Pith. sign in

REVIEW 3 major objections 6 minor 300 references

WrapFlow forecasts irregular multivariate time series in continuous time without grid preprocessing or training-time ODE solvers.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-31 19:47 UTC pith:CNRYIKUP

load-bearing objection Solid IMTSF engineering paper: typed continuous-time tokens plus residual flow matching beat most baselines, but the continuous-ODE story is thinner than the SOTA framing. the 3 major comments →

arxiv 2607.28035 v1 pith:CNRYIKUP submitted 2026-07-30 cs.LG

Enhancing Irregular Time Series Forecasting with Continuous-Time Modeling Framework

Tianen Shen , Zhengyu Li , Yutong Li , Xiangfei Qiu , Xingjian Wu , Bin Yang , Jilin Hu This is my paper
classification cs.LG
keywords irregular multivariate time seriescontinuous-time tokenizationresidual flow matchingtime series forecastingTransformergap-aware tokenssimulation-free training
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Irregular multivariate series—clinical monitors, motion sensors, climate stations—arrive at uneven times, and those gaps themselves carry meaning. Common practice either snaps the data onto a fixed grid, which erases that timing signal, or runs specialized differential-equation models that are slow to train. This paper argues both problems can be avoided at once. It encodes each raw observation, long silence, and training mask as a typed continuous-time token, lets an ordinary Transformer build memory over those tokens, then refines a simple base forecast by learning a residual vector field with flow matching that never simulates an ODE during training. At test time only a few fixed Euler steps are needed. On four standard benchmarks the method reports the best or near-best error, showing continuous-time forecasting need not mean heavy solvers or lost temporal semantics.

Core claim

WrapFlow is a unified continuous-time framework for irregular multivariate forecasting: Continuous-Time Tokenization preserves raw event values, identities, absolute and relative times, and long-gap semantics without discretization, while simulation-free Residual Flow Matching learns conditional residual corrections around a detached base prediction, yielding state-of-the-art MSE and MAE on PhysioNet, MIMIC, HumanActivity, and USHCN with only lightweight fixed-step rollouts at inference.

What carries the argument

Residual Flow Matching around a base anchor: the model first emits a deterministic prediction, then trains a conditional vector field on the linear path from noise to the residual (true value minus detached base); training regresses that field directly, and inference integrates a few Euler steps from zero residual to produce the final correction.

Load-bearing premise

A few fixed Euler steps along a straight residual path from a detached base forecast are enough to capture the continuous dynamics that actually improve irregular forecasts.

What would settle it

On the same four benchmarks, replace the residual flow with the base head alone or with many more/fewer rollout steps and check whether the reported gains over the strongest baselines disappear; if they do not, residual path geometry is not doing the claimed work.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Irregular forecasting pipelines can keep raw asynchronous events and still use a standard Transformer backbone.
  • Continuous residual refinement can be trained without numerical ODE simulation or backprop through a solver.
  • Long unobserved gaps become explicit model inputs rather than missing-data artifacts to impute away.
  • High-quality continuous forecasts are obtainable from a small fixed number of inference rollout steps.
  • The same residual-flow head can be supervised on both future targets and masked historical spans during training.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If residual flow is the main gain, similar anchor-plus-residual matching could improve other query-based irregular tasks such as imputation or online monitoring without changing the encoder.
  • Gap-type tokens suggest a natural extension to regimes where silence length itself is a clinical or physical event (e.g., seizure-free intervals or sensor dropouts).
  • Because training never differentiates through a solver, the method may scale more readily to longer lookbacks than classical Neural-ODE forecasters on the same hardware.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. WrapFlow proposes a continuous-time framework for irregular multivariate time series forecasting (IMTSF). On the input side, Continuous-Time Tokenization encodes raw asynchronous events as typed tokens (Event, GAP-aware Event, Mask GAP) with value, variable, absolute-time, and relative-interval features, then feeds them to a standard Transformer backbone. On the output side, a query-based decoder produces a deterministic base (anchor) prediction; Residual Flow Matching then learns a conditional residual vector field along linear noise-to-residual paths in a simulation-free training regime, with few fixed-step Euler rollouts at inference to correct the anchor. Training jointly optimizes direct anchor loss and future/history flow-matching losses. Experiments on PhysioNet, MIMIC, HumanActivity, and USHCN against 14 baselines report strong MSE/MAE results, with ablations and sensitivity studies in the main text and appendix.

Significance. If the claims hold, the paper offers a practical middle path between discretization pipelines that discard continuous-time semantics and ODE-native models that require specialized dynamics and solver-heavy training. Using a standard Transformer with typed continuous-time tokens and a simulation-free residual flow head is attractive for adoption. The empirical package is a genuine strength: four standard IMTS benchmarks, fourteen baselines spanning continuous-time, set, graph, and patch methods, five seeds with mean±std, no DropLast, plus ablations (Table 3), lookback/horizon studies, and sensitivity to D, L, and τ_gap. Code and data are promised in the supplement. The main open question is how much of the gain is truly continuous residual transport versus typed event encoding plus any residual corrector.

major comments (3)
  1. [§3.6–3.7, Eqs. (12)–(19)] §3.6–3.7, Eqs. (12)–(19): Residual Flow Matching is central to the continuous-time output claim, but train and inference paths are inconsistent and unanalyzed. Training samples x_0 ~ N(0,I) and regresses velocity r*−x_0 along the linear path; inference integrates from the deterministic state r_0=0. The manuscript does not show that the learned field is approximately straight, low-error under the zero start, or that few-step Euler faithfully realizes the trained path. Please either (i) analyze/justify this mismatch (e.g., path straightness, endpoint error vs. start distribution) or (ii) align training with the inference start, and state the Euler step count used for Table 1 with a short sensitivity study.
  2. [Table 3; §4.3] Table 3, “w/o Flow-Matching Refinement”: this ablation removes the entire residual head and keeps only the deterministic anchor, so it cannot attribute gains to continuous residual transport versus ordinary one-shot residual regression (e.g., predict r̂ and add it to the anchor). Given that CT Tokenization removal hurts far more than FM removal, and that HumanActivity MSE is not best in Table 1, the claim that residual path geometry and few-step ODE rollout drive continuous-time forecasting gains is under-supported. Add a control that replaces FM+ODE with direct residual regression under the same conditioning (context, time, anchor), and discuss attribution accordingly.
  3. [Abstract; §1; Table 1; §4.2] Abstract, §1, and §4.2 claim state-of-the-art performance, but Table 1 is mixed: on HumanActivity MSE, APN (0.0421) and several others beat WrapFlow (0.0449); WrapFlow is best or second-best on most other cells, with clearer gains on USHCN and MAE. Please temper the global SOTA language to match the table (e.g., best or competitive across datasets/metrics, with largest gains on USHCN/MAE) and avoid implying uniform dominance.
minor comments (6)
  1. [§3.3; A.6] §3.3: the gap threshold τ_gap is introduced without a default value or selection rule in the main text; appendix sensitivity helps, but main-text defaults (and whether τ_gap is tuned per dataset) should be stated once for reproducibility.
  2. [§3.7; §4.1] §3.6–3.7: λ_f, λ_h, the form of ℓ_flow / ℓ_dir (MSE vs MAE, etc.), and the number of Euler steps are free parameters of the method but not fully specified beside the high-level objective in Eq. (17). A short hyperparameter table would help.
  3. [Figure 1] Figure 1 is dense; the training vs inference paths for residual flow (noise interpolation vs r_0=0 rollout) would be clearer with an explicit small schematic or caption callout matching Eqs. (12) and (18).
  4. [§2.2] Related work §2.2 correctly positions CGFM and TSFlow; a one-sentence contrast on asynchronous variable-time queries vs sequence-level FM would further clarify novelty without overstating uniqueness.
  5. [Throughout; Table 1] Minor prose issues: missing spaces in several compounds (“Existingapproaches”, “continuous-timesemantics”, “WrapFlowdirectly”) and leading-zero omission in tables is fine if consistent, but bold/underline rules should be checked where ties or near-ties occur.
  6. [§3.1; §3.5] §3.1 notation: Ω_x / Ω_y and per-variable lengths are clear; briefly state how multi-step future queries share t_last vs previous query time when building Δs_j, to avoid ambiguity in multi-horizon settings.

Circularity Check

0 steps flagged

No significant circularity: supervised IMTS forecasting with held-out MSE/MAE; losses optimize predictors rather than redefine the test metric.

full rationale

WrapFlow is an engineering ML method paper. Its central claims are architectural (Continuous-Time Tokenization + Residual Flow Matching around a detached base anchor) and empirical (Table 1 SOTA vs external baselines on PhysioNet, MIMIC, HumanActivity, USHCN). Training objectives L_dir, L_f, L_h regress base estimates and conditional residual velocities against ground-truth targets on valid positions; residual targets are explicitly detached from the base head (Sec. 3.6). Inference adds a few-step Euler residual to the anchor and is scored by held-out MSE/MAE—not by a quantity fitted into the metric by construction. Self-citations (e.g., Qiu/Liu group TFB–TAB protocols, APN baseline, dataset splits) supply evaluation conventions and comparators; they do not supply a uniqueness theorem or fitted identity that forces the leaderboard result. No step reduces Eq. (prediction) to Eq. (input fit) or renames a known law as a derived first principle. Weaknesses (train/infer path mismatch, incomplete FM ablations) are correctness/attribution issues, not circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

Load-bearing content is mostly engineering inductive biases and hyperparameters, not physical axioms. Claims rest on standard ML math, domain assumptions about irregular sampling semantics, and paper-specific token/flow constructions whose value is justified empirically rather than independently measured outside these benchmarks.

free parameters (5)
  • λ_f, λ_h (FM loss weights)
    Scale future and history flow-matching losses against direct anchor loss in Eq. 17; chosen by the authors and not derived.
  • τ_gap (gap-aware threshold)
    Decides when an event becomes a GAP-aware token (§3.3); sensitivity studied in appendix, value is a design knob.
  • hidden dimension D and encoder depth L
    Capacity hyperparameters selected per dataset complexity (Fig. 2, Fig. 3); performance depends on these choices.
  • number of Euler rollout steps at inference
    Controls residual integration fidelity (§3.7); paper claims a small fixed number suffices but does not fix a universal value in the main text.
  • masked-history span sampling for Ω_h
    Training-time auxiliary supervision depends on how continuous historical spans are masked; procedure is described qualitatively.
axioms (5)
  • domain assumption Standard Transformer attention and MLP function approximation suffice to encode irregular event tokens and decode query contexts.
    Invoked by adopting a standard masked-attention backbone and shared query decoder (§3.4–3.5) without specialized continuous dynamics modules.
  • domain assumption Long unobserved intervals carry usable dynamic semantics distinct from ordinary short gaps and from artificially masked spans.
    Motivates GAP-aware vs Mask GAP token types in §1 and §3.3; supported by ablation but not independently proven.
  • domain assumption Conditional flow matching along linear residual paths (x_τ=(1-τ)x_0+τ r) yields a useful residual corrector when trained by velocity regression without solver simulation.
    Core of Residual Flow Matching (§3.6, Eqs. 12–15), imported from flow-matching literature and applied to IMTS residuals.
  • domain assumption Forecast targets at query times are well evaluated by MSE/MAE under the cited lookback/horizon protocols and 80/10/10 splits.
    Experimental setup §4.1 and Appendix A follow prior IMTSF benchmarks (APN, Hi-Patch, etc.).
  • standard math Linear algebra / calculus of ODE rollouts and expectation of ℓ_flow / ℓ_dir losses are valid optimization targets.
    Background mathematics for Eqs. 12–19; not re-proved.
invented entities (2)
  • Continuous-Time Tokenization (Event / GAP-aware Event / Mask GAP tokens) no independent evidence
    purpose: Encode raw irregular observations, long gaps, and masked recovery positions without grid discretization.
    Defined in §3.3 with type embeddings and absolute/relative time encoders; utility argued via ablations, not external measurement theory.
  • Residual Flow Matching refinement branch for IMTS queries no independent evidence
    purpose: Learn conditional residual vector fields around base anchor predictions with simulation-free training and few-step inference rollout.
    Formulated in §3.6–3.7 as future + masked-history FM; related to CGFM-style residual flows but specialized here to irregular query decoding.

pith-pipeline@v1.2.0-daily-grok45 · 20720 in / 3475 out tokens · 71505 ms · 2026-07-31T19:47:28.052007+00:00 · methodology

0 comments
read the original abstract

Irregular multivariate time series are widely encountered in applications such as healthcare monitoring, human activity recognition, and environmental sensing. Their core challenges stem from asynchronous observations, non-uniform sampling intervals, and the fact that temporal patterns themselves carry critical dynamic information. Existing approaches either rely on discretization-based preprocessing (e.g., interpolation, imputation, or aggregation), which disrupts the underlying continuous-time semantics, or adopt continuous-time modeling via ODE-based frameworks, which typically require specialized architectures and incur substantial computational overhead due to numerical solvers. To address these limitations, we propose WrapFlow, a continuous-time modeling framework for irregular time series forecasting. On the input side, WrapFlow introduces Continuous-Time Tokenization, which directly encodes raw observation events and explicitly models long unobserved intervals via gap-aware tokens. The resulting continuous-time tokens are then processed by a standard Transformer backbone to capture long-range temporal dependencies. On the output side, we develop a simulation-free training paradigm for Residual Flow Matching, which learns conditional residual vector fields around base predictions while avoiding numerical-solver simulation and backpropagation during training. This design enables high-quality continuous forecasting using only a small number of fixed rollout steps at inference. Extensive experiments on multiple real-world datasets demonstrate that WrapFlow achieves state-of-the-art performance.

Figures

Figures reproduced from arXiv: 2607.28035 by Bin Yang, Jilin Hu, Tianen Shen, Xiangfei Qiu, Xingjian Wu, Yutong Li, Zhengyu Li.

Figure 1
Figure 1. Figure 1: WrapFlow architecture. (a) Continuous-Time Tokenization and Backbone convert irregular observations into typed [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Parameter sensitivity of WrapFlow with respect to the hidden dimension [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Parameter sensitivity of WrapFlow on HumanActivity, USHCN, PhysioNet, and MIMIC. From top to bottom: hidden [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Forecasting performance with varying lookback lengths and fixed forecast horizons. From left to right, the datasets [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

300 extracted references · 35 linked inside Pith

  1. [1]

    S.; Januschowski, T.; and G \"u nnemann, S

    Bilo s , M.; Sommer, J.; Rangapuram, S. S.; Januschowski, T.; and G \"u nnemann, S. 2021. Neural flows: Efficient alternative to neural ODEs. Advances in neural information processing systems, 34: 21325--21337

    2021

  2. [2]

    Che, Z.; Purushotham, S.; Cho, K.; Sontag, D.; and Liu, Y. 2018. Recurrent neural networks for multivariate time series with missing values. Scientific reports, 8(1): 6085

    2018

  3. [3]

    T.; Rubanova, Y.; Bettencourt, J.; and Duvenaud, D

    Chen, R. T.; Rubanova, Y.; Bettencourt, J.; and Duvenaud, D. K. 2018. Neural ordinary differential equations. Advances in neural information processing systems, 31

    2018

  4. [4]

    R.; Li, J.; Zhang, X.; Hong, D.; Gupta, R

    Chowdhury, R. R.; Li, J.; Zhang, X.; Hong, D.; Gupta, R. K.; and Shang, J. 2023. Primenet: Pre-training for irregular multivariate time series. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, 7184--7192

    2023

  5. [5]

    De Brouwer, E.; Simm, J.; Arany, A.; and Moreau, Y. 2019. GRU-ODE-Bayes: Continuous modeling of sporadically-observed time series. Advances in neural information processing systems, 32

    2019

  6. [6]

    Easterling, D. 2002. United States Historical Climatology Network daily temperature and precipitation data (1871-1997). Technical report, ORNL Oak Ridge National Laboratory (US)

    2002

  7. [7]

    Gao, H.; Shen, W.; Qiu, X.; Xu, R.; Yang, B.; and Hu, J. 2025. SSD-TS : Exploring the potential of linear state space models for diffusion models in time series imputation. In SIGKDD 2025

    2025

  8. [8]

    Horn, M.; Moor, M.; Bock, C.; Rieck, B.; and Borgwardt, K. 2020. Set functions for time series. In ICML, volume 119, 4353--4363

    2020

  9. [9]

    Z.; Wang, S.; Zhang, J.; Li, Z.; and Chen, T

    Hu, Y.; Wang, X.; Ding, Z.; Wu, L.; Zhang, H.; Li, S. Z.; Wang, S.; Zhang, J.; Li, Z.; and Chen, T. 2024. FlowTS: Time Series Generation via Rectified Flow. arXiv preprint arXiv:2411.07506

    Pith/arXiv arXiv 2024

  10. [10]

    E.; Pollard, T

    Johnson, A. E.; Pollard, T. J.; Shen, L.; Lehman, L.-w. H.; Feng, M.; Ghassemi, M.; Moody, B.; Szolovits, P.; Anthony Celi, L.; and Mark, R. G. 2016. MIMIC-III, a freely accessible critical care database. Scientific data, 3(1): 1--9

    2016

  11. [11]

    Kidger, P.; Morrill, J.; Foster, J.; and Lyons, T. J. 2020. Neural Controlled Differential Equations for Irregular Time Series. In NeurIPS 2020

    2020

  12. [12]

    u dke, D.; Schwinn, L.; and G \

    Kollovieh, M.; Lienen, M.; L \" u dke, D.; Schwinn, L.; and G \" u nnemann, S. 2025. Flow Matching with Gaussian Process Priors for Probabilistic Time Series Forecasting. In The Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025 . OpenReview.net

    2025

  13. [13]

    Li, B.; Luo, Y.; Liu, Z.; Zheng, J.; Lv, J.; and Ma, Q. 2025. Hyperimts: Hypergraph neural network for irregular multivariate time series forecasting. arXiv preprint arXiv:2505.17431

    Pith/arXiv arXiv 2025

  14. [14]

    T.; Ben-Hamu, H.; Nickel, M.; and Le, M

    Lipman, Y.; Chen, R. T.; Ben-Hamu, H.; Nickel, M.; and Le, M. 2022. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747

    Pith/arXiv arXiv 2022

  15. [15]

    Liu, X.; Gong, C.; and Liu, Q. 2022. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003

    Pith/arXiv arXiv 2022

  16. [16]

    Liu, X.; Qiu, X.; Cheng, H.; Wu, X.; Guo, C.; Yang, B.; and Hu, J. 2026 a . ASTGI: Adaptive Spatio-Temporal Graph Interactions for Irregular Multivariate Time Series Forecasting. In ICLR

    2026

  17. [17]

    Liu, X.; Qiu, X.; Wu, X.; Li, Z.; Guo, C.; Hu, J.; and Yang, B. 2026 b . Rethinking irregular time series forecasting: A simple yet effective baseline. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, 23873--23881

    2026

  18. [18]

    Luo, Y.; Zhang, B.; Liu, Z.; and Ma, Q. 2025. Hi-Patch: Hierarchical Patch GNN for Irregular Multivariate Time Series. In ICML, volume 267, 41494--41519

    2025

  19. [19]

    Mercatali, G.; Freitas, A.; and Chen, J. 2024. Graph neural flows for unveiling systemic interactions among irregularly sampled time series. Advances in Neural Information Processing Systems, 37: 57183--57206

    2024

  20. [20]

    Paszke, A.; Gross, S.; Massa, F.; Lerer, A.; Bradbury, J.; Chanan, G.; Killeen, T.; Lin, Z.; Gimelshein, N.; Antiga, L.; et al. 2019. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32

    2019

  21. [21]

    S.; Sheng, Z.; and Yang, B

    Qiu, X.; Hu, J.; Zhou, L.; Wu, X.; Du, J.; Zhang, B.; Guo, C.; Zhou, A.; Jensen, C. S.; Sheng, Z.; and Yang, B. 2024. TFB: Towards Comprehensive and Fair Benchmarking of Time Series Forecasting Methods. In Proc. VLDB Endow. , 2363--2377

    2024

  22. [22]

    S.; and Yang, B

    Qiu, X.; Li, Z.; Qiu, W.; Hu, S.; Zhou, L.; Wu, X.; Li, Z.; Guo, C.; Zhou, A.; Sheng, Z.; Hu, J.; Jensen, C. S.; and Yang, B. 2025 a . TAB: Unified Benchmarking of Time Series Anomaly Detection Methods. In Proc. VLDB Endow. , 2775--2789

    2025

  23. [23]

    Qiu, X.; Wu, X.; Lin, Y.; Guo, C.; Hu, J.; and Yang, B. 2025 b . DUET : Dual Clustering Enhanced Multivariate Time Series Forecasting. In SIGKDD 2025, 1185--1196

    2025

  24. [24]

    Rasul, K.; Seward, C.; Schuster, I.; and Vollgraf, R. 2021. Autoregressive denoising diffusion models for multivariate probabilistic time series forecasting. In ICML, 8857--8868

    2021

  25. [25]

    Schirmer, M.; Eltayeb, M.; Lessmann, S.; and Rudolph, M. 2022. Modeling irregular time series with continuous recurrent units. In International conference on machine learning, 19388--19405

    2022

  26. [26]

    N.; and Marlin, B

    Shukla, S. N.; and Marlin, B. M. 2021. Multi-Time Attention Networks for Irregularly Sampled Time Series. In International Conference on Learning Representations

    2021

  27. [27]

    J.; Celi, L

    Silva, I.; Moody, G.; Scott, D. J.; Celi, L. A.; and Mark, R. G. 2012. Predicting in-hospital mortality of icu patients: The physionet/computing in cardiology challenge 2012. In 2012 computing in cardiology, 245--248

    2012

  28. [28]

    Tashiro, Y.; Song, J.; Song, Y.; and Ermon, S. 2021. Csdi: Conditional score-based diffusion models for probabilistic time series imputation. Advances in neural information processing systems, 34: 24804--24816

    2021

  29. [29]

    Tong, A.; Fatras, K.; Malkin, N.; Huguet, G.; Zhang, Y.; Rector-Brooks, J.; Wolf, G.; and Bengio, Y. 2023. Improving and generalizing flow-based generative models with minibatch optimal transport. arXiv preprint arXiv:2302.00482

    Pith/arXiv arXiv 2023

  30. [30]

    Wang, H.; Li, H.; Chen, X.; Gong, M.; Chen, Z.; et al. 2025. Optimal transport for time series imputation. In ICLR

    2025

  31. [31]

    Wu, X.; Qiu, X.; Gao, H.; Hu, J.; Yang, B.; and Guo, C. 2025. K ^2 VAE : A Koopman-Kalman Enhanced Variational AutoEncoder for Probabilistic Time Series Forecasting. In ICML

    2025

  32. [32]

    Xu, H.; Yu, R.; Wu, L.; Wang, X.; and Liu, Q. 2025. Bridging the last mile of prediction: Enhancing time series forecasting with conditional guided flow matching. arXiv preprint arXiv:2507.07192

    Pith/arXiv arXiv 2025

  33. [33]

    K.; Madhusudhanan, K.; Scholz, R.; Ahmed, N.; Burchert, J.; Jawed, S.; Born, S.; and Schmidt-Thieme, L

    Yalavarthi, V. K.; Madhusudhanan, K.; Scholz, R.; Ahmed, N.; Burchert, J.; Jawed, S.; Born, S.; and Schmidt-Thieme, L. 2024. Grafiti: Graphs for forecasting irregularly sampled time series. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 16255--16263

    2024

  34. [34]

    Yao, Z.; Bi, J.; and Chen, Y. 2018. Applying Deep Learning to Individual and Community Health Monitoring Data: A Survey. Int. J. Autom. Comput., 15(6): 643--655

    2018

  35. [35]

    Zhang, J.; Zheng, S.; Cao, W.; Bian, J.; and Li, J. 2023. Warpformer: A multi-scale modeling approach for irregular clinical time series. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 3273--3285

    2023

  36. [36]

    Zhang, W.; Yin, C.; Liu, H.; Zhou, X.; and Xiong, H. 2024. Irregular Multivariate Time Series Forecasting: A Transformable Patching Graph Neural Networks Approach. In ICML, volume 235, 60179--60196

    2024

  37. [37]

    Zhang, X.; Zeman, M.; Tsiligkaridis, T.; and Zitnik, M. 2022. Graph-Guided Network for Irregularly Sampled Multivariate Time Series. In ICLR

    2022

  38. [38]

    Zhou, Z.; Huang, Y.; Wang, Y.; Wu, Y.; Kwok, J.; and Liang, Y. 2026. Revitalizing canonical pre-alignment for irregular multivariate time series forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, 29115--29123

    2026

  39. [39]

    Proceedings of the National Academy of Sciences , volume=

    Hamiltonian systems and transformation in Hilbert space , author=. Proceedings of the National Academy of Sciences , volume=. 1931 , publisher=

    1931

  40. [40]

    An Introduction to the Kalman Filter , author=

  41. [41]

    , author=

    beta-vae: Learning basic visual concepts with a constrained variational framework. , author=. ICLR (Poster) , volume=

  42. [42]

    NeurIPS , volume=

    Koopa: Learning non-stationary time series dynamics with koopman predictors , author=. NeurIPS , volume=

  43. [43]

    NeurIPS , volume=

    Variational autoencoder for deep learning of images, labels and captions , author=. NeurIPS , volume=

  44. [44]

    Embedded systems programming , volume=

    Kalman filtering , author=. Embedded systems programming , volume=

  45. [45]

    Physica D: Nonlinear Phenomena , volume=

    Linearization in the large of nonlinear systems and Koopman operator spectrum , author=. Physica D: Nonlinear Phenomena , volume=. 2013 , publisher=

    2013

  46. [46]

    Journal of fluid mechanics , volume=

    Dynamic mode decomposition of numerical and experimental data , author=. Journal of fluid mechanics , volume=. 2010 , publisher=

    2010

  47. [47]

    International journal of forecasting , volume=

    DeepAR: Probabilistic forecasting with autoregressive recurrent networks , author=. International journal of forecasting , volume=

  48. [48]

    IEEE Transactions on Intelligent Transportation Systems , volume=

    Metaprobformer for charging load probabilistic forecasting of electric vehicle charging stations , author=. IEEE Transactions on Intelligent Transportation Systems , volume=

  49. [49]

    Applied Energy , volume=

    A deep generative model for probabilistic energy forecasting in power systems: normalizing flows , author=. Applied Energy , volume=

  50. [50]

    Science Advances , volume=

    Generative emulation of weather forecast ensembles with diffusion models , author=. Science Advances , volume=

  51. [51]

    High-dimensional multivariate forecasting with low-rank Gaussian Copula Processes , booktitle =

    David Salinas and Michael Bohlke. High-dimensional multivariate forecasting with low-rank Gaussian Copula Processes , booktitle =

  52. [52]

    Multivariate Probabilistic Time Series Forecasting via Conditioned Normalizing Flows , booktitle =

    Kashif Rasul and Abdul. Multivariate Probabilistic Time Series Forecasting via Conditioned Normalizing Flows , booktitle =

  53. [53]

    Seeger and Jan Gasthaus and Lorenzo Stella and Yuyang Wang and Tim Januschowski , title =

    Syama Sundar Rangapuram and Matthias W. Seeger and Jan Gasthaus and Lorenzo Stella and Yuyang Wang and Tim Januschowski , title =. NeurIPS , pages =

  54. [54]

    IJCAI , pages =

    Longyuan Li and Junchi Yan and Xiaokang Yang and Yaohui Jin , title =. IJCAI , pages =

  55. [55]

    Shibo Feng and Chunyan Miao and Ke Xu and Jiaxiang Wu and Pengcheng Wu and Yang Zhang and Peilin Zhao , title =

  56. [56]

    Matteson , title =

    Binh Tang and David S. Matteson , title =. NeurIPS , pages =

  57. [57]

    ICML , volume =

    Kashif Rasul and Calvin Seward and Ingmar Schuster and Roland Vollgraf , title =. ICML , volume =

  58. [58]

    NeurIPS , year =

    Kollovieh, Marcel and Ansari, Abdul Fatir and Bohlke-Schneider, Michael and Zschiegner, Jasper and Wang, Hao and Wang, Yuyang , title =. NeurIPS , year =

  59. [59]

    NeurIPS , pages =

    Yusuke Tashiro and Jiaming Song and Yang Song and Stefano Ermon , title =. NeurIPS , pages =

  60. [60]

    Management science , volume=

    Scoring rules for continuous probability distributions , author=. Management science , volume=

  61. [61]

    SIGKDD , year=

    FM4TS-Bench: A Comprehensive and Unified Benchmark of Foundation Models for Time Series Forecasting , author=. SIGKDD , year=

  62. [62]

    ICLR , year =

    Xinyao Fan and Yueying Wu and Chang Xu and Yuhao Huang and Weiqing Liu and Jiang Bian , title =. ICLR , year =

  63. [63]

    AAAI , volume=

    Causal recurrent variational autoencoder for medical time series generation , author=. AAAI , volume=

  64. [64]

    arXiv preprint arXiv:2111.08095 , year=

    Timevae: A variational auto-encoder for multivariate time series generation , author=. arXiv preprint arXiv:2111.08095 , year=

    Pith/arXiv arXiv

  65. [65]

    ICASSP , pages=

    Missing data in traffic estimation: A variational autoencoder imputation method , author=. ICASSP , pages=

  66. [66]

    Measurement , volume=

    Variational auto-encoders based on the shift correction for imputation of specific missing in multivariate time series , author=. Measurement , volume=

  67. [67]

    ICASSP , pages=

    Anomaly detection for time series using vae-lstm hybrid model , author=. ICASSP , pages=

  68. [68]

    WWW , pages=

    Revisiting VAE for Unsupervised Time Series Anomaly Detection: A Frequency Perspective , author=. WWW , pages=

  69. [69]

    WWW , pages=

    A semi-supervised vae based active anomaly detection framework in multivariate time series for online systems , author=. WWW , pages=

  70. [70]

    AAAI , volume=

    Ddg-da: Data distribution generation for predictable concept drift adaptation , author=. AAAI , volume=

  71. [71]

    Nikolaos Passalis and Anastasios Tefas and Juho Kanniainen and Moncef Gabbouj and Alexandros Iosifidis , title =

  72. [72]

    Progress in Environmental Science , volume=

    Nonstationary time series analysis and forecasting , author=. Progress in Environmental Science , volume=. 1999 , publisher=

    1999

  73. [73]

    energy , volume=

    Grey prediction with rolling mechanism for electricity demand forecasting of Turkey , author=. energy , volume=. 2007 , publisher=

    2007

  74. [74]

    1991 , publisher=

    Time series: theory and methods , author=. 1991 , publisher=

    1991

  75. [75]

    Ziquan Fang and Lu Pan and Lu Chen and Yuntao Du and Yunjun Gao , title =. Proc

  76. [76]

    ICML , year =

    Yong Liu and Haoran Zhang and Chenyu Li and Xiangdong Huang and Jianmin Wang and Mingsheng Long , title =. ICML , year =

  77. [77]

    ICML , year =

    Gerald Woo and Chenghao Liu and Akshat Kumar and Caiming Xiong and Silvio Savarese and Doyen Sahoo , title =. ICML , year =

  78. [78]

    Nguyen and Pankaj Dayama and Chandra Reddy and Wesley M

    Vijay Ekambaram and Arindam Jati and Nam H. Nguyen and Pankaj Dayama and Chandra Reddy and Wesley M. Gifford and Jayant Kalagnanam , title =. CoRR , volume =

  79. [79]

    Chronos: Learning the Language of Time Series , journal =

    Abdul Fatir Ansari and Lorenzo Stella and Ali Caner T. Chronos: Learning the Language of Time Series , journal =

  80. [80]

    ICML , year =

    Abhimanyu Das and Weihao Kong and Rajat Sen and Yichen Zhou , title =. ICML , year =

Showing first 80 references.