acceptodds
Withdrawn from ICLR 2027

S2S-Diff: Structure-to-Signal Diffusion Models for Time Series Synthesis

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Time-series synthesis is fundamental to simulation, data augmentation, and privacy-preserving data sharing. Despite recent advances in generative modeling, most approaches generate sequences directly in high-dimensional data or latent spaces without modeling an explicit structural representation, limiting interpretability and controllability. To address this limitation, we propose S2S-Diff, a structure-to-signal cascaded diffusion model that synthesizes time series through sparse structural composition. S2S-Diff represents each sequence as a sparse atom-set composition over a learned multi-scale structural dictionary, where the active atoms summarize the dominant structural characteristics of the sequence. Generative modeling is performed in this compact structural space, and sampled structural compositions condition a diffusion model for signal realization. By decoupling high-level structural composition from fine-grained signal realization, S2S-Diff reduces the complexity of generative modeling while introducing an explicit structural inductive bias and preserving stochastic flexibility. Experiments across diverse benchmarks show that S2S-Diff achieves strong synthesis quality while maintaining structural sparsity and controllability. These findings support structural abstraction as a principled basis for interpretable and controllable time-series synthesis.

Primary area:
learning on time series and dynamical systems
Keywords:
Time‑series synthesis, Diffusion model, Structure‑to‑signal generation, Multi‑scale structural dictionary, Sparse compositional modeling, Interpretable generative model, Controllable time‑series generation

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_s2EEOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

S2S-Diff proposes a two-stage generative model for time series. It first learns a multi-scale dictionary of temporal atoms, and then maps each training sequence to atom-matching activations. However, the claims of sparse structural composition, controllability, and improved long-range modeling remain insufficiently substantiated, while inconsistencies in the formulas and reported results limit confidence in reproducibility.

Critical strengths
  1. Making a learned atom representation an explicit sampled variable is an interesting way to separate pattern selection from signal realization.
  2. The paper covers nine datasets from varied domains, compares multiple generative model families, and includes several ablations studies.
Critical weaknesses
  1. The Introduction identifies a limitation of prior methods: they struggle to reconcile long-range structure with fine-grained dynamics. However, the paper does not clearly explain how the proposed atom-set representation captures long-range temporal organization.

  2. The two purportedly complementary activation channels are deterministically redundant. From Equations (4)-(6), for nonnegative distances and fixed scale factor epsilon_s, p_hat[s,m,d] = exp(-(epsilon_s * dist_min[s,m,d])^2). Thus p_hat is a strictly monotone transform of dist_min, and one contains no additional information about the matching distances beyond the other.

  3. Section 3.2 does not specify how atom values are initialized, how lengths are rounded for each dataset, or how the AttnFuse operation in Equation (7) is implemented. These choices affect the learned structural space and should be reported explicitly.

  4. Sections 3.3-3.4 describe a Transformer prior over activations and a conditional Transformer signal denoiser. However, the paper does not explicitly state whether the two denoisers share any parameters. Equation (12) also switches from epsilon_phi in the text to epsilon_psi in the objective.

  5. Section 3.5 establishes that S2S-Diff samples a structural variable from its own prior, whereas the cited alternatives use structural cues as guidance. This distinguishes model formulations, but does not by itself show that sampling structure separately is preferable.

  6. Figure 4 labels two groups as "class 1" and "class 2," but neither the caption nor Section 4.3 names the dataset, defines the classes, or explains how samples and atoms were selected.

  7. Algorithm 1 uses sqrt(alpha_k) and sqrt(1-alpha_k) to noise clean activations and signals at arbitrary step k, whereas Equation (16) uses cumulative alpha_bar_k.

  8. Section 4.1 states that Context-FID improves by 50% on all nine datasets. Relative to the strongest baseline in Table 1, C.T. improves from 0.385 to 0.228 (about 40.8%) and J.V. from 0.119 to 0.113 (about 5.0%). The claim should be recalculated or narrowed.

  9. Equation (8) defines lambda_1 for alignment, lambda_2 for atom L1 sparsity, and lambda_3 for diversity, with values 0.5, 0.001, and 0.1. Appendix H instead labels its sparsity sweep lambda_1 (centered at 0.001), diversity sweep lambda_2 (0.1), and alignment sweep lambda_3 (0.5). Please reconcile the notation so the reported configuration and sensitivity results are unambiguous.

Questions for authors

See Weaknesses.

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

This paper proposes S2S-Diff, a structure-to-signal cascaded diffusion framework for time series generation. It introduces multi-scale structural dictionary to extract structural activations through atom matching and attention-based fusion, with alignment, sparsity, and diversity regularization. A two-stage diffusion framework then models the distribution of structural activations and generates time series conditioned on the sampled structures. Experiments on nine public real-world datasets show improved generation quality over existing baselines.

Critical strengths
  • Intuitive hierarchical generation framework. The two-stage design separates structural representation generation from fine-grained signal synthesis, providing a clear and intuitive framework for incorporating structural information into the generative process.

  • Structural modeling for interpretability. By introducing learnable structural atoms and structural activations, the framework represents temporal patterns, offering a promising direction for understanding the role of structural information in diffusion-based generation.

Critical weaknesses
  • *Insufficient experiments"

    ---comparison with existing shapelet learning methods. The proposed structural matching and activation construction mechanisms share similarities with existing shapelet learning approaches, particularly in learning representative temporal subsequences. A more detailed discussion and comparison with shapelet learning methods would help better establish the novelty of the proposed structural representation.

    ---Insufficient analysis of experimental results. Although the paper presents extensive quantitative comparisons, the discussion of the results remains relatively limited. For instance, TSGDiff achieves comparable results to S2S-Diff on several metrics and datasets, butperforms noticeably worse in terms of Context-FID and Correlational scores on the EEG dataset. A more detailed analysis of such dataset-dependent differences could provide deeper insights into when and why the proposed approach is particularly effective.

  • Limited validation of the regularization objectives. The three regularization terms are designed to encourage alignment, sparsity, and diversity, respectively. However, the existing ablation studies mainly evaluate their impact on generation quality without directly examining whether they achieve their intended effects. Quantitative measurements or visualizations of atom sparsity, inter-atom similarity and alignment quality would provide stronger evidence for their effectiveness.

Questions for authors
  1. Temporal ordering in structural activations. The proposed structural activation representation appears to discard the explicit occurrence positions and temporal ordering of structural atoms through min/max pooling (Equations 5–7), potentially making subsequent signal realization more challenging. However, Figure 4 appears to show a correspondence between structural atom indices and their occurrence order in the time series. Does this correspondence emerge naturally from the learned representation? How does the model preserve or recover temporal ordering information during generation?

  2. Interpretability of local discriminative patterns. Figure 4 mainly highlights the most responsive structural atoms shared across samples of the same class, illustrating common structural characteristics. However, it remains unclear whether the learned dictionary can also capture subtle but discriminative local patterns. Could the authors provide additional visualizations or quantitative analyses to demonstrate whether such fine-grained discriminative patterns are effectively represented?

Reviewer_4f6tOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes S2S-Diff, a structure-to-signal diffusion framework for multivariate time series generation. The main idea is to introduce an explicit structural representation through a learnable multi-scale dictionary, where each time series is represented by a set of structural atom activations. The framework first learns a generative distribution over these structural activations and then uses them to condition a second diffusion model for signal generation. Experiments on nine time series datasets show improvements over several existing generative baselines, particularly in Context-FID and correlation-based metrics. The authors further claim that the proposed structural representation improves interpretability, controllability, and generative modeling efficiency.

Critical strengths

The idea of introducing explicit structural information into the generative latent space is interesting and makes sense for time series, where local patterns, trends, and recurring structures often play important roles. Separating high-level structural information from detailed signal generation is a reasonable approach, and the reported improvements in generation fidelity suggest that such structural conditioning can be useful.

Critical weaknesses
  • The proposed representation does not convincingly establish sparse structural composition. Although the paper repeatedly emphasizes sparse structural decomposition, the actual implementation is closer to a dictionary-based similarity representation than an explicit decomposition. The L1 regularization is applied to the dictionary atoms rather than their activations, so the sparsity of the resulting structural representation is not clearly established. Moreover, the min/max aggregation over temporal positions discards information about where and in what order the structural patterns occur, which seems particularly problematic for modeling temporal composition. The paper also provides little discussion of the identifiability or stability of the learned structural factors, and it is unclear whether different atom configurations can reliably represent distinct temporal structures.

  • The theoretical and experimental evidence does not sufficiently support the claimed benefits of structural decomposition. The current ablation studies mainly remove individual regularization terms or activation components, but they do not establish why the proposed sparse structural representation is necessary compared with a conventional dense latent representation of comparable size. More direct experiments on structural reconstruction, atom-level interventions, and controllability would be needed to demonstrate that the learned atoms are meaningful and can reliably control the generated signals. In addition, the structural activation space has dimension (2M\times D), which is (40D) under the default setting, compared with (48D) for the original input sequences. This is only a modest reduction, and the paper provides insufficient computational evidence to support its claim of reduced generative complexity.

Questions for authors

Can the authors provide a more direct evaluation of the proposed structural representation? For example, comparisons with a dense latent representation of comparable dimensionality, together with atom-level intervention or structure-consistency experiments, would help clarify whether the improvements come from meaningful structural decomposition or simply from additional learned conditioning features.

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

Related papers

Discussion (0)

Sign in to comment.