acceptodds
Under review as a conference paper at ICLR 2027

Causal Representation Learning with Instantaneous and Lagged Relations via Nonstationarity

Abstract

Causal representation learning for time-series data aims to identify latent states and their causal relations from observations. In this setting, an important challenge is to model both lagged causal relations across observation intervals and faster causal effects that appear as instantaneous relations within an interval, while accounting for nonstationarity in time-series data. However, methods that jointly handle these causal relations and nonstationarity remain limited. To address this gap, we establish sufficient conditions for identifying latent states up to component permutation and component-wise invertible transformations, and their instantaneous and lagged causal structures up to the same permutation, using an observed auxiliary variable, such as time or a condition label, associated with changes in transition-noise distributions. Based on these results, we propose iCReN, a framework that uses contrastive learning with discrete or continuous auxiliary variables to learn latent representations and estimate their instantaneous and lagged causal structures. Experiments demonstrate accurate recovery of latent states and both instantaneous and lagged causal structures on synthetic data and the utility of the learned representations for downstream forecasting on real-world data.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Causal Representation Learning, Causal Discovery, Time Series, Nonstationarity

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_1KFUOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work studies latent causal representation learning on non‑stationary time‑series. It targets the joint identification of instantaneous and lagged causal relations from purely observed data, where underlying latent states are unobservable. The paper develops identifiability results for latent states and causal graphs, and introduces the auxiliary‑variable‑driven contrastive‑learning framework iCReN for latent representation learning and causal‑structure inference. Empirical results verify its causal‑structure recovery performance on synthetic datasets and demonstrate its practical value in real‑world prediction tasks.

Critical strengths
  • By integrating instantaneous causal relations, lagged causal relations, and nonstationarity into a unified framework, the study addresses the practical needs of complex dynamical systems more comprehensively than approaches that investigate only one type of causal relation.
  • Building on the TCL/GCL contrastive losses under the IIA framework, the authors develop a complete pipeline and provide two variants—iCReN-T and iCReN-G—for discrete and continuous auxiliary variables, respectively, and present a complete inference procedure for extracting causal graphs from Jacobian matrices.
Critical weaknesses
  • The method relies on relatively strong assumptions that are difficult to verify in real-world applications. It rules out unobserved instantaneous confounding and sampling aliasing through its modeling assumptions, and relies on the full-rank condition for instantaneous variability and the separation condition for neighboring components to resolve directional ambiguities. Importantly, the paper itself shows that linear models with instantaneous edges and additive-separable inverse transition functions violate the full-rank condition for instantaneous variability, leaving instantaneous causal structure without theoretical identifiability guarantees. The synthetic experiments also show a substantial drop in instantaneous-structure F1 scores in linear settings. However, this critical limitation is mentioned only in an appendix remark and is not adequately qualified in the abstract or introduction, potentially leading readers to overestimate the method's applicability.

  • The theoretical analysis defines causal edges in terms of whether partial derivatives are identically zero or nonzero at the function level, whereas neural networks yield only numerically estimated Jacobian values subject to approximation and numerical errors. More importantly, the synthetic experiments employ an oracle strategy for graph extraction: the true number of parents of each target node is provided, and the top-(k) edges with the highest Jacobian scores are selected as predictions. This strategy relies on ground-truth information unavailable in practical applications, where the number of true parents is unknown. Consequently, the reported structural recovery performance may not reflect the method's practical ability to infer causal graphs.

Questions for authors
  • Identifiability conditions are sufficient but not necessary. From a practical perspective, how can users assess whether real-world datasets approximately satisfy the key assumptions, particularly sufficient variation induced by the auxiliary variable and the separation condition for neighboring latent components? Can the authors provide computable diagnostic metrics to assess the plausibility of these assumptions in practice?
  • The theoretical analysis relies on the assumption that the transformations are C2C^2 diffeomorphisms, whereas the implementation uses Leaky-ReLU activations, which are only piecewise smooth and do not satisfy the C2C^2 assumption. How do the authors reconcile this discrepancy between the theoretical assumptions and the practical implementation?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 1KFU — Part 1/2

Part 1/2 — Weakness 1 and Question 1

We thank the reviewer for the detailed comments on the assumptions and graph evaluation, which helped us clarify the relationship between theory and implementation and add empirical checks.

1. Scope and practical assessment of the sufficient conditions (Weakness 1; Question 1)

The method relies on relatively strong assumptions that are difficult to verify in real-world applications. ... this critical limitation ... is not adequately qualified in the abstract or introduction, potentially leading readers to overestimate the method's applicability.

Can the authors provide computable diagnostic metrics to assess the plausibility of these assumptions in practice?

We agree that assessing the plausibility of identification conditions, such as sufficient variability and separation, is important in practice. However, it is difficult to assess the plausibility of these conditions in advance because they involve unknown latent distributions and structures. We are therefore unable to provide computable diagnostic metrics for this purpose in this work. In practice, domain knowledge may help assess the relevance of the auxiliary variable to nonstationarity and the plausibility of the learned representations and inferred causal relations. We have revised Section 7 to note the difficulty of assessing the theoretical assumptions in real-world data.

We thank the reviewer for highlighting the need to clarify the scope of our identification guarantees. These guarantees apply to the stated discrete-time model without unmodeled common causes and under the identification conditions. As the reviewer points out, sampling can introduce aliasing, but analyzing the effects of time discretization and sampling intervals on causal structure is outside the scope of this work. As also noted in the review, when instantaneous causal relations are present, the sufficient-variability condition in Theorem 1 fails if the inverse transition function is additively separable in the current and past states; linear transition models are a special case, as discussed in Remark 1. We have clarified in the Introduction that Section 4 discusses the scope and limitations of our guarantees in detail, including the exclusion of linear transition models with instantaneous causal relations.

Continued in Part 2/2.

AuthorsOpenReview ↗

Response to Reviewer 1KFU — Part 2/2

Part 2/2 — Weakness 2 and Question 2

Continued from Part 1/2.

2. Numerical graph extraction and parent-count-based evaluation (Weakness 2)

More importantly, the synthetic experiments employ an oracle strategy for graph extraction: the true number of parents of each target node is provided, and the top-(k) edges with the highest Jacobian scores are selected as predictions.

We thank the reviewer for highlighting the gap between function-level identification and numerical graph extraction, as well as the use of ground-truth parent counts. As the reviewer notes, while the theoretical result characterizes causal edges through derivatives that are not identically zero as functions, graph extraction from a learned model relies on numerical scores and is subject to estimation error and threshold choice.

To address this concern empirically, we evaluated how accurately graphs could be recovered without ground-truth parent counts in the default setting with both latent and observation dimensions set to 5. We normalized the average absolute Jacobian scores and applied a prespecified threshold of 0.05 instead of selecting edges according to the true parent counts. Although F1 scores decreased relative to top-k selection, fixed-threshold selection achieved mean instantaneous and lagged F1 scores of 0.890 and 0.937, respectively, over 10 seeds without using ground-truth parent counts. We have added the normalization and thresholding procedure to Appendix C.7, along with a comparison of fixed-threshold and top-k results. We have also added a brief discussion of these results to Sections 6.1 and 7.

3. Smoothness assumptions and Leaky-ReLU implementations (Question 2)

The theoretical analysis relies on the assumption that the transformations are diffeomorphisms, whereas the implementation uses Leaky-ReLU activations, which are only piecewise smooth and do not satisfy the assumption. How do the authors reconcile this discrepancy between the theoretical assumptions and the practical implementation?

As the reviewer notes, our implementation does not strictly satisfy the diffeomorphism assumptions in our theory. Replacing Leaky-ReLU with smooth activations alone would not resolve this discrepancy, since smoothness does not guarantee invertibility of the learned mappings. Given the high recovery performance obtained with standard neural networks, we retain this implementation. We also conducted an experiment replacing only the activations in the true structural and observation functions with Smooth Leaky-ReLU, while leaving the estimator unchanged. Combined with the existing orthogonal observation matrices and additive-noise DAG structure, this replacement yields diffeomorphic observation and augmented transition mappings. The mean latent MCC and instantaneous and lagged graph F1 scores over 10 seeds were 0.951, 0.925, and 0.910, respectively. These results suggest that this discrepancy may not have a substantial impact on recovery performance in the evaluated setting. We have added this experiment to Appendix C.6, reported the results in Table 12, and clarified the relationship between theory and implementation in Appendix C.1.2.

Thank you for the authors' responses! I have no questions.

Thank you again for carefully considering our response and revised manuscript. We appreciate your constructive feedback and continued support.

Reviewer_2WCQOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents iCReN, a temporal causal representation learning method. Accounting for the non-stationarity of time series, iCReN captures both lagged and instantaneous causal relationships. iCReN leverages observable auxiliary variables that modulate the transition noise distribution to characterize non-stationarity, and devises dedicated contrastive learning strategies for discrete and continuous auxiliary variables respectively.

Critical strengths
  1. The motivation of this paper is well justified. The existing causal representation learning methods struggle to identify both instantaneous and lag effects under non-stationarity. To address this limitation, this paper introduces observable auxiliary variables associated with transition noise distribution shifts instead of real interventions to guide the learning of causal representations and latent causal relationships.
  2. This work enjoys solid theoretical support. Theorems 1 and 2 establish the identifiability of latent components and their causal relationships respectively, alongside the corresponding identification conditions.
  3. The experimental validation is relatively comprehensive. The benchmark suite includes representative baselines such as IDOL and iVAE. The experiments verify the sensitivity of iCReN to factors such as data dimension, sample size, and complexity of the generating function, and included ablation studies on the structural regularization term and reconstruction loss.
  4. The reproducibility is strong. The manuscript fully describes the data generation pipeline, hyperparameter configurations, and other experimental details, and releases the algorithm code.
Critical weaknesses
  1. The reliance on observable auxiliary variables introduces notable limitations. iCReN may fail to work in real-world scenarios where such auxiliary variables are unavailable.
  2. The treatment of non-stationarity is inadequate. Beyond non-stationarity arising from shifting noise distributions, time series can also become non-stationary due to variations in causal strength and causal structure. These more complex cases are not considered in this work.
  3. The experimental setup has a drawback. In the synthetic experiments, the observation dimension is set equal to the latent dimension. In real-world applications, however, the observation dimension is typically larger, often substantially larger, than the latent dimension. The authors do not evaluate the algorithm’s performance under this more realistic setting.
  4. Real-data experiments are insufficient to validate iCReN’s capability to recover causal representations and latent causal relationships.
Questions for authors
  1. Although the paper lists multiple examples of auxiliary variables for real-world problems, scenarios where such auxiliary variables are unavailable do exist in practice. Without auxiliary variables, can iCReN work properly?
  2. How does iCReN perform on higher-dimensional data? Specifically, what is its behavior when the observed dimension is substantially larger than the latent dimension?
  3. Could the authors report iCReN’s performance in recovering causal representations and latent causal relationships on real-world datasets? Alternatively, please present iCReN’s identification results on real data and analyze the plausibility of these results in practical applications.
3 comments
AuthorsOpenReview ↗

Response to Reviewer 2WCQ

We appreciate the reviewer’s comments on the applicability of our framework, which motivated the higher-dimensional observation experiments and helped us clarify the interpretation of the real-world results.

1. Requirement for an observed auxiliary variable (Weakness 1; Question 1)

Without auxiliary variables, can iCReN work properly?

The current formulation of iCReN requires an observed auxiliary variable associated with variation in transition-noise distributions. When such a variable is unavailable, the method cannot be applied as formulated. However, the auxiliary variable need not specify intervention targets. It may therefore be possible to construct and use auxiliary labels from available information, such as gait phase derived from foot-strike events or calendar-based labels for weather data, as illustrated by our real-world experiments. Extending the framework to unobserved auxiliary variables is a direction for future work, as discussed in Section 7.

2. Scope of the modeled nonstationarity (Weakness 2)

Beyond non-stationarity arising from shifting noise distributions, time series can also become non-stationary due to variations in causal strength and causal structure.

We agree that changes in causal strength and causal structure are also important sources of nonstationarity. Methods such as TDRL (Yao et al., 2022) and CtrlNS (Song et al., 2024) model nonstationarity through changes in structural functions. Our work focuses on a specific setting in which the structural functions, causal graph, and observation function remain invariant, while transition-noise distributions vary with an observed auxiliary variable. Extending our framework to handle broader forms of nonstationarity, including changes in structural functions, is an important direction for future work. In Section 7, we have clarified the scope of nonstationarity modeled by our framework, discussed TDRL and CtrlNS, and suggested comparing theoretical guarantees and empirical performance across methods as a direction for extending the framework.

3. Observations with higher dimension than the latent state (Weakness 3; Question 2)

Specifically, what is its behavior when the observed dimension is substantially larger than the latent dimension?

We thank the reviewer for highlighting the need to evaluate observations with higher dimension than the latent states. To address this concern, we conducted additional experiments with observation dimensions of 10, 100, and 1000, keeping the latent dimension fixed at 5. We expanded the representation using an orthogonal matrix before applying the final activation in the nonlinear observation mapping. The table reports means over 10 seeds for these additional settings and the original 5-dimensional observation setting.

Observation dimensionLatent MCCInstantaneous F1Lagged F1
5 (original)0.9510.9500.980
100.9531.0001.000
1000.9580.9750.980
10000.9580.9751.000

High recovery accuracy was observed across the evaluated observation dimensions. These results suggest that accurate recovery is possible even when the observation dimension substantially exceeds the latent dimension, for the nonlinear observation mappings tested here. We have added the experimental setup and results, including standard deviations, in Appendix C.4 and Table 8.

4. Interpretation of the real-world results (Weakness 4; Question 3)

Alternatively, please present iCReN’s identification results on real data and analyze the plausibility of these results in practical applications.

We agree that evaluating the validity of learned latent states and causal relations on real-world data is important. However, directly assessing their recovery is difficult without ground truth. Moreover, identifiability alone does not guarantee clear interpretations of the learned components and causal relations in real-world applications, so interpreting and validating these results may require additional evidence. We therefore evaluate downstream predictive utility in our real-world experiments, without claiming to validate causal recovery. Section 7 discusses this limitation and identifies validation using domain knowledge or interventional information as a direction for future work.

Reviewer_2WCQOpenReview ↗

Response to Authors

Thank you for your reply. I have two remaining concerns:

  1. While experiments with higher observational dimensions are helpful, the scalability of iCReN to larger latent dimensions remains unvalidated. When the true latent dimension is unknown, how should we set the model’s estimated latent dimension? Can iCReN learn meaningful causal representations under latent-dimension misspecification? Specifically, with an overspecified latent dimension, can iCReN prune redundant dimensions and recover the true latent structure and causal representations?
  2. Downstream predictive performance alone is insufficient to validate the learned causal representations, as strong predictive accuracy can be achieved without faithful causal representations. Even though ground-truth latent causal components are unavailable for real-world data, I encourage the authors to examine whether the inferred representations conform to domain priors as auxiliary evidence. For example, do observed variables with co-evolution patterns tend to be encoded into the same or a shared set of latent causal components by iCReN?
AuthorsOpenReview ↗

Response to Reviewer 2WCQ — Follow-up

Thank you for taking the time to consider our response and for your constructive suggestions for further evaluation. We appreciate your continued engagement and the opportunity to clarify the scope of our results and the remaining challenges.

1. Scalability to larger latent dimensions and latent-dimension overspecification

“… the scalability of iCReN to larger latent dimensions remains unvalidated.”

“Specifically, with an overspecified latent dimension, can iCReN prune redundant dimensions and recover the true latent structure and causal representations?”

In contrast to the results with higher-dimensional observations, scalability to higher latent dimensions remains a challenge, particularly for instantaneous-graph recovery. Both the evaluation in Table 7 of Appendix C.3 and the additional experiments reported in Table 11 of Appendix C.6 show lower recovery accuracy at higher latent dimensions. Addressing this difficulty to model more complex systems is discussed as future work in Section 7.

Our additional experiment suggests that reasonably accurate latent-state and graph recovery is possible even when the representation dimension exceeds the true latent dimension. We compared iCReN-T with eight estimated components against the existing five-dimensional models, keeping both the true latent and observation dimensions at five and using the same 10 seeds. For evaluation, we selected five of the eight estimated components using a one-to-one matching that maximizes the mean absolute Pearson correlation with the five ground-truth components on the validation set, and fixed this matching for test evaluation. We then evaluated MCC for the selected components and F1 scores among them using top-k edge selection. The table reports means and standard deviations over 10 seeds. Although F1 scores decrease, the selected components retain a similarly high mean MCC and reasonably high instantaneous and lagged F1 scores. However, the dimension of the learned representation is specified as a hyperparameter, and component selection in this evaluation relies on ground truth. We have added this evaluation to Table 16 in Appendix C.9 and discussed automatic dimension selection and removal of redundant components as topics for future work in Section 7.

Representation dimension dz^d_{\hat z}MCCInstantaneous F1Lagged F1
5 (original)0.951 ± 0.0110.950 ± 0.1580.980 ± 0.063
80.953 ± 0.0100.850 ± 0.2110.920 ± 0.079

2. Auxiliary validation using domain knowledge

Even though ground-truth latent causal components are unavailable for real-world data, I encourage the authors to examine whether the inferred representations conform to domain priors as auxiliary evidence.

For example, do observed variables with co-evolution patterns tend to be encoded into the same or a shared set of latent causal components by iCReN?

Following your suggestion, we visualized Pearson correlations between marker coordinates and learned latent components in the treadmill data to examine their associations with body motion. The model with the lowest pretraining validation loss on Slow and Fast among the 10 seeds was applied to all three speeds for the same six validation participants. Multiple marker coordinates correlated strongly with common latent components: for example, z^1\hat z_1 with the front–back positions of thigh and knee markers, z^3\hat z_3 with the up–down positions of ankle and heel markers, and z^5\hat z_5 with the left–right positions of knee, ankle, and foot markers. These broad patterns were similar across the three speeds, including Comfortable, which was excluded from pretraining. This analysis is limited to Pearson correlations from a single model and does not examine nonlinear associations between marker coordinates and latent components in detail. Although it does not validate the causal structure, we view the results as auxiliary evidence of associations between the motion of multiple markers and common latent components. We have added this analysis to Appendix D.1, with the visualization shown in Figure 7.

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

This paper proposes iCReN, which utilizes non-stationary transition noise and auxiliary variables to identify latent states in time series and their instantaneous and lagged causal relationships through contrastive learning.

Critical strengths
  1. The research problem is of research value. Simultaneously considering both instantaneous and lagged causal relationships among latent states better aligns with dynamic systems where mechanisms operating at different time scales coexist.
  2. The utilization of auxiliary information is flexible. The framework offers two implementations, TCL and GCL, which can leverage discrete environment labels and continuous time information, respectively, without requiring known intervention targets.
Critical weaknesses
  1. Limited novelty. Latent state identification draws on IDOL, transition noise identification relies on IIA, and causal edge recovery adopts existing Jacobian methods; the algorithm primarily combines contrastive learning with structural regularization. The paper needs to more clearly demonstrate what important, identifiable model categories are actually introduced compared to existing theories.
  2. Strong theoretical assumptions with a gap between theory and implementation. Identifiability relies on invertibility, sufficient variation, temporal separation, and constraints on the number of learned graph edges. However, standard neural networks and gated sparsity regularization do not directly guarantee that these conditions hold. Furthermore, the theory does not cover linear transition models with instantaneous relations, which limits its applicability.
  3. Graph recovery evaluation relies on the ground-truth number of parent nodes. In the experiments, the top-scoring edges equal to the true number of parent nodes are selected for each target. Therefore, this primarily measures edge ranking capability under a known in-degree, and does not fully reflect the actual discovery performance when the structural scale is unknown. W4. Insufficient scalability to high dimensions. The main experiments focus on 5-dimensional latent states; when the dimension increases to 16, the instantaneous F1 score drops from 0.950 to 0.140, indicating that the low-dimensional results are difficult to generalize to more complex systems.
  4. Insufficient empirical evidence on real-world data. Improvements in downstream prediction cannot verify the correctness of the learned latent states and causal structures. Moreover, the lack of comparisons with specialized forecasting models (e.g., Pathformer) and non-causal representation methods that use the same auxiliary information makes it difficult to isolate and clarify the specific added value of the causal modeling itself.
Questions for authors

See weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer muQu — Part 1/2

Part 1/2 — Weaknesses 1–3

We thank the reviewer for the detailed comments, which helped us clarify the scope of our theoretical guarantees and expand the experimental evaluation.

1. Theoretical contribution and relation to prior work (Weakness 1)

The paper needs to more clearly demonstrate what important, identifiable model categories are actually introduced compared to existing theories.

We thank the reviewer for pointing out the lack of clarity about the identifiable model class. Our theoretical contribution is to establish sufficient conditions for identifying latent-state components and their instantaneous and lagged causal structure in nonlinear models with nonstationary transition noise. Specifically, whereas IDOL (Li et al., 2025) assumes time-invariant transition-noise distributions, our formulation explicitly models nonstationarity in the data through variation in these distributions associated with an observed auxiliary variable. Theorems 1–2 provide identification guarantees in this setting by extending IIA’s transition-noise identification result (Morioka et al., 2021) to latent-state components and their causal structure, drawing on IDOL’s identification argument. As empirical support for the utility of this nonstationary formulation, Table 2 shows that both variants of iCReN achieved higher latent-state and graph recovery scores than IDOL under both discrete and continuous transition-noise variation. We have also clarified in the Introduction that Section 4 discusses the scope and limitations of these guarantees.

2. Scope of the identifiability results and relation to the implementation (Weakness 2)

However, standard neural networks and gated sparsity regularization do not directly guarantee that these conditions hold.

Furthermore, the theory does not cover linear transition models with instantaneous relations, which limits its applicability.

As the reviewer notes, our neural-network implementation does not guarantee that all the conditions in Theorem 1 are satisfied, including the diffeomorphism and edge-count conditions; gate-based sparsity regularization encourages sparsity rather than strictly enforcing the edge-count constraints. In addition, linear transition models with instantaneous causal relations do not satisfy the current sufficient-variability condition. This limits the scope of our theoretical guarantees, and Table 9(a) in Appendix C.6 also shows lower accuracy in recovering the instantaneous graph in the linear setting. We have clarified the limitations of the soft regularizers in Section 5. We have also clarified in the Introduction that Section 4 discusses cases outside the scope of our guarantees, including linear transition models with instantaneous causal relations.

3. Graph evaluation without ground-truth parent counts (Weakness 3)

Graph recovery evaluation relies on the ground-truth number of parent nodes.

We thank the reviewer for highlighting the need for graph evaluation without ground-truth parent counts. The original top-k evaluation selects as many edges for each target as its true number of parents. We therefore conducted an additional evaluation in the default setting with five latent and five observed variables, normalizing the average absolute Jacobian scores and applying a prespecified threshold of 0.05 instead of using the true parent counts. Although F1 scores decreased relative to top-k selection, fixed-threshold selection achieved mean instantaneous and lagged F1 scores of 0.890 and 0.937, respectively, over 10 seeds without using ground-truth parent counts. We have added the evaluation procedure to Appendix C.7, along with a comparison of fixed-threshold and top-k results. We have also added a brief discussion of these results to Sections 6.1 and 7.

Continued in Part 2/2.

AuthorsOpenReview ↗

Response to Reviewer muQu — Part 2/2

Part 2/2 — Weaknesses 4–5

Continued from Part 1/2.

4. Performance at higher latent dimensions (Weakness 4)

The main experiments focus on 5-dimensional latent states; when the dimension increases to 16, the instantaneous F1 score drops from 0.950 to 0.140, …

We thank the reviewer for highlighting the decline in instantaneous-graph recovery at higher latent dimensions shown in Table 7 in Appendix C.3. Motivated by the improvement from simplifying the observation function in the five-dimensional setting, we extended the comparison between one- and three-layer observation functions to 8, 12, and 16 latent dimensions. Instantaneous F1 improved at 8 and 12 dimensions but remained at 0.140 at 16 dimensions, despite improvements in latent MCC and lagged F1. Thus, simplifying the observation function alone does not resolve the difficulty at 16 dimensions. We have added this comparison to Table 11 in Appendix C.6 and discussed improving graph recovery at higher latent dimensions as a direction for future work in Section 7.

O3 and O1 denote three-layer and one-layer observation functions, respectively. Each cell reports the mean over 10 seeds in the order O3 / O1.

Latent dimensionMCC: O3 / O1Inst. F1: O3 / O1Lagged F1: O3 / O1
50.951 / 0.9570.950 / 1.0000.980 / 1.000
80.860 / 0.8670.743 / 0.9290.869 / 0.956
120.886 / 0.8960.527 / 0.6180.812 / 0.867
160.792 / 0.8920.140 / 0.1400.569 / 0.731

5. Real-world evaluation (Weakness 5)

Improvements in downstream prediction cannot verify the correctness of the learned latent states and causal structures.

Moreover, the lack of comparisons with specialized forecasting models (e.g., Pathformer) and non-causal representation methods that use the same auxiliary information makes it difficult to isolate and clarify the specific added value of the causal modeling itself.

We thank the reviewer for suggesting additional comparisons. First, we have added TCL (Hyvärinen and Morioka, 2016) to both real-world experiments as a non-causal representation-learning baseline using the same auxiliary information. Whereas iCReN-T estimates transition noise using IIA-TCL, the added TCL baseline learns representations by predicting auxiliary labels from observations at each time step. Both methods use the same gait-phase or calendar labels during pretraining, followed by frozen representations of the same dimensionality and a common MLP forecasting head. iCReN achieved lower mean prediction errors than TCL at all reported horizons on the treadmill data and in all eight weather regions. We also added a comparison with Pathformer (Chen et al., 2024) on the treadmill task, where both inputs and prediction targets are marker coordinates, and found that Pathformer achieved lower mean prediction errors than iCReN. Our representation-learning comparison evaluates the predictive utility of frozen five-dimensional representations using a common MLP forecasting head. Pathformer instead learns the entire forecasting model directly from observed coordinates; we report its lower prediction errors and the comparison conditions in the appendix as a direct forecasting reference. We have also clarified the scope of the experiments in the main text and explicitly noted Pathformer’s lower prediction errors.

We agree that validating the learned latent states and causal structures is important. However, directly assessing their recovery is difficult because ground truth is unavailable for the real-world datasets used here. Drawing on the evaluation approaches of prior work on temporal causal representation learning (e.g., Li et al., 2025; Song et al., 2024), we evaluate the utility of the learned representations through downstream tasks. As discussed in Section 7, validating the latent states and causal structures using additional evidence, such as domain knowledge or interventional information, remains an important direction for future work.

I thank the authors for their detailed response. I have no further questions and will raise my score accordingly.

Thank you again for your time and thoughtful consideration of our response and revised manuscript. We appreciate your positive reassessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.