acceptodds
Under review as a conference paper at ICLR 2027

CausalSR: Scale-Resolved Causal Discovery from Time Series with Heterogeneous Responses

Abstract

Causal discovery from time series has become increasingly important across a wide range of application domains. In practice, multivariate time series often exhibit heterogeneous responses across short, medium, and long time scales, causing causal effects at different temporal scales to overlap in the observed signals and making scale-specific causal relations difficult to separate, orient, and recover reliably. To address this issue, we propose CausalSR, a scale-resolved causal discovery framework that reorganizes multivariate time series in the time-frequency domain and learns causal structures at their corresponding temporal scales, instead of directly operating on the original time-domain observations. Specifically, the framework separates scale-specific temporal components, estimates interaction strength and directional evidence within each scale, suppresses indirect relations through regularized conditional-association screening, and jointly refines the resulting graphs using reconstruction consistency and acyclicity constraints, thereby enabling the recovery of more reliable scale-resolved structural information from complex time series for downstream temporal analysis. Theoretically, we prove that coherence-weighted phase aggregation preserves directional temporal ordering and enables scale-specific directed edge inference, thereby providing theoretical support for scale-resolved causal discovery. Experiments on synthetic benchmarks and real-world temporal datasets demonstrate the effectiveness of our method.

Primary area:
causal reasoning
Keywords:
Causal discovery, multivariate time series, heterogeneous temporal responses, time-frequency representation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ReCbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper aims to recover causal relations that operate at different temporal scales. It proposes CausalSR, which decomposes time series into frequency bands, combines interaction strength with phase information to orient candidate edges, and applies conditional screening and joint graph refinement. Experiments on synthetic and reference-based temporal datasets report improved structural recovery, with additional analyses of reconstruction quality and sensitivity to the chosen bands and thresholds.

Critical strengths
  1. The paper addresses the difficulty of separating interactions with different temporal responses. The scale-specific graph representation provides an interpretable way to examine relations that may be obscured in the original signals.
  2. The experiments include band-wise versions of existing methods and ablations of scale partitioning and joint refinement. These comparisons help distinguish the contribution of the proposed framework from that of decomposition alone.
Critical weaknesses

-- The theoretical analysis does not establish when the inferred directions represent direct causal relations. The phase result proves that averaging preserves an ordering already shared by the selected local phases. It does not show that this ordering identifies a causal direction or separates direct effects from responses to a common driver. The authors acknowledge that phase is only directional evidence, but the remaining steps do not close this gap. In particular, screening conditions on simultaneous scale components rather than their relevant histories, so delayed common causes can still produce misleading associations. The connection between the stated assumptions and reliable causal recovery therefore remains insufficiently justified.

-- The joint refinement does not clearly preserve the evidence used to construct the initial graphs. The initial stage selects edges through phase ordering and conditional screening, whereas refinement updates the graph weights using reconstruction and acyclicity. The description does not specify whether rejected edges remain excluded or whether the assigned directions are maintained. Moreover, reconstruction uses contemporaneous combinations of the scale components even though the target relations include delayed effects. Although the ablation demonstrates an average improvement and the authors describe reconstruction as a local consistency criterion, these results do not explain why the updated graphs retain the directional and conditional evidence used to select their candidates.

-- The evaluation does not clearly demonstrate recovery of the correct temporal scale. The synthetic generator provides a separate reference graph for each band, but the reported structural metrics do not specify how scale assignments enter the comparison. It is therefore unclear whether an otherwise correct edge placed in the wrong band is penalized, or how methods producing a single graph are compared with scale-specific outputs. The river visualization is informative, but the displayed distance ranges do not independently establish the temporal scale of each reference relation. This ambiguity limits the evidence for the central claim of recovering scale-resolved structure, beyond improving overall edge recovery.

-- The experimental settings are not sufficiently specified to assess the reported advantage. The synthetic construction describes the general process without giving the frequency boundaries or the distributions of delays and coefficients that determine how easily the components can be separated. The implementation discussion also leaves the actual baseline lag settings and the selection procedure for the default configuration unclear. The existing sensitivity analysis shows that the band partition affects recovery, making these omissions consequential. Without a reproducible account of these choices, it is difficult to determine whether the improvement reflects a broadly useful method or a particularly favorable match between the generated signals and the selected representation.

Questions for authors

See weaknesses.

1 comment

W1

Thank you for your careful reading. In our structural model, a parent has a direct effect if changing its value changes the target while the target’s other direct causes are held fixed [1]. In Eq. (46), increasing a lagged parent by Δ\Delta changes the target by aΔa\Delta when the other parents and the disturbance are held fixed.

For an established direct lagged relation, temporal precedence determines its direction [2]. Consider Y(t)=aX(t−ℓ)+U(t)Y(t)=aX(t-\ell)+U(t), where a>0a>0 and UU is independent of the source process XX. Under our cross-spectrum convention, the phase difference is wrap⁡(2πfℓ)\operatorname{wrap}(2\pi f\ell). When 0<fℓ<1/40<f\ell<1/4, its sign indicates that XX leads YY. T2 shows that circular aggregation preserves this order when the selected local phases lie in the same admissible quadrant.

CausalSR screens candidate edges using ridge partial correlation conditioned on the contemporaneous band components of the other observed variables (Eqs. (11)–(14)). This screen can remove an association induced by a common driver when the conditioning components capture the relevant driver history. For example, suppose X=f(Zt−a,U)X=f(Z_{t-a},U) and Y=g(Zt−b,V)Y=g(Z_{t-b},V), where UU and VV are mutually independent and the pair (U,V)(U,V) is independent of ZZ. If each lagged value of ZZ carries information beyond the other, both belong in the conditioning set. Conditioning on both blocks the common-driver path even when the response delays differ, giving X⊥Y∣(Zt−a,Zt−b)X\perp Y\mid(Z_{t-a},Z_{t-b}).

W2

We appreciate your thoughtful question. Our implementation includes a fixed-support option. For each band kk, a binary mask M(k)M^{(k)} records the directed edges retained after phase selection and conditional screening. With this option enabled, we impose A(k)=M(k)⊙B(k)A^{(k)}=M^{(k)}\odot B^{(k)} throughout optimization. A retained edge can weaken or drop out, but it cannot reverse direction, and a screened-out edge cannot reappear.

With the directions fixed, joint refinement uses information across bands to estimate the edge weights. Let uu be the source component of a retained edge and rr the current target residual after reconstruction across bands. Changing the edge weight by dd gives a squared error of ∥r−du∥2\|r-du\|^2. For nonzero uu, the minimizing adjustment is d=u⊤r/(u⊤u)d=u^\top r/(u^\top u), which reduces the reconstruction error whenever u⊤r≠0u^\top r\ne 0. We plan to study an end-to-end model that jointly learns band boundaries, lagged effects, and graph weights.

W3

Thank you for the opportunity to clarify. On synthetic data, each prediction is a triple (source, target, band). We compute F1 separately against each band’s reference graph and report the unweighted mean of the three scores.

For comparisons with methods that return a single graph, we evaluate all methods against the directed union of the reference graphs. For CausalSR, we likewise take the directed union of its predicted graphs. For example, suppose a true short-band edge is predicted in the correct direction only in the mid band. It counts as a false negative in the short band, a false positive in the mid band, and a true positive in the union.

The CausalRivers reference graph records directed connectivity between stations. At comparable propagation speeds, a longer river path implies a longer travel time. We therefore use along-river distance as a proxy for response time to interpret which station pairs appear in each band. We evaluate scale assignment on the synthetic benchmark, which provides a band label for each reference edge.

W4

Thank you for this suggestion. We choose the analysis bands based on the sampling interval and the expected response times in the system. In the matched synthetic setting, we align the analysis bands with the data-generating bands so that each scale has a unique reference graph. This lets us test whether CausalSR recovers the correct parents and edge directions at each scale. Fig. 7 examines sensitivity to this alignment by holding the analysis bands fixed and varying the data-generating band boundaries.

One limitation of CausalSR is that the response bands must be specified in advance. We plan to investigate how to learn their boundaries from data. Upon acceptance, we will release the exact frequency ranges, the sampling distributions for delays and coefficients, the lag windows used by the baselines, the tuning protocol, the random seeds, and the run commands alongside the code.

References

[1] Pearl, J. Causality: Models, Reasoning, and Inference. 2nd ed., Cambridge University Press, 2009.

[2] Granger, C. W. J. Testing for causality: A personal viewpoint. Journal of Economic Dynamics and Control 2, 329–352, 1980.

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

The paper proposes CausalSR, a method that learns one causal graph per temporal scale (short, mid, long) from multivariate time series. It applies a continuous wavelet transform and splits the coefficients into three fixed frequency bands. Within each band, it computes wavelet coherence for every pair of series. A pair becomes a candidate edge when the coherence-weighted interaction strength passes a threshold, and the direction comes from the sign of the coherence-weighted mean phase difference. Edges are then pruned by a ridge-regularized partial correlation screen. The graphs are refined with a reconstruction loss and a NOTEARS acyclicity penalty for each scale. Four theorems are given. They cover band reconstruction, preservation of phase ordering under aggregation, positive definiteness of the regularized correlation matrix, and the NOTEARS acyclicity property. Experiments use a new synthetic benchmark with scale-specific graphs, CausalRivers, fMRI, and CauseMe. The authors report higher F1 and lower SHD than PCMCI, DYNOTEARS, oCSE, pairwise Granger, and VAR-LiNGAM.

Critical strengths
  • The problem is well motivated, and the method is simple and cheap to run. The ablation includes useful controls (band-wise DYNOTEARS and band-wise PCMCI, no scale partition, no joint refinement), which suggest the scale split is the main source of gain.
  • The authors are open about weaker results. The long-scale panel in the river case study has six spurious edges, DirectLiNGAM retains F1 better on CauseMe, and the real-data graphs are described as reference graphs rather than causal truth.
Critical weaknesses
  • Edge direction comes from phase lead, and the paper never shows that this is causal. A common driver with different delays creates a phase lead between two series that have no direct link. Delays longer than a quarter period are dropped or flipped. The screening step and the same-time reconstruction loss carry no direction, so only phase and acyclicity set it.
  • The theory is weaker than claimed. Theorem 2 only says that an average of angles in (0, pi/2) stays in (0, pi/2). The other theorems are inverse transform linearity, ridge positive definiteness, and the known NOTEARS result. None of them gives identifiability, so the abstract overstates what is proved.
  • The synthetic benchmark favors the method. The generator appears to use the same bands as the method, and Fig. 7 shows a clear drop when the bands are misaligned. Baseline lag windows are not reported, and there is no wavelet or frequency-domain baseline in the main results.
Questions for authors
  • How are the short, mid, and long graphs scored against a single reference graph on CausalRivers and fMRI?
  • What lag windows and hyperparameters were used for each baseline, and was anything tuned on the synthetic data? How were the band cutoffs chosen, and are they the same as those used in the generator?
  • In Algorithm 3, is the reconstruction done at the same time and frequency with no lag? If so, what directional information does it add beyond acyclicity?
2 comments

W1

Thank you for raising this point. Under the structural definition [1], a source has a direct effect if intervening on it changes the target while the target's other direct causes are held fixed. In Eq. (46), changing a parent's lagged value by Δ\Delta changes the target by aΔa\Delta.

Consider a pure common driver, with X=f(Zt−a,U)X=f(Z_{t-a},U) and Y=g(Zt−b,V)Y=g(Z_{t-b},V). If UU and VV are mutually independent and jointly independent of ZZ, then X⊥Y∣{Zt−a,Zt−b}.X\perp Y\mid\{Z_{t-a},Z_{t-b}\}. Conditioning on both driver lags screens out the induced association even when the response delays differ.

For an established direct relation in a strictly lagged system, the cause precedes its effect [2]. For Y(t)=aX(t−ℓ)+U(t)Y(t)=aX(t-\ell)+U(t), with a>0a>0 and UU independent of the source process, the Fourier shift theorem [3] gives ϕ=wrap⁡(2πfℓ)\phi=\operatorname{wrap}(2\pi f\ell) under our cross-spectrum convention. When 0<fℓ<1/40<f\ell<1/4, the source's phase lead reflects this temporal order. T2 preserves it under circular aggregation when the selected local phases share the required quadrant.

W2

Thank you for raising the question of identifiability. The four theorems jointly support directed graph recovery from multiscale observations. T1 establishes that the band components sum to reconstruct the observed signal. Within each band, T2 and T3 support orientation and screening. When local phases reflect temporal order and satisfy the quadrant condition, T2 ensures that aggregation preserves that order. T3 guarantees an invertible regularized precision matrix, so we can score candidate associations after conditioning on other variables.

T4 constrains how the screened and oriented edges form a graph through the NOTEARS equivalence between h(A)=0h(A)=0 and acyclicity. We compare the resulting graphs with reference graphs to assess directed-edge recovery.

W3

Thank you for the suggestion. We apologize for the unclear rationale for band selection. Our synthetic benchmark tests recovery with known response bands. Suitable partitions help distinguish heterogeneous responses (Fig. 7). We currently use prior knowledge. Adaptive band selection and joint learning of scales and graphs are future directions.

MS-CASTLE learns instantaneous and lagged linear structures at stationary wavelet scales, which we map to our response bands. CMC estimates frequency-specific influences through nonlinear cross-mapping, from which we extract a directed graph. We used the authors' MS-CASTLE implementation and an independent CMC reproduction. The table reports F1 on variable-level summary graphs.

MethodF1
CausalSR0.630 ± 0.110
MS-CASTLE0.454 ± 0.211
CMC0.448 ± 0.115

Q1

Thank you for pointing this out. We apologize for the unclear scoring rule. CausalRivers provides a station-level reference graph and NetSim fMRI a connectivity matrix per simulated subject. We threshold each scale graph and take the directed union, counting each edge once. Single-graph baselines are scored directly against the same reference. This tests whether scale-resolved discovery improves variable-level edge recovery.

We evaluate scale assignment separately on synthetic data with scale-specific ground truth. An edge predicted only in the wrong band counts as a false positive there and a false negative in its true band.

Q2

Thank you for the careful reading. We apologize for the missing details. Lag windows determine the available history, while bands distinguish response scales. In the supplementary controls, PCMCI and DYNOTEARS use a maximum lag of 3, covering the generator's direct-effect delays of 1–3 steps.

We choose bands using the sampling interval and expected response scales, within the supported frequency range. Generation and analysis share band boundaries to test recovery at known scales. We use fixed settings for candidate reliability, conditional screening, matrix regularization, and graph sparsity. Upon acceptance, we will release the full configurations, selection records, seeds, and commands with the code.

Q3

Thank you for the question. Complex wavelet coefficients retain local amplitude and phase. For a narrowband delayed response, the phase difference at the same time–scale index reflects the time shift between signals [4].

Phase orientation supplies the initial edge direction. Inverse wavelet reconstruction retains these phases, so waveform error adds a phase-sensitive measure of graph fit alongside acyclicity.

References

[1] Pearl, J. Causality: Models, Reasoning, and Inference. 2nd ed., Cambridge University Press, 2009.

[2] Granger, C. W. J. Testing for causality: A personal viewpoint. Journal of Economic Dynamics and Control 2, 329–352, 1980.

[3] Smith, J. O. Mathematics of the Discrete Fourier Transform (DFT): With Audio Applications. 2nd ed., 2007.

[4] Grinsted et al. Application of the cross wavelet transform and wavelet coherence to geophysical time series. Nonlin. Processes Geophys. 11, 561–566, 2004.

Thank you for the response and the new baselines. My main concerns about novelty and soundness remain, and I will keep my original score.

  • Novelty. Each part of the method is established: cross-wavelet phase for direction, partial correlation screening, and NOTEARS. Multi-scale causal discovery also has prior work, as one of the other reviewers notes. The new MS-CASTLE and CMC comparison does not say which dataset it uses, and the paper does not yet position itself against this literature.
  • Soundness of direction. Phase lead is not a causal direction. The screening argument holds only if the common driver is observed and its lags are conditioned on. Delays longer than a quarter period are still not addressed, and Q3 is not answered.
  • Theory. The four theorems do not give identifiability, alone or together. The abstract should be revised.
  • Evaluation. The generator and the method share band boundaries, so the synthetic results favor CausalSR. Full configurations, selection records, seeds, and commands with the code are deferred until after acceptance.
Reviewer_YTdbOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes CausalSR, a time series causal discovery method that recovers causal graphs from three different frequency scales (short, mid, long). This is indeed an important problem as causal relationships in complex systems may exhibit different delays and propagate at different time scales. Studying this problem is therefore important for real world applications where such complex systems are commonplace. The proposed method is backed with theoretical analysis and extensively evaluated on synthetic and real-world data.

Critical strengths

Important problem. Multi-scale time series causal discovery is indeed an important problem, very much relevant in real world temporal complex systems.

Theory and datasets. The proposed CausalSR is backed with theory and the selected datasets are solid testbeds for this problem, especially CausalRivers due to its scale dependence relevance and real-world nature.

Critical weaknesses

The paper has completely ignored a directly relevant existing literature on multi-scale causal discovery. Causal inference and discovery from multiple time frequency scales is not new and multiple methods exist in this space (e.g. [1-6]).

The paper must be positioned properly within existing literature. This means inclusion of relevant work in at least sections 1 and 2 of the paper. The paper must also show how its ideas and the method compare to relevant existing methods. This means critical relevant baselines are missing from the experiments.

References

[1] Saberski, E., Lorimer, T., Carpenter, D., Deyle, E., Merz, E., Park, J., ... & Sugihara, G. (2024). The impact of data resolution on dynamic causal inference in multiscale ecological networks. Communications Biology, 7(1), 1442.

[2] Zhao, X., Sun, Y., Li, X., & Shang, P. (2018). Multiscale transfer entropy: Measuring information transfer on multiple time scales. Communications in Nonlinear Science and Numerical Simulation, 62, 202-212.

[3] D'Acunto, G., Di Lorenzo, P., & Barbarossa, S. (2022). Multiscale causal structure learning. arXiv preprint arXiv:2207.07908. https://openreview.net/forum?id=Ub6XILEF9x

[4] D’Acunto, G., Bonchi, F., Morales, G.D.F., Petri, G. (2024). Extracting the Multiscale Causal Backbone of Brain Dynamics. Proceedings of the Third Conference on Causal Learning and Reasoning, inProceedings of Machine Learning Research 236:265-295.

[5] Benkő, Z., Varga, B., Stippinger, M., & Somogyvári, Z. (2025). Detecting causality in the frequency domain with cross-mapping coherence. Physica D: Nonlinear Phenomena, 477, 134708.

[6] Ferdous, M. H., Stockwell, I., & Gani, M. O. (2026). DCD: Decomposition-based Causal Discovery from Autocorrelated and Non-Stationary Temporal Data. arXiv preprint arXiv:2602.01433.

Questions for authors
  1. How is the proposed CausalSR method different from existing multi-scale causal discovery methods? (theoretically and empirically)
  2. Do the performance gains come from the multi-scale representation or scale-specific causal discovery? It might be worth running other causal discovery methods on identified frequency scales to isolate the effect.
2 comments

Q1

Thank you for the references. We apologize for the incomplete literature discussion, but disagree that all six papers are essential missing baselines.

CausalSR addresses heterogeneous responses by returning separate short, mid, and long band graphs over the same variables. Each edge specifies a direct relation and its band; the graphs remain separate. Direct baselines should match these definitions and targets.

[1] includes direct and transitive dynamic effects. Its resource-to-secondary-consumer relation through primary consumers is a true influence there, but not an additional direct edge in our task. It varies temporal resolution and taxonomic aggregation, changing both nodes and what counts as correct recovery. We therefore do not treat this resolution study as a necessary direct baseline.

[2] measures predictive information transfer over coarse-graining windows. Such windows do not isolate exclusive frequency bands, and the resulting transfer profiles do not directly specify our band-specific structural graphs. Converting them requires new scale-mapping and graph-construction choices. Given these different scales and targets, we do not include MTE as a direct structural baseline.

[3], MS-CASTLE, is a close comparator. It estimates instantaneous and lagged linear SVAR coefficients at SWT scales. CausalSR lets users specify response bands through CWT and restricts initial directional evidence to coherent local regions. This gives control over the response ranges and local evidence used for discovery, without a full lag-coefficient tensor. The comparison below tests the complete pipeline.

[4] extracts a multiscale backbone shared across individuals. It uses MS-CASTLE for individual graphs, then selects shared edges. A true individual-specific edge can correctly be excluded from that backbone, whereas recovering it belongs to our task. Comparing MS-CASTLE covers the overlapping discovery stage. Population-backbone extraction is not an additional baseline for our task.

[5], CMC, measures coherence between observations and cross-mapped reconstructions, using shifted peaks for direction. CausalSR uses local wavelet phase evidence, multivariate screening and joint graph refinement. Its practical advantage is an integrated route from selected bands to network structure; CMC needs additional rules to turn pairwise spectral scores into those graphs. We compare against an independent CMC reproduction with an explicitly defined graph adapter.

The new comparison uses 10 shared synthetic datasets, 5 variables and 1,024 samples, with nominal frequency ranges aligned to SWT. Here, F1 is computed on the directed union of the three band graphs. Values are mean±SD.

MethodF1
CausalSR0.630±0.110
MS-CASTLE0.454±0.211
CMC0.448±0.115

[6], DCD, combines trend, seasonal and residual analyses into one graph. Trend/seasonal edges link a time proxy to variables; variable-to-variable mechanisms come from residuals. We instead recover variable-to-variable graphs in every band. DCD does not estimate this target and is not a direct baseline.

Q2

Thank you. We reproduce the controls reported in Appendix F.7/Fig.6 below.

MethodF1
CausalSR0.77
Without joint refinement0.71
Band-wise DYNOTEARS0.58
Band-wise PCMCI0.50
Without scale partitioning0.45

The reported controls show higher F1 for the complete pipeline on the tested configuration; decomposition alone does not determine the result. We accept fuller literature coverage, but not six mandatory comparisons across different targets. We directly compare the close methods and give specific reasons for excluding the others. With these clarifications and evidence, we respectfully ask you to reconsider the clear-rejection rating.

(Q1) I appreciate the commentary on the suggested literature and additional results, but as far as I can see none of these were added to the manuscript. It is critical to add these to the paper so the work is properly positioned within the literature. As the other reviewer rightly pointed out, the details of how these additional results were obtained are not provided.

(Q2) Thank you for pointing me to Figure 6, this is indeed what I had in mind.

I also share concerns with other reviewers around novelty. This is partly why I recommended that the authors look more closely into the existing literature because effectively this work re-packages existing results as a new method, making it very incremental.

Since the critical weakness has not been addressed in the paper at all (i.e. no PDF revision), I am also not updating my original score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.