acceptodds
Under review as a conference paper at ICLR 2027

ITSY: Causal Discovery From Irregular Time-Series Data

Abstract

Structural causal models for time series recover contemporaneous and lagged effects, but most methods require complete observation windows and become misspecified when samples are missing. We introduce ITSY, the first continuous-optimization method for causal discovery from irregular time series under a linear structural vector autoregressive model. ITSY reformulates the structural equation so that prediction uses the nearest available history rather than the possibly missing current slice, and jointly imputes missing values while learning both graphs. A weighted reconstruction objective corrects the noise transformation induced by this reformulation. Across synthetic regimes varying missingness, scale, graph density, and noise, and on a semi-synthetic fMRI benchmark, ITSY consistently improves graph recovery over representative SCM-based baselines. The results establish a focused solution for irregular linear first-order dynamics and clarify the assumptions required for nonlinear or higher-order extensions.

Primary area:
causal reasoning
Keywords:
temporal causal discovery

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.60 over 5 reviews, above 10% of the others.
Reviewer_8aamOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a time series causal discovery method for linear first-order dynamics where samples are missing completely at random to recover lagged and instantaneous causal relationships. The core idea of ITSY, the proposed method, is to combine the existing algorithm DYNOTEARS with an imputation mechanism via a reconstruction loss that are both optimised end-to-end. ITSY is compared to existing methods on synthetic and semi-synthetic data.

While the problem itself (causal discovery from irregular time series) is important and worth pursuing, the proposed method is very incremental and lacks strong guarantees. This, combined with critical experimental issues, lead me to recommend a clear rejection.

Critical strengths

Problem. Time series causal discovery from irregular data is an important and understudied problem worth pursuing to improve the performance of the methods on real world data which may involve irregular samples.

Motivation. The main motivation in Section 4 and Table 1 shows importantly the core issue, which is how data irregularities affect standard causal discovery approaches.

Critical weaknesses

This is a simple and largely unsupported idea. The proposed method is essentially DYNOTEARS plus model-based imputation. The theory behind it uses existing results and lacks guarantees that ITSY recovers the true graph under missingness.

The experimental work has critical gaps. Some simple alternatives as baselines are missing (impute then DYNOTEARS; drop missing indexes and run plain DYNOTEARS). Synthetic data seems to be not standardised, which is a critical design flaw as it may favour NOTEARS-like methods (see [1]), including DYNOTEARS. Figure 3 results do not show error bars despite 50 random runs performed. It is also unclear how irregular data was provided to the baselines (padded?), and the choice of this on methods’ performance.

References

[1] Reisach, A., Seiler, C., & Weichwald, S. (2021). Beware of the simulated dag! causal discovery benchmarks may be easy to game. Advances in Neural Information Processing Systems, 34, 27772-27784.

Questions for authors
  1. Can you provide guarantees that ITSY recovers a true causal graph under irregular samples?
  2. Exactly what input the baselines received? Was it zero-padded data? Or something else? Does the choice affect how baselines perform?
  3. Does ITSY take advantage of non-standardised data? Will standardising the data harm its performance?
  4. How does ITSY compare to naïve baselines, such as impute+DYNOTEARS, or DYNOTEARS on data with dropped missing indexes?
Reviewer_QJ3KOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The authors introduce ITSY, a continuous optimization method for causal discovery that learns a contemporaneous and lagged adjacency matrix from SVAR time series with potentially missing values. Their methodology is based on DYNOTEARS. It formulates an objective that minimizes the MSE between the observed values and the ones that are computed via the approximated graph and imputed missing values. The objective contains two regularizers: one promoting sparsity in the window graph and the other promoting acyclicity of the contemporaneous adjacency. The methodology is compared against existing baselines on causal discovery in a few synthetic scenarios and a semi-synthetic fmri dataset, showing improved performance over the baselines.

Critical strengths

The method is well motivated and useful in practical scenarios, where the dataset contains missing values. The methodology adopts a linear SVAR time series model including contemporaneous dependencies (compared to Granger methods that do not).

The methodology is easy to understand and intuitive. It is formulated via a constraint that can be efficiently implemented and perform sample imputation and graph learning in one go by using a masking matrix.

Critical weaknesses

The paper has weaknesses in novelty and experiments. Regarding novelty, even though the idea of introducing a mask and coupling graph learning and time-series imputation with a single objective is interesting, it is still somewhat incremental with regard to DYNOTEARS, in the sense that it gives the objective of DYNOTEARS as special case if there were no missing values. So the objective is heavily based on it.

The method itself only applies to SVARs with 1 lag, which is stated by the authors as limitation, but it is a significant one.

Secondly, the experiments are limited in my opinion. The authors generate synthetic data with very small graphs of 5-35 nodes and with only 1 outgoing edge per node. In essence these graphs are trees. It should be stated whether the methodology is applicable in that special case and change the scope of the paper.

The SHD metric in some cases is higher than half of the number of total edges. This shows failure and should be stated by the authors. For this purpose the SHD divided by the number of total edges could be used (normalized SHD).

Also, in the comparison with baselines, It is stated in the paper that default hyperparameters have been used. A proper comparison should finetune the hyperparameters of each method to perform well on this experimental setup. It is very likely that the results underestimate the baselines. The reader is also not aware of how DYNOTEARS (or any other methodology working on full available data) is executed. Did you impute the values at first, and then run DYNOTEARS?

I would also suggest in the synthetic experiments to compare against CUTS, CUTS+ in a scenario where only lagged edges are available.

Questions for authors

How did you compare against methods that require to impute first, like DYNOTEARS? Did you use zeros on the missing values? How did you impute the missing values first?

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

This paper proposes a continuous optimization framework that jointly imputes missing observations and learns contemporaneous and lagged causal structures from irregularly observed time series under a linear first-order SVAR model.

Critical strengths
  1. The paper addresses an important problem: recovering contemporaneous and lagged causal structures from irregularly observed time-series data with missing observations.

  2. The motivation is clearly presented, with a well-explained discussion.

Critical weaknesses
  1. Limited scope: Although the authors acknowledge this limitation, the study is restricted to stationary, causally sufficient, linear first-order SVARs under MCAR, which may limit its practical applicability. In particular, the restriction to first-order temporal dynamics is a significant limitation.

  2. Insufficient discussion of methodological novelty: The proposed method shares substantial similarities with DYNOTEARS, including the underlying SVAR formulation, causal graph optimization framework, and reduced-form transition transformation. The manuscript would benefit from a more explicit discussion of these similarities and a clearer distinction between the components inherited from DYNOTEARS and the novel methodological contributions of ITSY.

  3. Limited experimental validation: The empirical evaluation relies primarily on synthetic datasets and a single semi-synthetic fMRI benchmark. While these experiments provide evidence under controlled settings, they offer limited validation of the method's practical applicability.

Questions for authors
  1. The proposed method recursively propagates the last available state through the learned transition operator, as formulated in Equation 5, X^(t)=Pθ^iX(t−i).\hat{X}(t) = P_{\hat{\theta}}^{i} X(t-i). assuming the same transition matrix across the missing interval. Although Section 6.2 evaluates different overall random missing rates, it does not explicitly investigate the effect of consecutive missing-interval length on causal graph recovery. Since repeated application of the transition operator may accumulate prediction errors and introduce numerical instability, this needs dedicated discussion or separate reporting of the rate of consecutive missing intervals to help readers understanding. Ideally the authors may need to consider an additional experiment that systematically varies the length of consecutive missing blocks while controlling the overall missing rate? This would provide more direct empirical evidence of the method's robustness to different missingness patterns.

  2. Methodological novelty relative to DYNOTEARS: From my understanding, the reduced-form SVAR transformation introduced in Section 5.1, Xt=(I−A)−1BXt−1+(I−A)−1ϵt,X_t=(I-A)^{-1}BX_{t-1}+(I-A)^{-1}\epsilon_t, is mathematically equivalent, up to notation and matrix conventions, to the formulation already derived in Section 2.3 (Alternative Formulations) of DYNOTEARS [1], when restricted to a first-order autoregressive model. Furthermore, several components of the proposed framework appear closely related to DYNOTEARS, suggesting that the main methodological contribution is the extension to missing-data scenarios under first-order SVAR assumptions. Could the authors provide a more detailed comparison with DYNOTEARS, explicitly distinguishing the existing components from the novel contributions of ITSY? A dedicated discussion would help clarify the methodological novelty and significance of the proposed approach.

  3. Additional real-world experiments: While this may partly due to the restricted scope of the proposed method—stationary, causally sufficient, linear first-order SVARs under MCAR—the current experimental evaluation provides limited evidence of its practical applicability. The experiments rely primarily on synthetic datasets and a single semi-synthetic benchmark. Considering other cited baseline methods such as [1,2] include evaluations on real-world time-series datasets, such experiments would help establish whether the reported improvements generalize beyond data generated under the assumptions of the proposed model.

[1] DYNOTEARS: Structure Learning from Time-Series Data

[2] Rhino: Deep causal temporal relationship learning with history-dependent noise.

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

This paper introduces ITSY, which learns contemporaneous and lagged causal graphs from time series with missing observations. It recursively completes missing states using the transition operator and optimizes the objective, which consists of a weighted reconstruction loss and sparsity and acyclicity penalty terms. Experiments consider synthetic data and a semi-synthetic fMRI benchmark.

Critical strengths
  • The method provides a compact and intuitive way to combine missing-data imputation and graph learning. Using the masks, observed values are retained and missing values are generated using the same parameters that define the causal graphs, avoiding separate statistical imputation which might mislead causal signals.
  • Empirical verification is concrete. It covers several different settings, and provides handful of implementation details.
  • This paper clearly acknowledges its scope and limitations on its assumptions and interpretation of experiment results.
Critical weaknesses

(1) Across missing intervals, prediction errors accumulate past innovations, and multiplying it by I−AI-A doesn’t generally remove the correlation, as is also acknowledged in Appendix B. The scope of the whitening justification in Section 5.4 should therefore be clarified.

(2) Related, theoretical support for structural recovery seems limited. Theoretical analysis in the paper provides a Gauss-Markov argument, but does not establish that the proposed objective favors the correct causal graphs under missingness.

(3) The source of the performance gains is unclear. It is difficult to see how much improvement comes from structural weighting or differentiable completion.

Questions for authors

(1) Regarding Weaknesses 1 and 2, even under linear SVAR and MCAR observations, it appears that there could be a case of the proposed objective assigning a lower expected loss to a model with a spurious contemporaneous edge than to the true graph. Particularly, accumulated past innovations can introduce residual correlation that a false contemporaneous edge absorbs. Could the authors address this possibility and clarify the conditions under which the objective supports correct structural recovery?

(2) How does each baseline handle missing observations? Specifically, I believe setting \tau_\min = 1 for PCMCI+ excludes contemporaneous links. How are its outputs converted for contemporaneous evaluation?

(3) For the augmented Lagrangian, what is the criterion or factor used to control ρ\rho, tolerance for h(A)h(A) and final acyclicity residuals?

(4) Regarding Weakness 3, could the authors provide ablations comparing Eq (7) with Eq (13), and allowing versus stopping gradients through completion?

Reviewer_4tnMOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ITSY (Irregular Time-Series Causal DiscoverY), a continuous-optimization framework for causal structure learning from irregularly observed multivariate time-series data. The approach addresses an important limitation of existing structural causal model (SCM)-based methods, which generally assume complete observations when estimating contemporaneous and time-lagged causal relationships. ITSY reformulates a linear first-order Structural Vector Autoregressive (SVAR) model to predict missing observations using a learned transition operator. It integrates recursive, mechanism-aware imputation with causal graph learning through a differentiable optimization framework incorporating sparsity regularization and an acyclicity constraint. The authors further introduce a weighted reconstruction objective intended to compensate for the transformed innovation covariance arising from the SVAR reformulation.

The proposed method is evaluated using synthetic datasets under varying missingness rates, dimensionalities, graph densities, and noise conditions, together with a semi-synthetic fMRI benchmark. Comparisons with DYNOTEARS, NTS-NOTEARS, PCMCI+, Rhino, and CUTS+ demonstrate improvements in graph-recovery F1 scores across the evaluated incomplete-data settings, although competing methods achieve lower Structural Hamming Distance (SHD) in some cases. Overall, the work addresses a relevant problem and presents promising empirical results, but its theoretical justification, methodological novelty, and applicability beyond restrictive modeling assumptions require further scrutiny.

Critical strengths
  1. The paper addresses an important challenge in temporal causal discovery: recovering contemporaneous and lagged causal relationships from incomplete observations. Unlike conventional approaches that separate missing-value imputation from causal structure estimation, ITSY integrates both processes through a shared structural transition mechanism. This coupling is technically meaningful because independently imputed observations may distort causal dependencies. The proposed formulation provides a coherent approach to preserving temporal structural information while learning from incomplete data, making the contribution relevant to causal representation learning and time-series analysis.

  2. The experimental evaluation considers different missingness rates, sequence lengths, variable dimensions, graph structures, and noise distributions. The reported results generally demonstrate improved graph recovery compared with representative SCM-based baselines, particularly under substantial missingness. The semi-synthetic fMRI experiments provide additional evidence of effectiveness beyond the primary synthetic generator. These findings support the potential value of mechanism-aware imputation, although the experimental scope remains limited.

Critical weaknesses
  1. Insufficient theoretical justification of causal identifiability and the proposed weighted reconstruction objective.

The paper provides a mathematically motivated optimization framework, but the theoretical analysis does not establish sufficient conditions for identifiable and consistent recovery of contemporaneous and lagged causal graphs under irregular observations. The principal theoretical results rely mainly on established DAG invertibility properties and the Gauss–Markov theorem, rather than providing new recovery guarantees for the proposed estimator. More importantly, the weighted reconstruction objective in Equation (13) is described as an exact correction of the transformed noise structure, whereas Appendix B acknowledges that the correction does not fully recover the accumulated innovation terms across missing intervals. Consequently, the statistical validity of the proposed weighting under incomplete observations remains insufficiently established. This limitation is consequential because the claimed methodological advantage depends on correctly separating structural causal relationships from errors introduced by recursive imputation.

  1. Limited methodological novelty and insufficient experimental isolation of the proposed contributions.

The proposed framework largely combines established components, including linear SVAR reformulation, recursive state estimation, differentiable DAG optimization, sparsity regularization, and weighted least-squares reconstruction. Although their integration is potentially useful, the manuscript does not sufficiently demonstrate which components constitute substantive methodological advances over existing approaches. In particular, the experimental evaluation lacks systematic ablation studies isolating the contributions of mechanism-aware imputation, weighted reconstruction, and joint optimization. Furthermore, the evaluation is restricted primarily to stationary linear first-order SVARs under Missing Completely at Random (MCAR) assumptions, without examining informative missingness, latent confounding, or nonlinear causal mechanisms. The omission of a direct comparison with the closely related ReTimeCausal method, discussed in the related work, further limits the novelty assessment. These limitations weaken the evidence that the observed improvements arise from the proposed innovations rather than from the selected modeling assumptions or implementation choices.

Questions for authors

Q1: Clarification of theoretical validity and identifiability (High Priority):

Could the authors clarify the assumptions under which the proposed objective uniquely identifies the contemporaneous and lagged causal graphs from incomplete observations? Specifically, how is the claim of exact noise correction in Equation (13) reconciled with the discussion in Appendix B acknowledging accumulated innovations over missing intervals? Please distinguish the mathematical properties established by the current analysis from those that remain heuristic or empirically supported.

Q2: Clarification of methodological novelty and experimental evidence (High Priority):

Could the authors clarify the specific methodological contributions of ITSY beyond existing SVAR-based causal discovery, differentiable DAG optimization, and recursive imputation? In particular, do the existing experiments or implementation analyses provide evidence isolating the effects of mechanism-aware imputation and the weighted reconstruction objective? Please also explain why ReTimeCausal was not included as a direct comparison and clarify how this omission affects the assessment of novelty and performance.

These questions seek clarification of the submitted methodology and existing evidence; additional experiments are not required to address them during rebuttal.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.