Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Merlin: Multi-View Representation Learning for Robust Multivariate Time Series Forecasting with Unfixed Missing Rates

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 A training wrapper that aligns a student model's missing-rate views to a complete-data teacher and to each other makes standard forecasting models beat imputation-based pipelines at missing rates from 25% to 90% with one trained model.

desk verdict A solid, clearly-presented training wrapper for missing-rate robustness, but the headline claims rest on small MAE gaps with no reported variance. read the letter →

arxiv 2506.12459 v1 pith:IMFESWBB submitted 2025-06-14 cs.LG cs.AIstat.ML

Chengqing Yu , Fei Wang , Chuanguang Yang , Zezhi Shao , Tao Sun , Tangwen Qian , Wei Wei , Zhulin An
show 1 more author
Yongjun Xu
This is my paper
classification cs.LGcs.AIstat.ML
keywords multivariatetimeseriesforecastingunfixedmissingratesmulti-viewrepresentationlearningofflineknowledgedistillationcontrastivevaluerobustnessimputationbaselines
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

Merlin addresses a practical failure mode of multivariate time series forecasting: when sensors drop out, missing values destroy the periodicity and trend information a model needs and introduce spurious zeros, and the missing rate itself drifts over time. The paper's claim is that a training wrapper can fix this without touching the forecaster's architecture. Merlin trains a frozen teacher on complete windows and a student on the same windows masked at several rates; MSE distillation pulls the student's hidden states and forecasts toward the teacher's, and multi-view contrastive learning pulls the student's representations of the same window at different missing rates together. The paper reports that STID+Merlin gives the lowest MAE, RMSE, and MAPE on all four datasets at 25%-90% random point missingness, surpassing one-stage models and imputation-based two-stage pipelines, and that one training pass suffices for unfixed rates. The payoff, if true, is that robustness to changing missingness becomes a training-time property rather than a reason to adopt imputation or per-rate model banks.

What carries the argument

The machinery is a teacher-student training wrapper plus a pairwise contrastive loss. A teacher model with the same architecture as the student is first trained on complete observations only; the student then sees $m$ masked copies of each window (here, missing rates 25%, 50%, 75%, and 90%), and its total loss is a weighted sum of an L1 forecasting loss against ground truth, an MSE hidden-representation distillation loss to the teacher, an MSE forecast-result distillation loss to the teacher, and a multi-view contrastive loss. The contrastive loss is a pairwise InfoNCE-style cross-entropy (Eq. 8) computed on projected representations, with positive pairs formed by the same time window at different missing rates and negatives formed by other samples in the batch. Distillation supplies complete-data semantics, contrastive alignment supplies missing-rate invariance, and summing all losses in one stage, rather than training in stages, avoids information forgetting.

What would settle it

On a real or synthetic dataset with block-wise or informative missingness (for example, sensors that fail during high-congestion periods), compare Merlin-wrapped STID against STID with a strong imputation baseline under identical masking; if Merlin fails to beat or match the imputation pipeline at 50%-90% missing rates, the semantic-alignment claim is falsified in that regime. A second check: train the teacher with 25% or higher missingness and see whether Merlin's gain over the raw backbone disappears.

Watch

Extended reading notes

Core claim

The central claim is that semantic alignment is the right remedy for incomplete multivariate time series: a forecasting model should represent an incomplete observation the way it would represent the complete observation it came from, and should represent the same time window the same way regardless of which points are missing. Merlin operationalizes this with two losses layered on a supervised forecasting loss. Offline knowledge distillation uses a teacher trained on complete observations to guide a student trained on masked observations, minimizing MSE between their hidden representations and between their forecasts. Multi-view contrastive learning treats the same window at different missing rates as positive pairs and other windows in the batch as negatives, aligning the student's views across rates. The paper reports that this wrapper improves STID, MTGNN, TimeMixer, and DUET beyond their imputation-augmented counterparts, and that in the unfixed-rates setting STID+Merlin, trained once, beats baselines trained separately per rate.

Load-bearing premise

The method assumes missingness is random point-wise zeros, independent of the series values and of the forecast target, and that the teacher can be trained on essentially complete data; if missingness comes in blocks, correlates with the signal, or contaminates the training set heavily, the alignment targets may become misleading and the central claim can collapse.

Editorial extensions

If this is right

  • Forecasting backbones can gain missing-value robustness by a drop-in training wrapper, leaving their architecture and inference path unchanged.
  • A single Merlin-trained model replaces a family of per-missing-rate models; the paper reports it is trained once and still wins on all four datasets at unfixed rates.
  • Imputation-based two-stage pipelines become unnecessary, avoiding reconstruction error accumulation and separate training runs.
  • The wrapper transfers across heterogeneous backbones (STID, MTGNN, TimeMixer, DUET), so it targets a general weakness rather than one architecture.
  • Even when the teacher is trained on data with 5% missingness, Merlin retains most of its advantage over imputation baselines (Appendix B).

Reading between the lines

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

  • Treating each missing rate as a view turns the missing rate itself into a data augmentation; the same recipe could be tried for other corruptions such as sensor noise, temporal subsampling, or feature dropout, as long as positive pairs retain forecasting-relevant semantics.
  • The two losses likely have different value across the missing-rate spectrum: distillation may matter most at low rates where complete semantics are still recoverable, while contrastive alignment may matter most at high rates; a rate-dependent loss schedule is a natural but untested extension.
  • The method inherits a clean-teacher assumption; a self-distilling variant that bootstraps the teacher from the student's own cleaned views would remove the need for complete training data, at the cost of a new circularity risk.
  • Block-wise or sensor-correlated missingness would stress the positive-pair construction, since same-window views could then share missing blocks rather than losing random points independently; testing on such masks would delimit the method's range.
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 Merlin, a training wrapper for multivariate time series forecasting models that combines offline knowledge distillation from a complete-data teacher with multi-view contrastive learning across different missing-rate views of the same input. The student model is trained once on data with multiple missing rates (25%, 50%, 75%, 90%) and is claimed to outperform both one-stage forecasting models and two-stage imputation-based pipelines on four real-world datasets (METR-LA, PEMS04, China AQI, Global Wind). The paper also reports an unfixed-rate simulation, ablations, hyperparameter analyses, transferability to three other backbones, and an efficiency comparison.

Significance. If the empirical claims hold, Merlin is a simple and potentially practical plug-in that improves the missing-value robustness of several existing forecasting architectures without changing their inference-time behavior, and it avoids training separate models per missing rate. The manuscript has real strengths: the method is clearly described with explicit loss equations, the code is publicly available, the experimental coverage is broad (four datasets, multiple backbones, ablations, and a distinct unfixed-rate protocol), and the authors explicitly acknowledge the random-point-missing scope in Section 2.1 and include a degraded-teacher analysis in Appendix B. The significance is, however, conditional: the headline claim of consistent superiority over imputation baselines rests entirely on point estimates of MAE/RMSE/MAPE with no reported variance or significance testing, and several reported margins are small (e.g., 1–2% relative).

major comments (3)
  1. The central empirical claim, stated in Sec. 4.2 as "STID+Merlin can achieve the best forecasting results on all datasets and all missing rates," is supported only by point estimates averaged over five seeds. No standard deviations, confidence intervals, or significance tests are reported anywhere in the paper. Several margins are small relative to typical time-series seed noise: on METR-LA at 25% missing, STID+Merlin reports MAE 3.21 vs. 3.27 for STID+GATGPT (a 1.8% gap); on Global Wind at 25%, 5.76 vs. 5.82 (about 1%); at 90% on Global Wind, 6.08 vs. 6.15. With only five seeds, these differences may be within seed variance, especially given that all methods share the same zero-imputation input format. The same issue applies to Table 5 (unfixed-rate experiment), where the difference between STID+Merlin and STID+GATGPT (Separately) is 0.07 MAE on METR-LA. The manuscript should report variance or perform paired significance tests across the five seeds, and conclusions should be moderated accordingly.
  2. The paper explicitly restricts its protocol to random point-wise missing with zero substitution (Sec. 2.1: "typically conducted under the random point missing scenario"; Sec. 4.1: "We randomly assign mask points..."). The unfixed-rate simulation in Sec. 4.6 also uses random point masking. Yet the Introduction and Conclusion make general claims about "robust MTSF with unfixed missing rates" without this scope qualification. The contrastive and distillation objectives rely on the assumption that different missing-rate views share the same underlying semantics; this assumption is reasonable for random point missing but may fail for block-wise, sensor-correlated, or informative missingness. The authors should either test these scenarios or explicitly state in the abstract and conclusion that the demonstrated robustness is limited to random point missing.
  3. The paper attributes the method's success to "semantic alignment" between incomplete and complete observations and across missing rates. However, the alignment is measured by the very losses used for training (MSE to the teacher's hidden states and forecasts, and contrastive loss between missing-rate views). As such, the mechanism language is partly circular: the model is optimized to minimize those distances, so reporting that the distances decrease does not independently validate the semantic-alignment interpretation. I am not asking for new experiments, but the narrative should be softened to "enforced alignment" or should include an auxiliary probe (e.g., representation similarity on held-out complete/incomplete pairs, or visualization) that is not part of the training objective, if the authors wish to claim a specific mechanism.
minor comments (5)
  1. Equation (11) has a typo: the mean expression is written as Mean(|Y_tru - Y_M,i|,(11) with an unclosed parenthesis; it should read Mean(|Y_tru - Y_M,i|).
  2. The hyperparameter table lists "gamme 0.5", which should be "gamma 0.5".
  3. The sentence "Since neither the imputation model nor the teacher model is needed during the inference phase" is imprecise: the proposed method does not need the teacher, but the two-stage baselines do require their imputation models at inference time. Please clarify that the statement refers only to the proposed student model.
  4. Figure 1(a) is described qualitatively in the text, but the curves are not labeled in enough detail to identify which model corresponds to which line. Adding a legend or explicit annotation would improve readability.
  5. The hyperparameter sensitivity analysis is conducted only on PEMS04. Given that the optimal values may vary across datasets and that the method relies on several interacting loss weights, reporting at least one additional dataset would strengthen the claim that the chosen defaults are robust.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Merlin's claims are empirical and evaluated on held-out test sets; losses define the training objective rather than smuggling in the conclusions.

full rationale

The paper's central claims are empirical: that the Merlin training wrapper (offline knowledge distillation plus multi-view contrastive learning) improves forecasting accuracy of base models under random point missing rates from 25% to 90%, and that one model trained once handles unfixed missing rates. These claims are supported by held-out test-set MAE/RMSE/MAPE comparisons in Tables 2, 3, and 5, not by the training losses themselves. The losses in Eqs. (5), (6), (10), and (12) define the optimization objective; the paper does not present those losses as an independent prediction of downstream accuracy. The phrase 'semantic alignment' is used as a mechanism description for minimizing those same losses, but the headline evaluations are external forecasting metrics, so this is not a derivation that reduces to its inputs. No fitted parameter is renamed as a prediction: the only fitted quantities are model weights and hyperparameters, and the comparison targets are held-out forecasts. The paper invokes no uniqueness theorem, and its self-citations (STID as backbone, GinAR/DSformer as related work and baselines, Merlin's own prior context) are contextual rather than load-bearing; no central premise is justified solely by an unverified self-citation. The lack of variance or significance reporting on small MAE gaps is a legitimate evidence-quality concern, but it is a correctness/validity issue, not a circularity issue. The random-point-missing scope is explicitly acknowledged in Section 2.1 and Section 4.1, and the unfixed-rate protocol is defined under it; acknowledging a limitation is not circular. Overall, the derivation chain is self-contained in the sense required here: the empirical superiority claims are tested against data, not guaranteed by construction.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

All numbers the method depends on are standard hyperparameters tuned by hand or validation, not derived from theory. The main assumptions are the random point-missing scenario and complete training data, both explicitly adopted by the paper. No new unobserved entities are posited.

free parameters (7)
  • loss weight beta_1 (L1 prediction) = 2
    Selected by hand or validation; controls the supervised forecasting loss in Eq. (12).
  • loss weight beta_2 (distillation) = 2
    Selected by hand or validation; scales L_HD plus L_RD in Eq. (12).
  • loss weight beta_3 (contrastive) = 1
    Selected by hand or validation; scales L_CL in Eq. (12).
  • temperature tau = 1
    Temperature in the contrastive loss in Eq. (8); tuned in Sec. 4.5.
  • projection dimension of Z_E = 16
    Dimension of the FC projection for contrastive representations; tuned in Sec. 4.5.
  • batch size = 16
    Sets the number of negatives in contrastive pairs; tuned in Sec. 4.5.
  • mask rate set = 25%, 50%, 75%, 90%
    The four missing rates used to construct incomplete views; a design choice motivated by robustness range rather than derived from data.
assumptions (4)
  • domain assumption Training-phase observations are complete, with no missing values.
    Used to train the teacher model and to construct masks for the student views; stated in Sec. 2.1 and Sec. 3.2. Appendix B relaxes this to 5% missingness.
  • domain assumption Missing values are random point-wise and are encoded as zeros.
    Required for the positive-pair assumption that different missing-rate views of the same time point share semantics; stated in Sec. 2.1 and Sec. 4.1.
  • domain assumption Teacher representations and forecasts on complete data are valid targets that improve the student on incomplete data.
    Core premise of the knowledge distillation objective in Eqs. (5) and (6); not proven theoretically, only tested empirically through ablations.
  • domain assumption Samples from different time points within a batch are negative pairs for contrastive learning.
    Assumes distinct time windows do not share the same semantics; this is the standard contrastive construction in Eq. (8).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Merlin: Multi-View Representation Learning for Robust Multivariate Time Series Forecasting with Unfixed Missing Rates." pith.science (2026). https://pith.science/paper/IMFESWBB

@misc{pith2026250612459,
  author       = {Pith},
  title        = {Pith review of: Merlin: Multi-View Representation Learning for Robust Multivariate Time Series Forecasting with Unfixed Missing Rates},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IMFESWBB}},
  note         = {Machine review of arXiv:2506.12459}
}
read the original abstract

Multivariate Time Series Forecasting (MTSF) involves predicting future values of multiple interrelated time series. Recently, deep learning-based MTSF models have gained significant attention for their promising ability to mine semantics (global and local information) within MTS data. However, these models are pervasively susceptible to missing values caused by malfunctioning data collectors. These missing values not only disrupt the semantics of MTS, but their distribution also changes over time. Nevertheless, existing models lack robustness to such issues, leading to suboptimal forecasting performance. To this end, in this paper, we propose Multi-View Representation Learning (Merlin), which can help existing models achieve semantic alignment between incomplete observations with different missing rates and complete observations in MTS. Specifically, Merlin consists of two key modules: offline knowledge distillation and multi-view contrastive learning. The former utilizes a teacher model to guide a student model in mining semantics from incomplete observations, similar to those obtainable from complete observations. The latter improves the student model's robustness by learning from positive/negative data pairs constructed from incomplete observations with different missing rates, ensuring semantic alignment across different missing rates. Therefore, Merlin is capable of effectively enhancing the robustness of existing models against unfixed missing rates while preserving forecasting accuracy. Experiments on four real-world datasets demonstrate the superiority of Merlin.

Figures

Figures reproduced from arXiv: 2506.12459 by the authors.

Figure 1
Figure 1. Examples of MTSF with missing values on PEMS04. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The overall framework of our model. During the training phase, we first mask the raw data (complete observations) to [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Results of ablation experiments. w/o HD represents [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Results of hyperparameter experiments (PEMS04). [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Training time for each epoch of several models. [PITH_FULL_IMAGE:figures/full_fig_p009_5.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. Frequency-Aligned Knowledge Distillation for Lightweight Spatiotemporal Forecasting

    cs.LG 2025-06 reject novelty 4.0 of 10

    The proposed spectral distillation method is not actually evaluated in the experiments, and the abstract's headline error reductions contradict the reported tables.

Reference graph

Works this paper leans on

80 extracted references · 65 canonical work pages · cited by 1 Pith paper

  1. [1]

    Hyun Ahn, Kyunghee Sun, and Kwanghoon Pio Kim. 2022. Comparison of missing data imputation methods in time series forecasting.Computers, Materials & Continua70, 1 (2022), 767–779

    work page 2022

  2. [2]

    Cristian Challu, Kin G Olivares, Boris N Oreshkin, Federico Garza Ramirez, Max Mergenthaler Canseco, and Artur Dubrawski. 2023. Nhits: Neural hierarchi- cal interpolation for time series forecasting. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 37. 6989–6997

    work page 2023

  3. [3]

    Jatin Chauhan, Aravindan Raghuveer, Rishi Saket, Jay Nandy, and Balaraman Ravindran. 2022. Multi-Variate Time Series Forecasting on Variable Subsets. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 76–86

    work page 2022

  4. [4]

    Ling Chen, Donghui Chen, Zongjiang Shang, Binqing Wu, Cen Zheng, Bo Wen, and Wei Zhang. 2023. Multi-scale adaptive graph neural network for multivariate time series forecasting.IEEE Transactions on Knowledge and Data Engineering35, 10 (2023), 10748–10761

    work page 2023

  5. [5]

    Si-An Chen, Chun-Liang Li, Sercan O Arik, Nathanael Christian Yoder, and Tomas Pfister. 2023. TSMixer: An All-MLP Architecture for Time Series Forecast- ing.Transactions on Machine Learning Research(2023). https://openreview.net/ forum?id=wbpxTuXgm0

    work page 2023

  6. [6]

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. InInterna- tional conference on machine learning. PMLR, 1597–1607

    2020

  7. [7]

    Xinyu Chen, Zhanhong Cheng, HanQin Cai, Nicolas Saunier, and Lijun Sun. 2024. Laplacian Convolutional Representation for Traffic Time Series Imputation.IEEE Transactions on Knowledge and Data Engineering36, 11 (2024), 6490–6502

    work page 2024

  8. [8]

    Xiaodan Chen, Xiucheng Li, Bo Liu, and Zhijun Li. 2023. Biased Temporal Convolution Graph Network for Time Series Forecasting with Missing Values. InThe Twelfth International Conference on Learning Representations

    work page 2023

Show all 80 references
  1. [9]

    Yakun Chen, Xianzhi Wang, and Guandong Xu. 2023. Gatgpt: A pre-trained large language model with graph attention network for spatiotemporal imputation. arXiv preprint arXiv:2311.14332(2023)

    2023 arXiv

  2. [10]

    Fang Cheng and Hui Liu. 2024. Multi-step electric vehicles charging loads fore- casting: An autoformer variant with feature extraction, frequency enhancement, and error correction blocks.Applied Energy376 (2024), 124308

    2024

  3. [11]

    Yu Chengqing, Yan Guangxi, Yu Chengming, Zhang Yu, and Mi Xiwei. 2023. A multi-factor driven spatiotemporal wind power prediction model based on ensemble deep graph attention reinforcement learning networks.Energy263 (2023), 126034

    2023

  4. [12]

    Andrea Cini, Ivan Marisca, and Cesare Alippi. 2022. Filling the G_ap_s: Mul- tivariate Time Series Imputation by Graph Neural Networks. InInternational Conference on Learning Representations

    2022

  5. [13]

    Jinliang Deng, Xiusi Chen, Zipei Fan, Renhe Jiang, Xuan Song, and Ivor W Tsang

  6. [14]

    Jinliang Deng, Xiusi Chen, Renhe Jiang, Xuan Song, and Ivor W Tsang. 2022. A multi-view multi-task learning framework for multi-variate time series fore- casting.IEEE Transactions on Knowledge and Data Engineering35, 8 (2022), 7665–7680

    2022

  7. [15]

    Jinliang Deng, Xiusi Chen, Renhe Jiang, Du Yin, Yi Yang, Xuan Song, and Ivor W. Tsang. 2024. Disentangling Structured Components: Towards Adaptive, Inter- pretable and Scalable Time Series Forecasting.IEEE Transactions on Knowledge and Data Engineering36, 8 (2024), 3783–3800

    2024

  8. [16]

    Jiewen Deng, Renhe Jiang, Jiaqi Zhang, and Xuan Song. 2024. Multi-modality spatio-temporal forecasting via self-supervised learning. InProceedings of the Thirty-Third International Joint Conference on Artificial Intelligence. 2018–2026

    2024

  9. [17]

    Jinliang Deng, Feiyang Ye, Du Yin, Xuan Song, Ivor Tsang, and Hui Xiong. 2024. Parsimony or capability? decomposition delivers both in long-term time series forecasting.Advances in Neural Information Processing Systems37 (2024), 66687– 66712

    2024

  10. [18]

    Aimei Dong, Jian Liu, Guodong Zhang, Zhonghe Wei, Yi Zhai, and Guohua Lv

  11. [19]

    Jiaxiang Dong, Haixu Wu, Haoran Zhang, Li Zhang, Jianmin Wang, and Ming- sheng Long. 2024. Simmtm: A simple pre-training framework for masked time- series modeling.Advances in Neural Information Processing Systems36 (2024)

    2024

  12. [20]

    Wenjie Du, David Côté, and Yanf Liu. 2023. Saits: Self-attention-based imputation for time series.Expert Systems with Applications219 (2023), 119619

    2023

  13. [21]

    Jianping Gou, Liyuan Sun, Baosheng Yu, Shaohua Wan, and Dacheng Tao. 2023. Hierarchical multi-attention transfer for knowledge distillation.ACM Transac- tions on Multimedia Computing, Communications and Applications20, 2 (2023), 1–20

    2023

  14. [22]

    Jincai Huang, Yongjun Xu, Qi Wang, Qi Cheems Wang, Xingxing Liang, Fei Wang, Zhao Zhang, Wei Wei, Boxuan Zhang, Libo Huang, et al. 2025. Foundation models and intelligent decision-making: Progress, challenges, and perspectives. The Innovation(2025), 100948

    2025

  15. [23]

    Zhenhua Huang, Shunzhi Yang, MengChu Zhou, Zhetao Li, Zheng Gong, and Yunwen Chen. 2022. Feature map distillation of thin nets for low-resolution object recognition.IEEE Transactions on Image Processing31 (2022), 1364–1379

    2022

  16. [24]

    Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic opti- mization.arXiv preprint arXiv:1412.6980(2014)

    2014 arXiv

  17. [25]

    Mengzhang Li and Zhanxing Zhu. 2021. Spatial-temporal fusion graph neural networks for traffic flow forecasting. InProceedings of the AAAI conference on artificial intelligence, Vol. 35. 4189–4196

    2021

  18. [26]

    Yaguang Li, Rose Yu, Cyrus Shahabi, and Yan Liu. 2018. Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting. InInternational Conference on Learning Representations

    2018

  19. [27]

    Zhe Li, Zhongwen Rao, Lujia Pan, Pengyun Wang, and Zenglin Xu. 2023. Ti-mae: Self-supervised masked time series autoencoders.arXiv preprint arXiv:2301.08871 (2023)

    2023 arXiv

  20. [28]

    Ke Liang, Lingyuan Meng, Meng Liu, Yue Liu, Wenxuan Tu, Siwei Wang, Sihang Zhou, Xinwang Liu, Fuchun Sun, and Kunlun He. 2024. A survey of knowledge graph reasoning on graph types: Static, dynamic, and multi-modal.IEEE Trans- actions on Pattern Analysis and Machine Intelligenc...

    2024

  21. [29]

    Hui Liu, Chengqing Yu, Haiping Wu, Zhu Duan, and Guangxi Yan. 2020. A new hybrid ensemble deep reinforcement learning model for wind speed short term forecasting.Energy202 (2020), 117794

    2020

  22. [30]

    Jiexi Liu and Songcan Chen. 2024. Timesurl: Self-supervised contrastive learning for universal time series representation learning. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 13918–13926

    2024

  23. [31]

    Xinwei Liu, Muchuan Qin, Yue He, Xiwei Mi, and Chengqing Yu. 2021. A new multi-data-driven spatiotemporal PM2. 5 forecasting model based on an ensemble graph reinforcement learning convolutional network.Atmospheric Pollution Research12, 10 (2021), 101197

    2021

  24. [32]

    Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. 2024. iTransformer: Inverted Transformers Are Effective for Time Series Forecasting. InThe Twelfth International Conference on Learning Representations. https://openreview.net/forum?id=JePfAI8fah

    2024

  25. [33]

    Ivan Marisca, Cesare Alippi, and Filippo Maria Bianchi. 2024. Graph-based Fore- casting with Missing Data through Spatiotemporal Downsampling. InProceedings of the 41st International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 235). PMLR, 34846–34865

    2024

  26. [34]

    Ivan Marisca, Andrea Cini, and Cesare Alippi. 2022. Learning to reconstruct missing data from spatiotemporal graphs with sparse observations.Advances in neural information processing systems35 (2022), 32069–32082

    2022

  27. [35]

    Xiaoye Miao, Yangyang Wu, Jun Wang, Yunjun Gao, Xudong Mao, and Jianwei Yin. 2021. Generative semi-supervised learning for multivariate time series imputation. InProceedings of the AAAI conference on artificial intelligence, Vol. 35. 8983–8991

    2021

  28. [36]

    Alessio Monti, Angelo Porrello, Simone Calderara, Pasquale Coscia, Lamberto Ballan, and Rita Cucchiara. 2022. How many observations are enough? knowledge distillation for trajectory forecasting. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognitio...

    2022

  29. [37]

    Subhabrata Mukherjee and Ahmed Hassan Awadallah. 2020. XtremeDistil: Multi- stage Distillation for Massive Multilingual Models. InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics. 2221–2234

    2020

  30. [38]

    Jensen, Zhenli Sheng, and Bin Yang

    Xiangfei Qiu, Jilin Hu, Lekui Zhou, Xingjian Wu, Junyang Du, Buang Zhang, Chenjuan Guo, Aoying Zhou, Christian S. Jensen, Zhenli Sheng, and Bin Yang

  31. [39]

    Jensen, and Bin Yang

    Xiangfei Qiu, Zhe Li, Wanghui Qiu, Shiyan Hu, Lekui Zhou, Xingjian Wu, Zhengyu Li, Chenjuan Guo, Aoying Zhou, Zhenli Sheng, Jilin Hu, Christian S. Jensen, and Bin Yang. 2025. TAB: Unified Benchmarking of Time Series Anomaly Detection Methods. InProc. VLDB Endow

    2025

  32. [40]

    Xiangfei Qiu, Xingjian Wu, Yan Lin, Chenjuan Guo, Jilin Hu, and Bin Yang

  33. [41]

    Zezhi Shao, Tangwen Qian, Tao Sun, Fei Wang, and Yongjun Xu. 2025. Spatial- temporal large models: A super hub linking multiple scientific areas with artificial intelligence.The Innovation6, 2 (2025), 100763

    2025

  34. [42]

    Jensen, and Xueqi Cheng

    Zezhi Shao, Fei Wang, Yongjun Xu, Wei Wei, Chengqing Yu, Zhao Zhang, Di Yao, Tao Sun, Guangyin Jin, Xin Cao, Gao Cong, Christian S. Jensen, and Xueqi Cheng

  35. [43]

    Zezhi Shao, Zhao Zhang, Fei Wang, Wei Wei, and Yongjun Xu. 2022. Spatial- Temporal Identity: A Simple yet Effective Baseline for Multivariate Time Series Forecasting. InProceedings of the 31st ACM International Conference on Informa- tion and Knowledge Management. 4454–4458

    2022

  36. [44]

    Zezhi Shao, Zhao Zhang, Fei Wang, and Yongjun Xu. 2022. Pre-training enhanced spatial-temporal graph neural network for multivariate time series forecasting. InProceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining. 1567–1577

    2022

  37. [45]

    Zezhi Shao, Zhao Zhang, Wei Wei, Fei Wang, Yongjun Xu, Xin Cao, and Chris- tian S Jensen. 2022. Decoupled dynamic spatial-temporal graph neural network Merlin: Multi-View Representation Learning for Robust Multivariate Time Series Forecasting with Unfixed Missing Rates KDD ’25...

    2022

  38. [46]

    S Sridevi, S Rajaram, C Parthiban, S SibiArasan, and C Swadhikar. 2011. Impu- tation for the analysis of missing values and prediction of time series data. In 2011 international conference on recent trends in information Technology (ICRTIT). IEEE, 1158–1163

    2011

  39. [47]

    Exploring Progress in Multivariate Time Series Forecasting: Comprehensive Benchmarking and Heterogeneity Analysis.IEEE Transactions on Knowledge and Data Engineering37, 1 (2025), 291–305

    2025

  40. [48]

    Xianfeng Tang, Huaxiu Yao, Yiwei Sun, Charu Aggarwal, Prasenjit Mitra, and Suhang Wang. 2020. Joint modeling of local and global temporal dynamics for multivariate time series forecasting with missing values. InProceedings of the AAAI Conference on Artificial Intelligence, Vol...

    2020

  41. [49]

    Trang H Tran, Lam M Nguyen, Kyongmin Yeo, Nam Nguyen, Dzung Phan, Roman Vaculin, and Jayant Kalagnanam. 2023. An End-to-End Time Series Model for Simultaneous Imputation and Forecast.arXiv preprint arXiv:2306.00778(2023)

    2023 arXiv

  42. [50]

    Fei Wang, Di Yao, Yong Li, Tao Sun, and Zhao Zhang. 2023. AI-enhanced spatial- temporal data-mining technology: New chance for next-generation urban com- puting.The Innovation4, 2 (2023)

    2023

  43. [51]

    Kai Wang, Yu Liu, Qian Ma, and Quan Z Sheng. 2021. Mulde: Multi-teacher knowledge distillation for low-dimensional knowledge graph embeddings. In Proceedings of the Web Conference 2021. 1716–1726

    2021

  44. [52]

    Jing Tan, Hui Liu, Yanfei Li, Shi Yin, and Chengqing Yu. 2022. A new ensemble spatio-temporal PM2. 5 prediction method based on graph attention recursive networks and reinforcement learning.Chaos, Solitons & Fractals162 (2022), 112405

    2022

  45. [53]

    Gerald Woo, Chenghao Liu, Doyen Sahoo, Akshat Kumar, and Steven Hoi. 2021. CoST: Contrastive Learning of Disentangled Seasonal-Trend Representations for Time Series Forecasting. InInternational Conference on Learning Representations

    2021

  46. [54]

    Haixu Wu, Tengge Hu, Yong Liu, Hang Zhou, Jianmin Wang, and Mingsheng Long. 2023. TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis. InThe Eleventh International Conference on Learning Representations

    2023

  47. [55]

    Haixu Wu, Hang Zhou, Mingsheng Long, and Jianmin Wang. 2023. Interpretable weather forecasting for worldwide stations with a unified deep model.Nature Machine Intelligence(2023), 1–10

    2023

  48. [56]

    Xingjian Wu, Xiangfei Qiu, Hongfan Gao, Jilin Hu, Chenjuan Guo, and Bin Yang. 2025. K2VAE: A Koopman-Kalman Enhanced Variational AutoEncoder for Probabilistic Time Series Forecasting. InICML

    2025

  49. [57]

    Shiyu Wang, Haixu Wu, Xiaoming Shi, Tengge Hu, Huakun Luo, Lintao Ma, James Y Zhang, and JUN ZHOU. 2024. TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting. InInternational Conference on Learning Representations (ICLR)

    2024

  50. [58]

    Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, and Chengqi Zhang. 2019. Graph wavenet for deep spatial-temporal graph modeling. InProceedings of the 28th International Joint Conference on Artificial Intelligence. 1907–1913

    2019

  51. [59]

    Qing Xu, Zhenghua Chen, Mohamed Ragab, Chao Wang, Min Wu, and Xiaoli Li

  52. [60]

    Yi Xu, Armin Bazarjani, Hyung-gun Chi, Chiho Choi, and Yun Fu. 2023. Uncov- ering the Missing Pattern: Unified Framework Towards Trajectory Imputation and Prediction. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 9632–9643

    2023

  53. [61]

    Chuanguang Yang, Zhulin An, Linhang Cai, and Yongjun Xu. 2022. Mutual contrastive learning for visual representation learning. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 36. 3045–3053

    2022

  54. [62]

    Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, Xiaojun Chang, and Chengqi Zhang. 2020. Connecting the dots: Multivariate time series forecasting with graph neural networks. InProceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining. 753–763

    2020

  55. [63]

    Chuanguang Yang, Helong Zhou, Zhulin An, Xue Jiang, Yongjun Xu, and Qian Zhang. 2022. Cross-image relational knowledge distillation for semantic segmen- tation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 12319–12328

    2022

  56. [64]

    Kun Yi, Qi Zhang, Wei Fan, Hui He, Liang Hu, Pengyang Wang, Ning An, Long- bing Cao, and Zhendong Niu. 2023. FourierGNN: Rethinking Multivariate Time Series Forecasting from a Pure Graph Perspective. InThirty-seventh Conference on Neural Information Processing Systems

    2023

  57. [65]

    Chengming Yu, Ji Qiao, Chao Chen, Chengqing Yu, and Xiwei Mi. 2024. TFE- former: A new temporal frequency ensemble transformer for day-ahead photo- voltaic power prediction.Journal of Cleaner Production448 (2024), 141690

    2024

  58. [66]

    Chengqing Yu, Fei Wang, Zezhi Shao, Tangwen Qian, Zhao Zhang, Wei Wei, Zhulin An, Qi Wang, and Yongjun Xu. 2025. GinAR+: A Robust End-To-End Framework for Multivariate Time Series Forecasting with Missing Values.IEEE Transactions on Knowledge and Data Engineering(2025), 1–14

    2025

  59. [67]

    Chengqing Yu, Fei Wang, Zezhi Shao, Tangwen Qian, Zhao Zhang, Wei Wei, and Yongjun Xu. 2024. Ginar: An end-to-end multivariate time series forecasting model suitable for variable missing. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 3989–4000

    2024

  60. [68]

    Chuanguang Yang, Zhulin An, Helong Zhou, Fuzhen Zhuang, Yongjun Xu, and Qian Zhang. 2023. Online knowledge distillation via mutual contrastive learn- ing for visual recognition.IEEE Transactions on Pattern Analysis and Machine Intelligence45, 8 (2023), 10212–10227

    2023

  61. [69]

    Chengqing Yu, Fei Wang, Yilun Wang, Zezhi Shao, Tao Sun, Di Yao, and Yongjun Xu. 2024. MGSFformer: A Multi-Granularity Spatiotemporal Fusion Transformer for Air Quality Prediction.Information Fusion(2024), 102607

    2024

  62. [70]

    Zhihan Yue, Yujing Wang, Juanyong Duan, Tianmeng Yang, Congrui Huang, Yunhai Tong, and Bixiong Xu. 2022. Ts2vec: Towards universal representation of time series. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 36. 8980–8987

    2022

  63. [71]

    Xiaoxia Zhang, Shang Shi, HaiChao Sun, Degang Chen, Guoyin Wang, and Kesheng Wu. 2024. ACVAE: A novel self-adversarial variational auto-encoder combined with contrast learning for time series anomaly detection.Neural Networks171 (2024), 383–395

    2024

  64. [72]

    Yunhao Zhang and Junchi Yan. 2022. Crossformer: Transformer utilizing cross- dimension dependency for multivariate time series forecasting. InThe Eleventh International Conference on Learning Representations

    2022

  65. [73]

    Chuanpan Zheng, Xiaoliang Fan, Cheng Wang, Jianzhong Qi, Chaochao Chen, and Longbiao Chen. 2023. INCREASE: Inductive Graph Representation Learning for Spatio-Temporal Kriging. InProceedings of the ACM Web Conference 2023. 673–683

    2023

  66. [74]

    Chengqing Yu, Fei Wang, Zezhi Shao, Tao Sun, Lin Wu, and Yongjun Xu. 2023. Dsformer: A double sampling transformer for multivariate time series long-term prediction. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management. 3062–3072

    2023

  67. [80]

    Tian Zhou, Peisong Niu, Liang Sun, Rong Jin, et al . 2023. One fits all: Power general time series analysis by pretrained lm.Advances in neural information processing systems36 (2023), 43322–43355. A Compared with Different Loss Functions In terms of constructing the loss func...

    2023

  68. [2021]

    The pulse of urban transport: Exploring the co-evolving pattern for spatio- temporal forecasting.ACM Transactions on Knowledge Discovery from Data (TKDD)15, 6 (2021), 1–25

    2021

  69. [2022]

    Contrastive adversarial knowledge distillation for deep model compression in time-series regression tasks.Neurocomputing485 (2022), 242–251

    2022

  70. [2023]

    Momentum contrast transformer for COVID-19 diagnosis with knowledge distillation.Pattern Recognition143 (2023), 109732

    2023

  71. [2024]

    VLDB Endow.17, 9 (2024), 2363–2377

    TFB: Towards Comprehensive and Fair Benchmarking of Time Series Forecasting Methods.Proc. VLDB Endow.17, 9 (2024), 2363–2377

    2024

  72. [2025]

    In SIGKDD

    DUET: Dual Clustering Enhanced Multivariate Time Series Forecasting. In SIGKDD. 1185–1196

Pith tools

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