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