acceptodds
Under review as a conference paper at ICLR 2027

ChronoFlow: Hierarchical Flow Matching for Irregular Time Series Generation

Abstract

Recent advances in generative modeling have substantially improved time series generation, yet most existing methods either assume a regular temporal grid or focus on feature dynamics under a given sampling structure. This makes them illsuited for generating irregular time series in their native form, where a model must capture not only feature values, but also how many observations occur, when they occur, and which features are observed together. To address this heterogeneous generation problem, we propose ChronoFlow, a unified hierarchical flow matching framework organized by statistical granularity. Following a coarse-to-fine hierarchy, ChronoFlow first generates observation counts and feature-wise frequencies, then jointly generates observation times and feature co-observation patterns, and finally generates values conditioned on the realized pattern. This turns a complex joint generation problem into structurally aligned subproblems while preserving their dependencies. To evaluate complete irregular time series generation, we introduce complementary metrics spanning sample realism, sampling structure, value fidelity, and temporal and cross-feature dependencies, and validate them through controlled corruptions. Across five benchmarks, ChronoFlow achieves strong improvements in generation fidelity over existing baselines, while factorization studies support the proposed hierarchy. Our code is available at https://anonymous.4open.science/r/ChronoFlow.

Primary area:
learning on time series and dynamical systems
Keywords:
Irregularly Sampled Time Series, Time Series Generation, Flow Matching

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_MFpgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper tackles the problem of unconditional generation of irregular multivariate time series. The method consists of three conditional flow-matching modules: IntensityFlow, PatternFlow, and ValueFlow, which captures, respectively, observation count and feature-wise frequencies, joint observation time and features, and values conditioned on the generated sampling pattern. The paper also provides evaluation contributions through six complementary diagnostics, real-real references, and controlled corruption experiments. Across five clinical benchmarks, the proposed method achieves the best performance in 21 of 30 dataset-metric comparisons.

Critical strengths

One of the paper’s strengths is a coherent formulation of irregular data generation, an important and under-studied area in generative modeling. The proposed method appears to address a broader and more structurally heterogeneous task than previous proposed methods.

Another strength is the paper’s broad empirical evaluation with multiple metrics. The proposed method achieves the state-of-the-art performance on the majority of experiments across five datasets and evaluation metrics, which includes sampling structure, values, transitions, cross-feature relationships, and an overall discriminator. Corruption tests and oracle interventions also help interpret the metrics and model. Therefore evidence is not confined to a single favorable statistic or dataset.

Critical weaknesses

While the paper offers multiple datasets paired with multiple metrics, the comparisons offered do not necessarily establish advantage over the strongest relevant alternatives. For example, there is no direct comparison with a grid-and-mask generator, despite the central motivation of the paper contrasting native irregular generation with grid-based approaches. The paper also identifies TimEHR but does not evaluate it.

Another weakness of the paper is that there is not enough evidence to clearly isolate the central architecture’s benefits and structural consistency. The proposed hierarchy may work empirically, but the evidence does not establish which components are necessary or how reliably its stages remain mutually consistent. Factorization alternatives differ in capacity and conditioning. There is no isolated ablation of feature-frequency conditioning, nor time-then-panel or panel-then-time comparison. Downstream training uses true conditions, while generation uses model-produced conditions.

Questions for authors
  1. How does ChronoFlow compare with a strong resolution-matched grid-and-mask generator and a simpler panel-based generation baseline?
  2. what happens when feature frequency conditioning is removed, and do the principal hierarchy/coupling findings persist beyond eICU?
  3. how often do sampled frequencies disagree with decoded frequencies, panels require repair, or timestamp clipping create ties?
  4. What evidence shows preservation of dependencies between observation patterns and values, beyond their separate marginals and local summaries?
7 comments

We thank the reviewer for the precise questions and for recognizing (1) our coherent formulation of irregular time-series generation, an important, under-studied, and structurally heterogeneous problem, and (2) our broad evaluation, in which ChronoFlow leads on most dataset–metric pairs and corruption tests and oracle interventions keep the evidence from resting on a single favorable statistic. We address each concern below.


[W1] & [Q1] Comparison with resolution-matched grid-and-mask generators and a simpler panel-based baseline

We compared ChronoFlow on eICU, with the same event representation and six metrics, against two resolution-matched grid-and-mask generators, each given a fair chance, and a simpler panel-based baseline:

  1. TimEHR [1], a two-stage 2D CNN image GAN (masks, then values): all eICU entries are losslessly packed into its 128 × 128 input, and an extended budget (1,250 CWGAN / 320 Pix2Pix epochs, learning rates selected on validation) changed little (Set-Discr. 0.4895, Value-W1 0.3470).
  2. TimeDiff [2], a mixed Gaussian–multinomial diffusion jointly generating continuous channels and binary missingness indicators, trained at full eICU resolution with 1,000 reverse diffusion steps.
  3. Bernoulli-r: IntensityFlow, the generated times, and ValueFlow are kept, but each occasion's panel is drawn as independent Bernoulli(r_f) per feature, preserving the sampled rates but ignoring co-observation.
Model (eICU)ParamsSet-Discr. ↓CoOb ↓Pattern-MMD² ↓Trans-MMD² ↓Value-W1 ↓Set-Corr. ↓
TimEHR [1]55.74M0.4907 ± 0.00640.9857 ± 0.33450.0216 ± 0.01950.0546 ± 0.00920.3648 ± 0.05080.1067 ± 0.0051
TimeDiff [2]4.11M0.4982 ± 0.00110.7182 ± 0.07390.1904 ± 0.02890.4685 ± 0.01022.9703 ± 0.04770.1306 ± 0.0098
Bernoulli-r2.06M0.4218 ± 0.02150.4383 ± 0.01420.0005 ± 0.00020.0077 ± 0.00030.0598 ± 0.00370.0366 ± 0.0040
ChronoFlow2.06M0.0995 ± 0.00980.0380 ± 0.00470.0002 ± 0.00010.0045 ± 0.00020.0504 ± 0.00470.0385 ± 0.0003
  • Why grid-and-mask generators fall short. Both remain almost perfectly separable from real records despite 2–27× more parameters, because they generate the full time-by-feature grid and its mask: an eICU stay spans 15,840 cells at 1-minute resolution, only 0.86% observed, so almost all of their output is empty cells, and co-observation must emerge from a dense mask (for TimEHR, through convolutions that depend on an arbitrary feature order). ChronoFlow generates only the occasions a stay has (about 34) and each occasion's panel jointly, so its CoOb is 26× lower than TimEHR's.
  • Why per-feature rates are not enough. Bernoulli-r matches the sampled rates yet is almost perfectly separable and breaks co-observation, so modeling which features are measured together, PatternFlow's role, is essential.

We have added Bernoulli-r to Table 2 and Section 5.3, and we will add TimEHR and TimeDiff to Table 1.

[1] Karami et al. "TimEHR: Image-Based Time Series Generation for Electronic Health Records." IEEE Journal of Biomedical and Health Informatics 2025.
[2] Tian et al. "Reliable Generation of Privacy-Preserving Synthetic Electronic Health Record Time Series via Diffusion Models." JAMIA 2024.

[W2] & [Q2] Ablation studies of conditioning, Stage-2 ordering, and generalization beyond eICU

1. Which components are necessary, capacity, and mutual consistency. We removed one component at a time on eICU:

Variant (eICU)Set-Discr. ↓CoOb ↓Pattern-MMD² ↓Trans-MMD² ↓Value-W1 ↓Set-Corr. ↓
w/o IntensityFlow (2.18M)0.1178 ± 0.02670.0593 ± 0.00470.0020 ± 0.00040.0069 ± 0.00020.0555 ± 0.00630.0344 ± 0.0047
Joint TPV (size-matched, 2.06M)0.3019 ± 0.05080.0821 ± 0.00320.0007 ± 0.00000.0258 ± 0.00120.2915 ± 0.00490.0660 ± 0.0010
ChronoFlow (2.06M)0.0995 ± 0.00980.0380 ± 0.00470.0002 ± 0.00010.0045 ± 0.00020.0504 ± 0.00470.0385 ± 0.0003
  • Each of these components matters. Removing IntensityFlow degrades the pattern metrics, merging pattern and value generation collapses fidelity, independent per-feature panels break co-observation (Bernoulli-r, W1: CoOb 0.4383), and removing OT raises Set-Discr. to 0.4316 (Table 2).
  • Capacity does not explain the gains. Factorizations necessarily differ in conditioning, so we controlled capacity: Joint TPV fails at matched size, and the larger w/o IntensityFlow (2.18M), Value→Pattern (2.70M: 0.4511), and Joint All (2.21M: 0.2445) are also worse. With the same conditioning (no r), separate pattern and value generation gives Set-Discr. 0.0819 (M only, point 2) vs. 0.3019 for the size-matched Joint TPV.
  • The stages remain mutually consistent. Realized panel frequencies stay within 1.93 percentage points of the sampled ones and follow r almost one-to-one (slope 0.9927; Q3), and ChronoFlow preserves the pattern–value coupling closest to Real–Real (D_CV 0.0231 vs. 0.0185; next-best FlexTPP 0.0337; Q4).

2. Feature-frequency conditioning and time–panel order. We removed r (M only) and replaced the joint time–panel PatternFlow with sequential Time-first and Panel-first variants (eICU):

Variant (eICU)ParamsSet-Discr. ↓CoOb ↓Pattern-MMD² ↓Trans-MMD² ↓Value-W1 ↓Set-Corr. ↓
M only–0.0819 ± 0.01960.0301 ± 0.00840.0002 ± 0.00010.0049 ± 0.00040.0487 ± 0.00290.0399 ± 0.0042
Time-first2.13M0.0751 ± 0.01360.0324 ± 0.00690.0004 ± 0.00010.0039 ± 0.00030.0537 ± 0.00740.0388 ± 0.0032
Panel-first2.14M0.0733 ± 0.02560.0322 ± 0.00600.0002 ± 0.00000.0046 ± 0.00010.0523 ± 0.00120.0326 ± 0.0043
ChronoFlow (joint)2.06M0.0995 ± 0.00980.0380 ± 0.00470.0002 ± 0.00010.0045 ± 0.00020.0504 ± 0.00470.0385 ± 0.0003
  • r is designed for controllability. Removing it does not consistently change unconditional fidelity (better on some metrics, worse on others), since PatternFlow learns marginal rates implicitly. However, r lets a practitioner specify how often each feature is observed, which is impossible with M only: with M and the noise fixed, changing r moves the realized frequencies almost one-to-one (slope 0.9927).
  • The time–panel order is flexible. The sequential variants reach similar fidelity, so the hierarchy does not hinge on this order; what matters is generating each occasion's features together and the pattern before the values (point 1). The joint form needs a single flow, whereas the sequential variants split Stage 2 in two, with 3.3–3.8% more parameters and 7–8% longer training. We have revised Section 5.3 and added this comparison to Appendix G.1.

[W2] & [Q2], continued

3. Training on true conditions vs. sampling on generated ones. Conditioning crosses only two coarse-to-fine boundaries, not hundreds of autoregressive steps in which errors compound. Replacing generated conditions with ground-truth ones, and perturbing them:

Setting (eICU, Table A4)(M, r)(τ, B)Value-W1 ↓Set-Corr. ↓
Full generationsampledsampled0.0504 ± 0.00470.0385 ± 0.0003
+ ground-truth (M, r)ground truthsampled0.0481 ± 0.00370.0365 ± 0.0022
+ ground-truth (M, r), (τ, B)ground truthground truth0.0508 ± 0.01000.0358 ± 0.0019
Ground truth + 10% noiseperturbedperturbed0.0460 ± 0.00150.0378 ± 0.0049
Ground truth + 20% noiseperturbedperturbed0.0473 ± 0.00260.0364 ± 0.0050

Ground-truth patterns leave the value marginals essentially unchanged and values stay stable under 10–20% noise, while ground-truth (M, r) lowers Set-Discr. by 85% (0.0995 → 0.0153), so Stage 1 is the main source of the Set-Discr. gap. We have added mitigation paths (a better Stage-1 sampler, conditioning-noise augmentation, and self-distillation) to Section 5.4.

Q2. Removing feature-frequency conditioning, and findings beyond eICU. Removing r is answered in point 2. On the non-clinical Activity dataset, retraining PatternFlow without OT (w/o OT) is worse than ChronoFlow on all six metrics and merging all stages (Joint All) on five of six, so the hierarchy and OT findings persist beyond clinical EHR:

Variant (Activity)Set-Discr. ↓CoOb ↓Pattern-MMD² ↓Trans-MMD² ↓Value-W1 ↓Set-Corr. ↓
ChronoFlow0.1753 ± 0.02170.0739 ± 0.01000.0278 ± 0.00800.0100 ± 0.00150.1018 ± 0.00680.0474 ± 0.0067
w/o OT0.2292 ± 0.03760.1360 ± 0.00530.0355 ± 0.00310.0195 ± 0.00160.1067 ± 0.00620.0512 ± 0.0069
Joint All0.2691 ± 0.05520.1589 ± 0.00550.0200 ± 0.00220.0431 ± 0.01950.1438 ± 0.00990.0625 ± 0.0093
Reviewer_MF4mOpenReview ↗
Rating
2
Confidence
4(+1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ChronoFlow, a hierarchical flow matching framework for generating irregular time series. It decomposes generation into three stages: global observation statistics, joint observation-time and feature-panel generation, and conditional value generation. The paper also introduces a complementary evaluation protocol for full irregular time-series generation and conducts experiments on five benchmarks, showing competitive generation quality.

Critical strengths
  1. The proposed occasion–panel representation provides a structured way to model asynchronous multivariate observations and their co-occurrence relationships, which is important within irregular time series setting.
  2. The methodology is straightforward, and paper is easy to follow and reproduce.
  3. The experiments cover 5 benchmarks and multiple complementary fidelity metrics. The additional perturbation analyses and independently trained runs strengthen the empirical evaluation.
Critical weaknesses

1.Some incremental methodological novelty. The central coarse-to-fine generation strategy, in which global statistics or structural patterns are modeled before generating local observations and values, is well established in time series generative modeling. Related work such as MG-TSD [1], HDT [2] has explored multi-granularity guidance, while TimEHR [3] already adopts a pattern-first, value-second generation strategy for irregular clinical records. ChronoFlow extends these ideas by introducing observation-frequency conditioning and jointly modeling observation times and feature panels. However, all three stages employ standard flow matching objectives and established neural architectures. If both M and r can be directly computed from the final sampling pattern, the necessity of the IntensityFlow module remains unclear. The authors should consider including an ablation study w/o IntensityFlow to demonstrate its contribution and justify the importance of the proposed three-stage framework.

2 Missing comparisons with directly relevant generation baselines. The experimental evaluation lacks comparisons with several directly relevant generative baselines. In particular, TimEHR, which explicitly models missingness patterns and conditional values, should be included in the evaluation under irregular sampling scenarios. Additionally, KoVAE [4] is highly relevant to irregular time series generation and should be considered as an important baseline.

3 Insufficient evidence for computational efficiency. The computational efficiency claims in Section 5.5 are not sufficiently supported by the reported results. A comprehensive efficiency evaluation should include overall training and inference time, FLOPs, and memory consumption. Reporting only ms/sample is insufficient to fully characterize the computational overhead or substantiate the claimed efficiency advantages of the proposed method.

4 Lack of consistency between generated global statistics and local observation patterns. I found that the feature-wise frequency vector is defined as a deterministic statistic of the observation panel. However, the generated panels are not constrained to reproduce the sampled frequencies after thresholding. This raises questions about the statistical consistency of the hierarchical generation process. The authors should quantify the discrepancy between sampled global frequencies and realized panel frequencies

[1]. Fan, Xinyao, et al. "MG-TSD: Multi-granularity time series diffusion models with guided learning process." International Conference on Learning Representations. Vol. 2024. 2024.

[2]. Shibo, Feng, et al. "Hdt: Hierarchical discrete transformer for multivariate time series forecasting." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 39. No. 1. 2025.

[3]. Karami, Hojjat, et al. "Timehr: Image-based time series generation for electronic health records." IEEE Journal of Biomedical and Health Informatics 29.12 (2025): 9170-9180.

[4], Naiman, Ilan, et al. "Generative modeling of regular and irregular time series data via Koopman VAEs." International conference on learning representations. Vol. 2024. 2024.

Questions for authors

Please refer to critical weaknesses

5 comments

We thank the reviewer for the precise suggestions and for recognizing (1) our occasion–panel representation as a structured way to model asynchronous observations and their co-occurrence, (2) a straightforward, reproducible method, and (3) an evaluation over five benchmarks and complementary metrics, strengthened by perturbation analyses and independently trained runs. We address each concern below.


[W1] Novelty relative to MG-TSD, HDT, and TimEHR

Thank you for pointing us to MG-TSD [1] and HDT [2]; we have added a dedicated comparison in Section 2. Our contribution is a generative formulation of irregular records rather than a new network: ChronoFlow generates different random variables of the observation process in order of statistical granularity, and this order matters, since reversing it raises Set-Discr. from 0.0995 to 0.4511 (Table 2). It differs from these works and from TimEHR [3] as follows:

  1. What the hierarchy is built over (vs. MG-TSD and HDT). Both forecast fully observed, regular series, and their coarse-to-fine levels are coarser views of the same value signal: MG-TSD uses temporally smoothed targets as intermediate diffusion targets, and HDT generates VQ tokens of the long-term trend and then of the target. Neither models when observations occur, which features are observed together, or how many observations a record has. Each level of ChronoFlow instead generates a different random variable: global statistics (M, r) → occasion times and panels (τ, B) → values x. Since this only assumes variable-size sets of timestamped, marked observations, it applies beyond EHRs, e.g., to wearable-sensor recordings (Activity, Table 1).
  2. How records are represented (vs. TimEHR). TimEHR shares the pattern-then-value order but generates a padded 2D time-by-feature image (CWGAN-GP + Pix2Pix); an eICU stay spans 15,840 cells at 1-minute resolution, only 0.86% observed, so convolutions spend capacity on empty cells and make co-observation depend on an arbitrary feature order. ChronoFlow keeps only the measured occasions, their panels, and the observed values (about 34 occasions and 137 values per eICU stay), treats co-observation as an inductive bias (4.0 of 11 features per eICU occasion), and generates each set in parallel, yielding Set-Discr. 0.0995 vs. 0.4907 for TimEHR on eICU (W2 below).
  3. Standard building blocks, by design. We use standard flow-matching objectives and architectures so that the gains can be attributed to the formulation. What is new is the granularity-ordered factorization and what makes it trainable on irregular records: the occasion–panel representation, which turns asynchronous observations into variable-size sets, and the 1D optimal-transport coupling over occasion times, without which Set-Discr. rises to 0.4316 (Table 2). This also keeps the method straightforward and reproducible, as the reviewer noted.

[1] Fan et al. "MG-TSD: Multi-Granularity Time Series Diffusion Models with Guided Learning Process." ICLR 2024.
[2] Feng et al. "HDT: Hierarchical Discrete Transformer for Multivariate Time Series Forecasting." AAAI 2025.
[3] Karami et al. "TimEHR: Image-Based Time Series Generation for Electronic Health Records." IEEE Journal of Biomedical and Health Informatics 2025.

[W1, continued] Necessity of IntensityFlow when (M, r) is computable from B

(M, r) can be computed from B, but generating it first is what makes B easy to generate. It specifies how much a record is observed, the number of occasions and each feature's target frequency, so PatternFlow only has to decide when the occasions occur and which features are measured together. Its effect is large: the ground-truth (M, r) lowers CoOb from 0.0380 to 0.0128 (−66%) and Set-Discr. from 0.0995 to 0.0153 (−85%; Table A4).

As requested, we removed IntensityFlow (PatternFlow then generates slot presence, time, and panel over 128 fixed slots). Like Joint TPV in Table 2, which instead merges PatternFlow and ValueFlow, this removes one of ChronoFlow's two stage boundaries (eICU):

Model (eICU)ProcessSet-Discr. ↓CoOb ↓Pattern-MMD² ↓Trans-MMD² ↓Value-W1 ↓Set-Corr. ↓
w/o IntensityFlow (2.18M)(τ, B) → x0.1178 ± 0.02670.0593 ± 0.00470.0020 ± 0.00040.0069 ± 0.00020.0555 ± 0.00630.0344 ± 0.0047
Joint TPV (Table 2, 1.12M)M → (τ, B, x)0.2639 ± 0.00390.0867 ± 0.00330.0008 ± 0.00020.0275 ± 0.00130.3179 ± 0.00440.0698 ± 0.0101
ChronoFlow (2.06M)(M, r) → (τ, B) → x0.0995 ± 0.00980.0380 ± 0.00470.0002 ± 0.00010.0045 ± 0.00020.0504 ± 0.00470.0385 ± 0.0003

Removing either boundary degrades fidelity: without IntensityFlow, all three pattern metrics worsen despite more parameters (2.18M vs. 2.06M), and merging PatternFlow and ValueFlow raises Set-Discr. and Value-W1 several-fold. This supports both IntensityFlow and the three-stage framework, and we have added the w/o IntensityFlow row to Table 2. Explicit r also makes the observation process controllable: with M fixed, changing r moves the realized panel frequencies almost one-to-one (slope 0.9927 ± 0.0039; our response to Reviewer GEGe, W2), a handle lost without IntensityFlow.


[W2] Comparison with TimEHR and KoVAE

Following your recommendation, we compared against TimEHR [1] and KoVAE [2] on eICU. For TimEHR, all eICU entries are losslessly packed into its 128 × 128 input. KoVAE generates values on a regular grid without observation times or masks, so, as for LogNormMix + mTAN / TFM, we paired it with the LogNormMix pattern model, trained it with the official implementation on eICU's hourly grid, chose among its released configurations on validation (Appendix F.1), and assigned each event KoVAE's value for its hour.

Model (eICU)Set-Discr. ↓CoOb ↓Pattern-MMD² ↓Trans-MMD² ↓Value-W1 ↓Set-Corr. ↓
TimEHR [1]0.4907 ± 0.00640.9857 ± 0.33450.0216 ± 0.01950.0546 ± 0.00920.3648 ± 0.05080.1067 ± 0.0051
LogNormMix + KoVAE [2]0.4733 ± 0.00060.1226 ± 0.00490.0011 ± 0.00000.0251 ± 0.00120.3155 ± 0.00610.1070 ± 0.0018
ChronoFlow0.0995 ± 0.00980.0380 ± 0.00470.0002 ± 0.00010.0045 ± 0.00020.0504 ± 0.00470.0385 ± 0.0003

Both remain almost perfectly separable from real records (Set-Discr. 0.47–0.49), TimEHR also with an extended, validation-tuned budget (1,250 CWGAN / 320 Pix2Pix epochs: Set-Discr. 0.4895, CoOb 0.9165, Value-W1 0.3470), and LogNormMix + KoVAE's Value-W1 is 6.3× ChronoFlow's.

[1] Karami et al. "TimEHR: Image-Based Time Series Generation for Electronic Health Records." IEEE Journal of Biomedical and Health Informatics 2025.
[2] Naiman et al. "Generative Modeling of Regular and Irregular Time Series Data via Koopman VAEs." ICLR 2024.

[W3] Computational efficiency on common hardware

We measured ChronoFlow and the four original baselines of Table 1 on one NVIDIA RTX 6000 Ada GPU (FP32): training through each model's own loop (batch 256, projected to the 150,000-update budget) and end-to-end sampling (batch 16, 100 steps). FLOPs count matrix multiplications and attention, so they are lower bounds.

Model (eICU)Training time (GPU-h) ↓Training FLOPs (PFLOPs) ↓Peak training memory (GiB) ↓Sampling latency (ms/sample) ↓Set-Discr. (Table 1) ↓
Poisson + LatentODE365.839.32.64.20.4242
LogNormMix + mTAN12.421.233.544.80.4555
LogNormMix + TFM1.310.22.4181.70.4518
FlexTPP [1]¹81.01,00029.8551.20.3629
ChronoFlow8.618218.7114.20.0995

¹ With its exact chunked attention (chunk size 256), as in its training run; dense attention does not fit in memory at batch 256.

  • ChronoFlow is Pareto-optimal under every cost measure, and the only such model. No baseline is cheaper and more faithful on any measure, while every baseline is costlier and less faithful than ChronoFlow on at least one: FlexTPP on training time, FLOPs, memory, and latency; LogNormMix + mTAN on training time and memory; LogNormMix + TFM on latency; and Poisson + LatentODE on training time.
  • Training is cheap. Each flow is trained by flow matching on ground-truth upstream conditions, without solving an ODE (unlike Poisson + LatentODE) or generating token by token (unlike FlexTPP): ChronoFlow trains 42.5× faster than Poisson + LatentODE and 9.4× faster than FlexTPP (5.5× fewer FLOPs).
  • Sampling is parallel. Each stage generates all occasions or values at once, so the number of sequential network evaluations is set by the ODE steps rather than the number of events: ChronoFlow samples 4.8× faster than FlexTPP and 1.6× faster than LogNormMix + TFM (1,000 eICU records in under two minutes), and 50 steps keep fidelity (Set-Discr. 0.0912, Section 5.5).

We have added these measurements to Section 5.5 and Appendix G.3.

[1] Draxler et al. "Transformers for Mixed-type Event Sequences." NeurIPS 2025.

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

The paper proposes ChronoFlow, a unified hierarchical flow matchingframework organized by statistical granularity. Following a coarse-to-fine hierarchy, ChronoFlow first generates observation counts and feature-wise frequencies, then jointly generates observation times and feature co-observation patterns, and finally generates values conditioned on the realized pattern.

Critical strengths
  1. Well-motivated problem formulation. Treating “how many / when / which / what value” as one generative object is a real gap.
  2. Strong empirical results on the clinical datasets. We observe great improvements across many datasets.
Critical weaknesses
  1. the technical insights are not so strong. The pattern-then-value ordering has a direct precedent in TimEHR.

  2. The role of r is not clear. Ablate the conditioning variable: (i) M only, (ii) (M, r), (iii) neither, with M sampled from the empirical distribution.

  3. training and inference has strong drift, but no clear solutions are explored in this paper.

  4. baselines are all from the modification version by author, so it may not be strong enough to be valid baseline

Questions for authors

I think overall is a good paper, but some details still need improvements. If with more technical insights, this would be great to be accepted.

5 comments

We thank the reviewer for the constructive feedback and for recognizing (1) our problem formulation, which treats how many, when, which, and what value as a single generative object and fills a real gap, and (2) our strong empirical improvements across clinical datasets. We address each concern below.


[W1] Technical insights and distinction from TimEHR

We agree that the manuscript should state more explicitly why pattern-then-value generation in a sparse, non-autoregressive representation is an advance. TimEHR [1] is a valuable precedent for generating masks before values, but the two differ in how records are represented and generated:

  • TimEHR (dense 2D image) maps records onto a padded time-by-feature image (CWGAN-GP + Pix2Pix). An eICU stay at 1-minute resolution spans 1,440 × 11 = 15,840 cells, only 0.86% observed, so 2D convolutions spend capacity on empty cells and make co-observation depend on an arbitrary feature order.
  • ChronoFlow (sparse, non-autoregressive sets) keeps only the M occasions with a measurement (times τ), their feature panels B, and the N observed values (about 34 occasions and 137 values per eICU stay). IntensityFlow samples M and the feature frequencies r, PatternFlow generates all times and panels jointly in one flow (with 1D optimal-transport coupling over times), and ValueFlow generates the values with a permutation-equivariant Set Transformer [2], each stage producing its whole set in parallel.

Our ablation studies show which boundaries matter:

  1. Cardinality first. (M, r) sets the number of occasions and each feature's target frequency, so later stages generate variable-size sets in parallel; removing this stage degrades pattern fidelity despite more parameters (2.18M vs. 2.06M; CoOb 0.0593 vs. 0.0380, Trans-MMD² 0.0069 vs. 0.0045; W2).
  2. The pattern → value boundary matters; the order inside a stage does not. Reversing or merging the stages raises Set-Discr. from 0.0995 to 0.4511 (Value→Pattern) and 0.2639 (Joint TPV; Table 2), while joint, time-first, and panel-first time–panel generation all stay at 0.07–0.10 (our response to Reviewer MFpg, W2).
  3. Occasion–panel geometry. Without the OT coupling it enables, Set-Discr. rises to 0.4316 and Pattern-MMD² from 0.0002 to 0.0382 (Table 2).

Head-to-head on eICU, with all entries losslessly packed into TimEHR's 128 × 128 input:

Model (eICU)Set-Discr. ↓CoOb ↓Pattern-MMD² ↓Trans-MMD² ↓Value-W1 ↓Set-Corr. ↓
TimEHR [1]0.4907 ± 0.00640.9857 ± 0.33450.0216 ± 0.01950.0546 ± 0.00920.3648 ± 0.05080.1067 ± 0.0051
ChronoFlow0.0995 ± 0.00980.0380 ± 0.00470.0002 ± 0.00010.0045 ± 0.00020.0504 ± 0.00470.0385 ± 0.0003

An extended, validation-tuned TimEHR budget (1,250 CWGAN / 320 Pix2Pix epochs) changes little (Set-Discr. 0.4895, Value-W1 0.3470). We have clarified this distinction in Sections 2 and 3 and will add this comparison to Table 1.

[1] Karami et al. "TimEHR: Image-Based Time Series Generation for Electronic Health Records." IEEE Journal of Biomedical and Health Informatics 2025.
[2] Lee et al. "Set Transformer: A Framework for Attention-based Permutation-Invariant Neural Networks." ICML 2019.

[W2] Role of r and conditioning ablation study

We appreciate this concrete design. IntensityFlow models the number of occasions M, on whose shared timestamps all features are aligned, and each feature's frequency r_f, the fraction of occasions at which it is measured (M · r_f observations). PatternFlow then decides when occasions occur and which features co-occur, while r serves as an explicit, controllable interface. We evaluated the three requested setups on eICU with a shared frozen ValueFlow, and additionally removed Stage 1:

  • (i) M only: a learned count flow p(M) and a newly trained PatternFlow p(τ, B | M).
  • (ii) (M, r) (ChronoFlow): IntensityFlow p(M, r) conditioning p(τ, B | M, r).
  • (iii) Neither: M drawn from the training records' empirical counts, with no global conditioning vector (M enters only as the number of occasion tokens), a training-distribution reference.
  • No Stage 1: PatternFlow predicts slot presence, time, and panel over 128 fixed slots.
Conditioning of PatternFlow (eICU)Set-Discr. ↓CoOb ↓Pattern-MMD² ↓Trans-MMD² ↓Value-W1 ↓Set-Corr. ↓
No Stage 1 (M and r removed)0.1178 ± 0.02670.0593 ± 0.00470.0020 ± 0.00040.0069 ± 0.00020.0555 ± 0.00630.0344 ± 0.0047
(i) M only (r removed)0.0819 ± 0.01960.0301 ± 0.00840.0002 ± 0.00010.0049 ± 0.00040.0487 ± 0.00290.0399 ± 0.0042
(ii) (M, r) (ChronoFlow)0.0995 ± 0.00980.0380 ± 0.00470.0002 ± 0.00010.0045 ± 0.00020.0504 ± 0.00470.0385 ± 0.0003
(iii) Neither, M from training records0.0635 ± 0.01640.0297 ± 0.01450.0002 ± 0.00000.0054 ± 0.00070.0492 ± 0.00350.0360 ± 0.0035
  • Removing Stage 1 hurts. All three pattern metrics degrade despite more parameters (2.18M vs. 2.06M), supporting our statistical-granularity design.
  • ChronoFlow stays close to the training-distribution reference. Sampling M from real records naturally helps, yet ChronoFlow is close on CoOb, Pattern-MMD², and Value-W1 and better on Trans-MMD²; its Set-Discr. gap is the Stage-1 gap analyzed in W3.
  • r is designed for controllability. Removing r does not consistently change unconditional fidelity (M only is better on Set-Discr. and CoOb, worse on Trans-MMD² and Set-Corr.), since PatternFlow learns marginal rates implicitly. However, r is what makes the observation process controllable: a practitioner may want to specify how often each feature is observed and generate diverse sampling patterns that match those counts, which is impossible when only M is modeled.

To test this, we fixed M and the initial noise, varied the target rates r_A → r_B, and measured the realized panel frequencies (128 paired interventions per seed):

Model (eICU)Response slope (→ 1.0)Directional cosine (→ 1.0)Post-repair target-rate MAE ↓
(i) M only (negative control)0.0000 ± 0.0000— (no response by design)—
(ii) (M, r) (ChronoFlow)0.9927 ± 0.00390.9860 ± 0.00300.0186 ± 0.0013

r thus acts as an accurate, interpretable control knob that preserves unconditional fidelity; M only has no such handle.

[W3] Train–inference conditioning drift

ChronoFlow limits this drift by design: it treats co-observation as an inductive bias, since clinical features are typically recorded together (4.0 of 11 per eICU occasion), and generates each occasion with its whole panel, so conditioning crosses only two coarse-to-fine boundaries instead of hundreds of autoregressive steps in which errors compound. We measured the drift by replacing generated conditions with ground-truth ones and by perturbing the ground truth:

Setting (eICU, Table A4)(M, r)(τ, B)Value-W1 ↓Set-Corr. ↓
Full generationsampledsampled0.0504 ± 0.00470.0385 ± 0.0003
+ ground-truth (M, r)ground truthsampled0.0481 ± 0.00370.0365 ± 0.0022
+ ground-truth (M, r), (τ, B)ground truthground truth0.0508 ± 0.01000.0358 ± 0.0019
Ground truth + 10% noiseperturbedperturbed0.0460 ± 0.00150.0378 ± 0.0049
Ground truth + 20% noiseperturbedperturbed0.0473 ± 0.00260.0364 ± 0.0050
  1. Generated patterns barely affect value generation. With ground-truth (τ, B), the value marginals are essentially unchanged and cross-feature correlations change only slightly.
  2. Stage 1 is the main source of the Set-Discr. gap. Ground-truth (M, r) lowers Set-Discr. by 85% (0.0995 → 0.0153) and CoOb from 0.0380 to 0.0128.
  3. Values are robust, while patterns depend on accurate global statistics. Value-W1 stays at 0.0460–0.0473 under 10–20% noise, whereas 10% noise on (M, r) alone keeps Set-Discr. at 0.0113 but raises CoOb to 0.0616.

Mitigation should therefore target Stage 1. A better Stage-1 sampler directly improves the full pipeline with the later stages unchanged, e.g., more ODE steps or a higher-order solver for IntensityFlow, or discrete flow matching for counts (Appendix A). The boundary can also be regularized by conditioning-noise augmentation, which trains PatternFlow on perturbed (M, r), and by self-distillation, which fine-tunes the downstream stages on IntensityFlow's own samples so that training and inference conditions match. We have added these mitigation paths to Section 5.4.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

We thank all reviewers for their constructive feedback and are encouraged that they found the problem formulation well motivated (GEGe, MFpg), the occasion–panel representation well structured (MF4m), the method reproducible (MF4m), and the results strong across five benchmarks (GEGe, MF4m, MFpg).


How we addressed the main concerns (new experiments on eICU and Activity)

Is ChronoFlow more than coarse-to-fine or pattern-first generation? (GEGe, MF4m) Yes. Unlike MG-TSD/HDT (coarser views of the same values) and TimEHR (a dense image with 0.86% of cells observed), ChronoFlow generates different random variables in order of statistical granularity, and this order matters: reversing it raises Set-Discr. from 0.0995 to 0.4511.

Does it beat stronger baselines? (GEGe, MF4m, MFpg) Yes. TimEHR, TimeDiff, and LogNormMix + KoVAE stay almost perfectly separable (Set-Discr. 0.47–0.50 vs. 0.0995), and ChronoFlow is better on every metric than all three.

Is each component necessary, beyond capacity? (GEGe, MF4m, MFpg) Yes. Removing IntensityFlow hurts despite more parameters (2.18M vs. 2.06M), a size-matched Joint TPV collapses (0.3019), and independent panels break co-observation; an M-only variant shows that M drives fidelity and r enables control of observation rates (slope 0.9927), and the findings hold on Activity.

Are the stages consistent, and where does drift enter? (GEGe, MF4m, MFpg) Realized frequencies stay within 1.93 percentage points of the sampled ones, and ground-truth (M, r) lowers Set-Discr. by 85%, locating the gap at Stage 1, for which we give mitigation paths.

Is it efficient? (MF4m) Yes. ChronoFlow trains faster than three of the four baselines (28.4× vs. Poisson + LatentODE, 9.5× vs. FlexTPP) and samples 4.8× faster than FlexTPP, while every faster baseline remains almost fully separable from real records (Set-Discr. ≥ 0.42).

Are pattern–value dependencies preserved? (MFpg) Yes. ChronoFlow is closest to Real–Real on all three patient-level dependency metrics, while the LogNormMix baselines stay at the level of a marginal-preserving shuffle.


Revisions (in blue in the revised paper)

  • Section 2: MG-TSD, HDT, and KoVAE added, and the distinction from TimEHR stated; Section 3: each stage generates its whole set in parallel.
  • Section 4: which distribution each metric evaluates (the full sample, the sampling pattern, or the values given the pattern) and why each metric is needed.
  • Section 5.1: the rationale for each baseline and how it was tuned.
  • Table 2 and Section 5.3: new ablation studies, w/o IntensityFlow and Bernoulli-r panels, with the time–panel order and a size-matched Joint TPV in Appendix G.1 and the M-only variant in Appendix G.4.
  • Section 5.4: mitigation paths for the train–inference conditioning shift.
  • Section 5.5 and Appendix G.3: actual training time, training and inference FLOPs and memory, and sampling latency; Appendix G.2: consistency between sampled and realized frequencies.
  • TimEHR, TimeDiff, and LogNormMix + KoVAE will be added to Table 1.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.