Pith. sign in

REVIEW 5 major objections 5 minor 61 references

MuSiCNet: A Gradual Coarse-to-Fine Framework for Irregularly Sampled Multivariate Time Series Analysis

T0 review · 5 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read MuSiCNet claims that treating an irregularly sampled multivariate time series as a coarse-to-fine stack of relatively regular series, refined by frequency-domain correlation-weighted attention, is enough to match or beat task-specific…

desk verdict Genuinely new multi-scale architecture for irregular time series, but the 'consistent SOTA' claim rests on an interpolation comparison that is unverifiable as reported. read the letter →

arxiv 2412.01063 v1 pith:EHE7KJRB submitted 2024-12-02 cs.LG stat.ML

Jiexi Liu , Meng Cao , Songcan Chen This is my paper · ORCID
classification cs.LGstat.ML
keywords irregularlysampledtimeseriesmultivariatemulti-scalelearningLomb-Scargleperiodogramdynamicwarpingattentionmechanisminterpolationforecasting
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that irregular sampling in multivariate time series is relative, not absolute: by imposing lower, uniform sampling rates on an instance, the same data can be viewed as a stack of increasingly regular series, from coarse to fine. The paper's claim is that building representations by refining from coarse to fine, rather than treating the data as a regular grid with missing values, makes one general model competitive with task-specific state-of-the-art on classification, interpolation, and forecasting. If true, this gives practitioners a single imputation-free backbone for irregular multivariate series, with coarse scales supplying broad temporal context and fine scales preserving detail. The claim is carried by a shared encoder-decoder network whose inter-series attention is re-weighted by frequency-domain correlation matrices computed with Lomb-Scargle periodograms and dynamic time warping.

What carries the argument

The load-bearing object is the LSP-DTW correlation matrix $C_T$, whose entries are $c_{d_i d_j} = \min_\pi \sum_{(m,n)\in\pi} \left(\mathrm{LSP}(X_{d_i})[m] - \mathrm{LSP}(X_{d_j})[n]\right)^2$, i.e. the dynamic time warping distance between the Lomb-Scargle periodograms of two variables. The Lomb-Scargle periodogram converts unevenly spaced observations of each variable into a power spectrum without imputation, and DTW aligns the two spectra so that shared periodic structure shows up as a small distance and thus high correlation. This matrix re-weights inter-series attention inside the CorrNet encoder at every scale. Around it, the coarse-to-fine hierarchy $X^{(l)}_{\mathrm{mask}} = M^{(l)} \odot \mathrm{AvgPooling}_L(X)$ supplies broad temporal context, and two cross-scale losses, reconstruction adjustment and contrastive consistency, transfer that context into the finer representations.

What would settle it

On a synthetic irregular series with known inter-variable correlations and missingness at the level of P12, compare the ranking of pairwise correlations from LSP-DTW against ground truth; if the ranking is no better than chance at 88 percent missingness, or if replacing the LSP-DTW matrix with a fixed identity matrix on real datasets does not hurt AUROC, the central correlation mechanism is not load-bearing.

Watch

Extended reading notes

Core claim

MuSiCNet's central discovery is that the irregularity of an irregularly sampled multivariate time series is relative: applying average pooling at decreasing window sizes produces a hierarchy of relatively regular series, and this hierarchy is an asset rather than a nuisance. Within each scale, a multi-correlation attention network encodes the series at reference time points using time attention for intra-series information and a correlation matrix, computed by Lomb-Scargle periodogram followed by dynamic time warping (LSP-DTW), to re-weight inter-series attention; across adjacent scales, reconstruction-output adjustment and contrastive learning rectify the finer representation with the coarser one. The paper reports that this framework sets or matches state-of-the-art results on P12, P19, and PAM classification, reduces interpolation MSE on PhysioNet by a large margin (for example 0.918 versus 4.139 at 50% observed points), and stays within the top two on USHCN, MIMIC-III, and Physionet12 forecasting, with ablations showing that each component contributes.

Load-bearing premise

The load-bearing premise is that the frequency-based correlation fingerprints computed from data that is 78 to 95 percent missing are accurate enough to guide attention; if the fingerprints mostly reflect which times happened to be sampled, the re-weighting could hurt instead of help.

Editorial extensions

If this is right

  • If correct, imputation is unnecessary for representation learning on irregular multivariate series: coarse scales provide regular structure and LSP-DTW provides inter-variable structure directly from raw timestamps.
  • A single MuSiCNet backbone can serve classification, interpolation, and forecasting, so practitioners no longer need to pick a task-specific irregular-series model.
  • The coarse-to-fine refinement implies that broad-view temporal information is a usable prior for fine-grained series, and that adding coarse scales should help most on very sparse data.
  • The ablations imply that all three components (correlation matrix, reconstruction adjustment, contrastive learning) contribute, since removing all three drops P12 AUROC by about 1.9 points.
  • Because the correlation matrix is computed once per instance from raw data, the method adds no per-scale imputation overhead and can be applied at 78 to 95 percent missingness.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the coarse-to-fine prior is the main driver of the gains, the same hierarchy should transfer to tasks the paper does not test, such as anomaly detection and online event prediction, where long-term context and sparse sensors matter.
  • A natural testable extension is to make the correlation matrix learnable per scale, using LSP-DTW as an initialization; if the fixed matrix already captures the signal, a learned correction should change little.
  • For very short series such as Physionet12, which averages about 130 observations, periodogram estimates are high-variance; a stratified ablation by series length could reveal whether robustness there comes from the multi-scale prior or from LSP-DTW.
  • If the paper is right, then missingness is partly a modeling choice: by coarsening, one can always produce a regular view, which reframes the field's core difficulty from imputation to choosing the right scale at which to match signals.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The manuscript introduces MuSiCNet, a multi-scale and multi-correlation attention framework for irregularly sampled multivariate time series. The method first converts each instance into a hierarchy of relatively regular coarse-to-fine series via masked average pooling, then encodes each scale with an attention module whose inter-series attention scores are reweighted by a frequency-domain correlation matrix computed with Lomb-Scargle periodograms and dynamic time warping (LSP-DTW). Adjacent scales are tied together by a reconstruction-adjustment loss and a contrastive loss. The framework is evaluated on classification (P19, P12, PAM), interpolation (PhysioNet), and forecasting (USHCN, MIMIC-III, Physionet12), and the authors claim that it is consistently competitive with state-of-the-art across all three tasks.

Significance. The core idea, treating irregularity as relative and learning from coarse to fine through a hierarchy of regularized views of the same instance, is interesting and reasonably well motivated. The paper combines this with an original LSP-DTW correlation estimator and a broad comparison across three tasks, and the ablations isolate the main components. If the reported results can be reproduced under a clearly specified, shared evaluation protocol, the framework could serve as a useful general backbone for ISMTS analysis. However, no code or precise data-processing specification is provided, and the strongest quantitative evidence, the interpolation results, is currently not verifiable. The paper would be strengthened by a reproducible protocol and a more measured statement of the cross-task claim.

major comments (5)
  1. [Section 4.2, Table 2, Appendix C Eq. (10)] The interpolation table carries a disproportionate load in the paper's central claim of consistent state-of-the-art performance, but the reported 4-5x MSE advantage over mTAND-Full is difficult to accept without a precise protocol. The text says models reconstruct values for the entire dataset from a selected subset of observed points, while MuSiCNet's interpolation loss (Eq. 10 in Appendix C) is only the masked reconstruction at the finest scale. The paper does not specify how the observed-percentage masks are generated and applied to MuSiCNet versus the baselines, how test-time evaluation points are selected, or whether all methods use identical normalization constants. Please provide the exact protocol, release the mask-generation and preprocessing code, and, if needed, rerun the comparison under identical evaluation conditions.
  2. [Abstract; Section 4.1, Table 1] The abstract's claim of being competitive with state-of-the-art consistently across three tasks is not supported on P19. In Table 1, MuSiCNet obtains P19 AUROC 86.8 versus 89.2 for ViTST and P19 AUPRC 45.4 versus 53.1 for ViTST, a substantial gap relative to the reported standard deviations. Because the cross-task consistency claim is one of the paper's main contributions, the claim should be qualified, or additional evidence should be provided that this level of performance is considered competitive.
  3. [Section 3.2; Algorithm 1, line 6] Algorithm 1 line 6 passes X^(L) to fCorrE for every scale l, which as written would make all scales operate on the finest-scale series and contradict the coarse-to-fine construction described in Section 3.2 and Figure 2(a). Please clarify whether this is a typo and should read X^(l); if the implementation actually uses X^(l), the pseudo-code must be corrected, because the coarse-to-fine hierarchy is the central mechanism of the paper.
  4. [Section 4.4, Table 4] The ablation conclusions are stronger than the statistics support. Removing the correlation matrix changes P12 AUROC from 86.1 +/- 0.4 to 85.5 +/- 0.3 and AUPRC from 54.1 +/- 2.2 to 53.0 +/- 2.1, which is within one standard deviation; several other component removals show similarly small deltas. The text states that each component is crucial, but Table 4 does not establish that. Please report confidence intervals or significance tests, or soften the claim.
  5. [Section 4.5, Figure 3, Table 5] The claim that LSP-DTW 'accurately identifies correlations between variables' is not directly validated. Figure 3 is visual and Table 5 measures downstream classification accuracy, which is an indirect proxy. Given that the correlation matrices are computed from series with 88-95% missingness, the LSP-DTW output could be sensitive to the observation window rather than to the underlying signal. Please add a synthetic experiment with known inter-variable correlations or report a quantitative correlation-recovery metric to support the frequency-domain correlation claim.
minor comments (5)
  1. [Section 4.2, Table 6, Appendix D.2] The PhysioNet interpolation dataset is described in Section 4.2 and D.2 as containing all 8,000 instances, but Table 6 lists 4,000 samples; please correct this inconsistency.
  2. [Section 4.3, Table 6, Appendix D.3] Physionet12 is described as containing 12,000 patients with an 80.4% missing ratio, while Table 6 lists 5,333 samples and an 85.7% missing ratio; the differing numbers should be reconciled. Also, the PhysioNet challenge dataset in D.2 is cited as [Reiss and Stricker, 2012] but should be [Silva et al., 2012].
  3. [Appendix C, after Eq. (10)] The sentence stating that L_int 'essentially represents the reconstruction outcome at the finest scale as l^(L)_adj in Eq. (4)' is confusing because Eq. (4) defines the multi-scale data generation, not a loss; the reference should be to the masked reconstruction loss in Eq. (3) or a restated loss.
  4. [Table 5] The row labels '(4-6) DTW-based methods' are not actually listed in the table; the table jumps from 'Diag' to 'I-GAK, I-DTW, LSP-DTW'. Please label these rows explicitly for readability.
  5. [Appendix E.1] The hyperparameters lambda1, lambda2, and lambda3 are said to be selected from [1e-3, 1e-2, ..., 1e2], but the chosen values for each dataset are not reported; please include the selected values for reproducibility.

Circularity Check

2 steps flagged · score 6.0 of 10

Interpolation and forecasting evaluations leak the target into the LSP-DTW correlation matrix via Algorithm 1, making those 'predictions' partially self-referential; classification is not affected.

  1. other [Algorithm 1 lines 1-9; Eq. (10); Section 4.2]
    "CT ← Eq.(2) with X ... {X^(1), · · ·, X^(L)} ← Mask_r(AvgPoolingL(X)) ... r^(l) ← fCorrE(X^(L), CT, |τ^(L)|/2^(L-l)) ... ℓrecon ← ℓrecon + Eq.(3) with X^(l) and Xhat^(l) ... Lint = Σ_i ||M^(L) ⊙ ((Xhat^(L)_reco)_i − X^(L)_i)||^2_2"

    The LSP-DTW correlation matrix is computed from the full training instance X in Algorithm 1 line 1, before the interpolation mask is applied in line 3. The masked timestamps that form the reconstruction/interpolation target in Eq. (10) are therefore part of the X used to build CT. The encoder then uses CT to re-weight attention, so the interpolation target re-enters the predictor through the correlation statistic. At test time only the observed subset would be available for CT, so the training-time 'prediction' is not a function of the inputs a test-time predictor would receive; it is a function of the answer through CT.

  2. other [Algorithm 1 lines 1-9; Eq. (11); Section 4.3]
    "CT ← Eq.(2) with X ... r^(l) ← fCorrE(X^(L), CT, |τ^(L)|/2^(L-l)) ... Lfore = Σ_i ||(Mfore)_i ⊙ ((Xhat^(L)_fore)_i − (Xfore)_i)||^2_2 ... the model observes the first 36 hours in the series and predicts the next 3 time steps."

    Under the same Algorithm 1, CT is computed from the whole training series X, which in the forecasting setup includes Xfore, the next-3-step ground truth. The forecasting head predicts Xfore from representations r^(l) built using CT, so the future target values are fed back into the predictor through the LSP-DTW statistic. The forecast is therefore not a function of the observed past alone; it is partially a function of the future values it is supposed to predict. This is a structural target leak in the forecasting benchmark.

full rationale

MuSiCNet's classification experiments are self-contained: the model is trained on held-out splits, the LSP-DTW matrix is computed from the input series, and there are no self-citations or imported uniqueness theorems. The circularity appears in the two unsupervised benchmarks because Algorithm 1 computes CT = DTW(LSP(X_di), LSP(X_dj)) from the raw instance X before any masking or forecasting split; lines 3-9 then mask the same X to form reconstruction targets and feed CT into the encoder. For interpolation (Eq. 10), the masked target timestamps lie inside the LSP spectra that determine CT. For forecasting (Eq. 11), the next-time-step target Xfore lies inside the X used for CT, since the setup splits one series into observed past and future. At test time only the observed subset or past is available, so the training-time prediction is not the same function of inputs as a test-time predictor would use; it is partially a function of the answer through CT. The classification results and the ablations in Tables 4-5 remain independent evidence, but the abstract's 'three mainstream tasks consistently' claim depends on interpolation and forecasting, so the overall circularity score is 6. The Table 2 protocol underspecification and the 4-5x interpolation MSE gap are additional verifiability concerns, but they are correctness issues rather than the circular mechanism itself.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the multi-scale pooling assumption, the reliability of LSP-DTW for extremely sparse series, and the coherence prior between adjacent scales. All are modeling choices rather than independently verified facts; together they define the method's scope and failure modes.

free parameters (4)
  • Number of scales L per dataset = P12:6, P19:5, PAM:7, PhysioNet:6, USHCN:8, MIMIC-III:4, Physionet12:5
    Selected per dataset from observed timestamps; directly controls the multi-scale hierarchy that is the paper's central idea.
  • Loss weights lambda1, lambda2, lambda3 = searched in [1e-3, 1e-2, ..., 1e2] per task
    Chosen per task on validation; the reported results depend on these choices.
  • Maximum reference point count |tau| = 48/96/128 depending on dataset
    Sets the number of query time points at the finest scale; affects representation capacity.
  • Window size heuristic = 1/4 of length, halved until majority of windows have at least one observation
    Used to define scale-consistent sampling; a heuristic, not derived.
assumptions (3)
  • domain assumption Average pooling of an irregularly sampled series produces relatively regular coarse-grained series that preserve broad temporal structure
    Eq. (4) constructs X_multi by AvgPoolingL; if coarse scales alias or misrepresent the process, the rectification losses (Eqs. 5-6) propagate that bias downward.
  • domain assumption LSP-DTW gives a valid correlation matrix for variables with very high missingness
    Section 3.2 Correlation Extraction; no theoretical guarantee for series as short and sparse as P19 (94.9% missing).
  • ad hoc to paper Cross-scale representations should be similar (adjacent-scale coherence)
    The contrastive loss and reconstruction adjustment in Section 3.3 impose this prior; if adjacent scales legitimately disagree (e.g., coarse scale misses a transient), the model is forced toward a less accurate fine representation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MuSiCNet: A Gradual Coarse-to-Fine Framework for Irregularly Sampled Multivariate Time Series Analysis." pith.science (2026). https://pith.science/paper/EHE7KJRB

@misc{pith2026241201063,
  author       = {Pith},
  title        = {Pith review of: MuSiCNet: A Gradual Coarse-to-Fine Framework for Irregularly Sampled Multivariate Time Series Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EHE7KJRB}},
  note         = {Machine review of arXiv:2412.01063}
}
read the original abstract

Irregularly sampled multivariate time series (ISMTS) are prevalent in reality. Most existing methods treat ISMTS as synchronized regularly sampled time series with missing values, neglecting that the irregularities are primarily attributed to variations in sampling rates. In this paper, we introduce a novel perspective that irregularity is essentially relative in some senses. With sampling rates artificially determined from low to high, an irregularly sampled time series can be transformed into a hierarchical set of relatively regular time series from coarse to fine. We observe that additional coarse-grained relatively regular series not only mitigate the irregularly sampled challenges to some extent but also incorporate broad-view temporal information, thereby serving as a valuable asset for representation learning. Therefore, following the philosophy of learning that Seeing the big picture first, then delving into the details, we present the Multi-Scale and Multi-Correlation Attention Network (MuSiCNet) combining multiple scales to iteratively refine the ISMTS representation. Specifically, within each scale, we explore time attention and frequency correlation matrices to aggregate intra- and inter-series information, naturally enhancing the representation quality with richer and more intrinsic details. While across adjacent scales, we employ a representation rectification method containing contrastive learning and reconstruction results adjustment to further improve representation consistency. MuSiCNet is an ISMTS analysis framework that competitive with SOTA in three mainstream tasks consistently, including classification, interpolation, and forecasting.

Figures

Figures reproduced from arXiv: 2412.01063 by the authors.

Figure 1
Figure 1. Comparative visualization of multi-scale time series data [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of MuSiCNet framework, shown in (a), containing three main components for better representation learning, including [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Visualization of various methods to extract the correlation matrix from P12 dataset. The darker the color, the more similar the [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

61 extracted references · 52 canonical work pages

  1. [1]

    Modelling irregularly sampled time series without imputation

    [Agarwal et al., 2023] Rohit Agarwal, Aman Sinha, Dilip K Prasad, Marianne Clausel, Alexander Horsch, Mathieu Constant, and Xavier Coubez. Modelling irregularly sampled time series without imputation. arXiv preprint arXiv:2309.08698,

    arXiv 2023

  2. [9]

    Time-aware multi-scale rnns for time series modeling

    [Chen et al., 2021] Zipeng Chen, Qianli Ma, and Zhenxi Lin. Time-aware multi-scale rnns for time series modeling. In IJCAI, pages 2285–2291,

    work page 2021

  3. [10]

    Nonstationary tem- poral matrix factorization for multivariate time series fore- casting

    [Chen et al., 2022] Xinyu Chen, Chengyuan Zhang, Xi-Le Zhao, Nicolas Saunier, and Lijun Sun. Nonstationary tem- poral matrix factorization for multivariate time series fore- casting. arXiv preprint arXiv:2203.10651,

    arXiv 2022

  4. [12]

    Fast global alignment kernels

    [Cuturi, 2011] Marco Cuturi. Fast global alignment kernels. In Proceedings of the 28th international conference on ma- chine learning (ICML-11), pages 929–936,

    work page 2011

  5. [14]

    Saits: Self-attention-based imputation for time series

    [Du et al., 2023] Wenjie Du, David Cˆot´e, and Yan Liu. Saits: Self-attention-based imputation for time series. Expert Systems with Applications, 219:119619,

    work page 2023

  6. [15]

    Multiscale vision transformers

    [Fan et al., 2021] Haoqi Fan, Bo Xiong, Karttikeya Man- galam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6824–6835,

    work page 2021

  7. [16]

    Dynamic nonlinear matrix comple- tion for time-varying data imputation

    [Fan, 2022] Jicong Fan. Dynamic nonlinear matrix comple- tion for time-varying data imputation. In AAAI, March

    work page 2022

  8. [17]

    Physiobank, physiotoolkit, and physionet: components of a new re- search resource for complex physiologic signals

    [Goldberger et al., 2000] Ary L Goldberger, Luis AN Ama- ral, Leon Glass, Jeffrey M Hausdorff, Plamen Ch Ivanov, Roger G Mark, Joseph E Mietus, George B Moody, Chung-Kang Peng, and H Eugene Stanley. Physiobank, physiotoolkit, and physionet: components of a new re- search resource for complex physiologic signals. circu- lation, 101(23):e215–e220,

    work page 2000

Show all 61 references
  1. [20]

    Multivariate time series forecasting with dynamic graph neural odes

    [Jin et al., 2022] Ming Jin, Yu Zheng, Yuan-Fang Li, Siheng Chen, Bin Yang, and Shirui Pan. Multivariate time series forecasting with dynamic graph neural odes. IEEE Trans- actions on Knowledge and Data Engineering,

    2022

  2. [21]

    Mimic-iii, a freely accessible critical care database sci

    [Johnson et al., 2016] AE Johnson, Tom J Pollard, Lu Shen, L-w H Lehman, Mengling Feng, Mohammad Ghassemi, Benjamin Moody, Peter Szolovits, L Anthony Celi, and Roger G Mark. Mimic-iii, a freely accessible critical care database sci. Data, 3(1):1,

    2016

  3. [23]

    Neural controlled differential equations for irregular time series

    [Kidger et al., 2020] Patrick Kidger, James Morrill, James Foster, and Terry Lyons. Neural controlled differential equations for irregular time series. NeurIPS, 33:6696– 6707,

    2020

  4. [24]

    Time series as images: Vision transformer for irregularly sampled time series

    [Li et al., 2023] Zekun Li, Shiyang Li, and Xifeng Yan. Time series as images: Vision transformer for irregularly sampled time series. InThirty-seventh Conference on Neu- ral Information Processing Systems,

    2023

  5. [25]

    Pyraformer: Low-complexity pyramidal attention for long-range time series modeling and forecasting

    [Liu et al., 2021] Shizhan Liu, Hang Yu, Cong Liao, Jian- guo Li, Weiyao Lin, Alex X Liu, and Schahram Dust- dar. Pyraformer: Low-complexity pyramidal attention for long-range time series modeling and forecasting. In ICLR,

    2021

  6. [26]

    Least-squares frequency analysis of unequally spaced data

    [Lomb, 1976] Nicholas R Lomb. Least-squares frequency analysis of unequally spaced data. Astrophysics and space science, 39:447–462,

    1976

  7. [30]

    Phased lstm: Accelerating recurrent network training for long or event-based sequences

    [Neil et al., 2016] Daniel Neil, Michael Pfeiffer, and Shih- Chii Liu. Phased lstm: Accelerating recurrent network training for long or event-based sequences. NeurIPS, 29,

    2016

  8. [31]

    Introducing a new benchmarked dataset for activity moni- toring

    [Reiss and Stricker, 2012] Attila Reiss and Didier Stricker. Introducing a new benchmarked dataset for activity moni- toring. In 2012 16th international symposium on wearable computers, pages 108–109. IEEE,

    2012

  9. [33]

    Latent ordinary differential equa- tions for irregularly-sampled time series

    [Rubanova et al., 2019] Yulia Rubanova, Ricky TQ Chen, and David K Duvenaud. Latent ordinary differential equa- tions for irregularly-sampled time series. NeurIPS, 32,

    2019

  10. [34]

    Studies in astronomical time series analysis

    [Scargle, 1982] Jeffrey D Scargle. Studies in astronomical time series analysis. ii-statistical aspects of spectral analy- sis of unevenly spaced data. Astrophysical Journal, Part 1, vol. 263, Dec. 15, 1982, p. 835-853., 263:835–853,

    1982

  11. [36]

    Scaleformer: Itera- tive multi-scale refining transformers for time series fore- casting

    [Shabani et al., 2022] Mohammad Amin Shabani, Amir H Abdi, Lili Meng, and Tristan Sylvain. Scaleformer: Itera- tive multi-scale refining transformers for time series fore- casting. In The Eleventh ICLR,

    2022

  12. [37]

    Interpolation-prediction networks for irreg- ularly sampled time series

    [Shukla and Marlin, 2018] Satya Narayan Shukla and Ben- jamin Marlin. Interpolation-prediction networks for irreg- ularly sampled time series. In ICLR,

    2018

  13. [38]

    A survey on principles, models and methods for learning from irregularly sampled time series

    [Shukla and Marlin, 2020] Satya Narayan Shukla and Ben- jamin M Marlin. A survey on principles, models and methods for learning from irregularly sampled time series. arXiv preprint arXiv:2012.00168,

    2020 arXiv

  14. [39]

    Multi-time attention networks for irregu- larly sampled time series

    [Shukla and Marlin, 2021] Satya Narayan Shukla and Ben- jamin Marlin. Multi-time attention networks for irregu- larly sampled time series. In ICLR,

    2021

  15. [40]

    Predict- ing in-hospital mortality of icu patients: The phys- ionet/computing in cardiology challenge

    [Silva et al., 2012] Ikaro Silva, George Moody, Daniel J Scott, Leo A Celi, and Roger G Mark. Predict- ing in-hospital mortality of icu patients: The phys- ionet/computing in cardiology challenge

    2012

  16. [41]

    In 2012 Computing in Cardiology, pages 245–248. IEEE,

    2012

  17. [42]

    Te-esn: Time encoding echo state network for predic- tion based on irregularly sampled time series data

    [Sun et al., 2021] Chenxi Sun, Shenda Hong, Moxian Song, Yen-Hsiu Chou, Yongyue Sun, Derun Cai, and Hongyan Li. Te-esn: Time encoding echo state network for predic- tion based on irregularly sampled time series data. In Zhi- Hua Zhou, editor, IJCAI, pages 3010–3016. Internatio...

    2021

  18. [43]

    Time pat- tern reconstruction for classification of irregularly sampled time series

    [Sun et al., 2024] Chenxi Sun, Hongyan Li, Moxian Song, Derun Cai, Baofeng Zhang, and Shenda Hong. Time pat- tern reconstruction for classification of irregularly sampled time series. Pattern Recognition, 147:110075,

    2024

  19. [44]

    Csdi: Conditional score-based diffusion models for probabilistic time series imputation

    [Tashiro et al., 2021] Yusuke Tashiro, Jiaming Song, Yang Song, and Stefano Ermon. Csdi: Conditional score-based diffusion models for probabilistic time series imputation. NeurIPS, 34,

    2021

  20. [45]

    Understanding the lomb–scargle periodogram

    [VanderPlas, 2018] Jacob T VanderPlas. Understanding the lomb–scargle periodogram. The Astrophysical Journal Supplement Series, 236(1):16,

    2018

  21. [46]

    Attention is all you need

    [Vaswani et al., 2017] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30,

    2017

  22. [47]

    Optimal transport: old and new, volume

    [Villani and others, 2009] C´edric Villani et al. Optimal transport: old and new, volume

    2009

  23. [49]

    Con- necting the dots: Multivariate time series forecasting with graph neural networks

    [Wu et al., 2020] Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, Xiaojun Chang, and Chengqi Zhang. Con- necting the dots: Multivariate time series forecasting with graph neural networks. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discov- ery...

    2020

  24. [50]

    Dy- namic gaussian mixture based deep generative model for robust forecasting on sparse multivariate time series

    [Wu et al., 2021] Yinjun Wu, Jingchao Ni, Wei Cheng, Bo Zong, Dongjin Song, Zhengzhang Chen, Yanchi Liu, Xuchao Zhang, Haifeng Chen, and Susan B Davidson. Dy- namic gaussian mixture based deep generative model for robust forecasting on sparse multivariate time series. In AAAI,...

    2021

  25. [51]

    Grafiti: Graphs for forecasting irregu- larly sampled time series

    [Yalavarthiet al., 2024] Vijaya Krishna Yalavarthi, Kiran Madhusudhanan, Randolf Scholz, Nourhan Ahmed, Jo- hannes Burchert, Shayan Jawed, Stefan Born, and Lars Schmidt-Thieme. Grafiti: Graphs for forecasting irregu- larly sampled time series. InProceedings of the AAAI Con- fe...

    2024

  26. [52]

    Gain: Missing data imputation using gen- erative adversarial nets

    [Yoon et al., 2018] Jinsung Yoon, James Jordon, and Mi- haela Schaar. Gain: Missing data imputation using gen- erative adversarial nets. In International conference on machine learning, pages 5689–5698. PMLR,

    2018

  27. [53]

    Imputation with inter-series infor- mation from prototypes for irregular sampled time series

    [Yu et al., 2024] Zhihao Yu, Xu Chu, Liantao Ma, Yasha Wang, and Wenwu Zhu. Imputation with inter-series infor- mation from prototypes for irregular sampled time series. arXiv preprint arXiv:2401.07249,

    2024 arXiv

  28. [54]

    Are transformers effective for time series fore- casting? In Proceedings of the AAAI conference on artifi- cial intelligence, volume 37, pages 11121–11128,

    [Zeng et al., 2023] Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. Are transformers effective for time series fore- casting? In Proceedings of the AAAI conference on artifi- cial intelligence, volume 37, pages 11121–11128,

    2023

  29. [55]

    Life: Learning individ- ual features for multivariate time series prediction with missing values

    [Zhang et al., 2021c] Zhao-Yu Zhang, Shao-Qun Zhang, Yuan Jiang, and Zhi-Hua Zhou. Life: Learning individ- ual features for multivariate time series prediction with missing values. In 2021 IEEE International Conference on Data Mining (ICDM), pages 1511–1516. IEEE,

    2021

  30. [56]

    Multi-scale group transformer for long sequence modeling in speech separation

    [Zhao et al., 2021] Yucheng Zhao, Chong Luo, Zheng-Jun Zha, and Wenjun Zeng. Multi-scale group transformer for long sequence modeling in speech separation. In IJCAI, pages 3251–3257,

    2021

  31. [57]

    Informer: Beyond efficient transformer for long sequence time-series forecasting

    [Zhou et al., 2021] Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wan- cai Zhang. Informer: Beyond efficient transformer for long sequence time-series forecasting. In Proceedings of the AAAI conference on artificial intelligence, volume 35, p...

    2021

  32. [58]

    Fedformer: Fre- quency enhanced decomposed transformer for long-term series forecasting

    [Zhou et al., 2022] Tian Zhou, Ziqing Ma, Qingsong Wen, Xue Wang, Liang Sun, and Rong Jin. Fedformer: Fre- quency enhanced decomposed transformer for long-term series forecasting. In International conference on machine learning, pages 27268–27286. PMLR,

    2022

  33. [59]

    The interpolation task can be obtained by removing the projection head fcls and the classification loss term Lcls from the total loss in line #17

    A Pseudo Code for MuSiCNet The Pseudo Code is provided using classification as an exam- ple. The interpolation task can be obtained by removing the projection head fcls and the classification loss term Lcls from the total loss in line #17. While in the case of forecasting task...

    2019

  34. [60]

    #Avg. obs

    P19 dataset [Reyna et al., 2020] comprises data from 38, 803 patients, each monitored by 34 irregularly sampled sensors, including 8 vital signs and 26 laboratory values. The origi- nal dataset contained 40, 336 patients, but we excluded those Table 6: Statistics of the ISMTS ...

    2020

  35. [61]

    This dataset features multivariate time series from 36 sensors collected during the first 48 hours of ICU stay

    P12 [Goldberger et al., 2000] includes data from 11, 988 pa- tients after removing inappropriate 12 samples as explained in [Horn et al., 2020]. This dataset features multivariate time series from 36 sensors collected during the first 48 hours of ICU stay. Each patient has a b...

    2000

  36. [1976]

    A theory for multireso- lution signal decomposition: the wavelet representation

    [Mallat, 1989] Stephane G Mallat. A theory for multireso- lution signal decomposition: the wavelet representation. IEEE transactions on pattern analysis and machine intel- ligence, 11(7):674–693,

    1989

  37. [1982]

    Modeling irregular time series with continuous recurrent units

    [Schirmer et al., 2022] Mona Schirmer, Mazin Eltayeb, Ste- fan Lessmann, and Maja Rudolph. Modeling irregular time series with continuous recurrent units. In Interna- tional Conference on Machine Learning , pages 19388– 19405. PMLR,

    2022

  38. [1989]

    [Menne et al., 2015] Matthew J Menne, CN Williams Jr, and Russell S V ose. United states historical climatology net- work daily temperature, precipitation, and snow data.Car- bon Dioxide Information Analysis Center, Oak Ridge Na- tional Laboratory, Oak Ridge, Tennessee,

    2015

  39. [1994]

    Neural flows: Efficient alternative to neural odes

    [Biloˇs et al., 2021] Marin Bilo ˇs, Johanna Sommer, Syama Sundar Rangapuram, Tim Januschowski, and Stephan G ¨unnemann. Neural flows: Efficient alternative to neural odes. Advances in neural information processing systems, 34:21325–21337,

    2021

  40. [2000]

    Set func- tions for time series

    [Horn et al., 2020] Max Horn, Michael Moor, Christian Bock, Bastian Rieck, and Karsten Borgwardt. Set func- tions for time series. In International Conference on Ma- chine Learning, pages 4353–4363. PMLR,

    2020

  41. [2009]

    Deep learning for multivariate time series imputation: A survey

    [Wang et al., 2024] Jun Wang, Wenjie Du, Wei Cao, Keli Zhang, Wenjia Wang, Yuxuan Liang, and Qingsong Wen. Deep learning for multivariate time series imputation: A survey. arXiv preprint arXiv:2402.04059,

    2024 arXiv

  42. [2011]

    Gru-ode-bayes: Con- tinuous modeling of sporadically-observed time series

    [De Brouwer et al., 2019] Edward De Brouwer, Jaak Simm, Adam Arany, and Yves Moreau. Gru-ode-bayes: Con- tinuous modeling of sporadically-observed time series. NeurIPS, 32,

    2019

  43. [2012]

    Early prediction of sepsis from clinical data: the physionet/computing in cardiology challenge

    [Reyna et al., 2020] Matthew A Reyna, Christopher S Josef, Russell Jeter, Supreeth P Shashikumar, M Brandon West- over, Shamim Nemati, Gari D Clifford, and Ashish Sharma. Early prediction of sepsis from clinical data: the physionet/computing in cardiology challenge

    2020

  44. [2015]

    Hierarchical trans- formers are more efficient language models.arXiv preprint arXiv:2110.13711,

    [Nawrot et al., 2021] Piotr Nawrot, Szymon Tworkowski, Michał Tyrolski, Łukasz Kaiser, Yuhuai Wu, Christian Szegedy, and Henryk Michalewski. Hierarchical trans- formers are more efficient language models.arXiv preprint arXiv:2110.13711,

    2021 arXiv

  45. [2016]

    Time2vec: Learning a vector represen- tation of time

    [Kazemi et al., 2019] Seyed Mehran Kazemi, Rishab Goel, Sepehr Eghbali, Janahan Ramanan, Jaspreet Sahota, San- jay Thakur, Stella Wu, Cathal Smyth, Pascal Poupart, and Marcus Brubaker. Time2vec: Learning a vector represen- tation of time. arXiv preprint arXiv:1907.05321,

    2019 arXiv

  46. [2017]

    Using dynamic time warping to find patterns in time series

    [Berndt and Clifford, 1994] Donald J Berndt and James Clif- ford. Using dynamic time warping to find patterns in time series. In Proceedings of the 3rd international conference on knowledge discovery and data mining, pages 359–370,

    1994

  47. [2018]

    Neural ordinary dif- ferential equations

    [Chen et al., 2018] Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary dif- ferential equations. NeurIPS, 31,

    2018

  48. [2019]

    Recurrent neural networks for multivariate time series with missing values

    [Che et al., 2018] Zhengping Che, Sanjay Purushotham, Kyunghyun Cho, David Sontag, and Yan Liu. Recurrent neural networks for multivariate time series with missing values. Scientific reports, 8(1):1–12,

    2018

  49. [2020]

    Exit: Extrapolation and interpolation-based neural controlled differential equa- tions for time-series classification and forecasting

    [Jhin et al., 2022] Sheo Yon Jhin, Jaehoon Lee, Minju Jo, Seungji Kook, Jinsung Jeon, Jihyeon Hyeong, Jayoung Kim, and Noseong Park. Exit: Extrapolation and interpolation-based neural controlled differential equa- tions for time-series classification and forecasting. In Pro- c...

    2022

  50. [2021]

    Msgnet: Learning multi- scale inter-series correlations for multivariate time series forecasting

    [Cai et al., 2024] Wanlin Cai, Yuxuan Liang, Xianggen Liu, Jianshuai Feng, and Yuankai Wu. Msgnet: Learning multi- scale inter-series correlations for multivariate time series forecasting. In Proceedings of the AAAI Conference on Ar- tificial Intelligence, volume 38, pages 111...

    2024

  51. [2022]

    Primenet: Pre-training for irregular multi- variate time series

    [Chowdhury et al., 2023] Ranak Roy Chowdhury, Jiacheng Li, Xiyuan Zhang, Dezhi Hong, Rajesh K Gupta, and Jingbo Shang. Primenet: Pre-training for irregular multi- variate time series. InProceedings of the AAAI Conference on Artificial Intelligence , volume 37, pages 7184–7192,

    2023

  52. [2023]

    Patient subtyp- ing via time-aware lstm networks

    [Baytas et al., 2017] Inci M Baytas, Cao Xiao, Xi Zhang, Fei Wang, Anil K Jain, and Jiayu Zhou. Patient subtyp- ing via time-aware lstm networks. In Proceedings of the 23rd ACM SIGKDD international conference on knowl- edge discovery and data mining, pages 65–74,

    2017

  53. [2024]

    Improving missing data imputation with deep generative models

    [Camino et al., 2019] Ramiro D Camino, Christian A Ham- merschmidt, and Radu State. Improving missing data imputation with deep generative models. arXiv preprint arXiv:1902.10666,

    2019 arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.