption.

Weaknesses

The main limitation I see is the following: For example, if the causal effect changes by the day of the week but the samples are collected in every hour, this periodicity is hard to represent in the presented framework. Given the process inherent timescale is hourly, we cannot just subsample the data as this would induce transitive dependencies. However in the current form the method will try to fit a new causal mechanism in every hour (until the period is reached of course). It would be good to know what happens if this type of data was fed to the algorithm, as this is hard to avoid. To avoid the identification of the transitive dependencies as genuine we are motivated to work as fine timescale as available. Similar question applies for the "sudden" mechanism change. In real life most likely the sudden change will transform to a smooth change as the sampling frequency increases (for example seasonal effect in a daily scale). Most likely the effect of the Atlantic air temperature on the Central Pacific air temperature does not disappears suddenly.

Questions

l257: **Following the previous work in Huang et al. [2020], F1 score, Adjacency Precision, and Adjacency Recall are used to measure the performance** Please specify in details how you compute these metrics. Classical methods reconstruct a single causal mechanism per variable, while the ground truth has multiple. l259: **The standard error of the averaged statistics, displayed either by color filling or by error bars is usually too small to be observed.** You mean here the standard error of the mean estimate? You can use the sample variance instead. Fig 2a, In horizontal axis $\omega_{max}$ is shown. Has the code been executed with the same $\omega_{ub}$ for all cases, or the $\omega_{ub}$ was smaller for smaller $\omega_{max}$ trials? In general what hyper parameters (upper bounds, significance thresholds) were selected? What CI test was used? In the Case Study, if you have monthly data, and Central Pacific has 3 parent sets, it seems to me the causal effect is different in every month say the time partitions are {January,April, July, .. } {February,May, August, ..} {March, June, September, ...}, how you come to the conclusion that: **"the causal effect from the tropical Atlantic air temperature to the Central Pacific air temperature would disappear every quarter of a year"*** ? I am not sure I understand this point. Minor comment: I would suggest to use "discrete valued" and "continuous valued" terminology instead of "discrete time series" / "continuous time series", as it can be mistaken with discrete and continuous time. How the sampling frequency influences the reconstruction of the SCM? (See weaknesses part.) **Rebuttal**: I have read the author's rebuttal and we exchanged further comments. I still see the paper having potential interest for the audience of this conference. I had one main open question, the result of the case study. Its result is unfortunately quite implausible, therefore I cannot further increase my score.

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

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

The author points out some limitation of the method. For example the higher sample complexity of the method relative to vanilla PCMCI. It is not clear how realistic the present semi-stationarity assumption.

Reviewer YnE65/10 · confidence 3/52023-07-05

Summary

## Summary The paper addresses the problem of discovering causal relations in semi-stationary time series data, where a finite number of causal mechanisms occur sequentially and periodically across time. The authors propose a constraint-based, non-parametric algorithm called $PCMCI_\Omega$ to handle this type of time series data. The main contributions are 1. The proposed algorithm, PCMCIΩ, is an extension of the PCMCI algorithm designed for stationary time series data. It leverages the PCMCI algorithm to systematically discover the superset and use CI test to find out the correct periodicity and remove unnecessary parents. 2. The soundness of the PCMCIΩ algorithm is demonstrated through a theoretical analysis, showing that it can recover the true causal graph under specific assumptions. 3. The authors validate the effectiveness of the PCMCIΩ algorithm through experiments on both continuous and discrete synthetic time series data. Although I think the baselines can be stronger.

Strengths

Strengths of the paper: 1. Originality and significance: The paper proposes a novel algorithm, PCMCIΩ, to address the challenging problem of causal discovery in semi-stationary time series data. Fewer work have considered such settings and most of the previous method assume stationarity. This is a step towards generalising causal discovery to non-stationary time series. 2. Soundness: This work provides a theoretical guarantees on the soundness of the proposed method, and validate it with experiments.

Weaknesses

The main weakness of this paper are 1. Limited scope of experiments: Although the authors have performed experiments on synthetic continuous and discrete time series data, as well as a real-world climate dataset. However, the baseline selection is too simple. For example, VARLinGaM and DYNOTEARS are both linear models with stationarity assumptions. It is expected to outperform those two methods. I suggests including more stronger baselines, such as Granger causality baselines [1], and SCM-based method [2]. 2. The paper is hard to understand when I first read it. There are many definitions defined in section 2, without knowing why they are important. The graphical illustration helps a little bit, but still confuses me on why we have to define them. For example, I am not sure I understand why Illusory Parent sets requires a subscript. Like why we need $Pa_2(X_7^1)$ and $Pa_3(X_7^1)$. Isn't it the same as $Pa_1(X_8^1)$? Since the following theoretical analysis extensively uses these definitions, I recommend the author to be very clear on these definitions and should make them easier to understand for the readers. 3. How does your method performs when there are mismatches in the assumptions? For example, the ground-truth is stationary/non-stationary. When the lag upper bound and period upper bound are shorter than the truth. etc. More ablation study would be helpful. 4. What are the potential limitations of the proposed method? Some discussion should be added. For example, instantaneous effect can be important when aggregation happens. Computational complexity with dimensionality, etc. [1] Khanna, S., & Tan, V. Y. (2019). Economy statistical recurrent units for inferring nonlinear granger causality. arXiv preprint arXiv:1911.09879. [2] Gong, W., Jennings, J., Zhang, C., & Pawlowski, N. (2022). Rhino: Deep causal temporal relationship learning with history-dependent noise. arXiv preprint arXiv:2210.14706.

Questions

1. I am a bit confused on line 123, why minimum periodicity has the same symbol as the number of causal mechanisms in $V$? 2. In definition 2.3, eq.8, the subscript $k$ is not showing in the latter equations. What does this k means? 3. Can you make the definition of Illusory parent sets easier to understand and explain the intuition behind it? 4. For figure 1, why $Z_2^1$ is identical to $Z_1^7$? I thought $Z_2$ defines the Markov chain for second time series? Also, you use superscript in $X$ to indicate time series number, I think $Z$ should follow the same pattern to avoid confusion.

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

2 fair

Contribution

3 good

Limitations

The author does not explicitly provide a discussion on the limitations. Hence, I suggest the author to include a limitation section.

Reviewer YnE62023-08-16

Thanks for the additional experiments provided by the authors. They managed to address my concerns, and hope the revised version to have a more clear presentation. I will keep my current score.

Authorsrebuttal2023-08-16

We appreciate your feedback, and we're pleased that the experimental results, guided by your insightful suggestions, have effectively addressed your concerns. We are going to refine the presentation to enhance the clarity of definitions in the forthcoming version. Thank you once again!

Reviewer nsZ65/10 · confidence 3/52023-07-07

Summary

The authors propose a causal discovery algorithm for a class of periodic random processes called semi-stationary process. The algorithm is built upon an existing algorithm assuming stationarity. Consistency results of the algorithm is provided.

Strengths

1. Consistency results of the algorithm was shown even though the algorithm is relatively complicated. 2. The algorithm shows superior performance against the baselines.

Weaknesses

1. The motivation for generalizing PCMCI to handle semi-stationary processes is not clear, especially from a theoretical perspective. The description of PCMCI is completely moved to the supplementary. I think including a brief description and the key procedures (e.g., how the superset is estimated) is helpful for understanding why PCMCI is good fit for the considered setting. 2. As mentioned in Section F. from Appendix, the proposed algorithm does not yield good performance in finite samples until the heuristic method $turning\text{ } points$ is introduced. To support the theoretical results, it should be shown that the algorithm starts to work when the sample size is sufficiently large. typo: Definition 2.3 is not correct: $\Pi_{k}^{j}(T)$ depends on $k$, but its definition in (8) does not depend on $k$.

Questions

Since PCMCI assumes stationarity, in Algorithm 1, why PCMCI is used to estimate the superset $\widehat{SPa}(X_{t}^{j})$ for semi-stationary processes?

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

2 fair

Contribution

2 fair

Limitations

No potential negative societal impact.

Reviewer ZpAB2023-08-11

Reply to Rebuttal

I would like to thank the authors for their time to answering my (and other reviewers') questions. I am mostly satisfied with the answers and explanations provided. I only mention here the questions where I have some additional comment or follow up questions. **Q1:** Indeed, my worry was exactly the diminishing sample size per CI test. **Q3:** So I assume if a method cannot handle non-stationarity ($\hat{\Omega} =1$), the reconstructed mechanism is repeated $\Omega$ times and compared to the ground truth. Am I correct? **Q5:** I misunderstood this in the original paper but indeed it was a mistake from my side. However, now I see that indeed this finding is highly implausible. Are you aware of any assumption violation that would produce this kind of artifact ? The provided new results in the PDF demonstrate that the method also seem to have a level of robustness on datasets where the assumptions are violated.

Authorsrebuttal2023-08-11

We appreciate your insightful comments, which are truly inspiring. **A1**: We recognize that there's room for more effective sample utilization in the real-world application of our algorithm when dealing with periodic causal mechanisms that allow for the repetition of the same causal mechanism. When there are only a few instances of the same causal mechanism, such as 2, reducing the samples per CI test by half might be reasonable in certain scenarios. However, if there's a substantial number of repeated causal mechanisms, such as 200 instances, it may be more appropriate to treat this issue as a change point detection problem, which is currently a topic of our ongoing research. If the repetition falls somewhere between being small and significantly large, we don't have a definitive solution at this point, but it presents an intriguing open question worth exploring. Your insightful question is greatly appreciated, and we intend to incorporate it into the discussion of limitations within our paper. **A3**: Yes. **A6**: We deeply appreciate your valuable input on the case study. We claimed in the paper that the significance of these results is under-explored, making your comment essential. There are specific assumptions that may be violated, as follows: **Hard Mechanism Change combined with the content in Q1**: If a soft mechanism change occurs in $X_{t-1}^{cp}$, the reliability of the CI test between $X_{t-1}^{ta}$ and $X_t^{cp}$ given $X_{t-1}^{cp}$, will be impacted by the skewed distribution of $X_{t-1}^{cp}$. It is further impacted by the fact that the sample size here is 900/3=300. **No Contemporaneous Causal Effects**: There's a possibility of potential causal effects from $X^{ta}_t$ to $X^{cp}_t$ at the same time point $t$ that we're unable to capture in our analysis. Thank you for raising these crucial points; we will include these comments in our case study section. In conclusion, we wholeheartedly invite further discussions.

Reviewer ztgZ2023-08-11

Read your rebuttal

Thank you, I went through your rebuttal question by question or lets say answer by answer. I found many answers being very useful to clarify aspects that in the original submission were not that clear to me. However, I'm still not satisfied by answers to lack of baselines, I understand the author's point but I would have expected a more involved discussion on this aspect. Furthermore, if you describe experiments on real data where you do not have ground truth and you do not have knowledge about the domain I still miss how you can validate your findings. Then, also the answer to Q6 is puzzling, indeed if you consider continuous variables you always have that the variable X takes on a specific continuous value x with probability zero, so I miss the point.

Authorsrebuttal2023-08-12

Thank you for your insightful comments and valuable questions. **Baselines**: Sorry we didn't write this before due to the limited space. Here is our detailed discussion. Indeed, several baseline methods are designed to handle non-stationary temporal data, such as CD-NOD[1] and JIT-LiNGAM[2]. However, it's important to note that the outputs of these baselines are **summary graphs**, allowing them to identify the causal relationship between time series $X^j$ and $X^i$ without pinpointing the specific time lag $\tau$. In contrast, our method enables the detection of the precise variable $X^j_{t-\tau} \in X^j$ that causes the behavior in $X^i_t$, presenting a distinguishing feature. These approaches highlight distinct aspects from ours. While summary graphs offer a degree of insight, they are inherently less informative about the underlying causal structure in our specific case compared to the detailed window graphs produced by our method. To showcase our method's capacity to handle periodicity in causal mechanisms with time-lag effect, a capability lacking in other approaches, we will introduce a related baseline called "Regime-PCMCI" in our new experiment. It's worth noting that this algorithm necessitates a linear model (while our method is non-parametric) and additionally mandates precise foreknowledge of the number of causal mechanisms inherent within the time series. Remarkably, based on several experiment results we have, our algorithm's superior performance is sustained even when the accurate count of causal mechanisms is supplied to the "Regime-PCMCI" method. Given the considerable time requirement of the "Regime-PCMCI" method, we regret to say that the result plot is currently unavailable. However, we intend to incorporate the said plot into the paper later. The significance of the comparison depicted in Fig. 3(a) with the baselines we have is apparent when considering the following points: 1. Our algorithm excels in the reduction of False Positive edges, as indicated by the higher Precision, particularly when compared to the stationary algorithm employed in a semi-stationary setting. This outcome effectively validates **Lemma 3.4**, providing a solid foundation for our approach. 2. With the expansion of the sample size, a notable reduction in False Negative edges becomes evident, as evidenced by the recall rate approaching that of PCMCI, which stands out as the best-performing among the baseline methods. 3. In cases where the ground truth involves periodic causal mechanisms, the misuse of a stationary algorithm leads to a significant amount of error, emphasizing the critical importance of employing an appropriate method, such as ours, to account for such complexities. As far as we know, our method is the first to address the semi-stationary Structural Causal Model (SCM) with the capability to accommodate periodicity. We hope that our approach can establish itself as a potential reference for individuals interested in periodic causal mechanisms. **Case Study**: Thank you for the comment. We will add the following discussion in the camera-ready paper to avoid potential misunderstanding regarding our conclusion in the case study. We could not comment on whether the result of the case study is significant. We open a door for the related experts; if assumptions A1-A8 are satisfied, the stationary assumption may not hold in this real-world data, and such periodicity exists. However, if the finding is not correct from an expert's viewpoint, the following assumptions may be violated: 1. **Hard Mechanism Change** combined with limited power of CI tests: if there is a soft mechanism change in the dataset, the reliability of the CI test of two variables given their parents will be influenced by the skewed distribution of the parent variables. This effect will be exacerbated by the fact that the sample size will be shrunk by $\omega$. 2. **No Contemporaneous Causal Effects**: There's a possibility of potential causal effects from $X^{ta}_t$ to $X^{cp}_t$ that we're unable to capture in our analysis under this assumption. Our method provides a sound and robust (shown as figures in the Rebutall pdf file) algorithm for experts in various fields who are interested in validating the presence of periodicity within the causal mechanisms specific to their domain. **More details about $p(V)\neq 0$**: Suppose Y is a continuous random variable. p(Y) here is the probability density, which is always positive in the domain of Y. For example, this would be a case for any non-degenerate jointly Gaussian distribution. Thank you again and we wholeheartedly invite further discussions. [1]Huang, Biwei, et al. "Causal discovery from heterogeneous/nonstationary data." The Journal of Machine Learning Research 21.1 (2020): 3482-3534. [2]Fujiwara, Daigo, et al. "Causal Discovery for Non-stationary Non-linear Time Series Data Using Just-In-Time Modeling." 2nd Conference on Causal Learning and Reasoning. 2023.

Reviewer ztgZ2023-08-12

yes but in the rebuttal you wrote P(X) and not p(X), but now that you specified, all is clear

yes but in the rebuttal you wrote P(X) and not p(X), but now that you specified, all is clear

Authorsrebuttal2023-08-14

We sincerely appreciate your dedication and time. We wholeheartedly welcome further discussions. Should our rebuttal effectively address the concerns, we kindly hope you consider it for acceptance.

Reviewer nsZ62023-08-15

Reply to the rebuttle

Q3: It is good to see that the algorithm does not heavily rely on turning points. This addresses my major concern of the algorithm. Based on this, I would raise my score slightly. Q1: I am still quite skeptical about using PCMCI for estimating $\hat{SPa}(X_{t}^{j})$ in finite samples, since PCMCI is proposed for stationary processes rather than semi-stationary processes. The proofs of Lemma 3.1 and 3.2 assume consistent CI tests, so I think the theoretical guarantees are quite limited. But I get that it is quite challenging to obtain stronger results. Overall, I think the studied setting is interesting and method has some novelty. But since semi-stationary processes defined in this paper are not studied in the literature, it is not clear whether it is general enough to model real world periodic random processes. Therefore, experiments on real data with ground truth will make the method more convincing.

Authorsrebuttal2023-08-15

**Q3**: Thank you for the reconsideration. **Q1**: Thank you for these valuable comments. The CI tests in both PCMCI and our algorithm are assumed to be consistent given i.i.d. samples. We do not assume the consistency of CI tests with respect to semi-stationary data. Therefore any CI tests that maintain consistency with i.i.d. samples can be seamlessly integrated into our algorithm. This applies not only during the initial PCMCI phase but also in the subsequent step of our algorithm, where Momentary Conditional Independence (MCI) tests come into play. For instance, CI tests with proven consistency, such as Chi-square tests or Fisher's exact tests, can be employed in both PCMCI and our algorithm. This is feasible because, in both cases, variables are conditioned on parent sets or supersets of parent sets, ensuring the independence of samples due to the independence of exogenous noise terms. By choosing a correct $\omega$, conditioning on a superset of the parent set can yield identical samples from the same causal mechanism. An incorrect $\omega$ results in samples from a mixture distribution, as outlined in Eq.50 within the supplementary material. Hence it may introduce more dependent relationships in the mixture distribution, supported by **Lemma 3.2**. Once again, we express our heartfelt appreciation for your insightful feedback. We aspire that our algorithm serves as an innovative tool, harnessing expertise from various domains to unveil the latent periodic patterns within causal mechanisms across real-world situations.

Authorsrebuttal2023-08-16

Your commitment to reviewing our submission and sharing invaluable insights is deeply appreciated. Thank you for investing your time and effort. We also want to express our gratitude for the reconsideration.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC