acceptodds
Withdrawn from ICLR 2027

Unmixing Causal Dynamics: Learning Latent Mechanisms from Aggregate Time Series

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Observed time series are often superpositions of multiple causal mechanisms that evolve concurrently over the same variables, so the causal relations of individual mechanisms are entangled in the aggregate signal. At each time step, each latent mechanism produces its own state vector, but we observe only the elementwise sum across mechanisms, together with sparse measurements that occasionally reveal mechanism states. We introduce UCD (nmixing ausal ynamics) to recover mechanism-specific causal structure from such aggregated observations. UCD models each mechanism with a temporal structural equation model (SEM) with independent additive noise and performs particle-based inference over the unobserved mechanism trajectories. It maintains a set of particles, each representing a plausible joint history of all mechanisms. Between measurements, particles propose and reweight next-step mechanism states that are consistent with the observed aggregate; when sparse state measurements arrive, particles are corrected to match the revealed states and reweighted by their plausibility. We then learn each mechanism's causal graph and noise scales by maximizing the likelihood of the residuals between predicted and revealed mechanism states under the assumed noise model. Experiments show that UCD reliably recovers mechanism-specific causal structure from aggregated time series given sparse state measurements.

Primary area:
causal reasoning
Keywords:
Causal Discovery, Time Series

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_F7jCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies temporal causal discovery from additive mixtures of concurrently evolving latent mechanisms. Each mechanism follows a linear temporal SEM, while the learner observes their aggregate states and occasionally their complete individual states. The goal is to recover mechanism-specific instantaneous and lagged causal structures. The authors propose UCD, a particle-based framework that estimates latent histories under aggregate constraints and updates model parameters using predictive likelihood with sparsity and acyclicity regularization. Experiments on synthetic systems and a CausalRivers-based task demonstrate improvements over existing baselines, supported by ablation studies and runtime analyses.

Critical strengths

a. The paper considers concurrent additive causal mechanisms and proposes a particle-based framework that explicitly enforces aggregate observation constraints. The use of importance weighting and measurement assimilation provides a coherent approach to estimating latent mechanism states.

b. UCD consistently outperforms the adapted baselines on synthetic benchmarks. The ablation studies and particle-count analysis further demonstrate the contributions of individual components and their computational trade-offs.

Critical weaknesses

a. The learning procedure fixes the inferred latent histories when updating model parameters, although the filtering distribution itself depends on these parameters. Consequently, the update may not correspond to optimizing the stated predictive likelihood, and a clearer justification of this approximation would strengthen the methodological soundness.

b. The default setting provides complete mechanism states at 40% of the time steps, offering substantial direct supervision. A comparison using only these revealed states would help establish the additional benefit of aggregate observations, while clarification of the unsupervised ablation objective is needed since the aggregate constraint is already enforced exactly.

Questions for authors

a. Could the authors clarify the justification for treating latent histories as fixed during parameter updates, given their dependence on model parameters? A comparison with a more direct predictive-likelihood optimization on a small system would help validate this approximation.

b. Could the authors include a baseline using only the revealed mechanism states while preserving the original temporal indices? Additionally, what is the exact objective used in the no-supervision ablation, and is its reconstruction loss computed before or after enforcing the aggregate constraint?

c. Could the authors clarify which assumptions in Appendix A.6 are guaranteed by the available observations and which are additional requirements? In particular, how does the identifiability result apply to the aggregate and sparsely observed mechanism states considered in the experiments?

Reviewer_357eOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to recover individual causal structures from time series generated by the superposition of multiple underlying causal mechanisms.

The authors propose a method called UCD (Unmixing Causal Dynamics). The method assumes that each latent mechanism follows an independent linear temporal structural equation model (Temporal SEM), while the observed signal is obtained by taking the element-wise sum of the states of all latent mechanisms. In addition, complete observations of all latent mechanism states are available at a subset of time points.

UCD consists of two main components:

1.Particle-based latent-state inference: The method maintains multiple possible histories of the latent mechanisms through particle filtering. It uses the observed aggregated signals to constrain the particle states, while importance sampling and resampling are employed to retain plausible latent-state estimates.

2.Likelihood-based parameter optimization: At time points where the true latent states are observed, the method uses model prediction residuals and their noise likelihoods to update the causal graph and noise parameters associated with each mechanism.

The authors adopt an alternating optimization strategy, iteratively updating the latent-state estimates and causal model parameters.

The experiments cover different numbers of latent mechanisms, time-series lengths, observation ratios, graph sizes, and noise levels. The authors also construct semi-synthetic experiments using the CausalRivers dataset. The results show that UCD outperforms the compared baseline methods under several experimental settings.

Critical strengths

S1. The research problem is interesting and has some novelty.

Existing time-series causal discovery methods typically learn a single causal graph directly from observed signals, making it difficult to distinguish causal relationships arising from multiple simultaneously evolving latent mechanisms. This paper considers a setting where multiple underlying mechanisms coexist and jointly generate the observed signals. By combining latent-state inference with causal structure learning, the authors provide a new perspective on this problem.

S2. The experimental evaluation is relatively comprehensive.

The authors conduct experiments under different settings, including varying numbers of latent mechanisms, time-series lengths, observation ratios, graph sizes, and noise levels. The ablation studies also remove particle-based inference, importance weighting, and sparse state supervision separately. These experiments provide some evidence of the contributions of different components of the proposed method.

Critical weaknesses

W1. The method relies heavily on complete latent-state supervision.

UCD requires complete observations of all latent mechanism states at certain time points. In the synthetic experiments, complete state observations are provided for 40% of the time points, while the CausalRivers experiments use a ratio of 30%.

The ablation results further show that removing sparse state supervision leads to a substantial drop in performance. This suggests that the method relies heavily on this type of supervision. However, in practical applications, it is often difficult to directly observe the complete states of all underlying mechanisms. This limits the applicability of the proposed method in real-world scenarios.

W2. The alternating optimization procedure lacks sufficient theoretical justification.

During the parameter optimization stage, the authors fix the latent histories inferred by the particle-based procedure and then update the causal model parameters using gradient-based optimization. However, the latent-history distribution approximated by the particles also depends on the model parameters.

Therefore, it remains unclear whether updating the model parameters while keeping the inferred latent histories fixed can adequately approximate the predictive likelihood objective formulated in the paper. The authors should provide further justification for this approximation and clarify whether it could affect the final parameter estimates.

W3. The method assumes that the number of latent mechanisms is known in advance.

The method explicitly specifies the number of latent mechanisms (K) in its formulation, and the experiments evaluate different values of (K) separately.

However, in practical applications, the number of underlying mechanisms is often unknown. The paper does not sufficiently explain how (K) should be determined, nor does it demonstrate whether the model can automatically identify the number of latent mechanisms when this information is unavailable.

This limits the applicability of the method to more general latent mechanism discovery problems.

W4. The real-world experiments provide limited evidence of practical applicability.

In the CausalRivers experiments, the authors construct a mixture of multiple latent mechanisms by taking the element-wise sum of two real river time series.

Although this approach allows the model to be evaluated using real measurement data, it is not equivalent to recovering causal structures from naturally occurring systems involving multiple interacting or superimposed mechanisms. Therefore, the current experiments provide limited evidence that the proposed method can handle the real-world mixed-mechanism scenarios emphasized in the paper.

W5. Some details of the experimental evaluation protocol remain unclear.

In the mechanism-level evaluation (Match Setting), the recovered causal graphs must be matched with the ground-truth graphs of the underlying mechanisms before computing the evaluation metrics.

However, the paper does not sufficiently explain the specific graph-matching procedure. It is also unclear whether the number of mechanisms (K) is always provided to the model as prior information. These details are important for interpreting the experimental results and ensuring reproducibility.

Questions for authors

Q1. In the alternating optimization procedure, the latent-history distribution approximated by the particles depends on the model parameters. Could the authors further explain why updating the model parameters while keeping the inferred latent histories fixed provides a reasonable approximation to the original predictive likelihood objective? Could this approximation affect the final parameter estimates?

Q2. Does the proposed method require the number of latent mechanisms (K) to be known in advance? In the mechanism-level evaluation, how exactly are the recovered causal graphs matched with the ground-truth graphs of the underlying mechanisms?

Q3. Beyond complete and noise-free observations of latent states, can UCD handle partially observed or noisy latent mechanism states? Furthermore, how sensitive is the method to deviations between the actual noise distribution and the noise distribution assumed by the model?

Reviewer_yWo3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors introduce UCD (Unmixing Causal Dynamics), a method for causal discovery in time series where the observed signal is element-wise sum of multiple unobserved mechanisms--each evolving under its own temporal SEM with a non Gaussian structural noise; here, sparse, occasional measurements reveal the true states of all latent mechanisms. UCD maintains a set of particles each representing a hypothesized history of all latent mechanism states consistent with the aggregate observations. Between measurement times, particles are propagated forward under the current causal model, constrained to match the aggregate observation through a clever proposal distribution, and reweighted by the plausibility of the implied noise under its family structure. At measurement time, particles are clamped to the revealed states and their residuals contribute to a plausibility likelihood used to update the causal parameters for each mechanism. The system alternates between particle based latent state inference and gradient-based updates to the causal graphs (with sparsity and acyclicity regularization). The authors find that UCD recovers mechanism specific causal structure more accurately than extended baselines across varying no of mechanisms, seq lengths, observation rates, size of graphs, noise scales, and include ablations for each design choice. They also report results on a semi synthetic, additive mixture task constructed from the CausalRivers benchmark.

Critical strengths
  • S1. The paper identifies a real gap in time series causal discovery. Prior methods fit a single causal graph to the observed signal, ignoring the possibility that the observation is an aggregation of more than one evolving latent mechanisms, each has its own causal structures. UCD decomposes the problem into "particle-based" latent history inference and "plausibility" parameter optimization.

  • S1. Ablations isolate the contribution of each component. Table 3 removes multiple particle inference, importance weighting, and sparse supervision, showing that each component contributes to overall performance. There are also sensitivity experiments across seq length, observed rate, graph size and noise scale.

Critical weaknesses
  • W1. The assumption of sparse state measurements is a strong one. In the experiments, 30-40% of time steps come with full ground-truth measurements for every latent mechanism. This limits real world deployment as true underlying latent measurements are not often observable. The ablation (UCD w/o supervision, Table 3) shows that the method relies on this supervision, and collapses without the sparse measurements. As I it; this makes UCD function more like a ``semi supervised learner'' than actually a latent mechanism discovery approach.

  • W2. The also doesn't address whether partial/noisy measurements can be handled. Partial or noisy sparse measurements are a more realistic scenario with potentially broader applicability than the assumption of clean measurements. Experiments evaluating how UCD performs under these environments could help strengthen the paper.

  • W3. The real experiment tests UCD under artificial additive mixture. CausalRivers constructs the additive mixture by ``artificially" summing 2 river seqs and treating them as latent mechanisms. This means that the seqs don't share a ground truth, which does not align with the paper's motivation and weakens its conclusions. A more convincing real world test would use a domain where there is a naturally occurring, or intrinsic, mixture (e.g., EEG signals--here, multiple neural sources sum up to data at the sensor level). These make it not clear whether the gains from UCD transfer to the scenarios that the formulation is motivated by.

  • W4. Reference issues: L569: Michael Jordan is not an author of the paper: Shohei Shimizu, Patrik O Hoyer, Aapo Hyvärinen, Antti Kerminen, and Michael Jordan. A linear non-gaussian acyclic model for causal discovery. JMLR 7(10), 2006 --- L533: The "ERDdS and R&wi" reference citing Erdős-Rényi random graphs is garbled text -- `"PARW ERDdS and A R&wi" should be "Paul Erdős and Alfréd Rényi."

Questions for authors
  • In the 'match' evaluation setting, how are the K recovered mechanism graphs aligned to the K ground truth mechanism graphs before scoring? Is the number of mechanisms K assumed to be known? It would be interesting to see how performance varies across sequence length, observation rates, etc., on the CausalRivers dataset, to see whether the performance shown on the synthetic dataset transfers.

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

Related papers

Discussion (0)

Sign in to comment.