DisMS-TS: Eliminating Redundant Multi-Scale Features for Time Series Classification

Binwu Wang; Bin Zhang; Changsheng Zhang; Peibo Duan; Qi Chu; Xuan Tang; Yongsheng Huang; Zhipeng Liu

REVIEW 3 major objections 9 minor 60 references

DisMS-TS: Eliminating Redundant Multi-Scale Features for Time Series Classification

T0 review · 3 major / 9 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Multi-scale time-series classifiers gain accuracy by separating shared from scale-specific features instead of fusing all scales directly.

desk verdict Modest but honest gains in multi-scale time series classification; the disentanglement story doesn't survive contact with the math, but the empirical core is worth a referee. read the letter →

arxiv 2507.04600 v2 pith:LIWYLKFW submitted 2025-07-07 cs.AI

Zhipeng Liu , Peibo Duan , Binwu Wang , Xuan Tang , Qi Chu , Changsheng Zhang , Yongsheng Huang , Bin Zhang This is my paper

classification cs.AI

keywords TimeseriesclassificationMulti-scaleanalysisFeaturedisentanglementRedundancyeliminationScale-sharedfeaturesScale-specificRepresentationlearningDeep

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 argues that a major obstacle in multi-scale time-series classification is duplication: downsampled views of the same series carry a large common component, and existing methods fuse all scales without removing it. DisMS-TS instead learns to split each scale's representation into a shared part and a scale-specific part, then enforces consistency among shared parts and disparity among specific parts through two regularizers. If the claim is right, the standard recipe for multi-scale classifiers should change from 'weight and fuse all scale embeddings' to 'remove the common component, then combine common and unique information separately.' The paper reports accuracy gains up to 9.71 percentage points over seven baselines on six datasets, with statistically significant improvements on five of the six.

What carries the argument

The load-bearing object is the per-scale decomposition $F_s = F_{s,\mathrm{sha}} + F_{s,\mathrm{spe}}$ implemented by the temporal disentanglement module: an MLP produces a mask $M_s$, and complementary sigmoid gates $M_{s,\mathrm{sha}} = \sigma(M_s/\tau)$ and $M_{s,\mathrm{spe}} = \sigma(-M_s/\tau)$ multiply the scale embedding $F_s$ to produce the two views. The two regularizers, Eq. (9) and Eq. (10), then pull all pairs of shared views toward cosine similarity 1 and all pairs of specific views toward cosine similarity 0. Prediction averages the shared views and concatenates the specific views, so the classifier consumes a single common pattern plus non-redundant scale-specific residuals. The stated theoretical motivation is Eq. (3): fused predictions apply one weight to the whole shared component, so any deviation of the fused weight from the ideal weight causes over- or under-focusing on shared features.

What would settle it

Construct a synthetic time-series classification task in which the class label depends only on the finest-scale component and all coarser scales are noise; if DisMS-TS cannot match a single-scale model trained on the finest scale, then the shared/specific regularization has discarded label-relevant information. Averaged over several random seeds, this comparison would settle whether the method removes redundancy or removes signal.

Watch

Extended reading notes

Core claim

The paper's central claim is that existing multi-scale classifiers are limited not by too little multi-scale information but by redundant scale-shared features, and that explicitly separating each scale's embedding into a shared part $F_{s,\mathrm{sha}}$ and a scale-specific part $F_{s,\mathrm{spe}}$ removes this redundancy. DisMS-TS does this per scale, averages the shared parts, concatenates the specific parts, and optimizes a classification loss together with a similarity loss that drives shared parts toward cosine similarity 1 and a disparity loss that drives specific parts toward cosine similarity 0. Across six benchmarks the model outperforms seven baselines on accuracy, F1, and MCC in the aggregate, with accuracy gains up to 9.71%; on the AWR dataset it trails the strongest multi-scale Transformer baseline by 0.16% in accuracy.

Load-bearing premise

The load-bearing premise is that the shared/specific regularization removes redundancy without throwing away the information needed for classification; the losses constrain geometry, not information content.

Editorial extensions

If this is right

Replacing scale-wise fusion with shared/specific disentanglement improves accuracy by up to 16.19 percentage points over one fusion baseline and 7.77 over another, so the gain comes from the disentanglement rather than from multi-scale encoding alone.
The learned scale-shared representations reach cosine similarities around 0.96 or higher and the learned scale-specific representations stay near zero correlation, so the regularizers behave as intended on real data.
The optimal number of downsampling scales varies by dataset, and too many coarse scales hurts performance, which means the redundancy itself is data-dependent and should be tuned rather than assumed.
The method also improves Matthews correlation coefficient substantially on class-imbalanced datasets, so the benefit is not confined to balanced, multivariate benchmarks.

Reading between the lines

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

A natural extension is to carry the same shared/specific split into forecasting or anomaly detection, where separating a common trend from scale-specific residuals could serve interpretability, but the paper's evidence covers classification only.
Because the regularizers are defined on cosine similarity, 'redundancy elimination' is being measured geometrically; an information-theoretic check, such as mutual information between shared representations and labels, would show whether the reduction is about information content rather than vector geometry.
The complementary sigmoid masks in Eq. (7) have a positive pointwise product, so shared and specific masks are not strictly disjoint; a direct comparison with exactly complementary or additive masks would reveal how much of the gain depends on the mask construction versus the two losses.
A synthetic generative dataset with known shared and scale-specific latent signals could test whether DisMS-TS recovers the true decomposition rather than merely a classification-friendly embedding.

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.

Desk Editor's Note

read the letter

The one thing you should know: this is a decent empirical paper wrapped in a theoretical claim that does not hold up. The architecture is new as far as the cited literature goes—soft masks to split each scale's representation into shared and specific parts, plus cross-scale similarity and disparity losses—and the experiments are careful: five runs per method, standard deviations, t-tests, ablations, and consistent wins on five of six datasets. The code is public, which earns real credit.

Now the soft spot, and it is load-bearing. Section 3.6 claims that because M_sha = sigmoid(M/tau) and M_spe = sigmoid(-M/tau), the resulting F_sha and F_spe are uncorrelated. That is false. sigmoid(-x) = 1 - sigmoid(x), so the two masks sum to one elementwise. For every unit i, (F_sha)_i * (F_spe)_i = m_i(1 - m_i)(F_i)^2 >= 0, with equality only in degenerate cases. The within-scale cosine similarity is generally positive, not zero. The losses L_sim and L_dis only push shared representations across scales toward each other and specific representations across scales toward orthogonality; they do not enforce within-scale separation. So the paper's central interpretation of its own gains—that it eliminates redundant shared features via this disentanglement module—is unsupported. The stress-test note is right.

The motivational derivation in Eq. (3) is also a reparameterization rather than a proof: writing scale-wise fusion as a weighted sum and then interpreting the coefficient sum as over- or under-focusing is definitional, not probative. Minor issues: final settings for S, lambda, and tau are not reported for all datasets; disentanglement baselines (e.g., generic feature-disentanglement methods) are missing; and the "first study" claim is unverified.

That said, I would not dismiss the empirical results. The ablations show that the TDM plus regularizers beat scale-wise fusion with the same encoder, and the gains on FM and NASDAQ are not trivial. The flaw is in the explanation, not necessarily in the method. The fix is direct and empirical: compute and report the actual cosine similarity between F_sha and F_spe within each scale, and either show the mechanism works or reframe the contribution as a regularizer that helps even without strict disentanglement.

Verdict: this deserves a serious referee. I would send it to review with the expectation of major revision on the theoretical section. The empirical core is honest enough to be salvageable.

Referee Report

3 major / 9 minor

Summary. The paper proposes DisMS-TS, a multi-scale framework for time series classification. The input series is downsampled into S coarser scales via average pooling (Eq. 4), each scale is encoded with a CNN patch-projection and a GRU aggregator (Eqs. 5-6), and a Temporal Disentanglement Module uses complementary sigmoid masks to split each scale representation into scale-shared and scale-specific parts (Eq. 7). The model is trained with cross-entropy plus a similarity loss that pulls cross-scale shared representations together and a disparity loss that pushes cross-scale specific representations apart (Eqs. 9-11); prediction uses the average shared representation and the concatenated specific representations (Eq. 8). Experiments on six datasets against seven baselines, with ablations and hyperparameter studies, are reported.

Significance. The empirical study has genuine strengths: five independent runs per method, standard deviations, t-tests at alpha = 0.01, ablations that replace the disentanglement module with two scale-wise fusion variants, public code, and honest reporting of the AWR exception. If the proposed mechanism is genuinely responsible for the gains, the idea of enforcing cross-scale consistency of shared features and cross-scale disparity of specific features would be a useful contribution to multi-scale time series classification. The reported wins on five of six datasets are consistent and nontrivial. However, the stated mechanism in Eq. (7) is not supported as written, and the paper's central interpretation of its own gains therefore needs additional evidence before the claim can be accepted.

major comments (3)

[Sec. 3.6, Eq. (7)] The claim that complementary sigmoid masks make F_s,sha and F_s,spe uncorrelated is false. Because sigmoid(-x) = 1 - sigmoid(x), the masks satisfy M_s,sha + M_s,spe = 1 elementwise. For every unit i, (F_s,sha)_i (F_s,spe)_i = m_i(1 - m_i)(F_s,i)^2 >= 0, so the within-scale inner product is a sum of nonnegative terms and is positive except in the degenerate case where all retained components are zero. Moreover, F_s,sha + F_s,spe = F_s, so the module is a soft reweighting of the original representation rather than a mechanism that removes shared features. The losses in Eqs. (9)-(10) constrain only cross-scale pairs, not within-scale separation. The paper's central statement that DisMS-TS 'eliminates redundant shared features' is therefore not supported. Please remove or rephrase the uncorrelatedness claim, reformulate the module as soft gating, and add a direct empirical check, for example reporting the within-scale cosine similarity between F_s,sha and F_s,spe for each scale, possibly compared against a random-mask baseline.
[Sec. 3.2, Eqs. (2)-(3)] The 'theoretical analysis' in Eqs. (2)-(3) is a reparameterization rather than a proof. Equation (2) defines alpha as the coefficient on F_sha in an ideal model, and Eq. (3) shows that a scale-wise fusion model has coefficient sum_s alpha_s on F_sha. The conclusion that the model over- or under-focuses on shared features whenever sum_s alpha_s != alpha is a restatement of definitions, and no mechanism is given by which a concrete attention-based fusion method would violate or satisfy the equality. Please reframe this passage as motivation or intuition, or state and prove a formal claim about a specific fusion procedure.
[Sec. 1 contribution bullet; Table 2, AWR row] The contribution bullet states that experimental results 'demonstrate that DisMS-TS outperforms all its counterparts,' but Table 2 reports that Pathformer achieves higher ACC, F1, and MCC on AWR (98.43/98.41/0.981 versus 98.27/98.23/0.976), and the text in Sec. 4.2 says DisMS-TS is suboptimal on that dataset. This is an internal inconsistency in a central empirical claim. Please qualify the contribution statement to 'five of six datasets' or otherwise state the AWR exception wherever the claim appears.

minor comments (9)

[Sec. 4.2, Table 2] The 'Improvements' rows report relative percentage improvements rather than absolute percentage-point gains; for example, the 9.71% ACC improvement on FM is 61.00 versus 55.60, which is 5.40 points absolute. Please state this explicitly to avoid misreading.
[Sec. 4.2] The t-test is not described as paired or unpaired, one- or two-sided, and no multiple-comparison correction is mentioned. Please specify, especially since n = 5 per condition.
[Sec. 3.1, Eq. (1)] Equation (1) frames the task as predicting a classification probability at timestamp T+1, but all experiments classify the whole time series. Please align the notation with the task.
[Sec. 4.1.4] 'PyCharm 3.8' appears to mean Python 3.8; PyCharm is an IDE, not a language version. Please correct this.
[Table 2, NASDAQ row] The F1 value 32.69 ± 0.833 has an extra decimal place compared with the other entries; please format consistently.
[Table 3, ISRUC-S3 row] The MCC for 'Replaced with MAGNN's SWF' is reported as 0.778 ± 0.14; MCC is bounded in [-1,1] and this is likely a typo for 0.014, matching the precision of the other rows.
[Figure 5] The colorbar ranges from 0.0 to 1.0, but the displayed correlation values are negative. Please use a symmetric color scale, such as [-1, 1], so that negative values are visually meaningful.
[References, [27]] The reference for Pyraformer contains the placeholder '#PLACEHOLDER_PARENT_METADATA_VALUE#' as its venue. This must be fixed before publication.
[Eq. (7), Sec. 4.4] The temperature tau in Eq. (7) is a hyperparameter, but no default value or sensitivity experiment is reported. Please state the value used in all experiments and consider adding tau to the hyperparameter analysis.

Circularity Check

2 steps flagged · score 2.0 of 10

Minor definitional/self-referential steps in the motivation and RQ4 verification, but the central accuracy claims rest on external benchmarks and are not circular.

self definitional [Section 3.2, Eq. (3)]
"However, an ideal accurate prediction can be achieved if and only if Σ_{s=0}^S α_s = α. When Σ_{s=0}^S α_s > α, the model will over-focus on the scale-shared features, while when Σ_{s=0}^S α_s < α, the model will under-focus on these scale-shared features."

The 'theoretical analysis' is an algebraic restatement of the paper's own decomposition. Once F_s is defined as F_sha + F_{s,spe}, the coefficient on F_sha in any scale-wise weighted sum is necessarily Σ_{s=0}^S α_s. Comparing this coefficient to α (the coefficient in the assumed ideal decomposition of Eq. 2) is a definitional identity, not an empirical derivation of an over- or under-focusing failure mode. The claimed limitation of existing methods is true by construction of the decomposition, so this motivation does not independently support the method; the external ablations and benchmark comparisons carry that burden.
fitted input called prediction [Section 4.5.1 (RQ4), Eqs. (9)-(10)]
"To further verify whether DisMS-TS can effectively disentangle F into F_sha and F_spe, we conduct correlation analysis on them, respectively. ... From Figure 3, we can observe that for any pair of scale-shared representations F_{i∈{0:S},sha} and F_{j∈{0:S},sha} (i ≠ j), they exhibit an extremely high correlation. Moreover, we can also observe that for any pair of scale-specific representations F_{i∈{0:S},spe} and F_{j∈{0:S},spe} (i ≠ j), they are highly uncorrelated."

The reported correlation matrices measure exactly the cosine similarities that L_sim (Eq. 9) drives to 1 and L_dis (Eq. 10) drives to 0. Showing near-1 and near-0 values after training confirms that the regularizers were minimized; it is not an independent test of redundancy elimination. The RQ4 conclusion that DisMS-TS 'can eliminate redundant shared features' is therefore forced by the training objective by construction, rather than being independently verified. The benchmark accuracy claims remain externally supported, so this is partial circularity in the mechanistic validation, not in the headline empirical result.

full rationale

The paper's headline performance claims (Tables 2-3, statistical tests against seven baselines on six datasets) are supported by external benchmarks and ablations, not by fitting a target and then calling it a prediction. There is no load-bearing self-citation chain: the authors' prior works [28-30] appear only as domain citations in the introduction and are not used to justify the disentanglement mechanism or to forbid alternatives. The two flagged steps are real but minor. First, the 'theoretical analysis' in Sec. 3.2 is an algebraic identity: once F_s is defined as F_sha + F_{s,spe}, any scale-wise weighted sum necessarily has coefficient Σα_s on F_sha, so comparing to α restates the assumed ideal decomposition; it motivates the method but does not derive an empirical limitation. Second, the RQ4 correlation analysis checks the exact cosine similarities optimized by Eqs. (9)-(10), so it confirms the regularizers were minimized rather than independently confirming redundancy elimination. Separately, the claim in Sec. 3.6 that complementary sigmoid masks make F_sha and F_spe 'uncorrelated' is mathematically unsupported (sigmoid(-M/τ) = 1 - sigmoid(M/τ), so within-scale inner products are nonnegative), but that is a correctness gap, not a circularity. Overall circularity score: 2.

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

The method is empirically evaluated, but the motivation relies on several unproven modeling assumptions: additive decomposition of scale features, linear scale-weighted prediction for baselines, and equivalence between cosine targets and redundancy removal. S and lambda are tuned per dataset. No invented physical entities are introduced.

free parameters (3)

Number of temporal scales S = S = 3, 5, 5, 6 for HAR, ISRUC-S3, NASDAQ, Epilepsy; values for AWR and FM not reported
Selected via validation accuracy in the hyperparameter analysis of Figure 4a. S directly controls model capacity and the claimed multi-scale benefit.
Regularization weight lambda (lambda1 = lambda2 = lambda) = Final values not reported for the main Table 2 runs; tested over {0.001, 0.005, 0.01, 0.05, 0.1, 0.5, 1.0} on HAR and…
Balances the classification, similarity, and disparity losses in Eq. (11). The paper shows accuracy varies with lambda, so this choice matters.
Temperature tau in sigmoid masks = Not reported
Controls the sharpness of the mask split in Eq. (7). It is not analyzed or disclosed, despite being part of the disentanglement module.

assumptions (4)

domain assumption Multi-scale features decompose additively into one shared component and one specific component per scale: F_s = F_sha + F_spe.
Introduced in Section 3.2 and used in Eq. (2). No justification is given for why a single shared component exists uniformly across all scales.
domain assumption Existing multi-scale methods make predictions as a fixed linear weighting of per-scale embeddings, sum_s alpha_s F_s.
Used in Eq. (3) to conclude that existing methods over- or under-focus on shared features. Attention and gating mechanisms can express more flexible weightings, so this does not describe all baselines.
ad hoc to paper Complementary sigmoid masks make F_sha and F_spe uncorrelated.
Stated after Eq. (7). Mathematically, sigmoid(z) and sigmoid(-z) are not orthogonal functions, and no derivation is provided for the representation-level uncorrelatedness.
ad hoc to paper Driving cross-scale cosine similarities to 1 and 0 removes redundant shared features while preserving predictive information.
The losses in Eqs. (9)-(10) enforce these targets, but the paper does not prove that the resulting representations are exactly non-redundant or that useful scale-specific information is not discarded.

how reviews work

0 comments

Cite this review

Pith. "Pith review of DisMS-TS: Eliminating Redundant Multi-Scale Features for Time Series Classification." pith.science (2026). https://pith.science/paper/LIWYLKFW

@misc{pith2026250704600,
  author       = {Pith},
  title        = {Pith review of: DisMS-TS: Eliminating Redundant Multi-Scale Features for Time Series Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LIWYLKFW}},
  note         = {Machine review of arXiv:2507.04600}
}

read the original abstract

Real-world time series typically exhibit complex temporal variations, making the time series classification task notably challenging. Recent advancements have demonstrated the potential of multi-scale analysis approaches, which provide an effective solution for capturing these complex temporal patterns. However, existing multi-scale analysis-based time series prediction methods fail to eliminate redundant scale-shared features across multi-scale time series, resulting in the model over- or under-focusing on scale-shared features. To address this issue, we propose a novel end-to-end Disentangled Multi-Scale framework for Time Series classification (DisMS-TS). The core idea of DisMS-TS is to eliminate redundant shared features in multi-scale time series, thereby improving prediction performance. Specifically, we propose a temporal disentanglement module to capture scale-shared and scale-specific temporal representations, respectively. Subsequently, to effectively learn both scale-shared and scale-specific temporal representations, we introduce two regularization terms that ensure the consistency of scale-shared representations and the disparity of scale-specific representations across all temporal scales. Extensive experiments conducted on multiple datasets validate the superiority of DisMS-TS over its competitive baselines, with the accuracy improvement up to 9.71%.

Figures

Figures reproduced from arXiv: 2507.04600 by the authors.

**Figure 2.** The framework of DisMS-TS. 𝑆 is assumed to be 2 for the sake of visualization. Disentanglement Module (TDM) is proposed to disentangle the temporal representations at each scale, generating both scale-shared and scale-specific temporal representations. Finally, the model is optimized utilizing similarity and disparity losses, combined with classification loss, to enable the predictions based on both scaleshared an… view at source ↗

**Figure 5.** Correlation analysis of scale-specific representa [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Hyperparameter analysis. 4.4.2 The Value of 𝜆1 and 𝜆2 (RQ4). We further discuss the sensitivity of the hyperparameters 𝜆1 and 𝜆2 in Eq.11, which balance the trade-offs between classification, similarity and disparity losses. In this study, we set 𝜆1 and 𝜆2 to be equal…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

60 extracted references · 30 canonical work pages

[1]

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 Engineering 35, 10 (2023), 10748–10761

2023
[2]

Peng Chen, Yingying Zhang, Yunyao Cheng, Yang Shu, Yihang Wang, Qing- song Wen, Bin Yang, and Chenjuan Guo. 2024. Pathformer: Multi-scale trans- formers with adaptive pathways for time series forecasting. arXiv preprint arXiv:2402.05956 (2024)

arXiv 2024
[3]

Wei Chen and Ke Shi. 2021. Multi-scale attention convolutional neural network for time series classification. Neural Networks 136 (2021), 126–140

work page 2021
[4]

Zipeng Chen, Qianli Ma, and Zhenxi Lin. 2021. Time-Aware Multi-Scale RNNs for Time Series Modeling.. In IJCAI. 2285–2291

work page 2021
[5]

Mingyue Cheng, Qi Liu, Zhiding Liu, Zhi Li, Yucong Luo, and Enhong Chen
[6]

Yunyao Cheng, Peng Chen, Chenjuan Guo, Kai Zhao, Qingsong Wen, Bin Yang, and Christian S Jensen. 2023. Weakly guided adaptation for robust time series forecasting. Proceedings of the VLDB Endowment 17, 4 (2023), 766–779

work page 2023
[7]

Emadeldeen Eldele, Mohamed Ragab, Zhenghua Chen, Min Wu, Chee Keong Kwoh, Xiaoli Li, and Cuntai Guan. 2021. Time-series representation learning via temporal and contextual contrasting. arXiv preprint arXiv:2106.14112 (2021)

arXiv 2021
[8]

Hongbo Guo, Xinzi Xu, Hao Wu, and Guoxing Wang. 2023. Multi-Scale and Multi-Modal Contrastive Learning Network for Biomedical Time Series. arXiv preprint arXiv:2312.03796 (2023)

work page Pith review arXiv 2023

Show all 60 references

[9]

Isabelle Guyon and André Elisseeff. 2003. An introduction to variable and feature selection. Journal of machine learning research 3, Mar (2003), 1157–1182

2003
[10]

Ronghang Hu, Amanpreet Singh, Trevor Darrell, and Marcus Rohrbach. 2020. Iterative answer prediction with pointer-augmented multimodal transformers for textvqa. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 9992–10002

2020
[11]

Qihe Huang, Lei Shen, Ruixin Zhang, Shouhong Ding, Binwu Wang, Zhengyang Zhou, and Yang Wang. 2023. Crossgnn: Confronting noisy multivariate time series via cross interaction refinement. Advances in Neural Information Processing Systems 36 (2023), 46885–46902

2023
[12]

Qihe Huang, Zhengyang Zhou, Kuo Yang, and Yang Wang. 2025. Exploiting Language Power for Time Series Forecasting with Exogenous Variables. In Pro- ceedings of the ACM on Web Conference 2025 . 4043–4052

2025
[13]

Siteng Huang, Donglin Wang, Xuehan Wu, and Ao Tang. 2019. Dsanet: Dual self-attention network for multivariate time series forecasting. In Proceedings of the 28th ACM international conference on information and knowledge management . 2129–2132

2019
[14]

Ming Jin, Huan Yee Koh, Qingsong Wen, Daniele Zambon, Cesare Alippi, Geof- frey I Webb, Irwin King, and Shirui Pan. 2024. A survey on graph neural networks for time series: Forecasting, classification, imputation, and anomaly detection. IEEE Transactions on Pattern Analysis an...

2024
[15]

Ming Jin, Shiyu Wang, Lintao Ma, Zhixuan Chu, James Y Zhang, Xiaoming Shi, Pin-Yu Chen, Yuxuan Liang, Yuan-Fang Li, Shirui Pan, et al . 2023. Time-llm: Time series forecasting by reprogramming large language models. arXiv preprint arXiv:2310.01728 (2023)

2023 arXiv
[16]

Tung Kieu, Bin Yang, Chenjuan Guo, Razvan-Gabriel Cirstea, Yan Zhao, Yale Song, and Christian S Jensen. 2022. Anomaly detection in time series with robust variational quasi-recurrent autoencoders. In 2022 IEEE 38th International Conference on Data Engineering (ICDE) . IEEE, 1342–1354

2022
[17]

Diederik P Kingma. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

2014 arXiv
[18]

Guokun Lai, Wei-Cheng Chang, Yiming Yang, and Hanxiao Liu. 2018. Modeling long-and short-term temporal patterns with deep neural networks. In The 41st international ACM SIGIR conference on research & development in information retrieval. 95–104

2018
[19]

Dongha Lee, Seonghyeon Lee, and Hwanjo Yu. 2021. Learnable dynamic temporal pooling for time series classification. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 35. 8288–8296

2021
[20]

Tianfu Li, Zhibin Zhao, Chuang Sun, Ruqiang Yan, and Xuefeng Chen. 2021. Hierarchical attention graph convolutional network to fuse multi-sensor signals for remaining useful life prediction. Reliability Engineering & System Safety 215 (2021), 107878

2021
[21]

Shengsheng Lin, Weiwei Lin, Wentai Wu, Haojun Chen, and Junjie Yang. 2024. Sparsetsf: Modeling long-term time series forecasting with 1k parameters. arXiv preprint arXiv:2405.00946 (2024)

2024 arXiv
[22]

Chenxi Liu, Kethmi Hirushini Hettige, Qianxiong Xu, Cheng Long, Shili Xiang, Gao Cong, Ziyue Li, and Rui Zhao. 2025. ST-LLM+: Graph Enhanced Spatio- Temporal Large Language Models for Traffic Prediction. IEEE Transactions on Knowledge and Data Engineering (2025)

2025
[23]

Chenxi Liu, Hao Miao, Qianxiong Xu, Shaowen Zhou, Cheng Long, Yan Zhao, Ziyue Li, and Rui Zhao. 2025. Efficient Multivariate Time Series Forecasting via Calibrated Language Models with Privileged Knowledge Distillation. In41th IEEE International Conference on Data Engineering

2025
[24]

Chenxi Liu, Qianxiong Xu, Hao Miao, Sun Yang, Lingzheng Zhang, Cheng Long, Ziyue Li, and Rui Zhao. 2025. Timecma: Towards llm-empowered multivariate time series forecasting via cross-modality alignment. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39....

2025
[25]

Chenxi Liu, Sun Yang, Qianxiong Xu, Zhishuai Li, Cheng Long, Ziyue Li, and Rui Zhao. 2024. Spatial-Temporal Large Language Model for Traffic Prediction. In 25th IEEE International Conference on Mobile Data Management . 31–40

2024
[26]

Minhao Liu, Ailing Zeng, Muxi Chen, Zhijian Xu, Qiuxia Lai, Lingna Ma, and Qiang Xu. 2022. Scinet: Time series modeling and forecasting with sample convolution and interaction. Advances in Neural Information Processing Systems 35 (2022), 5816–5828

2022
[27]

Shizhan Liu, Hang Yu, Cong Liao, Jianguo Li, Weiyao Lin, Alex X Liu, and Schahram Dustdar. 2022. Pyraformer: Low-complexity pyramidal at- tention for long-range time series modeling and forecasting. In # PLACE- HOLDER_PARENT_METADATA_V ALUE#

2022
[28]

Zhipeng Liu, Peibo Duan, Qi Chu, Levin Kuhlmann, Changsheng Zhang, Wenwei Yue, Xuan Tang, and Bin Zhang. 2025. An Attributed Multiplex Network Enabled GNN-based Stock Predictor with Observable and Non-observable Information. Expert Systems with Applications (2025), 129018

2025
[29]

Zhipeng Liu, Peibo Duan, Mingyang Geng, and Bin Zhang. 2025. A Distillation- based Future-aware Graph Neural Network for Stock Trend Prediction. InICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 1–5

2025
[30]

Zhipeng Liu, Peibo Duan, Xiaosha Xue, Changsheng Zhang, Wenwei Yue, and Bin Zhang. 2025. A dynamic hypergraph attention network: Capturing market-wide spatiotemporal dependencies for stock selection. Applied Soft Computing 169 (2025), 112524

2025
[31]

Hao Miao, Ronghui Xu, Yan Zhao, Senzhang Wang, Jianxin Wang, Philip S Yu, and Christian S Jensen. 2025. A Parameter-Efficient Federated Framework for Streaming Time Series Anomaly Detection via Lightweight Adaptation. TMC (2025)

2025
[32]

Hao Miao, Yan Zhao, Chenjuan Guo, Bin Yang, Kai Zheng, Feiteng Huang, Jian- dong Xie, and Christian S Jensen. 2024. A unified replay-based continuous learning framework for spatio-temporal prediction on streaming data. In 2024 IEEE 40th International Conference on Data Enginee...

2024
[33]

Mohammad Amin Morid, Olivia R Liu Sheng, and Joseph Dunbar. 2023. Time series prediction using deep learning methods in healthcare. ACM Transactions on Management Information Systems 14, 1 (2023), 1–29

2023
[34]

Michael C Mozer. 1991. Induction of multiscale temporal structure. Advances in neural information processing systems 4 (1991)

1991
[35]

Yuqi Nie, Nam H Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. 2022. A time series is worth 64 words: Long-term forecasting with transformers. arXiv preprint arXiv:2211.14730 (2022)

2022 arXiv
[36]

Yao Qin, Dongjin Song, Haifeng Chen, Wei Cheng, Guofei Jiang, and Garrison Cottrell. 2017. A dual-stage attention-based recurrent neural network for time series prediction. arXiv preprint arXiv:1704.02971 (2017)

2017 arXiv
[37]

Xiangfei Qiu, Jilin Hu, Lekui Zhou, Xingjian Wu, Junyang Du, Buang Zhang, Chenjuan Guo, Aoying Zhou, Christian S Jensen, Zhenli Sheng, et al. 2024. Tfb: Towards comprehensive and fair benchmarking of time series forecasting meth- ods. arXiv preprint arXiv:2403.20150 (2024)

2024 arXiv
[38]

Xiangfei Qiu, Xingjian Wu, Yan Lin, Chenjuan Guo, Jilin Hu, and Bin Yang
[39]

Rajat Sen, Hsiang-Fu Yu, and Inderjit S Dhillon. 2019. Think globally, act locally: A deep neural network approach to high-dimensional time series forecasting. Advances in neural information processing systems 32 (2019)

2019
[40]

Kamile Stankeviciute, Ahmed M Alaa, and Mihaela van der Schaar. 2021. Confor- mal time-series forecasting. Advances in neural information processing systems 34 (2021), 6216–6228

2021
[41]

Sandeep Subramanian, Ronan Collobert, Marc’Aurelio Ranzato, and Y-Lan Boureau. 2020. Multi-scale Transformer Language Models. arXiv preprint arXiv:2005.00581 (2020)

2020 arXiv
[42]

Xiao Teng, Xiang Zhang, and Zhigang Luo. 2022. Multi-scale local cues and hier- archical attention-based LSTM for stock price trend prediction. Neurocomputing 505 (2022), 92–100

2022
[43]

Binwu Wang, Yudong Zhang, Xu Wang, Pengkun Wang, Zhengyang Zhou, Lei Bai, and Yang Wang. 2023. Pattern expansion and consolidation on evolving graphs for continual traffic prediction. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 222...

2023
[44]

Chengsen Wang, Zirui Zhuang, Qi Qi, Jingyu Wang, Xingyu Wang, Haifeng Sun, and Jianxin Liao. 2023. Drift doesn’t matter: dynamic decomposition with diffusion reconstruction for unstable multivariate time series anomaly detection. Advances in Neural Information Processing Syste...

2023
[45]

Huiqiang Wang, Jian Peng, Feihu Huang, Jince Wang, Junhui Chen, and Yifei Xiao
[46]

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. arXiv preprint arXiv:2405.14616 (2024)

2024 arXiv
[47]

Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. 2021. Pyramid vision transformer: A versatile back- bone for dense prediction without convolutions. In Proceedings of the IEEE/CVF international conference on computer vi...

2021
[48]

In The eleventh international conference on learning representations

Micn: Multi-scale local and global context modeling for long-term series forecasting. In The eleventh international conference on learning representations
[49]

Yucheng Wang, Yuecong Xu, Jianfei Yang, Min Wu, Xiaoli Li, Lihua Xie, and Zhenghua Chen. 2024. Graph-Aware Contrasting for Multivariate Time-Series Classification. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 38. 15725–15734

2024
[50]

Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. 2021. Autoformer: De- composition transformers with auto-correlation for long-term series forecasting. Advances in neural information processing systems 34 (2021), 22419–22430

2021
[51]

Yuxuan Wang, Haixu Wu, Jiaxiang Dong, Yong Liu, Mingsheng Long, and Jianmin Wang. 2024. Deep time series models: A comprehensive survey and benchmark. arXiv preprint arXiv:2407.13278 (2024)

2024 arXiv
[52]

Haoyu Zhang, Wentao Zhang, Hao Miao, Xinke Jiang, Yuchen Fang, and Yifan Zhang. 2025. STRAP: Spatio-Temporal Pattern Retrieval for Out-of-Distribution Generalization. arXiv preprint arXiv:2505.19547 (2025)

2025
[53]

Yunhao Zhang and Junchi Yan. 2023. Crossformer: Transformer utilizing cross- dimension dependency for multivariate time series forecasting. In The eleventh international conference on learning representations

2023
[54]

Weiwei Ye, Zhuopeng Xu, and Ning Gui. 2025. Non-stationary Diffusion For Probabilistic Time Series Forecasting. arXiv preprint arXiv:2505.04278 (2025)

2025 arXiv
[55]

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

2022
[56]

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 systems 36 (2023), 43322–43355

2023
[57]

Shuhan Zhong, Sizhe Song, Weipeng Zhuo, Guanyao Li, Yang Liu, and S-H Gary Chan. 2023. A multi-scale decomposition mlp-mixer for time series analysis. arXiv preprint arXiv:2310.11959 (2023)

2023 arXiv
[60]

Chenglong Zhu, Xueling Ma, Weiping Ding, and Jianming Zhan. 2023. Long-term time series forecasting with multi-linear trend fuzzy information granules for LSTM in a periodic framework. IEEE Transactions on Fuzzy Systems (2023)

2023
[2023]

In Proceedings of the ACM Web Conference 2023

Formertime: Hierarchical multi-scale representations for multivariate time series classification. In Proceedings of the ACM Web Conference 2023 . 1437–1445

2023
[2024]

arXiv preprint arXiv:2412.10859 (2024)

Duet: Dual clustering enhanced multivariate time series forecasting. arXiv preprint arXiv:2412.10859 (2024)

2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

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 Engineering 35, 10 (2023), 10748–10761

2023

[2] [2]

Peng Chen, Yingying Zhang, Yunyao Cheng, Yang Shu, Yihang Wang, Qing- song Wen, Bin Yang, and Chenjuan Guo. 2024. Pathformer: Multi-scale trans- formers with adaptive pathways for time series forecasting. arXiv preprint arXiv:2402.05956 (2024)

arXiv 2024

[3] [3]

Wei Chen and Ke Shi. 2021. Multi-scale attention convolutional neural network for time series classification. Neural Networks 136 (2021), 126–140

work page 2021

[4] [4]

Zipeng Chen, Qianli Ma, and Zhenxi Lin. 2021. Time-Aware Multi-Scale RNNs for Time Series Modeling.. In IJCAI. 2285–2291

work page 2021

[5] [5]

Mingyue Cheng, Qi Liu, Zhiding Liu, Zhi Li, Yucong Luo, and Enhong Chen

[6] [6]

Yunyao Cheng, Peng Chen, Chenjuan Guo, Kai Zhao, Qingsong Wen, Bin Yang, and Christian S Jensen. 2023. Weakly guided adaptation for robust time series forecasting. Proceedings of the VLDB Endowment 17, 4 (2023), 766–779

work page 2023

[7] [7]

Emadeldeen Eldele, Mohamed Ragab, Zhenghua Chen, Min Wu, Chee Keong Kwoh, Xiaoli Li, and Cuntai Guan. 2021. Time-series representation learning via temporal and contextual contrasting. arXiv preprint arXiv:2106.14112 (2021)

arXiv 2021

[8] [8]

Hongbo Guo, Xinzi Xu, Hao Wu, and Guoxing Wang. 2023. Multi-Scale and Multi-Modal Contrastive Learning Network for Biomedical Time Series. arXiv preprint arXiv:2312.03796 (2023)

work page Pith review arXiv 2023