acceptodds
Under review as a conference paper at ICLR 2027

Graph2TS: Structure-Conditioned Time Series Generation from Quantile Graphs

Abstract

Time-series generation requires capturing recurring temporal structure while preserving stochastic variation across its realizations. We approach this problem from a structure–residual perspective, representing a time series through a structural condition while modeling the remaining variation stochastically. Based on this perspective, we introduce , a structure-conditioned time-series generator that uses quantile graphs as structural representations. A quantile graph compactly summarizes empirical one-step transitions between globally shared amplitude-quantile states, providing a fixed-size, lossy representation of distributional and local transition structure. Graph2TS conditions generation on this representation and uses a latent variable to model variation not specified by the graph. All graphs and normalization statistics are derived from training data, and generation requires no test-specific conditioning information. Experiments against GAN, diffusion, and flow-matching baselines on four real-world datasets show particularly strong performance on ECG and EEG signals, while remaining competitive on Sunspot and Electricity with dataset-dependent trade-offs. Dataset analysis further suggests that these benefits are most pronounced for volatile signals with weak or irregular periodicity. Additional Electricity experiments up to length 256 show stronger relative performance over DiffusionTS at longer sequence lengths, although coverage decreases as sequence length increases. These results support the structure–residual perspective and suggest graph-based structural representations as a promising direction for time-series generation.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Time Series Generation, Quantile Graph, Structure-Controlled Generation, Cross-Modal Generation, Structure-Residual Perspective

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_gs35OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Graph2TS, a structure-conditioned time series generation framework based on quantile graphs. The method represents temporal structure through first-order transitions between value-based quantile states and employs a conditional VAE to generate time series from graph embeddings and stochastic latent variables. A combined loss is used for training Graph2TS model, including reconstruction, KL regularization, contrastive alignment, and distribution-matching losses. Experiments on four datasets demonstrate good performance, particularly on ECG and CHB-MIT. Additional experiments investigate the effects of sequence length, structural conditioning, stochasticity, and quantile resolution.

Critical strengths
  1. Incorporating quantile graphs as structural conditions for time series generation is an interesting and potentially novel direction, although converting raw time series to quantile has been used in the time series literature for a long time.

  2. The proposed framework is easy to follow, with additional theoretical analysis providing a useful interpretation of the method.

Critical weaknesses
  1. Insufficient motivation. The choice of quantile graphs is not sufficiently justified, and their advantages over alternative structural representations remain unclear. Especially it is a concern that the information lost in quantile conversion from time series, including the long-term dependency, could cause a lot of negative impact.

  2. Limited evaluation. The experiments lack VAE-based baselines, cover relatively few and less common datasets, and do not evaluate the downstream task performance of generated data.

Questions for authors
  1. it is a concern that the information lost in quantile conversion from time series, including the long-term dependency, could cause a lot of negative impact. Can you analyze it both qualitatively and using experiments?

  2. Downstream task performance.It would be valuable to include a evaluation to assess the downstream task performance of the generated data beyond statistical fidelity.

  3. Dataset selection. The evaluation covers only four datasets, some of which are less commonly used in time series generation. I wonder whether the conclusions would hold on more widely adopted datasets as those used in TSFlow or DiffusionTS.

  4. VAE-based baselines. Given the VAE-based architecture, I would like to see comparisons with representative VAE-based time series generators, such as TimeVAE [1].

[1] Desai, Abhyuday, et al. "Timevae: A variational auto-encoder for multivariate time series generation." arXiv preprint arXiv:2111.08095 (2021).

1 comment

We thank the reviewer for recognizing the potential novelty of quantile-graph conditioning and the theoretical interpretation of Graph2TS. We address the concerns below.

  1. Motivation for quantile graphs and information loss (W1, Q1)

Quantile graphs are chosen because they provide a compact, fixed-size representation of empirical temporal transitions in a globally shared amplitude state space. Importantly, the information loss introduced by quantile conversion is intentional rather than an overlooked limitation. As discussed in Section 4.4, Graph2TS follows a structure–residual perspective: the graph captures selected transition characteristics, while the stochastic conditional generator models variations not specified by the graph, including potentially longer-range dependencies. This does not guarantee complete recovery of lost information, but avoids requiring the graph itself to preserve the entire sequence. The existing experiments also provide empirical evidence. Table 3 shows that removing graph conditioning increases ACF error from 0.051 to 0.088 and PSD discrepancy from 0.0059 to 0.0179. Furthermore, Table 5 demonstrates relatively stable performance across quantile resolutions from Q = 5 to 20. These results support the usefulness of the proposed structural representation despite its lossy nature.

  1. Downstream task evaluation (W2, Q2).

Our study focuses on generative fidelity and distributional coverage rather than downstream task augmentation. Accordingly, we evaluate marginal distributions, temporal correlations, spectral characteristics, representativeness, and coverage using nine complementary metrics. These evaluations directly address the stated objective of generating realistic and diverse time series. While downstream task evaluation provides an additional perspective, its results depend on the choice of auxiliary models, tasks, and training configurations, which may limit their representativeness and generalizability. We will consider broader downstream task evaluations in future work.

  1. Dataset selection (W2, Q3).

The four datasets cover distinct temporal characteristics, including physiological signals (ECG and EEG), solar activity (Sunspot), and electricity consumption (Electricity). This diversity allows us to examine how structural conditioning behaves across different signal characteristics. While broader benchmark coverage would further strengthen the generalizability assessment, the current experiments provide evidence of the method’s effectiveness across heterogeneous time-series characteristics.

  1. VAE-based baselines (W2, Q4).

Although Graph2TS adopts a conditional VAE architecture, its primary contribution lies in structural conditioning rather than the VAE backbone itself. Our evaluation includes representative GAN-, diffusion-, and flow-matching-based generators, particularly TSFlow (ICLR 2025), which represents a more recent generative modeling paradigm than conventional VAE approaches. Furthermore, Table 3 directly isolates the contribution of graph conditioning through a controlled ablation under the same architecture. We believe these comparisons provide meaningful evidence of Graph2TS’s effectiveness against both established and recent generative approaches.

In summary, Graph2TS deliberately uses a lossy structural representation to separate selected transition information from residual stochastic variation. The theoretical formulation, controlled ablations, and evaluation across diverse datasets collectively support this design choice and its effectiveness within the investigated settings.

Reviewer_LLNWOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Graph2TS, where from the training data, a corpus of quantile-based graphs is extracted as structured representations, and the training process for generative modeling involves training a decoder that generates a time series conditioned on the structured representation. During inference, the structured representation is sampled from the training corpus, and the decoder is invoked with the sampled graph as input to generate the time series. Due to this inductive bias, the results suggest that the approach naturally works well for datasets with solid structural representations, like ECG, and the approach is also competitive to SOTA baselines on other datasets.

Critical strengths
  1. The paper is clearly written, the approach is sound, and it works as intuitively expected.

  2. The quantitative results are impressive

Critical weaknesses
  1. By using a transition matrix, the approach assumes that the future can be represented by conditioning on the present values. This seems particularly restrictive in the time series domain, where the time series processes can be non-Markovian in nature. I understand that this non-Markovian part can be modeled as the “residual”. However, a fundamental question to clarify would be the following: What is the contribution of the structured graph representation vs residual modeling for datasets where the time series processes have Markovian transitions v.s. Non-Markovian transitions.

  2. Similarly, can the authors explain how the approach handles non-stationarity? From my understanding, the one-step transition based on magnitude ignores the non-stationarity aspect, and in the Graph2TS modeling paradigm, it will be offloaded to the residual modeling part. Is this the reason why the approach performs well in datasets such as ECG, where there is less non-stationarity?

Questions for authors

Check the critical weaknesses. Additionally,

  1. For generative time series tasks, multiple papers typically report the discriminative and predictive scores proposed in the TimeGAN paper. Can those scores be added?

  2. How can this be extended to conditional generation where conditions are continuous? You would need a graph generator as well, and it is not a simple extension.

I am willing to update my score towards acceptance if these questions are suitably answered.

1 comment

We appreciate the insightful questions and provide the following clarifications.

  1. Markovian vs. non-Markovian dependencies (W1). Graph2TS does not assume that the underlying time-series process is first-order Markovian. The quantile graph summarizes empirical one-step transitions rather than defining a Markov generative process. It serves as a lossy structural condition, while the conditional decoder learns the distribution of complete sequences from paired training data. Therefore, higher-order dependencies are not explicitly encoded in the graph but can be captured by the learned conditional distribution. The structure–residual formulation is designed precisely to accommodate information not specified by the graph.

  2. Non-stationarity (W2). Graph2TS does not assume stationarity. The quantile graph can capture transition characteristics associated with non-stationary dynamics through its weighted edges, although aggregating transitions within each window does not preserve their complete temporal evolution. The conditional generator further models variations not explicitly represented by the graph. Moreover, Appendix A.1 shows that ECG exhibits substantially heavier-tailed increments than Sunspot , indicating pronounced local fluctuations. This suggests that capturing local transition structure may contribute to Graph2TS’s strong performance on ECG.

  3. Discriminative and predictive scores (Q1). Discriminative and predictive scores, as introduced in TimeGAN, provide useful task-based evaluations. However, they depend on auxiliary classifiers or predictors and their training configurations. Our evaluation instead covers four complementary dimensions: marginal distribution fidelity (Wasserstein and KS), temporal and spectral characteristics (ACF and PSD), representativeness (prototype error and MDR), and distributional coverage. These metrics directly assess different properties of the generated sequences and provide a comprehensive evaluation without relying on auxiliary predictive models.

  4. Continuous conditional generation (Q2).

Graph2TS can be naturally extended to continuous conditional generation through two approaches. First, interpolation between training-derived quantile graphs can provide continuously varying structural conditions. Second, a conditional graph generator can be introduced to learn the mapping from continuous conditions to graph representations. Both approaches are compatible with the existing Graph2TS architecture without modifying the time-series decoder. As noted in Section 4.2, graph interpolation and generation are already supported conceptually by our formulation, although their empirical evaluation is beyond the scope of this work.

In summary, Graph2TS does not rely on Markovian or stationary assumptions. Instead, it uses quantile graphs as compact structural summaries while learning the remaining sequence variability through conditional generative modeling. We believe this distinction addresses the central concerns regarding the applicability and interpretation of the proposed framework.

Reviewer_MXfbOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents Graph2TS, a structure-conditioned generator for univariate time series generation. The authors argue that matching the distribution of values does not guarantee realistic temporal structure, and that realistic individual samples do not guarantee good coverage of the data. Graph2TS uses a conditional VAE to create a time-series from its quantile graph and a random latent variable. On four real datasets with 32-step windows, the model outperforms on ECG and EEG, and shows a mixed results on sunspot and electricity dataset.

Critical strengths
  1. The quantile graph is computed in one pass over each window, and its size depends only on the number of quantile bins, not on the length of the time-series.
  2. The model is light and generates synthetic time-series in one step. Sampling runs the graph encoder and decoder once, without the iterative steps that diffusion and flow-matching models need.
  3. I appreciate the honesty of the authors for reporting the mixed results, on Sunspot and Electricity dataset, and the drop in coverage for longer time-series. It also states that its proposition does not guarantee latent disentanglement.
  4. The paper evaluates the value distribution, temporal and spectral structure, and how well the samples cover the real data.
Critical weaknesses
  1. The work is incremental. Graph2TS is a conditional VAE [1] whose condition is the quantile-bin transition from earlier quantile graph work [2].
  2. All parts of the model are MLPs over flattened inputs, so it has no temporal parts in the model itself. A new model needs to be trained for each sequence length.
  3. No metrics checks whether the model learns the distribution or copies its training data. I would suggest to use the Authenticity metrics from Alaa et al. paper [3].
  4. The ablation runs only on the ECG dataset and without the graph the prototype error and Coverage value barely changes, so I am wondering regarding the necessity of the graphs.
  5. In the experiments, the author only uses univariate time-series data.
  6. For the long-sequenced generation, only one dataset has been tested. Longer sequence generation with multiple dataset would benefit the paper.

References

[1] Sohn, Kihyuk, Honglak Lee, and Xinchen Yan. "Learning structured output representation using deep conditional generative models." Advances in neural information processing systems 28 (2015).

[2] Campanharo, Andriana SLO, et al. "Duality between time series and networks." PloS one 6.8 (2011): e23378.

[3] Alaa, Ahmed, et al. "How faithful is your synthetic data? sample-level metrics for evaluating and auditing generative models." International conference on machine learning. PMLR, 2022.

Questions for authors
  1. How close are the generated samples to their nearest training window, compared with the distance between test windows and their nearest training window?
  2. Which units does each metric use, and how can Diffusion-TS on ECG reach a Wasserstein distance of 0.114 while its prototype error is 596.0?
1 comment

Thank the reviewer for the detailed assessment and recognition of Graph2TS’s efficiency and comprehensive evaluation. We would like to clarify several points that we believe are important for evaluating the contributions of this work.

  1. Novelty (W1).

While Graph2TS builds upon existing CVAE and quantile graph techniques, its contribution extends beyond their direct combination. We introduce a structure–residual formulation that separates structural conditioning from stochastic variation, instantiated through a probabilistic generative framework. Unlike graph-based reconstruction approaches, Graph2TS learns a conditional distribution and generates diverse realizations from training-derived graph conditions without requiring test-specific information.

  1. Temporal modeling (W2).

The use of MLPs does not imply the absence of temporal information. Quantile graphs explicitly encode transitions between successive amplitude states, providing temporal structural information to the decoder. Although this representation does not preserve complete temporal ordering, it captures local transition dependencies without requiring recurrent or attention-based architectures.

  1. Memorization (W3, Q1).

The reviewer raises a valid question regarding sample-level authenticity. However, memorization assessment and generative quality evaluation address distinct objectives. Our experiments evaluate generated samples against held-out test data using distributional, temporal, spectral, and coverage metrics. Importantly, Graph2TS uses only training-derived graph conditions, with no test-specific information during generation. While these results do not establish the absence of memorization, they demonstrate generalization in the evaluated distributional and structural metrics.

  1. Graph ablation (W4).

We respectfully disagree that the ablation results suggest limited utility of graph conditioning. Table 3 shows that removing graph conditioning increases Wasserstein distance from 0.014 to 0.049, ACF error from 0.051 to 0.088, and PSD discrepancy from 0.0059 to 0.0179. These correspond to approximately 3.5×, 1.7×, and 3.0× larger errors, respectively. Although prototype error and coverage change less substantially, the consistent improvements across all six reported metrics support the contribution of graph conditioning, particularly to distributional and temporal fidelity.

  1. Experimental scope (W5, W6).

Graph2TS focuses on univariate time-series generation, which is the explicitly defined scope of this study. Regarding sequence length, Table 2 demonstrates that Graph2TS outperforms DiffusionTS on all nine reported metrics at T = 256, despite decreasing absolute coverage. These results provide evidence of effectiveness beyond the standard 32-step setting, although broader long-sequence generalization remains to be established.

  1. Metric clarification (Q2).

The Wasserstein distance and prototype error measure different aspects of generative quality and are not directly comparable in magnitude. Their precise definitions and computation procedures are provided in Appendix B.7. They therefore operate on different scales.

In summary, Graph2TS provides a structure-conditioned generative framework that explicitly incorporates quantile-transition information while modeling residual stochastic variation. The experimental results demonstrate its effectiveness, particularly on ECG and EEG, and the ablation results provide direct evidence of the contribution of graph conditioning.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.