acceptodds
Under review as a conference paper at ICLR 2027

SEAM: Recovering Causal Reversals Across Recurring Regimes

Abstract

Many systems switch between a few recurring states whose causal relations partly differ: some relations stay the same, while others reverse direction. From one multivariate time series with unknown state boundaries, we ask when the state changes, which causal relations change, and what an intervention would do in each state. We introduce SEAM (Shared Equations Across Mechanisms), which jointly learns the recurring states, a causal graph for each, and which structural equations the states share. Separate causal orders let an effect reverse; estimating the unchanged equations from both states improves precision. One description-length score decides between no change, a change in noise levels only, and a change in the causal mechanisms, and a resampling test reruns the entire search on surrogate series to assess the evidence for a change. We give conditions under which the decision is consistent, both for Gaussian VAR series without a change and for mechanism changes that no change in noise levels can mimic, and we show that a row-permutation variant of the test controls its level at every sample size for exchangeable observations without lags. On five-variable systems with a reversal, SEAM's error on the reversed effects is 0.016 at the larger effect size, compared with 0.406–1.000 for five regime-aware baselines, and over all 640 series it finds the changed equations with F1 0.951; on 1400 series without a change, it reports a mechanism change in 0.36% of them. On five public daily river series, SEAM recovers the seasons in all 3 four-year windows, and in 2 of them the fitted influence between the Columbia and the Willamette reverses with the season, as hydrological studies describe.

Primary area:
causal reasoning
Keywords:
causal discovery, time series, regime switching, structural causal models, mechanism change, change detection, causal reversal, integer programming

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_nXrpOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper aims to recover recurring regimes and changing causal relations from a multivariate time series with unknown boundaries. It proposes SEAM, which learns a separate causal graph for each regime while sharing unchanged structural equations. A description-length score and a resampling procedure distinguish no change, changes in noise levels, and changes in mechanisms. Theoretical analysis and experiments examine change detection, graph recovery, and intervention effects, with an application to seasonal river dynamics.

Critical strengths
  1. The method addresses the useful distinction between changing causal directions and changing noise levels. The analysis also identifies a concrete failure of sharing equations without accounting for sign changes.
  2. The experiments include several regime-aware baselines, different innovation distributions, and informative ablations. The supplementary results openly report cases where the proposed method loses accuracy or another method performs better.
Critical weaknesses

-- The error guarantees do not cover the full claim of distinguishing mechanism changes from noise changes. The implemented resampling procedure generates data from a single fitted regime, whereas unchanged mechanisms can still produce different distributions when their noise levels change. The theoretical null result concerns a stationary Gaussian process, and the finite-sample guarantee applies to a different procedure without temporal dependence. The variance-change experiment is useful, but it examines a restricted setting in which only one root variable changes its noise level. Consequently, the reliability of mechanism attribution under more general noise changes remains uncertain, despite the low false-report rate on stationary series.

-- The paper does not establish when its reported causal directions and intervention effects are uniquely determined. The theory shows when a mechanism change can be distinguished from a scale change, but this does not establish that the selected graph is the only causal explanation of the observations. The paper acknowledges observational equivalence, yet reports individual graphs and effects without indicating which conclusions remain ambiguous. This matters in the river application: the stage measurements are closely related, and the documented reversal disappears when the Astoria series is removed. Although this sensitivity is reported transparently, agreement with the seasonal pattern alone does not establish the reliability of the inferred direction.

-- The reversal experiments provide limited evidence beyond a particularly favorable graph structure. Most of the main comparisons use the same reversal pattern, with additional variables that help distinguish the regimes and orient the focal relations. Increasing the system size retains this pattern within a shared background graph. The experiments on other methods' benchmarks broaden the evaluation, but the dense-graph results also show substantially poorer graph recovery than CASTOR. The current evidence therefore leaves unclear how much of the reported advantage depends on the distinguishing variables in the chosen reversal design, especially when their influence is weak or most equations change.

-- The benefit of sharing is not fully established when the regimes must also be learned. The sharing ablation supplies the true regime labels, and the efficiency result assumes the correct labels and structural description. These results support improved estimation once the groups are known. However, sharing also influences the score used to learn those groups, so incorrect assignments can change which equations are pooled. The existing comparison does not determine whether sharing improves the complete procedure or propagates segmentation errors. This distinction is particularly relevant because weaker changes already cause missed regimes and poorer localization.

Questions for authors

See weaknesses.

Reviewer_s1GROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper addresses causal discovery in nonstationary time series where a system alternates between recurring regimes and the causal structure may differ across regimes. Its contribution is SEAM, which combines regime detection with regime specific DAG learning, thereby allowing causal directions to reverse across regimes. The method attempts to exploit invariance by pooling structural equations judged to be shared across regimes, distinguishes structural changes from innovation-scale changes, and provides a resampling-based procedure for deciding whether a change should be reported. The paper seems to contain theoretical analysis and extensive experiments, including an application to river data. The main ideas are interesting and the empirical results appear strong. However, the presentation in the main paper is difficult to follow: many definitions, fitting details, restrictions, and qualifications required to understand what the method actually does are deferred to the appendices.

Critical strengths

In this method, the distinction between mechanism changes and changes in innovation scale is practically meaningful and is accompanied by an explicit no change testing procedure. Also pooling unchanged structural equations across regimes is an interesting idea that can improve statistical efficiency, supported by both theoretical and empirical results. The paper also contains several useful theoretical analyses of the limitations of competing model classes and of SEAM itself, particularly the single order lower bound, the reflection/sign-alignment issue, and the distinction between detecting a distributional change and identifying it as a mechanism change.

Critical weaknesses

The presentation of the method in the main text is not self contained. Several quantities, modelling restrictions, and algorithmic steps are introduced without enough information to determine exactly what is being fitted or how the procedure operates. See quesitons below for some examples. The necessary details are scattered across the appendices, and in several places the main text does not provide a sufficiently specific reference to where those details can be found. For example, understanding the precise equation classes and their fitting, the enforcement of DAG constraints, the restrictions on regime lengths and switches, the construction and validity of the surrogate test, and the complete decision procedure requires consulting Appendices A and B. In fact, the complete procedure only becomes clear from Algorithm 1 in Appendix B. It is recommended to make Sections 2–3 more self contained and provide clearly forward references wherever technical details are deferred to the appendix.

Questions for authors

P3, line 109, "...the same-time graph G_r, a directed acyclic graph (DAG).": From eq (1), no acyclicity can be read out. This should be clarified. P3, lines 111-112, "...in runs of at least m=5d(1+p) rows with at most eight switches...": Are the minimum run length m=5d(1+p) and the maximum of eight switches modelling assumptions, or just computational restrictions imposed by the search algorithm? Please justify these particular choices and discuss their effect on the method. P3, line 123, "\theta_{i\to j}^{(r)}=[(I-B_r)^{-1}]_{ij}": If DAG is clarified in the first place, the invertibility is clearer.

P3, lines 134-140, "What can change. We compare the two regimes one structural equation at a time...": This paragraph is difficult to follow. Please rewrite it to define more clearly what constitutes an equation and to distinguish among shared, graded, and regime-specific equations.

P3, line 135, "...its coefficients...": Which coefficients are being referred to here: the same-time coefficients, the lagged coefficients, or both? P3, lines 135-136, "...its innovation variance. It is shared if it is the same in both regimes, and it is then estimated once, from the rows of both.": What specifically does "it" refer to in these sentences? Also, what is meant by "the rows of both"? Please specify the data or design-matrix rows being pooled across the two regimes. These may be answered later but at least show where to find them.

P3, lines 145-146, "Let L_\star be the expected per-row Gaussian log-likelihood of the true model at the true labels, and L_1\leq L_0\leq L_\star its best values within \mathcal{H}1 and \mathcal{H}0.": With respect to which random variables is the expectation defining L\star taken? Please provide a precise population-level definition of L\star, L_0, and L_1. The gap \Delta=L_\star-L_1, \kappa=L_\star-L_0 is assumed to be nonnegative?

P4, lines 175-178, "For labels z with J(z) switches and a description M with k(M) changed equations...": Please define J(z) and k(M) explicitly before using them in the score. In particular, how is a "changed equation" counted when an equation is graded or shared only up to the gauge?

1 comment

Thanks for the responses. My score remains the same.

Reviewer_NFt7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2(+1)
Soundness
3(+1)
Read review
Paper summary

This paper proposes SEAM, a method that takes one multivariate time series switching between two recurring regimes at unknown times and learns, all together, the regime labels, a separate causal DAG for each regime (so an edge can reverse direction), and which structural equations stay the same across regimes. A single description-length score decides between no change, a change in noise levels only, and a change in the causal mechanisms, and a resampling test that reruns the whole search on surrogate series judges whether a change actually happened. The authors prove consistency results for this decision and show that any method with a single causal order has at least 1/2 normalized error on reversed effects. They report lower error than five regime-aware baselines on synthetic linear systems, and on Columbia–Willamette river data they recover the seasons and a seasonal reversal in two of three windows.

Critical strengths
  • The problem is well motivated. Recurring regimes with different causal orders are an important regime of real-world applications that need proper and further research.

  • The sign gauge result is new and very interesting. Theorem 3 shows that sharing equations in data's coordinates can choose the wrong graph. The results seem very promising.

  • The result seems to have a large improvement over baseline.

Critical weaknesses
  • The scope of the paper is very narrow and the novelty claim is overstated. “What these methods share is that none of them tests whether a change occurred” is contradicted by Rabel & Runge (2025), whose Lemma 5.10 is a homogeneity test with type-I error control.

  • Some relevant papers are not cited, notably Balsells-Rodas et al. (ICML 2024), whose identifiability results for switching dynamical systems the FlowMSM baseline builds on, and Balsells-Rodas et al. (ICML 2025), which learns state-dependent causal graphs from a single time series with hidden states, a setting close to SEAM’s.

  • The paper has an unusually long Appendix section containing some of the most of the important theorems and findings that are not on the main pages. The main text should make the practical scope of the guarantees more explicit, particularly the initialization and surrogate conditions in Appendix A.9 and the lag-support restrictions affecting the river analysis. A condensed version of Table 2 would help.

  • The finite-sample guarantee (Theorem 11) is for a row-permutation test that the paper never runs, and it holds only for data without lags. The implemented surrogate scheme is called Hansen’s bootstrap but is a variant of it, so Hansen’s results don’t carry over without further justification.

  • The guarantee that SEAM detects real changes (Theorem 1(ii)) assumes the search already starts near the right answer, which the paper never shows for its own initializations, and it proves that the significance test behaves only in special cases, the main one of which excludes the river analysis. The guarantee against false alarms (Theorem 1(i)) only holds when the true model is among those SEAM searches, and whenever d − 1 + dp > 16, as in the river data, SEAM leaves out any model where a variable depends on more than two past values.

[On the Identifiability of Switching Dynamical Systems Carles Balsells-Rodas, Yixin Wang, Yingzhen Li Proceedings of the 41st International Conference on Machine Learning, PMLR 235:2639-2672, 2024.]

[Causal Discovery from Conditionally Stationary Time Series Carles Balsells-Rodas, Xavier Sumba, Tanmayee Narendra, Ruibo Tu, Gabriele Schweikert, Hedvig Kjellstrom, Yingzhen Li Proceedings of the 42nd International Conference on Machine Learning, PMLR 267:2715-2741, 2025.]

Questions for authors
  • How does SEAM’s test differ from Rabel & Runge’s link-wise homogeneity test (Lemma 5.10) and from SPACETIME’s mechanism-equality tests? Would you narrow the claim to “a calibrated test over the whole search for mechanism and orientation changes”?

  • How much of the detected mechanism change comes from the anchor edges (c→b, d→a) rather than the reversal itself?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.