acceptodds
Under review as a conference paper at ICLR 2027

DistCUTS: Distribution-Aware Causal Discovery from Incomplete Multivariate Time Series

Abstract

Discovering causal relationships from incomplete multivariate time series is challenging because missing observations obscure temporal dependencies. Existing approaches typically address this challenge by alternating between missing-value reconstruction and causal graph learning. However, their structure-learning objectives primarily assess whether a source variable’s history improves point predictions of a target’s future values. As a result, they may overlook causal edges whose effects manifest primarily in the target’s conditional variability rather than its conditional mean, yielding incomplete causal structures. Omitting these edges can further impair missing-value reconstruction by excluding relevant causal information, with the resulting errors propagating into subsequent graph estimation. To address this limitation, we propose DistCUTS, a distribution-aware framework that complements point prediction with history-conditioned residual distribution matching to capture both mean-driven and scale-driven dependencies. Specifically, DistCUTS estimates history-dependent residual scales from graph-constrained source histories and uses these scales to parameterize conditional residual distributions. A maximum mean discrepancy (MMD) objective compares samples from these distributions with the corresponding observed residuals, encouraging the model to capture residual variability supported by the observations. As auxiliary components, we demean observed residuals to reduce sensitivity to predicted-mean offsets and use the learned residual scales to guide missing-value reconstruction. Experiments on synthetic and real-world datasets demonstrate that DistCUTS outperforms the evaluated baselines.

Primary area:
causal reasoning
Keywords:
Distribution-aware causal discovery, Incomplete time series, Conditional residual distribution matching, Maximum mean discrepancy

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_EAF5OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DistCUTS, a distribution-aware causal discovery method designed for incomplete multivariate time series. DistCUTS use graph-constrained recurrent prediction framework as main body and predict both conditional mean and the residual scale. To identify distribution-dependent causal edges, DistCUTS propose a history-conditioned residual distribution matching module by maximum mean discrepancy (MMD). The proposed method use a scale-aware strategy for missing-value reconstruction and alternating from missing reconstruction and causal structure learning. Evaluations on three synthetic datasets, CausalRiver and DREAM3 are provided to verify the effectiveness of DistCUTS.

Critical strengths
  1. DistCUTS combines incomplete time series and distribution-based causal relationship, providing a solution for a certain subfield in causal discovery. This paper is meaningful and provides a new perspective on causal structure learning under incomplete time series.
  2. The main algorithm design is reasonable. The history-conditioned residual distribution matching module provides additional learning signals for identifying distribution-dependent edges. Experiments are designed widely on multiple synthetic and real-world datasets across different settings.
Critical weaknesses
  1. The weights estimation of scale-aware missing-value reconstruction module may not capture real dynamics. DistCUTS calculates the blending weights using both the temporal interpolation and model's prediction. However, as a smoothing method, low overall interpolation error does not guarantee reliable overall estimates particularly around abrupt transitions. In comparison, the prediction may shows a lower overall prediction performance but can capture the real dynamic trends.
  2. DistCUTS use Gaussian as modeling distribution, but the rationality is not clearly justified. The proposed method use Gaussian for residual sampling and matching, but not fully justify the correctness with different noise families. Although authors provide mismatch distribution result in synthetic analysis, but most results are only evaluated by AUROC for causal graph recovery. A comprehensive metrics comparison (AUPRC, SHD, etc) and theoretical analysis can help to clarify the issue.
Questions for authors
  1. Could the authors provide analysis of the scale-aware strategy for missing-value reconstruction? What’s more, comparisons of alternative interpolation methods can also helps to identify the effectiveness.
  2. Could the authors report additional causal discovery metrics, such as AUPRC and SHD, alongside AUROC?
  3. For method including NGC, DYNOTEARS, PCMCI, and RHINO, different preprocessing or imputation method can affect results. Could the authors specify how the missing value is imputed before causal discovery?
  4. Could the authors provide further practical or theoretical analysis for rationality of Gaussian base residual sampling and matching?
2 comments

Thank you for the response. I have no more questions.

We sincerely thank the reviewer for the constructive feedback and thoughtful discussion. We appreciate your time and consideration, and are pleased that our responses have addressed your questions.

Reviewer_o8doOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DistCUTS, which discovers mean- and variance-driven causal dependencies from incomplete multivariate time series via residual distribution matching and scale-aware reconstruction.

Critical strengths
  1. Focusing on dependencies that alter only the conditional variance without changing the conditional mean addresses the limitation of point-prediction-based causal discovery methods, which might otherwise overlook such relationships.
  2. Labeling and evaluating mean-driven and scale-driven edges separately allows for a direct assessment of the impact of distributional supervision on the recovery of target dependencies.
Critical weaknesses
  1. Limited distributional modeling capacity. The implementation employs a fixed Gaussian basis distribution and models residual variations via history-dependent scales, which primarily captures heteroscedastic dependencies but struggles to accommodate more general distributional shifts, such as skewness. Furthermore, residual de-meaning cannot eliminate history-dependent mean prediction biases, potentially causing the scale branch to misinterpret mean model errors as conditional variance changes.
  2. Insufficient coverage of relevant baselines. The paper lacks comparisons with several methods highly relevant to the task. For instance, ReTimeCausal also investigates alternating imputation and structure learning in irregular time series; the authors should either include this comparison or clarify why it is inapplicable. Additionally, the missing value reconstruction experiments only compare against CUTS+, omitting specialized time series imputation methods as baselines, which makes it difficult to fully assess the effectiveness and competitiveness of the proposed reconstruction strategy.
  3. Ablation studies do not fully isolate the contributions of individual components. The "No demeaning" ablation simultaneously removes both residual de-meaning and mean gradient truncation, making it impossible to disentangle their respective effects. Similarly, the "No scale" ablation reverts to pure mean reconstruction, which fails to distinguish the benefits derived from the interpolation fusion mechanism itself versus those from the scale-adaptive weighting.
  4. Experimental settings are somewhat idealized. The synthetic data relies on regular topologies where mean parents and scale parents do not overlap, and it only considers Missing Completely At Random (MCAR) missingness. The method's robustness under complex topologies, block-wise missingness, and Missing Not At Random (MNAR) scenarios remains unverified. Moreover, real-world datasets lack ground-truth annotations for scale-driven edges, meaning they can only support evaluations of overall structure recovery or reconstruction utility, rather than directly validating the core mechanism.
  5. Insufficient statistical and efficiency evidence. The main results tables lack measures of result fluctuations (e.g., standard deviations) or confidence intervals for performance differences, and sparse graph recovery is evaluated solely via AUROC. Furthermore, the additional computational overhead of generating multiple residual samples per target and computing pairwise kernel values is not accompanied by an evaluation of runtime, GPU memory consumption, or scalability.
Questions for authors

See weaknesses.

Reviewer_HWwyOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a causal discovery algorithm (DistCUTS) for incomplete MTS, DistCUTS tackled noise additive SEM with only time lagged relationships, but assume that causal parents can influence the scale also in the SEM, which is known in the field by historical dependent noise.

Critical strengths

The paper targets an interesting and challenging task; SEM with historical dependent noise. The scale driven dependency is more realistic in many cases as fMRI and finance.

Critical weaknesses

The authors should be clearer about the assumptions they make, like additive SEM with only time-lagged relationships.

For me, the first two contribution points are the same, and we cannot state them as separate contributions: the success of the first one relies on the MMD relaxation.

The authors did not cite or compare to important baselines. Fantom [1] models instantaneous and time-lagged links for stationary and non-stationary MTS with heteroscedastic noise (more general than history-dependent noise). Scotch [2] uses SDEs for causal discovery and tackles exactly the same SEM as this paper while assuming complete MTS.

Novelty is limited, IMHO, as they used Cuts graph-constrained recurrent models and did not push it to instantaneous links. They tried to generalize it to history-dependent noise by adding an MMD relaxation, but I have some concerns about this (see the following points).

The paper feels like separate pieces to me. My first impression is that it was done by taking Cuts and Cuts+ and brainstorming with an LLM agent, which is totally fine, and I understand that research is changing nowadays. However, I feel everything that was achieved through an AI discussion should somehow be motivated, explained and fit into the storytelling of the paper, which for me has not been done in this paper. The MMD regularization is not motivated at all; it feels like something the authors found to work and added to Cuts and Cuts+ without trying to explain why it works and what it does. This relaxation was first introduced by OTM [3]. In that model, the authors motivated it by optimal transport matching, and in the IID setting this is a fair relaxation, because the definition of particles is clear and no time matching should be considered. Also, having it as a regularization was understandable because the authors in OTM showed that their constrained minimization problem could be optimized by adding a penalty to the objective. Their idea was first introduced, shown and proved in WAE [4]. In this setting, the intuition could be used but needs careful thought, which the authors ignore and do not discuss, motivate or prove. We are in an MTS setting, and optimal transport should be time-dependent or at least stated as ignored. The definition of particles is not clear, and the two distributions we want to match are neither clear nor stated.

Mathematical considerations: the paper assumes only time-lagged relationships, hence a consistency theorem should be feasible to show mathematically that your model could converge to what you aim for. The authors ignore mathematical soundness and treatment; they did not even state assumptions clearly, like causal stationarity, causal Markov factorization, causal sufficiency (for latent confounders)...

Assumptions about missingness are not stated or discussed, and also not treated in the experiments: MCAR, MAR and MNAR. For example, MNAR settings are not identifiable and cannot be recovered because the missing values depend on unobserved causes [5]. All the mathematical soundness was neglected in this paper.

pϵp_{\epsilon} is not tested using synthetic data if you do not assume any clear distribution. Otherwise, it should be clearly stated.

Let’s talk about the experiments section. First of all, reporting F1 score, SHD, precision and recall is important in causal discovery, and I think they should be reported in this paper. Second, the comparison is not fair because you compare with models that assume complete MTS (Rhino, Dynotears) without providing them with an imputer (PSW [6], TimeMixer [7]...). Third, the scenario is also not adequate for these models, as you assume only time-lagged links but these models take care of instantaneous links. A synthetic data comparison with these models in their settings is necessary, both when data is complete and when it has missing values, and a clear, transparent discussion is important for the community. A failure mode of your model is never a problem; there is no perfect model, and a clear discussion and diversified experiments open the door for more research and innovation. A comparison with Fantom in a heteroscedastic setting could also be interesting (but not mandatory), while a comparison with SCOTCH is very important as they assume the same SEM as your paper.

Related work. A more detailed section about the different causal structure learning algorithms in stationary and non-stationary MTS, complete and incomplete, is important. Also, it is really important for this paper to discuss MTS imputation and its relationship with causal discovery.

I have to flag your paper for LLM-generated references. You can read your references and you will find many weird generated citations:

  • Yuxiao Cheng, Lianglong Li, Tingxiong Xiao, Zongren Li, Jinli Suo, Kunlun He, and Qionghai Dai. CUTS+: high-dimensional causal discovery from irregular time-series. In Michael J. Wooldridge, Jennifer G. Dy, and Sriraam Natarajan (eds.), Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI 2024, Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence, IAAI 2024, Fourteenth Symposium on Educational Advances in Artificial Intelligence, EAAI 2024, February 20-27, 2024, Vancouver, Canada, pp. 11525–11533. AAAI Press, 2024.
  • Erdun Gao, Ignavier Ng, Mingming Gong, Li Shen, Wei Huang, Tongliang Liu, Kun Zhang, and Howard Bondell. Missdag: Causal discovery in the presence of missing data with continuous additive noise models. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh (eds.), Advances in Neural Information Processing Systems, volume 35, pp. 5024–5038. Curran Associates, Inc., 2022. doi: 10.52202/068431-0363. ....

References related to my text:

[1] 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. (NeurIPS 2025)

[2] Wang, Benjie, Joel Jennings, and Wenbo Gong. "Neural structure learning with stochastic differential equations." International Conference on Learning Representations. Vol. 2024. 2024. (ICLR 2024)

[3] Vo V, Zhao H, Le T, Bonilla EV, Phung D. Optimal transport for structure learning under missing data. arXiv preprint arXiv:2402.15255. 2024 Feb 23.

[4] Tolstikhin I, Bousquet O, Gelly S, Schoelkopf B. Wasserstein auto-encoders. arXiv preprint arXiv:1711.01558. 2017 Nov 5.

[5] Little RJ, Rubin DB. Statistical analysis with missing data. John Wiley & Sons; 2019 Apr 23.

[6] Wang H, Li H, Chen X, Gong M, Chen Z. Optimal transport for time series imputation. In International Conference on Learning Representations 2025 May 1 (Vol. 2025, pp. 828-852).

[7] Wang S, Wu H, Shi X, Hu T, Luo H, Ma L, Zhang J, Zhou J. Timemixer: Decomposable multiscale mixing for time series forecasting. InInternational conference on learning representations 2024 May 31 (Vol. 2024, pp. 38626-38652).

Questions for authors

see weaknesses

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.