acceptodds
Withdrawn from ICLR 2027

Regime-Specific Causal Discovery from a Variance-based Markov-Switching-Model

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Temporal causal discovery typically assumes that a single causal structure and set of mechanisms govern an entire time series. In many systems, however, the process repeatedly switches among unobserved regimes, and different regimes may have different causal mechanisms and lagged parent sets. Pooling observations across such regimes may therefore obscure genuine causal relations. We study the recovery of regime-specific lagged causal structure when the regimes themselves are latent and recurrent. We propose the Variance-based Markov-Switching-Model (VB-MSM), in which a shared latent Markov state selects regime-specific nonlinear conditional mechanisms. We measure the contribution of each candidate parent through its complete edge function, i.e., an edge is considered active when its contribution to the regime-conditional mean varies with its input. Thus our method is invariant of the specific parameterization. We show that removing a true parent changes the regime-conditional mean under common causal assumptions. VB-MSM combines nonlinear mechanism learning with exact forward–backward inference over the finite-state regime process in an EM-style update. On nonlinear switching benchmarks, it accurately recovers regime-specific graphs, latent regimes, and intervention effects. Comparisons with baseline methods like iMSM and FlowMSM+RHINO show its strengths across different mechanisms. Experiments with added noise also retain strong recovery at native and reduced sample lengths.

Primary area:
causal reasoning
Keywords:
temporal causal discovery, nonstationary time series, latent regimes, Markov-switching models, nonlinear structural causal models, sparse graph learning, amortized inference

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_fwbuOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a variance-based Markov-switching model for multi-regime temporal causal discovery, where a shared latent Markov state selects regime-specific causal models. Specifically, VBMSM adopts an additive structural equation framework and learns independent edge functions for each candidate edge to quantify its contribution. The core insight of VBMSM is that edge functions corresponding to true causal edges exhibit non-zero variance. Building on this observation, VBMSM introduce a novel penalty term that enables the model to jointly infer causal regimes and regime-specific causal relationships.

Critical strengths
  1. This paper correctly identifies the positive-variance property of edge functions under additive structural equation models. It further innovatively proposes a variance-based penalty term over edge functions to infer causal regime and regime-specific causal relationships.
  2. VBMSM has a reasonable model design. The introduced additive edge function adhere to the assumptions of the additive structural equation models and facilitate the model to independently quantify the contribution of candidate causal edges.
  3. This paper is clearly written. It elaborates on model parameters, experimental settings, and releases the algorithm codes, leading to strong reproducible.
Critical weaknesses
  1. The motivation of this paper is insufficiently clear. Multi-regime temporal causal discovery and Markov-switching models have both been extensively investigated in prior literature. While the authors introduce a variance-based penalty term, it remains unclear what unique advantages or distinguishing properties this penalty brings to VBMSM.
  2. VBMSM only accounts for lagged causal relationships, which constitutes a notable limitation of the method.
  3. The paper proves that edge functions corresponding to true causal edges must exhibit positive variance. Nevertheless, the penalty term in Equation (13) sums the variances across all candidate edge functions. There lacks a rigorous theoretical guarantee that minimizing this objective correctly recovers causal relationships and causal regimes. In my view, the penalty in Eq. (13) may merely serve as an alternative to standard sparse regularization. Worse still, it imposes extra constraints such as the additive model assumption and appears more sensitive to noise.
  4. The scalability of the proposed algorithm is not validated. The experiments are limited to datasets with only 8 variables and first-order lags, with no results reported for larger-scale systems or time series with longer lag orders.
  5. The experimental baselines are incomplete. Existing methods targeting multi-regime time-series causal discovery, such as FANTOM and RPCMCI, are omitted from comparisons.
  6. The benefits of the proposed variance-based penalty are not substantiated. The ablation study only includes the case with \lambda=0, and VBMSM achieves only marginal performance gains. As noted above, the variance-based penalty appears to be merely a substitute for sparse regularization. The authors should add a baseline where the variance penalty is replaced by standard sparse regularization and report corresponding experimental results.
Questions for authors
  1. What advantages or distinguishing properties does the variance-based penalty offer for multi-regime temporal causal discovery?
  2. How scalable is VBMSM when applied to higher-dimensional data?
  3. Is there rigorous theory proving that minimizing the variance-based penalty enables the model to recover causal regimes and regime-specific causal relationships?
Reviewer_TbfBOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper tackles regime specific causal discovery using a variance based Markov Switching model. It proposes one theorem with a learning algorithm.

Critical strengths

The problem considered is interesting.

Critical weaknesses

The paper is written very poorly, and it is hard to understand what the authors want to show and present to readers. The paper lacks several very important references: Spacetime [1], J(oint)-PCMCI+ [2], Fantom [3], Lin [4] and PCMCI omega [5]. I believe these papers should be cited and also compared against. Fantom takes into account instantaneous links, handles non-Gaussian noise, and also supports heteroscedastic noise. PCMCI omega tackles the same problem as the authors, and citing it would give readers a better view of the state-of-the-art models.

The authors do not properly discuss what has been done in this area, where they stand, what they add, and what the limitations of this paper are.

The motivation of the paper is not clear. Why do we need a variance-based Markov switching model instead of Spacetime, Lin, CASTOR, Fantom, PCMCI omega, or CDNOD?

Limiting the model to Gaussian additive noise is not stated clearly and feels implicit in the paper. Excluding instantaneous links is a huge limitation and should be discussed and motivated.

Notation and mathematical symbols are not clearly introduced and are used in a confusing way: the authors use non-bold uppercase letters for scalars, vectors, and matrices alike. Using a regular lowercase font for scalars (xx), bold lowercase for vectors (x\boldsymbol{x}), and bold uppercase for matrices (X\boldsymbol{X}) would make the paper easier to read.

[1] Mameche, Sarah, et al. "Spacetime: Causal discovery from non-stationary time series." Proceedings of the AAAI conference on artificial intelligence. Vol. 39. No. 18. 2025.

[2] Günther, Wiebke, Urmi Ninad, and Jakob Runge. "Causal discovery for time series from multiple datasets with latent contexts." Uncertainty in Artificial Intelligence. PMLR, 2023.

[3] Rahmani, Abdellah, and Pascal Frossard. "Flow based approach for Dynamic Temporal Causal models with non-Gaussian or Heteroscedastic Noises." Advances in Neural Information Processing Systems 38 (2026): 8300-8349.

[4] Liu, Chenxi, and Kun Kuang. "Causal structure learning for latent intervened non-stationary data." International Conference on Machine Learning. PMLR, 2023.

[5] Gao, Shanyun, et al. "Causal discovery in semi-stationary time series." Advances in Neural Information Processing Systems 36 (2023): 46624-46657.

Questions for authors

I believe the paper needs extensive work, and I suggest the following experiments:

A fair comparison with CASTOR and SpaceTime in their settings, where you introduce instantaneous links and also use only one MTS with regimes varying from 300 to 500 timestamps, to see how your model performs with few data points. Also scale to 40 or 50 nodes: Eq. (6) shows that you use an estimate of the function f for every edge e in every regime k, so I think a scalability experiment with more than 5 regimes, more than 8 lags and 50 nodes is important.

Running time and complexity will add more clarity to the paper.

A fair comparison with Fantom in non-Gaussian and heteroscedastic settings (one time series with multiple regimes) is also important.

A fair comparison with LIN in non-Gaussian with intervention setting is also important.

Experiments in the data standardization case could show the importance of your method, even though this problem occurs when you have acyclicity constraints, which you avoid by assuming only time-lagged links.

Questions:

Why only time-lagged links? Why only first-order Markov? Why Gaussian additive noise? Why would I use your model in the future instead of RPCMCI, CASTOR, FANTOM, SpaceTime, PCMCI omega, or LIN?

For example, the introduction presents your paper as if you were defining and introducing this research direction, because you do not state what others have done, how you are different, or why you want to tackle this problem from this new perspective. Also, why did you make such strong assumptions as only time-lagged links, Gaussian additive noise, and homoscedastic noise? Having limiting assumptions is not the real reason for rejection; not motivating them is.

I also believe the paper requires another round of writing, where a notation paragraph and the definition of any unusual concepts are important.

Reviewer_Lq9eOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes VB-MSM, a nonlinear Markov-switching framework for recovering regime-specific lagged causal graphs when the underlying regimes are latent and recurrent. The method models regime-dependent conditional means using additive edge-specific neural functions, performs latent-state inference via forward–backward recursions, and uses the variance of each fitted edge function for structural regularization and graph recovery. The theoretical analysis connects conditional-mean dependence, deletion-based support, and functional variance under explicit identifiability, additivity, and support assumptions.

Experiments on synthetic switching systems and empirical-innovation benchmarks evaluate graph recovery, regime identification, forecasting, and intervention-effect estimation. The results demonstrate strong performance under several nonlinear mechanisms, with comparisons against iMSM, FlowMSM+RHINO, and other baselines.

Critical strengths

Interesting problem formulation. The paper addresses regime-specific causal discovery with latent and recurrent regimes, which is a relevant and challenging problem.

Interpretable theoretical framework. The variance-based functional support criterion provides a parameterization-invariant way to identify causal dependencies under the stated assumptions.

Comprehensive experimental evaluation. The paper includes comparisons with existing methods, ablation studies, and intervention-effect evaluations, demonstrating strong performance on several synthetic benchmarks.

Critical weaknesses

Limited methodological novelty. The proposed framework largely combines existing Markov-switching models, additive neural mechanisms, and functional regularization. Its methodological contribution beyond prior work is not sufficiently clear.

Marginal benefit of the proposed regularization. The zero-penalty baseline achieves nearly identical graph recovery and regime identification performance (Table 2), raising questions about the necessity of the proposed regularization.

Strong dependence on initialization. The nonlinear feature-HMM initializer already achieves very high graph recovery accuracy. The experiments do not adequately establish whether the method remains effective with less informative initialization.

Sensitivity to graph thresholding. Appendix H shows that different readout and calibration choices lead to substantially different Hard F1 scores despite similar AUPRC, making the reported graph recovery advantages less convincing.

Restrictive theoretical assumptions. The causal interpretation relies on regime identifiability, causal sufficiency, additive mechanisms, and product-support assumptions, limiting the applicability of the theoretical guarantees.

Limited experimental generality. Most experiments consider small synthetic systems with known regime counts and lag order one. Evaluations with unknown regime counts, higher-order dependencies, and non-additive interactions are missing.

Poor robustness to outliers and lack of real-world validation. Under 5% gross shocks, VB-MSM's graph recovery and regime identification performance deteriorate dramatically (Table 14). The absence of validation on real-world causal discovery tasks further limits the practical significance of the results.

Questions for authors

Methodological novelty: What are the key methodological advantages of VB-MSM over existing nonlinear Markov-switching models beyond the proposed variance-based support criterion?

Effectiveness of regularization: Given the nearly identical performance of VB-MSM and its zero-penalty variant in Table 2, what is the main empirical benefit of the proposed functional regularization?

Initialization sensitivity: How sensitive is VB-MSM to the quality of the nonlinear feature-HMM initialization? Does the method remain effective under inaccurate initial regime assignments?

Robustness: What explains the substantial performance degradation under 5% gross shocks in Table 14? Which component of VB-MSM is primarily responsible for this failure?

Generalizability: How would the proposed method perform when the number of regimes is unknown, the lag order is greater than one, or the underlying causal mechanisms contain non-additive interactions?

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

Related papers

Discussion (0)

Sign in to comment.