acceptodds
Under review as a conference paper at ICLR 2027

HiGen: Hierarchical Generation of Irregular Multivariate Time Series

Abstract

Irregular multivariate time series are widely observed in real-world systems, where variables are recorded asynchronously and with varying frequencies. Generating such data is useful for data augmentation, data sharing, and downstream task training. Existing methods usually regularize irregular observations into aligned sequences, discretize them into predefined fixed time bins, or model them with a single smooth continuous-time state evolution that struggles to characterize irregular sampling rhythms. These designs change the original sampling process and struggle to preserve realistic temporal dynamics and asynchronous variable interactions. We propose HiGen, a hierarchical generation framework for irregular multivariate time series. HiGen directly models the original sample-point sample and organizes generation at the sample, window, and point levels. This hierarchy is used to model temporal dynamics through real time intervals at window and point level, and to model variable interactions through sample, window, and point levels. In this way, the model preserves both irregular sampling rhythms and asynchronous variable relationships. Experiments on irregular multivariate time series datasets show that HiGen achieves better downstream utility and generation fidelity than state-of-the-art baselines.

Primary area:
learning on time series and dynamical systems
Keywords:
irregular time series, data generation

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_hLgpOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2(+1)
Presentation
3
Soundness
2(+1)
Read review
Paper summary

This paper introduces HiGen, a hierarchical framework for generating irregular multivariate time series. By directly modeling the original observations through a Sample–Window–Point hierarchy, HiGen captures global characteristics, temporal dependencies across windows, and local interactions among asynchronously observed variables. To account for irregular sampling intervals, the framework further incorporates a learnable time-bin decoder. Extensive experiments on P12, Activity, and CesNet demonstrate promising results in downstream classification, generation fidelity, and the preservation of temporal irregularity, particularly in low-data regimes.

Critical strengths
  1. Meaningful problem and well-motivated framework. The paper addresses the important challenge of irregular multivariate time series generation by directly modeling the original observations, thereby avoiding potential information loss associated with temporal alignment and interpolation.

  2. Promising experimental results. The proposed Sample–Window–Point hierarchy offers a principled approach to capturing temporal dynamics and interactions among asynchronously observed variables. HiGen demonstrates strong empirical performance across three datasets, particularly in data-scarce settings.

Critical weaknesses
  1. Serious concerns regarding reproducibility. Based on the authors' linked code repository, missing the train.py file referenced in the README, along with three training entry points required by the low-data scripts. Furthermore, the released PRDC results (outputs/p12_lstm_ae_prdc.csv) report precision and recall values of approximately 0.9076 and 0.9107, respectively, which differ substantially from the corresponding values of 0.6430 and 0.6388 reported in Table 2, with no explanation provided for this discrepancy. Additional inconsistencies are evident in the experimental configurations: the checkpoint metadata specifies num_time_bins=8, whereas Appendix C states that K=10 is used for P12. Similarly, the low-data scripts default to training data proportions of 50%/70%/90%, rather than the 5%/10%/15% reported in the paper. Collectively, these discrepancies raise significant concerns about the reproducibility and reliability of the reported results, especially given the unusually large performance improvements.

  2. Insufficient experimental validation. The experimental evaluation lacks comparisons with recent time series generation methods, including Diffusion-TS (ICLR 2024) [1], L2D-Diff (ICLR 2026) [2], K-ProtoDiff (AAAI 2026) [3], and MA-TSD (ICML 2025) [4]. While some of these methods may require adaptation to handle irregularly sampled data, their omission limits the comprehensiveness of the evaluation. Moreover, the experiments are restricted to three relatively small datasets, with no systematic investigation of performance under varying sampling densities or degrees of temporal irregularity. The existing evaluation metrics also provide limited insight into the model's ability to preserve long-range temporal dependencies and cross-variable correlations. Additional metrics inspired by TimeGAN [5], such as discriminative and predictive scores, could offer complementary evidence if appropriately adapted to irregularly sampled data. Finally, the small font sizes in Figures 5 and 6 hinder readability.

[1] Yuan, X., & Qiao, Y. (2024). Diffusion-TS: Interpretable Diffusion for General Time Series Generation. ICLR 2024. [2] Shen, L., et al. (2026). Latent-to-Data Cascaded Diffusion Models for Unconditional Time Series Generation. ICLR 2026. [3] Duan, Y., Lin, L., & Wu, X. (2026). K-ProtoDiff: Key Prototypes-Guided Diffusion for Time Series Generation. AAAI 2026. [4] Wang, C., et al. (2025). A Non-isotropic Time Series Diffusion Model with Moving Average Transitions. ICML 2025. [5] Yoon, J., Jarrett, D., & van der Schaar, M. (2019). Time-series Generative Adversarial Networks. NeurIPS 2019.

Questions for authors
  1. Based on the authors' linked code repository, several training scripts appear to be missing. Could the authors provide the missing files and complete reproduction instructions?

  2. Could the authors clarify the discrepancies between the released PRDC results and Table 2, as well as the inconsistencies in num_time_bins and low-data training proportions?

  3. Why were recent methods, such as Diffusion-TS (ICLR 2024), L2D-Diff (ICLR 2026), K-ProtoDiff (AAAI 2026), and MA-TSD (ICML 2025), excluded? Could the authors provide additional comparisons under appropriately adapted protocols?

  4. Based on the authors' linked code repository, the evaluation covers only three datasets, two with fewer than 500 sequences. Could the authors include larger datasets or evaluate robustness across different sampling densities, missingness levels, and degrees of irregularity?

  5. Could the authors further evaluate long-range temporal dependencies and cross-variable correlations, potentially using discriminative and predictive scores adapted from TimeGAN (NeurIPS 2019)?

4 comments
AuthorsOpenReview ↗

Response to Reviewer hLgp (Part Ⅰ)

Thank you for your thoughtful comments and constructive suggestions. They have helped us identify points that need clearer explanations and stronger experimental support. We address your concerns below and describe the corresponding revisions.

Response to Q1: Missing Training Files and Complete Reproduction Instructions

We organize the materials into a complete reproduction checklist. We provide the train.py file referenced in the README and the training entry points required by the low-data scripts for the three datasets.


Response to Q2/W1: Discrepancies in PRDC, the Number of Time Components, and Low-Data Training Proportions

Thank you for pointing out these inconsistencies.

PRDC naming. The precision/recall values of 0.9076/0.9107 come from hard-threshold PRDC. Table 2 reports 0.6430/0.6388 using a PRDC version based on continuous neighborhood affinity. These are different metrics. All PRDC results in the paper use the affinity-based version. The paper does not mix the two metrics. We have corrected the metric in the code repository.

K. We verified that the released P12 checkpoint uses K=8. The corresponding configuration in Appendix C should be changed to the actual value. This was a typo, and we have corrected it.

Low-data training proportions. The released scripts incorrectly used 50%/70%/90% as their default training proportions. These defaults were not used in the actual training runs. The experiments reported in the paper used 5%/10%/15%, corresponding to r5/r10/r15. We have corrected it.


Response to Q3: Comparisons with Recent Generative Methods

We added comparisons with Diffusion-TS, K-ProtoDiff, and MA-TSD after adapting them to our task. The table below shows synthetic-only LSTM results at r10. Each cell reports AUROC / AUPRC. Our model outperforms all these new baselines.

Table R16. Recent Generative Baselines at r10.

Method and RatioP12ActivityCesNet
Diffusion-TS0.6570 ± 0.0372 / 0.2233 ± 0.02420.8740 ± 0.0377 / 0.7699 ± 0.04650.9637 ± 0.0080 / 0.8873 ± 0.0165
K-ProtoDiff0.5934 ± 0.0173 / 0.1919 ± 0.01730.8448 ± 0.0465 / 0.7058 ± 0.03110.9700 ± 0.0069 / 0.9086 ± 0.0164
MA-TSD0.6184 ± 0.0139 / 0.1990 ± 0.01000.8448 ± 0.0232 / 0.7675 ± 0.03980.6563 ± 0.0505 / 0.3691 ± 0.0418
HiGen0.7186 ± 0.0161 / 0.2350 ± 0.01280.9011 ± 0.0199 / 0.8072 ± 0.04050.9946 ± 0.0037 / 0.9834 ± 0.0091

We have not obtained a verifiable official implementation of L2D-Diff, so we did not reproduce it in this round.


Response to Q4: Larger Datasets and Controlled Irregular Sampling

To address the concern about larger datasets, we added UEA LSST, which contains 4,925 sequences, 14 classes, and 6 variables. We used its official train/test split of 2,459/2,466 sequences.

To evaluate robustness across different sampling densities and missingness levels, we separately sampled each variable’s observations at the densities of 20%, 40%, 60%, and 80% as the training data, giving rise to 80%, 60%, 40%, and 20% missing rates.

Table R17. Synthetic-Only LSTM Results on LSST at r10.

DensityModelAUROCAUPRC
20%TimeDiff0.51980.0747
20%TimEHR——
20%KoVAE0.47220.0751
20%GT-GAN0.51480.0855
20%HiGen0.63260.1701
40%TimeDiff0.53190.0914
40%TimEHR——
40%KoVAE0.50030.0842
40%GT-GAN0.49780.0845
40%HiGen0.60470.1548
60%TimeDiff0.51830.0928
60%TimEHR0.54440.0971
60%KoVAE0.45310.0705
60%GT-GAN0.49930.0863
60%HiGen0.61120.1518
80%TimeDiff0.51600.0866
80%TimEHR0.59060.1094
80%KoVAE0.47770.0768
80%GT-GAN0.49620.0878
80%HiGen0.69860.1917

HiGen has the highest mean AUROC/AUPRC under the listed conditions. The "—" entries for TimEHR mean that it could not generate valid synthetic samples because the input data were scarce. This experiment broadens the coverage of domains and classes while controlling sampling density.

AuthorsOpenReview ↗

Response to Reviewer hLgp (Part Ⅱ)

Response to Q5: Long-Range Temporal Dependencies and Cross-Variable Correlations

Following your suggestion, we added Discriminative and Predictive Scores and a Lagged Value Correlation MAE on all three datasets. These evaluations examine overall sequence fidelity, temporal predictive relationships over different time gaps, and cross-variable correlations, respectively.

To evaluate overall sequence fidelity, the Discriminative Score measures how well an independent sequence classifier distinguishes real from synthetic sequences. Its inputs include observed values, masks, and actual time intervals. A lower score means that real and synthetic sequences are harder to distinguish.

To evaluate temporal dependencies over different forecast horizons, we train a multivariate predictor on synthetic data and test it on held-out real data. It uses past observations of all variables to predict future values at short, medium, and long horizons. The Predictive Score averages the prediction errors; lower values indicate better transfer of temporal predictive relationships.

To directly evaluate cross-variable correlations across time, we compare Pearson correlations between different variables at different time lags, without interpolation. The Lagged Value Correlation MAE measures the average absolute difference between real and synthetic correlations; lower values indicate closer recovery.

Table R18. Sequence Fidelity, Multi-Horizon Prediction, and Lagged Cross-Variable Correlations at r10.

DatasetModelDiscriminative Score ↓Predictive Score ↓Lagged Value Correlation MAE ↓
P12HiGen0.2883 ± 0.01880.6006 ± 0.00100.1888
P12TimEHR0.4929 ± 0.00070.5989 ± 0.00250.3510
P12TimeDiff0.5000 ± 0.00002.8231 ± 0.00810.2236
P12GT-GAN0.5000 ± 0.00000.6514 ± 0.00210.4983
P12KoVAE0.4996 ± 0.00070.6411 ± 0.00430.3912
ActivityHiGen0.3556 ± 0.01270.6228 ± 0.00190.2558
ActivityTimEHR0.5000 ± 0.00000.7884 ± 0.01210.2896
ActivityTimeDiff0.5000 ± 0.00000.8755 ± 0.00340.2076
ActivityGT-GAN0.5000 ± 0.00000.6629 ± 0.00210.2039
ActivityKoVAE0.5000 ± 0.00000.7135 ± 0.00350.3723
CesNetHiGen0.3750 ± 0.02780.6520 ± 0.00080.0752
CesNetTimEHR0.4907 ± 0.00800.8256 ± 0.00500.4692
CesNetTimeDiff0.5000 ± 0.00007.9308 ± 0.22940.0833
CesNetGT-GAN0.4491 ± 0.08820.7712 ± 0.00770.2498
CesNetKoVAE0.4907 ± 0.00800.7474 ± 0.00510.3393

Overall, HiGen performs favorably across these complementary metrics, supporting its ability to preserve temporal predictive relationships and lagged cross-variable correlations.


Response to W2: Readability of Figures 5 and 6

We increased the font sizes of axis labels, tick labels, and legends in Figures 5 and 6. We also adjust panel spacing and legend positions, and prioritize scalable vector versions. We check the figures at the normal reading size in the paper, rather than relying on enlarged screenshots. The numerical tables in this response also provide directly readable results and reduce reliance on small text in figures.


We hope our clarifications and additional experiments address your concerns. We would be happy to discuss any further questions!

Reviewer_hLgpOpenReview ↗

Response to authors

Thank you for the further clarifications and additional results. I particularly appreciate the comparisons with recent baselines provided in your response, which I consider essential for a comprehensive evaluation of the work. Most of my concerns have been addressed, and I have increased my score accordingly.

AuthorsOpenReview ↗

Response to Reviewer hLgp

Thanks again for your insightful comments and positive feedback on our rebuttal. We truly appreciate the time and careful effort you put into reading our response and re-evaluating this work.

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

This paper proposes HiGen, a hierarchical framework for irregular multivariate time series generation. To preserve irregular sampling patterns and asynchronous variable dependencies, HiGen directly models time–variable–value observations through a sample–window–point hierarchy. Specifically, it employs a variational global representation, window-level recurrent memory, and point-level autoregressive generation with learnable time bins and structure-aware attention. Experiments on three datasets demonstrate improvements over selected baselines in downstream classification utility and generation fidelity, particularly under data-scarce settings.

Critical strengths
  1. The paper addresses irregular multivariate time series generation, an important problem with applications in healthcare and other real-world domains.

  2. The hierarchical architecture is intuitively motivated and reasonably well organized, with helpful illustrations in Figures 1 and 2.

Critical weaknesses
  1. The proposed framework mainly combines existing techniques. For example, the point-level autoregressive generation is closely related to the event time and type modeling in Neural Hawkes Process, while the learnable time-bin decoder resembles the existing conditional log-normal mixture model.

  2. The baselines omit recent relevant approaches, particularly Diff-MN [1], which supports both irregular-to-regular and irregular-to-continuous time series generation. Comparisons on common numerical generation and downstream tasks are feasible under unified evaluation protocols.

  3. The proposed time decoder closely resembles existing conditional log-normal mixture models. Since the component mean is parameterized as (brb_r+μr\mu_r), the learnable bin centers may be redundant with the predicted residual means. Comparisons with standard log-normal mixtures, fixed bin centers, and constrained residual models are needed.

  4. In Eq. (11), variable prediction is not explicitly conditioned on the currently generated time interval or synchronization decision. This factorization may limit the modeling of dependencies between observation times and variable occurrences. A time-conditioned variable decoder should be investigated.

  5. Although HiGen preserves exact observation timestamps, it still relies on fixed-width window partitioning. The authors should clarify how window boundaries affect generation, how out-of-window points are prevented, and how sensitive performance is to the windowing strategy.

[1] Diff-MN: Diffusion Parameterized MoE-NCDE for Continuous Time Series Generation with Irregular Observations.

  1. Only three datasets are evaluated, with Activity and CesNet containing relatively few samples. The authors should consider additional classification benchmarks, such as UCR/UEA datasets with controlled irregular sampling.

  2. Downstream utility is evaluated using only an LSTM classifier. Stronger and more diverse classification architectures, particularly CNN-based models, should be included under consistent preprocessing protocols to verify whether the improvements generalize across strong classifiers.

  3. Lack of systematic evaluation across sequence lengths and irregularity levels. The 5%, 10%, and 15% settings vary the number of training samples rather than sequence length or sampling irregularity. Experiments with different sequence lengths, observation horizons, sampling densities, and irregularity levels are needed to evaluate robustness.

  4. Although the ablation results demonstrate the effectiveness of individual components, they do not establish the advantages of the proposed designs over simpler or existing alternatives. Comparisons with flat autoregressive models, simpler hierarchical architectures, and standard log-normal mixture decoders would further strengthen the evaluation. Moreover, the ablations lack evaluations under varying missingness rates and more comprehensive generation-quality metrics, similar to those considered in Diff-MN, to assess the contributions of individual components to generation fidelity and robustness.

  5. The submission does not provide source code, making it difficult to verify the implementation details of the proposed method.

Questions for authors

Please refer to the Weaknesses.

7 comments
AuthorsOpenReview ↗

Response to Reviewer UJXA (Part Ⅰ)

Thank you for your thoughtful comments and constructive suggestions. They have helped us identify points that need clearer explanations and stronger experimental support. We address your concerns below and describe the corresponding revisions.


Response to W1: Relation to Neural Hawkes Process and Conditional Log-Normal Mixture Models

HiGen addresses unequal time gaps, asynchronous variable observations, and alternating sparse and dense sampling through (1) how history is organized to generate observations and (2) how history determines the next time interval.

1. How history is organized to generate observations.

  • The Neural Hawkes Process summarizes past events into a state in a continuous-time LSTM and uses this state to generate the next event’s time and type. The state determines the conditional intensity of each event type, which governs when the next event occurs and which type occurs. The generated event then updates the historical state. (Mei and Eisner, 2017).
  • HiGen uses a sample-level representation, window memory, and a point state to separate sequence-wide characteristics, cross-window evolution, and local observation history. It generates the interval, advances the point state, and then generates the variable and value. Each observation updates the point state, while each completed window updates window memory.

2. How history determines the next time interval.

  • A standard conditional log-normal mixture directly predicts the locations, scales, and weights of its log-normal distributions from history. These distributions are the mixture components. During generation, it selects a component according to the predicted weights and samples a positive interval from that distribution. (Shchur et al., 2020)
  • HiGen adjusts shared temporal prototypes using history-dependent residuals, rather than predicting mixture locations directly. Each location is the sum of a shared learnable prototype and a predicted residual; history also determines scales and weights. The decoder samples a continuous positive interval, not a fixed bin center.

Response to W2: Comparison with Diff-MN

We followed this suggestion and added Diff-MN results at r10. Each cell below reports AUROC / AUPRC.

Table R7. Comparison with Diff-MN at r10.

MethodP12ActivityCesNet
Diff-MN0.5282 ± 0.0315 / 0.1771 ± 0.01020.8130 ± 0.0411 / 0.6631 ± 0.05070.8645 ± 0.1227 / 0.7195 ± 0.1847
HiGen0.7186 ± 0.0161 / 0.2350 ± 0.01280.9011 ± 0.0199 / 0.8072 ± 0.04050.9946 ± 0.0037 / 0.9834 ± 0.0091

Our model achieves better performance.


Response to W3: Redundancy between Learnable Bin Centers and Predicted Residual Means

Please refer to our response to W1 for the distinction between HiGen and standard conditional log-normal mixtures. As suggested, we compared our learnable bins with three alternatives. Standard Lognormal Mixture directly predicts component locations, scales, and weights. Fixed Centers freezes the initial bin centers and retains conditional residuals. Constrained Residual limits residuals to the range allowed by gaps between adjacent log-centers.

Table R8. Time Decoder Alternatives.

ConfigurationP12 r10Activity r10CesNet r10
Learnable Bins0.7186 / 0.23500.9011 / 0.80720.9946 / 0.9834
Standard Lognormal Mixture0.6620 / 0.24870.6205 / 0.48710.9879 / 0.9612
Fixed Centers0.6850 / 0.24240.8296 / 0.71780.9807/ 0.9515
Constrained Residual0.6701 / 0.21050.8486 / 0.77280.9921 / 0.9773

The centers provide reference time scales shared across sequences, whereas the residuals adjust them to the current history. Replacing this combination with directly predicted locations reduces performance across all three settings, supporting the usefulness of the shared centers beyond history-dependent residual prediction.

AuthorsOpenReview ↗

Response to Reviewer UJXA (Part Ⅱ)

Response to W4: Conditioning Variable Prediction on the Current Time Interval and Synchronization Decision

HiGen already models dependencies between observation times and variable occurrences through the time-conditioned state used by the variable decoder. Eq. (11) omits this intermediate state. We make the dependency explicit as follows: p(sj∣qj−1)=p(rj∣qj−1)p(δj∣rj,qj−1)p(vj∣q~j)p(xj∣vj,q~j).p(s_j\mid q_{j-1}) = p(r_j\mid q_{j-1}) p(\delta_j\mid r_j,q_{j-1}) p(v_j\mid\widetilde q_j) p(x_j\mid v_j,\widetilde q_j). Here, a positive interval advances the previous state, while a synchronous observation retains it at zero interval. Variable and value prediction therefore already depend on the current interval and synchronization decision.


Response to W5: Fixed-Width Window Partitioning and Its Sensitivity

how window boundaries affect generation, when the next candidate timestamp reaches or exceeds the current window’s right boundary, the model stops generating within that window and proceeds to the next window.

how out-of-window points are prevented, out-of-window points are excluded before acceptance. Any candidate at or beyond the right boundary is discarded and is not added to the generated sequence.

Sensitivity to the windowing strategy, the window-width experiments in Section 4.7 show that HiGen is generally stable across the tested widths. The results suggest that performance does not depend on a single carefully chosen window width.


Response to W6: Additional UCR/UEA Classification Benchmarks

We added UEA LSST and retained its official split of 2,459 training samples and 2,466 test samples. It has 14 classes, 6 dynamic variables, and 36 original positions per variable. The r10 subset is fixed at 246 samples. All four sampling densities use the same subset IDs. All methods share the same observation mask within each condition.

We constructed irregular views by retaining original positions separately for each variable. We did not reindex positions after deletion or interpret them as physical hours. The number of retained observations per variable is rounded down. The target and actual sampling densities are therefore as follows.

Table R9. Target and Actual Sampling Densities on LSST.

Target DensityObservations per VariableObservation Points per SampleActual DensityActual Missingness Rate
20%74219.44%80.56%
40%148438.89%61.11%
60%2112658.33%41.67%
80%2816877.78%22.22%

The table below shows the synthetic-only LSTM results at r10 with the full sequence length. The values are means across three downstream classifier seeds. We do not claim statistical significance from these means.

Table R10. Synthetic-Only LSTM Results on LSST at r10.

Sampling DensityModelAUROCAUPRC
20%TimeDiff0.51980.0747
20%TimEHR——
20%KoVAE0.47220.0751
20%GT-GAN0.51480.0855
20%HiGen0.63260.1701
40%TimeDiff0.53190.0914
40%TimEHR——
40%KoVAE0.50030.0842
40%GT-GAN0.49780.0845
40%HiGen0.60470.1548
60%TimeDiff0.51830.0928
60%TimEHR0.54440.0971
60%KoVAE0.45310.0705
60%GT-GAN0.49930.0863
60%HiGen0.61120.1518
80%TimeDiff0.51600.0866
80%TimEHR0.59060.1094
80%KoVAE0.47770.0768
80%GT-GAN0.49620.0878
80%HiGen0.69860.1917

HiGen has the highest mean AUROC and AUPRC under all reported conditions. The "—" entries for TimEHR at 20% and 40% density mean that the model could not generate valid synthetic samples because the input data were scarce.

AuthorsOpenReview ↗

Response to Reviewer UJXA (Part Ⅲ)

Response to W7: CNN-Based Classifiers

We added ResNet1D and InceptionTime to test whether the downstream utility gains depend only on LSTM. All three classifiers use the same train/test splits, variable mapping, and inputs of values, masks, and time intervals within each comparison. We do not interpolate a separate dataset for the CNN-based models.

Table R11. Synthetic-Only Results with Three Classifiers.

DatasetMethodResNet AUROCResNet AUPRCInceptionTime AUROCInceptionTime AUPRC
P12HiGen0.7480 ± 0.01860.2550 ± 0.02400.7806 ± 0.01040.2713 ± 0.0093
P12TimEHR0.6938 ± 0.01280.2360 ± 0.01660.7161 ± 0.00410.2452 ± 0.0095
P12TimeDiff0.3845 ± 0.01800.1013 ± 0.00370.4754 ± 0.06600.1195 ± 0.0188
P12GT-GAN0.4923 ± 0.05830.1730 ± 0.01980.4858 ± 0.04250.1303 ± 0.0212
P12KoVAE0.6550 ± 0.05460.2145 ± 0.03300.6832 ± 0.02260.2711 ± 0.0120
ActivityHiGen0.9587 ± 0.00830.8988 ± 0.00460.9723 ± 0.00180.9186 ± 0.0038
ActivityTimEHR0.4067 ± 0.08400.3512 ± 0.07030.5415 ± 0.16650.4747 ± 0.1428
ActivityTimeDiff0.5681 ± 0.08570.4358 ± 0.05560.4007 ± 0.06090.3437 ± 0.0442
ActivityGT-GAN0.5786 ± 0.23490.4836 ± 0.17610.4804 ± 0.03910.3603 ± 0.0614
ActivityKoVAE0.5349 ± 0.16570.4399 ± 0.09430.8457 ± 0.04160.7678 ± 0.0935
CesNetHiGen0.9915 ± 0.00830.9823 ± 0.01260.9993 ± 0.00050.9977 ± 0.0016
CesNetTimEHR†0.6725 ± 0.02810.4182 ± 0.06240.6871 ± 0.03480.3771 ± 0.0553
CesNetTimeDiff0.5396 ± 0.05970.2940 ± 0.01920.4805 ± 0.06660.2416 ± 0.0572
CesNetGT-GAN0.8310 ± 0.03170.6040 ± 0.08300.7665 ± 0.03050.5749 ± 0.0340
CesNetKoVAE0.7693 ± 0.12200.5667 ± 0.18390.8099 ± 0.01710.6713 ± 0.0588

In this comparison using fixed synthetic pools, HiGen has higher mean AUROC/AUPRC than the four original baselines with all three classifiers. This supports that the utility gains are not specific to LSTM.

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

This paper investigates the generation of irregular multivariate time series data, aiming to simultaneously generate observation time, variable identity, and observation values ​​while preserving asynchronous sampling structure. HiGen divides the generation process into three levels: sample, window, and observation point. Sample-level variational representation provides global conditions, window memory models cross-window dependencies, and a point-level autoregressive decoder generates specific observations. Time intervals are modeled through synchronous decision-making and learnable time binning. Experiments cover P12, Activity, and CesNet, evaluating synthetic data training, data augmentation, distribution fidelity, and observation structure recovery under different training data proportions. Overall, HiGen outperforms the compared generative models.

From my personal perspective, current research on irregular time series data seems primarily focused on prediction tasks, with relatively little work on irregular generation for generative tasks. Therefore, a highly reliable and effective method for generating irregular time series data could potentially further stimulate enthusiasm for research in this field.

Critical strengths
  1. The overall architecture design of HiGen is good, effectively considering observation time, observed variables, and specific information simultaneously.
  2. The authors conducted ablation experiments from multiple perspectives on different functional modules within HiGen.
  3. The eleven-year baseline of this paper is also quite comprehensive, with a focus on evaluating generation under conditions of data scarcity.
Critical weaknesses
  1. The configuration of the authors' ablation experiments seems to differ somewhat from that of the baseline experiments. The baseline experiments focused on scenarios with scarce data, but the ablation experiments were conducted on complete datasets. Should additional ablation experiments be added for scenarios with scarce data?
  2. Regarding Equation 11, why did the authors use this autoregressive generation method? Is it actually implemented this way? Will the model's performance change significantly after incorporating the current time interval as a condition? Given historical states and global latent variables, it cannot be ruled out that there is a conditional independence assumption between the current time interval and the variables and values.
  3. The current baseline covers a variety of generative models, but lacks an event sequence model that more closely approximates the point-level autoregressive generation strategy presented in this paper. In particular, while the paper cites SynEHRgy, it fails to provide experimental comparisons. Furthermore, it is recommended to consider recent methods such as Diff-MN (ICML 2026) or ImagenI2R (NeurIPS 2025), and to clarify their comparability with this paper's generation objectives and evaluation protocols.
  4. This paper provides some implementation hyperparameters, but the model size, total training budget, checkpoint selection, and evaluation network configuration are still insufficient to fully reproduce the experiments based on the paper and to determine whether each method has achieved comparable training and hyperparameter tuning budgets.
  5. Appendix B provides an overview of CesNet and summary statistics after preprocessing. However, the official CESNET-TimeSeries24 release contains series aggregated over fixed time windows, including 10-minute, 1-hour, and 1-day resolutions. Could the authors clarify which data subset and temporal resolution were used, and provide the procedure for constructing the irregular multivariate samples? In particular, please explain the source of nonuniform observation times, how missing or empty aggregation windows were handled, and whether different variables are observed asynchronously.
  6. I suggest the authors add corresponding ablation experiments for the pre-training objectives in Equation 16, or further supplement with more detailed parameter sensitivity experiments, especially for the KL divergence loss. Based on my current experience in time series generation, sometimes removing the KL divergence loss seems to improve performance; overly strong regularization is often detrimental. Most importantly, the authors don't seem to provide the specific weight values ​​of the loss parameters for each dataset in the experiments. We can only infer their maximum and minimum ranges from Appendix G. Crucially, I think setting the KL divergence weight to 0.01 is too large; I suggest trying values ​​on the order of 1e-4 and 1e-6.
Questions for authors

See the Critical Weaknesses section for specific Questions.

6 comments
AuthorsOpenReview ↗

Response to Reviewer kHZJ (Part Ⅰ)

Thank you for your thoughtful comments and constructive suggestions. They have helped us identify points that need clearer explanations and stronger experimental support. We address your concerns below and describe the corresponding revisions.


Response to W1: Ablation Studies in Data-Scarce Settings

Given the time constraints, we added module ablations at r10, as suggested. They examine the sample-level random latent variable, cross-window memory, time decoder, and structure-aware attention biases. We do not rely only on ablation findings from complete datasets. Each cell below reports synthetic-only AUROC / AUPRC.

Table R3. Module Ablations at r10.

ConfigurationP12ActivityCesNet
Full0.7186 / 0.23500.9011 / 0.80720.9946/ 0.9834
w/o Var. sample rep0.6499 / 0.22690.8677 / 0.75860.9809 / 0.9259
w/o Window memory0.6644 / 0.21740.8493 / 0.75770.9863 / 0.9813
w/o Time bins0.6716 / 0.20350.8714 / 0.73470.9929 / 0.9832
w/o Attn. bias0.7075 / 0.23240.9008 / 0.76540.9900 / 0.9823

These experiments extend the module evaluation to the data-scarce settings emphasized in the paper. They confirm that the ablation findings also hold in these settings.


Response to W2: The Implementation of Eq. (11) and the Current Time Interval

Yes, we implement this autoregressive generation order. And, it does not assume that variable identity and value are independent of the current interval given the previous history and global latent variable. We generate time first, then variable identity and value, so that the current interval can inform which variable is observed and what value it takes.

The ambiguity comes from Eq. (11), which omits an intermediate state. In the implementation, a positive interval first advances the historical state through the transition module; a synchronous observation retains the state at zero interval. Variable identity and value are then predicted from this intermediate state. Therefore, the current interval already conditions both predictions through the state transition.

The dependency can be written as p(sj∣qj−1)=p(rj∣qj−1) p(δj∣rj,qj−1) p(vj∣q~j) p(xj∣vj,q~j).p(s_j\mid q_{j-1}) = p(r_j\mid q_{j-1}) \,p(\delta_j\mid r_j,q_{j-1}) \,p(v_j\mid\widetilde q_j) \,p(x_j\mid v_j,\widetilde q_j). Here, a positive interval advances the state, while a synchronous observation retains it at zero interval. Thus, the current interval affects both variable and value prediction through the state transition.

We have revised Eq. (11) to distinguish the states before and after the transition and clarify the synchronization branch.


Response to W3: Event Sequence Models, SynEHRgy, Diff-MN, and ImagenI2R

We added the event-based generative models SV-MTPP-CI and CEG-CDDM, as well as SynEHRgy, Diff-MN, and ImagenI2R. This addresses the focus on regularized representations in the original comparisons. The table below shows the synthetic-only LSTM results at r10. Each cell reports AUROC / AUPRC.

Table R4. Additional Generative Baselines at r10.

MethodP12ActivityCesNet
SV-MTPP-CI (event-based)0.5844 ± 0.0187 / 0.1692 ± 0.01340.7317 ± 0.0145 / 0.5936 ± 0.01700.9644 ± 0.0043 / 0.8829 ± 0.0045
CEG-CDDM (event-based)0.5381 ± 0.0093 / 0.1690 ± 0.01610.8835 ± 0.0201 / 0.7579 ± 0.04380.9606 ± 0.0134 / 0.8656 ± 0.0507
SynEHRgy0.6340 ± 0.0488 / 0.2006 ± 0.00970.8700 ± 0.0280 / 0.8108 ± 0.01110.9437 ± 0.0102 / 0.8270 ± 0.0471
ImagenI2R0.5812 ± 0.0329 / 0.1948 ± 0.01410.8098 ± 0.0342 / 0.6825 ± 0.02130.8605 ± 0.0202 / 0.7080 ± 0.0249
Diff-MN0.5282 ± 0.0315 / 0.1771 ± 0.01020.8130 ± 0.0411 / 0.6631 ± 0.05070.8645 ± 0.1227 / 0.7195 ± 0.1847
HiGen0.7186 ± 0.0161 / 0.2350 ± 0.01280.9011 ± 0.0199 / 0.8072 ± 0.04050.9946 ± 0.0037 / 0.9834 ± 0.0091

HiGen has higher mean scores in most table entries.

AuthorsOpenReview ↗

Response to Reviewer kHZJ (Part Ⅱ)

Response to W4: Model Size, Total Training Budget, Checkpoint Selection, and Evaluation Network Configuration

We provide the dataset-specific HiGen training settings in Appendix C.1, Table 5. These include K, batch size, optimizer, learning rate, training budget, and KL settings. Table R5 summarizes these configurations below.

Table R5: Original HiGen training settings.

SettingP12ActivityCesNet
Positive-time components K8810
Batch size321632
OptimizerAdamWAdamWAdamW
Learning rate1e-35e-41e-3
Weight decay1e-61e-61e-6
Gradient-norm clipping111
Optimizer-update budget30,00020,00020,000
Generator seed424242
KL weight0.10.0010.1
KL warm-up (updates)11,00008,000

The downstream evaluation configuration is detailed in Appendix C.2. The verified evaluation configuration is an LSTM with hidden size 64, one layer, dropout 0, and a fixed training budget of 30 epochs. It uses AdamW with learning rate 0.001, weight decay 1e-5, batch size 32, and gradient clipping at 1. The inputs include values, masks, and time intervals. Each comparison group uses the same data splits and evaluation implementation.


Response to W5: The CesNet Source and Construction of Irregular Multivariate Samples

Thank you for requesting these details. We clarify the data source and preprocessing below.

1. Which data subset and temporal resolution were used?

The official CESNET-TimeSeries24 release provides regularly aggregated series at 10-minute, 1-hour, and 1-day resolutions. We did not directly use those files. Instead, we used the numerical CESNET version released by Time-IMM because its multivariate numerical observations and nonuniform timestamps match our target task. We used its published timestamps rather than selecting or reconstructing a fixed-resolution grid. (Time-IMM repository)

2. Where do the nonuniform observation times come from?

Nonuniform timestamps are already present in the Time-IMM files. For example, consecutive records from one entity have time-of-day values of 23:23:10, 01:08:31, and 01:24:02. Our preprocessing preserves these timestamps. Time-IMM describes its CESNET component under scheduling jitter and delay in Appendix D.8. (Time-IMM paper)

3. How were the multivariate samples constructed?

We converted timestamps into elapsed hours from a common reference date and extracted 48-hour samples with a 24-hour stride. Within each sample, time is the original timestamp minus the window start. The selected 8,619 source rows became 17,236 rows after overlapping-window expansion.

4. How were missing values and empty windows handled?

We did not reconstruct or fill absent fixed aggregation windows. Gaps in the input remain elapsed-time gaps. A 48-hour extraction window containing no source records is skipped. Missing variable values are stored as NaNs and excluded when constructing observation triples, rather than treated as observed zeros. A row with all dynamic variables missing therefore contributes no observation points.

5. Are different variables observed asynchronously?

Yes, different variables in our constructed CESNET samples are observed asynchronously. This asynchrony is introduced by our preprocessing, not by Time-IMM. The Time-IMM input records all 10 variables jointly at each timestamp. We independently retain each variable observation with probability 0.30, so one variable may remain observed at a timestamp while miss some at other timestamps. This produces asynchronous observations across variables without changing the retained timestamps or values.

AuthorsOpenReview ↗

Response to Reviewer kHZJ (Part Ⅲ)

Response to W6: The KL Divergence Weight

Due to time constraints, we only added comparisons of KL weights. The table below shows the available synthetic-only LSTM results.

Table R6. Effect of the KL Weight on Synthetic-Only LSTM Performance.

DatasetKL WeightSynthetic-only AUROCSynthetic-only AUPRC
P120.10.7186±.01610.2350±.0128
P121.00E-040.6618 ± 0.01890.2399 ± 0.0330
P121.00E-060.5568 ± 0.01410.1630 ± 0.0071
P1200.5488 ± 0.03260.1569 ± 0.0057
Activity0.10.8258 ± 0.03390.7275 ± 0.0149
Activity1.00E-040.9013 ± 0.01750.8316 ± 0.0341
Activity1.00E-060.9123 ± 0.02010.8511 ± 0.0396
Activity00.9117 ± 0.02410.8086 ± 0.0214
CesNet0.10.9946 ± 0.00370.9834 ± 0.0091
CesNet1.00E-040.9847 ± 0.00650.9559 ± 0.0149
CesNet1.00E-060.9857 ± 0.00430.9580 ± 0.0120
CesNet00.9939 ± 0.00490.9845 ± 0.0098

The effect of the KL weight is dataset-dependent. Activity achieves its highest synthetic-only AUROC and AUPRC with a KL weight of 1e-6, whereas CesNet achieves its highest scores with a KL weight of zero.


We hope our clarifications and additional experiments address your concerns. We would be happy to discuss any further questions!

Reviewer_x3cmOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HiGen, a hierarchical generative framework for irregular multivariate time series. Instead of regularizing observations into aligned sequences, HiGen directly generates observation points represented by timestamps, variable identities, and values. The framework consists of three levels: a variational sample-level representation for global conditioning, a window-level recurrent memory for modeling long-range temporal dependencies, and a point-level autoregressive decoder with learnable time bins for generating individual observations. Experiments on three irregular time-series datasets show that HiGen consistently outperforms existing baselines in downstream classification utility, generation fidelity, and the preservation of irregular sampling patterns.

Critical strengths
  • A well-motivated problem setting. The authors provide a clear motivation for directly modeling irregular observations rather than converting them into regular time series. Preserving the original timestamps and asynchronous variable occurrences is particularly meaningful for applications such as clinical monitoring, where observation patterns themselves may contain useful information.

  • A reasonable hierarchical design. The sample–window–point hierarchy provides a natural way to separate global sequence characteristics, long-range temporal evolution, and local observation dynamics. In particular, combining window-level memory with point-level time interval generation is a reasonable approach to handling both sparse observation periods and locally dense sampling.

Critical weaknesses
  • Insufficient discussion and comparison with event-based generative models. A major motivation of HiGen is to directly generate irregular observation points without temporal alignment, but this idea is closely related to existing research on marked temporal point processes and event-based sequence generation. Relevant works, such as Decoupled Marked Temporal Point Process using Neural ODEs (ICLR 2024), Probabilistic Modeling for Sequences of Sets in Continuous-Time (AISTATS 2024), and Conditional Generative Modeling for High-dimensional Marked Temporal Point Processes (KDD 2025), are not sufficiently discussed. More importantly, the current baselines mainly rely on regularized or segment-aligned representations, making it difficult to determine whether the improvements come from the proposed architecture or simply from preserving the original event representation. I would be very willing to raise my score if the authors could include a strong event-based generative baseline, adapted to the same generation task, and demonstrate that HiGen still provides substantial improvements.

  • The actual contribution and necessity of hierarchical modeling remain unclear. Although the overall framework is reasonable, most of its individual components, including variational global conditioning, recurrent memory, attention biases, and mixture-based time interval generation, are based on established techniques. The current ablation studies show that removing certain modules reduces performance, but do not clearly demonstrate why the proposed three-level hierarchy is necessary compared with a simpler event-level autoregressive model. A comparison with a flat event-based generator using the same input representation and a comparable model capacity would help establish the actual benefit of the hierarchical design and better support the methodological contribution.

Questions for authors

Comparison with event-based generation. Have the authors considered adapting existing marked temporal point process or event-based generative models to directly generate time–variable–value observations? How would HiGen compare with these models when they use the same original observation representation, particularly in terms of temporal irregularity and downstream utility?

3 comments
AuthorsOpenReview ↗

Response to Reviewer x3cm

Thank you for your thoughtful comments and constructive suggestions. They have helped us identify points that need clearer explanations and stronger experimental support. We address your concerns below and describe the corresponding revisions.


Response to W1 / Q1: Comparison with Event-Based Generative Models

We added two event-based generative baselines, SV-MTPP-CI and CEG-CDDM. We adapted them to the complete generation task, including times, variables, continuous values, and labels.

The table below shows the synthetic-only LSTM results at r10. Each cell reports AUROC / AUPRC. The "±" values are sample standard deviations across three downstream classifier seeds.

Table R1. Event-Based Generative Baselines at r10.

MethodP12ActivityCesNet
SV-MTPP-CI0.5844 ± 0.0187 / 0.1692 ± 0.01340.7317 ± 0.0145 / 0.5936 ± 0.01700.9644 ± 0.0043 / 0.8829 ± 0.0045
CEG-CDDM0.5381 ± 0.0093 / 0.1690 ± 0.01610.8835 ± 0.0201 / 0.7579 ± 0.04380.9606 ± 0.0134 / 0.8656 ± 0.0507
HiGen0.7186 ± 0.0161 / 0.2350 ± 0.01280.9011 ± 0.0199 / 0.8072 ± 0.04050.9946 ± 0.0037 / 0.9834 ± 0.0091

HiGen achieves higher AUROC and AUPRC than both event-based baselines on all three datasets. This comparison suggests that preserving an event-based representation alone is not sufficient to achieve the same generation quality. Unlike the event-level recurrent histories in SV-MTPP-CI and CEG-CDDM, HiGen uses a sample-level representation to capture sequence-wide characteristics and a separate window memory to model evolution across observation periods, explicitly distinguishing changes across sparse and dense observation periods from local event dynamics.


Response to W2: The Benefit of the Three-Level Hierarchy over Established Components

HiGen combines sample-level global conditioning, window-level temporal organization, and point-level generation. We compare it with three simpler variants:

  • w/o window level (sample + point retained): A single point-level state generates the entire sequence under sample-level conditioning, without window representations or cross-window memory.
  • w/o stochastic sample representation (window + point retained): We replace the stochastic sample representation with its deterministic mean and remove the KL term.
  • Flat autoregressive model (point level only): We remove both sample-level conditioning and window-level modeling. A single point-level state is carried through the sequence.

As shown in Table R2, the full model achieves higher AUROC and AUPRC than all three variants on every dataset. These comparisons support the benefit of the three-level hierarchy.

Table R2. Ablation of Sample-Level and Window-Level Modeling. Each cell reports AUROC / AUPRC.

ConfigurationP12ActivityCesNet
Matched Full0.7186 / 0.23500.9011 / 0.80720.9946 /0.9834
w/o window level0.6644 / 0.21740.8221 / 0.74290.9881 / 0.9644
w/o sample level0.6499 / 0.22690.8677 / 0.75860.9809 / 0.9259
Flat autoregressive model0.6044 / 0.19450.8840 / 0.75150.7580 / 0.5118

We hope our clarifications and additional experiments address your concerns. We would be happy to discuss any further questions!

Thanks for the detail rebuttal from the authors. It solved some of my concerns and I raised the score.

AuthorsOpenReview ↗

Response to Reviewer x3cm

Thanks again for your insightful comments and positive feedback on our rebuttal. We truly appreciate the time and careful effort you put into reading our response and re-evaluating this work.

Comments on the paper

1 comment

The author has responded to everyone's comments. Please reply when it is convenient for you.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.