Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Cross-Domain Conditional Diffusion Models for Time Series Imputation

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Diffusion model borrows related-domain patterns to impute gaps.

desk verdict A genuinely new task formulation with a plausible method, but the DA baseline adaptation is under-specified, the CDCA loss has a definition bug, and there are no error bars; worth refereeing after fixes. read the letter →

arxiv 2506.12412 v1 pith:UEG53U4M submitted 2025-06-14 cs.LG stat.ML

Kexin Zhang , Baoyu Jing , K. Selçuk Candan , Dawei Zhou , Qingsong Wen , Han Liu , Kaize Ding This is my paper
classification cs.LGstat.ML
keywords cross-domaintimeseriesimputationconditionaldiffusionmodelsfrequencymixupdomainadaptationmissingdataFMixupconsistencyalignment
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

The paper targets cross-domain time series imputation: filling missing values in a target domain by borrowing knowledge from a related source domain, a setting where single-domain imputation fails under high missing rates and standard domain adaptation fails because missing data makes distribution alignment unreliable. The proposed CD2-TSI combines three mechanisms: a frequency-domain interpolation (FMixup) that mixes low-frequency amplitudes across domains while keeping the target's high-frequency and phase structure; a conditional diffusion model with shared representations but domain-specific denoising networks; and a bounded consistency alignment loss that applies adaptation only when cross-domain prediction disagreement is moderate. The paper reports that CD2-TSI outperforms both families of baselines on three real-world datasets, with RMSE improvements of 4.41%, 4.37%, and 4.13% over the best imputation baselines on Air Quality, Hydrology, and Electricity, respectively, and an average 1.92% RMSE improvement over the best domain adaptation baselines. If these results hold, cross-domain transfer becomes a viable remedy for severely incomplete target series.

What carries the argument

The load-bearing object is FMixup, a Fourier-domain interpolation with a binary low-frequency mask $M$. On each domain's conditional observations, it takes the FFT, blends the low-frequency amplitude spectra as $A_{\text{Src}\to\text{Tgt}} = A_{\text{Tgt}}*(1-M) + (\lambda A_{\text{Tgt}} + (1-\lambda) A_{\text{Src}})*M$, keeps the target's phase spectrum $P_{\text{Tgt}}$ and high-frequency amplitudes, and applies the inverse FFT to fill original missing values. This provides the missing targets with a prior that carries shared low-frequency patterns without flattening the target's local structure. The second mechanism is CDCA, a thresholded alignment loss that is zero below $\tau_l$, linear between $\tau_l$ and $\tau_h$, and capped above $\tau_h$, applied to the difference between source and target denoising networks' noise predictions on target samples.

What would settle it

Take a source-target pair with no shared periodic or trend structure and compare CD2-TSI with a target-only diffusion imputer: the paper's stated mechanism predicts the cross-domain gain should vanish or become negative, since FMixup has nothing shared to blend and CDCA would push the target toward irrelevant source structure. If the cross-domain version still wins by the reported margins, the claimed mechanism is not what produces the gains.

Watch

Extended reading notes

Core claim

On its own terms, the central discovery is that cross-domain imputation can be made to work by refusing to align domains uniformly. The paper claims that low-frequency spectral amplitudes encode shared temporal structure (trends, periodicities), while high-frequency amplitudes and phase encode domain-specific detail. FMixup uses this split to construct a better prior for missing values than zero-filling or linear interpolation. The diffusion model then separates shared feature extraction from domain-specific attention and denoising, and the consistency loss CDCA regularizes only output discrepancies in a middle band, avoiding both noise amplification from over-alignment and source overfitting from rigid alignment. The reported outcome is consistent gains in RMSE and MAE across the Air Quality, Hydrology, and Electricity settings.

Load-bearing premise

The method assumes related domains share their low-frequency spectral content while differing in high-frequency detail and phase; if that split fails for a given source-target pair, FMixup can contaminate the target prior with source-specific artifacts.

Editorial extensions

If this is right

  • Over the strongest domain adaptation baselines, CD2-TSI reports average gains of 1.92% in RMSE and 1.34% in MAE across all settings.
  • Over the best single-domain imputation baselines, it reports RMSE gains of +4.41%, +4.37%, and +4.13% on Air Quality, Hydrology, and Electricity, respectively.
  • Ablation shows that removing FMixup degrades performance more than replacing it with linear interpolation, and removing CDCA increases error, especially under block missing patterns.
  • Sensitivity experiments indicate the method remains effective as the target missing rate rises from 10% to 50% on Air Quality, and under mismatched training and testing masking patterns.

Reading between the lines

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

  • A testable extension is whether FMixup helps non-diffusion imputers; since it only changes the priors that fill missing targets, its benefit can be isolated from the diffusion denoiser by plugging it into a VAE- or GAN-based imputation pipeline.
  • The spectral-separation premise implies the method's gain should shrink as source and target domains share less low-frequency structure; a synthetic study with controlled spectral overlap could map where cross-domain imputation stops paying off.
  • CDCA's thresholds are absolute errors on noise predictions and may need rescaling per dataset; normalizing them relative to target noise level would make the alignment rule more portable.
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

3 major / 5 minor

Summary. The paper proposes CD2-TSI, a cross-domain conditional diffusion model for multivariate time series imputation. The method has three components: (i) FMixup, a frequency-domain interpolation that blends low-frequency amplitude spectra between source and target domains while preserving the target's high-frequency amplitude and phase, used to construct priors for originally missing values; (ii) a diffusion imputation model with a shared convolutional representation layer and separate source/target denoising networks to capture both shared and domain-specific temporal dependencies; and (iii) a cross-domain consistency alignment (CDCA) loss that selectively penalizes output-level discrepancies between the two denoising networks. Experiments on Air Quality, Hydrology, and Electricity datasets compare against domain-adaptation baselines (CORAL, CDAN, DIRT-T, AdvSKM, CotMix) and time-series-imputation baselines (M-RNN, BRITS, GRIN, SPIN, SPIN-H, CSDI, SSSD, PriSTI, MTSCI), reporting RMSE and MAE improvements. The authors state that CD2-TSI consistently outperforms all baselines, with average gains of 1.92% RMSE and 1.34% MAE over the strongest DA baselines and larger gains over TSI baselines; ablations indicate that FMixup, CDCA, and the model design each contribute. Code is publicly available.

Significance. If the empirical claims are substantiated, the paper addresses a genuinely underexplored problem—cross-domain time series imputation under high missing rates and domain shift—and proposes a coherent combination of frequency-domain priors, diffusion-based generation, and selective domain alignment. The three components are clearly motivated and the ablation study attempts to isolate their contributions. The availability of code is a concrete strength that aids reproducibility. However, the current evidence has important gaps: the adaptation protocol for DA baselines is not described, the CDCA loss definition is internally inconsistent, and no variance estimates or significance tests are provided for the headline performance claims. These issues are load-bearing because the central claim is empirical superiority over baselines, and some reported margins are very small. With revisions addressing these points, the contribution could be a useful step for data-centric cross-domain imputation.

major comments (3)
  1. [Section 5.1, Table 2] The adaptation of the DA baselines to imputation is not specified. CORAL, CDAN, DIRT-T, AdvSKM, and CotMix are designed for classification or forecasting with fully observed data; they do not natively produce missing-value reconstructions. The paper only lists them as "Baselines for DA" without describing how each was modified to perform imputation (e.g., an imputation head, integration with a generative model, or application of their alignment losses to a reconstruction objective). This is not a presentation detail: Section 5.2's central claim of average 1.92% RMSE and 1.34% MAE improvements over the strongest DA baselines is unverifiable and not reproducible without this information. Several margins are small (e.g., Electricity point RMSE 0.635 vs. 0.643), so the comparison could be driven by weak baseline adaptation rather than the proposed components. Please specify the exact adaptation protocol for each DA baseline, or omit the DA comparison and reframe the claim accordingly.
  2. [Section 4.5, Eq. (11)] The text and the equation for the CDCA loss are inconsistent. The text states that for τ_l ≤ Δ ≤ τ_h the penalty is Δ−τ_l, and for Δ > τ_h the penalty is capped at min(Δ−τ_l, τ_h). However, Eq. (11) defines L_align = min(Δ−τ_l, τ_h) for all Δ ≥ τ_l. In the range τ_h < Δ < τ_l + τ_h, this expression equals Δ−τ_l, so the penalty continues to increase beyond the supposed upper threshold and is not actually capped at τ_h; τ_h functions as a cap on the excess Δ−τ_l rather than as an upper threshold on Δ. Please reconcile the piecewise description with the implemented loss, and clarify whether the reported hyperparameter study of τ_h refers to a threshold or a cap value.
  3. [Tables 2 and 3] No error bars, confidence intervals, or significance tests are reported for any result. Several of the claimed improvements are small in absolute terms (e.g., Electricity point MAE 0.378 vs. 0.380, RMSE 0.635 vs. 0.643), and the ablation differences in Table 3 are also modest (e.g., Hydrology point RMSE 46.852 vs. 47.288 without CDCA). Without multiple seeds or a statistical test, the claim that CD2-TSI "consistently outperforms" baselines is not supported. Please report means and standard deviations over at least three runs, and where possible a paired significance test (e.g., Wilcoxon signed-rank across datasets or bootstrapped confidence intervals).
minor comments (5)
  1. [Section 5.1, Tables 2-4] The evaluation section lists CRPS as a metric, but no CRPS results appear in any table or figure. Either provide CRPS values or remove the metric from the experimental setup.
  2. [Section 5.4, Table 1] The hyperparameter study reports sensitivity curves for λ, τ_l, and τ_h but does not state the final selected values for these parameters or for μ_align, α, and T. Without these values, the experimental setup is not fully reproducible; please include a hyperparameter configuration table.
  3. [Section 5.2] The reported average improvements of 1.92% (RMSE) and 1.34% (MAE) over the strongest DA baselines do not recompute straightforwardly from Table 2; depending on whether one averages over the five point/block columns or over the three datasets, the numbers are closer to 1.88%/1.73% or 1.86%/1.80%. Please clarify the aggregation rule used for the averages.
  4. [Section 4.3] The FMixup design rests on the assumption that related domains share low-frequency amplitude spectra while high-frequency amplitude and phase are domain-specific. This assumption is stated as intuition and is not validated (e.g., by spectral coherence plots or analysis of the learned priors). A brief empirical check or a discussion of when the assumption could fail would strengthen the method's generalizability claim.
  5. [Throughout] Please fix minor presentation issues: Table 2 uses "Coral" while Section 5.1 uses "CORAL"; Eq. (3) is typeset ambiguously for σ_t^2; Figure 1's caption contains unexplained abbreviations "FF w/ BP" and "FF w/o BP"; and the sentence in Section 5.2 claiming gains over "training solely on the target domain" is not directly supported by any target-only baseline in Table 2.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the diffusion equations are standard DDPM, FMixup is a fixed input transformation, and CDCA is a heuristic regularizer; reported gains come from external benchmark comparisons.

full rationale

No significant circularity found. The forward/reverse process (Eqs. 1-3) and the noise-prediction loss (Eq. 4) are standard DDPM equations imported from Ho et al., not fitted to the target result. FMixup (Eqs. 5-7) is a fixed, parameterized spectral interpolation that combines source and target amplitude spectra while retaining the target phase; it acts as an input transformation that fills original missing values, whereas the artificial-missing targets used for training and evaluation are the originally observed values, so the imputation of those targets is not equal to the FMixup output by construction. The CDCA loss (Eqs. 10-11) is a heuristic regularizer with thresholds tuned on validation data, and the overall loss (Eq. 12) simply combines the domain-specific denoising losses with this regularizer; none of these equations re-imports the reported RMSE/MAE or a fitted parameter as a 'prediction.' The only self-citations ([41], [42]) appear in the introduction/related work as contextual survey references and do not carry the central derivation. The paper does leave the DA baseline adaptations unspecified in Section 5.1, which is a reproducibility concern, but that is not circular reasoning.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The central claim rests on domain assumptions about spectral composition and alignment proxies, plus several undisclosed or partially disclosed hyperparameters. No new physical or conceptual entities are introduced.

free parameters (6)
  • α (FMixup low-frequency ratio) = 0.003
    Proportion of low-frequency amplitude spectrum blended across domains; set empirically (Section 5.1).
  • λ (frequency mix ratio) = sampled in [0.0, 1.0]
    Controls balance between source and target amplitude spectra in Eq. (6); sampled during training or tuned in hyperparameter study (Sections 4.3, 5.4).
  • τ_l (CDCA lower threshold) = not disclosed
    Threshold below which alignment is skipped; tuned per dataset in hyperparameter study (Sections 4.5, 5.4).
  • τ_h (CDCA upper threshold) = not disclosed
    Threshold beyond which alignment penalty is capped; tuned per dataset (Sections 4.5, 5.4).
  • μ_align (alignment loss weight) = not disclosed
    Multiplier for L_align in Eq. (12); noted as 're-weighted by parameters μ_align' but no value is given (Section 4.6).
  • Diffusion steps T = not specified
    Number of noising steps in DDPM forward process; T is referenced but its value is not reported (Section 4.2).
assumptions (3)
  • domain assumption Related domains share low-frequency spectral components, while high-frequency components are domain-specific.
    Motivates FMixup in Section 4.3; if violated, cross-domain frequency blending harms target imputation.
  • domain assumption Source and target time series have equal length L and feature count K.
    The model requires X_d ∈ R^{K×L} for both domains; stated in Problem Definition (Section 3).
  • domain assumption Output discrepancy Δ between source and target denoising networks on target samples reflects domain shift magnitude.
    The CDCA loss (Section 4.5) gates alignment on Δ; there is no empirical validation that this proxy is reliable under high missingness.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Cross-Domain Conditional Diffusion Models for Time Series Imputation." pith.science (2026). https://pith.science/paper/UEG53U4M

@misc{pith2026250612412,
  author       = {Pith},
  title        = {Pith review of: Cross-Domain Conditional Diffusion Models for Time Series Imputation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UEG53U4M}},
  note         = {Machine review of arXiv:2506.12412}
}
read the original abstract

Cross-domain time series imputation is an underexplored data-centric research task that presents significant challenges, particularly when the target domain suffers from high missing rates and domain shifts in temporal dynamics. Existing time series imputation approaches primarily focus on the single-domain setting, which cannot effectively adapt to a new domain with domain shifts. Meanwhile, conventional domain adaptation techniques struggle with data incompleteness, as they typically assume the data from both source and target domains are fully observed to enable adaptation. For the problem of cross-domain time series imputation, missing values introduce high uncertainty that hinders distribution alignment, making existing adaptation strategies ineffective. Specifically, our proposed solution tackles this problem from three perspectives: (i) Data: We introduce a frequency-based time series interpolation strategy that integrates shared spectral components from both domains while retaining domain-specific temporal structures, constructing informative priors for imputation. (ii) Model: We design a diffusion-based imputation model that effectively learns domain-shared representations and captures domain-specific temporal dependencies with dedicated denoising networks. (iii) Algorithm: We further propose a cross-domain consistency alignment strategy that selectively regularizes output-level domain discrepancies, enabling effective knowledge transfer while preserving domain-specific characteristics. Extensive experiments on three real-world datasets demonstrate the superiority of our proposed approach. Our code implementation is available here.

Figures

Figures reproduced from arXiv: 2506.12412 by the authors.

Figure 1
Figure 1. Architecture of CD2 -TSI. FMixup is utilized to interpolate the original missing areas (blue), while artificial missing values (red) are retained to construct the missing targets Xe 0 . These targets are then transformed into the noisy targets Xe t to train the denoising network, with the help of conditional information Xco. The framework is optimized using a combination of denoising loss and consistency alignment l… view at source ↗
Figure 2
Figure 2. Sensitivity Analysis for Air Quality, Hydrology, Electricity Datasets [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Hyperparameter study on three key parameters of CD2 -TSI [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Efficiency analysis on Air Quality, Hydrology and Electricity datasets. Efficiency Study. We illustrate the total training time of DA models trained on all three datasets, and the experiments are conducted on an NVIDIA RTX [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Trustworthy AI in Digital Health: A Comprehensive Review of Robustness and Explainability

    cs.AI 2026-08 conditional novelty 3.0 of 10

    A survey of robustness and explainability methods for digital health AI, proposing a taxonomy and illustrating known XAI tools, without new empirical or theoretical results.

Reference graph

Works this paper leans on

45 extracted references · 34 canonical work pages · cited by 1 Pith paper

  1. [1]

    arXiv preprint arXiv:2208.09399 (2022)

    Alcaraz, J.M.L., Strodthoff, N.: Diffusion-based time series imputation and fore- casting with structured state space models. arXiv preprint arXiv:2208.09399 (2022)

    arXiv 2022

  2. [2]

    Advances in neural information processing systems31 (2018)

    Cao, W., Wang, D., Li, J., Zhou, H., Li, L., Li, Y.: Brits: Bidirectional recurrent imputation for time series. Advances in neural information processing systems31 (2018)

    work page 2018

  3. [3]

    Scientific reports (2018)

    Che, Z., Purushotham, S., Cho, K., et al.: Recurrent neural networks for multi- variate time series with missing values. Scientific reports (2018)

    work page 2018

  4. [4]

    arXiv preprint arXiv:2108.00298 (2021)

    Cini, A., Marisca, I., Alippi, C.: Filling the g_ap_s: Multivariate time series im- putation by graph neural networks. arXiv preprint arXiv:2108.00298 (2021)

    arXiv 2021

  5. [5]

    IEEE Transactions on Neural Networks and Learning Systems (2021)

    Ding, K., Shu, K., Shan, X., Li, J., Liu, H.: Cross-domain graph anomaly detection. IEEE Transactions on Neural Networks and Learning Systems (2021)

    work page 2021

  6. [6]

    Expert Systems with Applications (2023)

    Du, W., Côté, D., Liu, Y.: Saits: Self-attention-based imputation for time series. Expert Systems with Applications (2023)

    work page 2023

  7. [7]

    arXiv preprint arXiv:2406.12747 (2024)

    Du,W.,Wang,J.,Qian,L.,etal.:Tsi-bench:Benchmarkingtimeseriesimputation. arXiv preprint arXiv:2406.12747 (2024)

    arXiv 2024

  8. [8]

    IEEE Transactions on Artificial Intelligence5(3), 1185–1194 (2023) 16 K

    Eldele, E., Ragab, M., Chen, Z., Wu, M., Kwoh, C.K., Li, X.: Contrastive domain adaptation for time-series via temporal mixup. IEEE Transactions on Artificial Intelligence5(3), 1185–1194 (2023) 16 K. Zhang et al

    work page 2023

Show all 45 references
  1. [9]

    IEEE Transactions on Artificial Intelli- gence (2024)

    Eldele, E., Ragab, M., Chen, Z., Wu, M., Kwoh, C.K., Li, X.: Label-efficient time series representation learning: A review. IEEE Transactions on Artificial Intelli- gence (2024)

    2024

  2. [10]

    In: International conference on artificial intelligence and statistics

    Fortuin, V., Baranchuk, D., Rätsch, G., Mandt, S.: Gp-vae: Deep probabilistic time series imputation. In: International conference on artificial intelligence and statistics. pp. 1651–1661. PMLR (2020)

    2020

  3. [11]

    arXiv preprint arXiv:2410.13338 (2024)

    Gao, H., Shen, W., Qiu, X., et al.: Diffimp: Efficient diffusion model for proba- bilistic time series imputation with bidirectional mamba backbone. arXiv preprint arXiv:2410.13338 (2024)

    2024 arXiv

  4. [12]

    In: International confer- ence on machine learning

    He, H., Queen, O., Koker, T., Cuevas, C., Tsiligkaridis, T., Zitnik, M.: Domain adaptation for time series under feature and label shifts. In: International confer- ence on machine learning. pp. 12746–12774. PMLR (2023)

    2023

  5. [13]

    Advances in neural information processing systems33, 6840–6851 (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems33, 6840–6851 (2020)

    2020

  6. [14]

    In: International Conference on Machine Learning

    Jin, X., Park, Y., Maddix, D., Wang, H., Wang, Y.: Domain adaptation for time series forecasting via attention sharing. In: International Conference on Machine Learning. pp. 10280–10297. PMLR (2022)

    2022

  7. [15]

    In: International Conference on Machine Learning

    Kim, S., Kim, H., Yun, E., Lee, H., Lee, J., Lee, J.: Probabilistic imputation for time-series classification with missing data. In: International Conference on Machine Learning. pp. 16654–16667. PMLR (2023)

    2023

  8. [16]

    arXiv preprint arXiv:2503.01875 (2025)

    Kong, Y., Yang, Y., Hwang, Y., Du, W., Zohren, S., Wang, Z., Jin, M., Wen, Q.: Time-mqa: Time series multi-task question answering with context enhancement. arXiv preprint arXiv:2503.01875 (2025)

    2025 arXiv

  9. [17]

    Konrad, C.P., Anderson, S.W., Restivo, D.E., et al.: Network analysis of usgs streamflow gages (2022)

    2022

  10. [18]

    John Wiley & Sons (2019)

    Little, R.J., Rubin, D.B.: Statistical analysis with missing data. John Wiley & Sons (2019)

    2019

  11. [19]

    In: 2023 IEEE 39th International Conference on Data Engineering (ICDE)

    Liu, M., Huang, H., Feng, H., Sun, L., Du, B., Fu, Y.: Pristi: A conditional diffu- sion framework for spatiotemporal imputation. In: 2023 IEEE 39th International Conference on Data Engineering (ICDE). pp. 1927–1939. IEEE (2023)

    2023

  12. [20]

    In: IJCAI (2021)

    Liu, Q., Xue, H.: Adversarial spectral kernel matching for unsupervised time series domain adaptation. In: IJCAI (2021)

    2021

  13. [21]

    Advances in neural information processing systems31(2018)

    Long, M., Cao, Z., Wang, J., Jordan, M.I.: Conditional adversarial domain adap- tation. Advances in neural information processing systems31(2018)

    2018

  14. [22]

    Advances in neural information pro- cessing systems35, 32069–32082 (2022)

    Marisca, I., Cini, A., Alippi, C.: Learning to reconstruct missing data from spa- tiotemporal graphs with sparse observations. Advances in neural information pro- cessing systems35, 32069–32082 (2022)

    2022

  15. [23]

    In: Proceedings of the AAAI con- ference on artificial intelligence

    Miao, X., Wu, Y., Wang, J., Gao, Y., Mao, X., Yin, J.: Generative semi-supervised learning for multivariate time series imputation. In: Proceedings of the AAAI con- ference on artificial intelligence. pp. 8983–8991 (2021)

    2021

  16. [24]

    In: Inter- national Conference on Learning Representations (2022)

    Miyaguchi, K., Katsuki, T., Koseki, A., Iwamori, T.: Variational inference for dis- criminative learning with generative modeling of feature incompletion. In: Inter- national Conference on Learning Representations (2022)

    2022

  17. [25]

    Springer (1982)

    Nussbaumer, H.J., Nussbaumer, H.J.: The fast Fourier transform. Springer (1982)

    1982

  18. [26]

    IEEE Transactions on Neural Networks and Learning Systems35(1), 1341–1351 (2022)

    Ragab, M., Eldele, E., Chen, Z., Wu, M., Kwoh, C.K., Li, X.: Self-supervised au- toregressive domain adaptation for time series data. IEEE Transactions on Neural Networks and Learning Systems35(1), 1341–1351 (2022)

    2022

  19. [27]

    ACM Transactions on Knowledge Discovery from Data17(8), 1–18 (2023) Cross-Domain Conditional Diffusion Models for Time Series Imputation 17

    Ragab, M., Eldele, E., Tan, W.L., Foo, C.S., Chen, Z., Wu, M., Kwoh, C.K., Li, X.: Adatime: A benchmarking suite for domain adaptation on time series data. ACM Transactions on Knowledge Discovery from Data17(8), 1–18 (2023) Cross-Domain Conditional Diffusion Models for Time Se...

    2023

  20. [28]

    Water Resources Research (2017)

    Read, E.K., Carr, L., De Cicco, L., et al.: Water quality data for national-scale aquatic research: The water quality portal. Water Resources Research (2017)

    2017

  21. [29]

    arXiv preprint arXiv:1802.08735 (2018)

    Shu, R., Bui, H.H., Narui, H., Ermon, S.: A dirt-t approach to unsupervised domain adaptation. arXiv preprint arXiv:1802.08735 (2018)

    2018 arXiv

  22. [30]

    In: Computer vision–ECCV 2016 workshops

    Sun,B.,Saenko,K.:Deepcoral:Correlationalignmentfordeepdomainadaptation. In: Computer vision–ECCV 2016 workshops. Springer (2016)

    2016

  23. [31]

    Advances in neural information processing systems34, 24804–24816 (2021)

    Tashiro, Y., Song, J., Song, Y., Ermon, S.: Csdi: Conditional score-based diffusion models for probabilistic time series imputation. Advances in neural information processing systems34, 24804–24816 (2021)

    2021

  24. [32]

    Advances in neural information pro- cessing systems30(2017)

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, Ł., Polosukhin, I.: Attention is all you need. Advances in neural information pro- cessing systems30(2017)

    2017

  25. [33]

    In: Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining

    Wilson, G., Doppa, J.R., Cook, D.J.: Multi-source deep domain adaptation with weak supervision for time-series sensor data. In: Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining. pp. 1768–1778 (2020)

    2020

  26. [34]

    arXiv preprint arXiv:2210.02186 (2022)

    Wu, H., Hu, T., Liu, Y., et al.: Timesnet: Temporal 2d-variation modeling for general time series analysis. arXiv preprint arXiv:2210.02186 (2022)

    2022 arXiv

  27. [35]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Xu, Q., Zhang, R., Zhang, Y., Wang, Y., Tian, Q.: A fourier-based framework for domain generalization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14383–14392 (2021)

    2021

  28. [36]

    Advances in Neural Information Processing Systems 37, 52595–52623 (2024)

    Yang, X., Sun, Y., Chen, X., et al.: Frequency-aware generative models for multi- variate time series imputation. Advances in Neural Information Processing Systems 37, 52595–52623 (2024)

    2024

  29. [37]

    In: Proceedings of the 25th international joint conference on artificial intelligence (2016)

    Yi, X., Zheng, Y., Zhang, J., Li, T.: St-mvl: filling missing values in geo-sensory time series data. In: Proceedings of the 25th international joint conference on artificial intelligence (2016)

    2016

  30. [38]

    In: International conference on machine learning (2018)

    Yoon, J., Jordon, J., Schaar, M.: Gain: Missing data imputation using generative adversarial nets. In: International conference on machine learning (2018)

    2018

  31. [39]

    IEEE Transactions on Biomedical Engineering66(5), 1477–1490 (2018)

    Yoon, J., Zame, W.R., Van Der Schaar, M.: Estimating missing data in temporal data streams using multi-directional recurrent neural networks. IEEE Transactions on Biomedical Engineering66(5), 1477–1490 (2018)

    2018

  32. [40]

    In: Proceedings of the AAAI con- ference on artificial intelligence

    Yue, Z., Wang, Y., Duan, J., Yang, T., Huang, C., Tong, Y., Xu, B.: Ts2vec: Towards universal representation of time series. In: Proceedings of the AAAI con- ference on artificial intelligence. pp. 8980–8987 (2022)

    2022

  33. [41]

    arXiv preprint arXiv:2410.19265 (2024)

    Zhang, K., Liu, S., Wang, S., Shi, W., Chen, C., Li, P., Li, S., Li, J., Ding, K.: A sur- vey of deep graph learning under distribution shifts: from graph out-of-distribution generalization to adaptation. arXiv preprint arXiv:2410.19265 (2024)

    2024 arXiv

  34. [42]

    In: Proceedings of the 17th ACM International Conference on Web Search and Data Mining

    Zhang, K., Wang, Y., Li, X., Tang, R., Zhang, R.: Incmsr: An incremental learning approach for multi-scenario recommendation. In: Proceedings of the 17th ACM International Conference on Web Search and Data Mining. pp. 939–948 (2024)

    2024

  35. [43]

    Advances in neural infor- mation processing systems35, 3988–4003 (2022)

    Zhang, X., Zhao, Z., Tsiligkaridis, T., Zitnik, M.: Self-supervised contrastive pre- training for time series via time-frequency consistency. Advances in neural infor- mation processing systems35, 3988–4003 (2022)

    2022

  36. [44]

    In: Proceed- ings of the AAAI conference on artificial intelligence

    Zhou, H., Zhang, S., Peng, J., Zhang, S., Li, J., Xiong, H., Zhang, W.: Informer: Beyond efficient transformer for long sequence time-series forecasting. In: Proceed- ings of the AAAI conference on artificial intelligence. pp. 11106–11115 (2021)

    2021

  37. [45]

    In: Proceedings of the 33rd ACM International Conference on Information and Knowledge Management

    Zhou, J., Li, J., Zheng, G., Wang, X., Zhou, C.: Mtsci: A conditional diffusion model for multivariate time series consistent imputation. In: Proceedings of the 33rd ACM International Conference on Information and Knowledge Management. pp. 3474–3483 (2024)

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.