Pith. sign in

REVIEW 4 major objections 6 minor 46 references

FreRA: A Frequency-Refined Augmentation for Contrastive Learning on Time Series Classification

T0 review · 4 major / 6 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 Frequency-domain augmentation preserves time-series labels, the paper argues.

desk verdict A well-engineered frequency augmentation with an overclaimed theoretical guarantee; the empirical work deserves peer review, but the proof needs to be fixed. read the letter →

arxiv 2505.23181 v1 pith:ICZGHEQK submitted 2025-05-29 cs.LG cs.AI

Tian Tian , Chunyan Miao , Hangwei Qian This is my paper
classification cs.LGcs.AI
keywords timeseriesclassificationcontrastivelearningfrequency-domainaugmentationself-supervisedautomaticmutualinformationFouriertransformrepresentation
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 proposes FreRA, a trainable augmentation for contrastive learning on time series that works entirely in the frequency domain. FreRA learns a single vector that scores how much each Fourier component matters for the downstream label, leaves the high-scoring 'critical' components unchanged, and distorts the low-scoring 'unimportant' ones to create positive-view variance. The central theoretical claim is that this operation is semantic-preserving: the augmented view $A_s(x)$ carries exactly the same mutual information about the label $y$ as the original series, $\operatorname{MI}(A_s(x);y)=\operatorname{MI}(x;y)$. If correct, contrastive learners can generate views that add variation without sacrificing the information that classification needs, which would remove the main weakness of hand-picked time-domain augmentations. The paper reports consistent gains over existing augmentations and contrastive frameworks on 135 benchmark datasets.

What carries the argument

The load-bearing object is the Fourier decomposition of the input series and the single trainable importance vector $s$ that FreRA learns over it. Because the Fourier basis is orthogonal, FreRA can treat critical and unimportant components independently: $\mathcal{F}^{-1}(w_{\mathrm{crit}}\odot x_f)$ preserves the global semantics while $\mathcal{F}^{-1}(w_{\mathrm{dist}}\odot x_f)$ injects controlled distortion into every timestamp. The Gumbel-Softmax reparameterization makes the discrete critical mask differentiable, the stop-gradient on $w_{\mathrm{dist}}$ prevents distortion from feeding back into the separator, and the $\ell^1$ regularization enforces compactness so the model does not trivially mark all components critical. Parseval's theorem and the conjugate symmetry of the DFT justify the compact and lightweight structure: energy concentrates in few components, and only about half the Fourier coefficients need be manipulated.

What would settle it

On a dataset where the unimportant Fourier components are correlated with critical ones (for example, a periodic signal whose phase encodes the class and whose low-amplitude harmonics are phase-locked to it), estimate $\operatorname{MI}(A_s(x);y)$ and $\operatorname{MI}(x;y)$ from the learned augmentation; if the former is noticeably lower, Proposition 3's semantic-preservation claim is falsified in that setting.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the frequency domain offers three properties—global reach, independence across components, and compact energy—that let augmentation be both automatic and semantic-preserving. FreRA encodes each frequency component's importance in a lightweight trainable vector $s\in\mathbb{R}^F$, produces a binary mask $w_{\mathrm{crit}}$ via Gumbel-Softmax to keep critical components intact, and applies a self-adaptive distortion $w_{\mathrm{dist}}$ to unimportant components. The augmented view is $A_s(x)=\mathcal{F}^{-1}(w_{\mathrm{crit}}\odot x_f+w_{\mathrm{dist}}\odot x_f)$. Proposition 3 proves $\operatorname{MI}(A_s(x);y)=\operatorname{MI}(x;y)$ under the assumption that the unimportant Fourier components are independent of the label, and an $\ell^1$ penalty on $w_{\mathrm{crit}}$ pushes the selection to the minimal set of components that still carries all label information. The paper argues this satisfies the established definition of an optimal view generator: minimal but sufficient information about the label.

Load-bearing premise

The proof that FreRA preserves all label information assumes the frequency components it labels unimportant are independent of both the label and the critical components; if the learned separator mixes dependent components, the equality $\operatorname{MI}(A_s(x);y)=\operatorname{MI}(x;y)$ can fail.

Editorial extensions

If this is right

  • FreRA can be plugged into any contrastive framework that uses InfoNCE or similar losses; the paper shows gains with SimCLR, TS2Vec, TS-TCC, SoftCLT, and BYOL.
  • Because the importance vector is learned per dataset, the same augmentation recipe applies to heterogeneous archives (UCR, UEA) without hand-picking transformations or bandwidth thresholds.
  • The theoretical semantic-preservation result implies that the learned views do not reduce the information available for downstream classification, so downstream accuracy is bounded by the original data's label information rather than by augmentation loss.
  • The compactness penalty gives a principled way to select how many frequency components to keep, replacing the heuristic low-pass/high-pass choices of earlier frequency augmentations.
  • The same pre-trained encoder transfers to unseen domains and to anomaly detection, suggesting semantic-preserving augmentation improves generalization beyond classification.

Reading between the lines

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

  • If the mutual-information identity holds beyond the paper's benchmarks, then frequency-domain semantic separation is a general inductive bias: any contrastive learner should prefer augmentations that modify only label-independent components, and this could be tested by measuring $\operatorname{MI}(A_s(x);y)$ directly on held-out datasets.
  • The proof's assumption that unimportant components are independent of critical components is stronger than label-independence alone; on signals where phase and harmonics carry correlated label information, the learned separator may need a multivariate or band-aware mask rather than a per-component score.
  • FreRA's threshold rule (mean of $s$ vs. zero) is a pragmatic choice; the paper's ablation shows mean works best, but a learned or per-sample threshold could improve datasets where critical components are not the lowest-frequency ones.
  • Combining FreRA with data-dependent selection of $\lambda$ could tune the compactness of the critical set automatically, making the method fully parameter-free at deployment.
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

4 major / 6 minor

Summary. The paper proposes FreRA, a frequency-domain augmentation method for contrastive learning on time series classification. FreRA learns a lightweight importance vector s, uses a Gumbel-Softmax relaxation to derive a mask w_crit over frequency components, applies identity modification to the identified critical components and a self-adaptive distortion w_dist to the unimportant ones, and integrates the resulting transformation into contrastive learning. The central theoretical claim is Proposition 3, which states that the augmentation is semantic-preserving, i.e., MI(A_s(x);y)=MI(x;y). The paper reports extensive experiments on the UCR and UEA archives, three large HAR datasets, an anomaly-detection dataset, and a transfer-learning benchmark, showing consistent improvements over ten baselines and across several contrastive frameworks.

Significance. If the semantic-preservation guarantee were valid, FreRA would offer a principled and automatic augmentation strategy with a formal information-theoretic justification, and its plug-and-play design would be attractive for time-series self-supervised learning. The empirical evaluation is a clear strength: it spans 135 datasets, multiple contrastive backbones, and additional tasks such as anomaly detection and transfer learning, and the code is released. However, the theoretical claim is the linchpin of the paper's framing, and the proof provided in the appendix is not sound. The main load-bearing result is therefore unsupported, which substantially weakens the contribution's significance.

major comments (4)
  1. [Appendix A.3, Proposition 1] The stated equality H(x)=H(x_f) is not valid for continuous random vectors under the DFT. For a bijective linear map F, the density transforms as p(x_f)=p(F^{-1}x_f)/|det F|, so the differential entropy changes by log|det F|; the proof's claim that p(x)=p(x_f) holds only for discrete variables with counting measure. Proposition 2 (MI conservation) may still be true because the Jacobian term cancels in I(x;y)=H(x)-H(x|y), but the proof as written does not establish the entropy conservation that is later used in Proposition 3.
  2. [Appendix A.3, Proposition 3 proof, Eqs. (15)-(17)] The proof requires xcrit_f and xdist_f to be statistically independent of each other and of the label. These are deterministic functions of the same random vector x_f through complementary masks; orthogonality of the Fourier basis vectors does not imply independence of the random coefficients. For a simple counterexample, with L=2 and independent Rademacher entries, the DFT coefficients X(0)=x0+x1 and X(1)=x0-x1 are dependent. The equalities H(x_f)=H(xcrit_f)+H(xdist_f) and H((wcrit+wdist)⊙x_f)=H(xcrit_f)+H(wdist⊙x_f) are therefore unsupported, and Proposition 3's conclusion MI(A_s(x);y)=MI(x;y) is not established. This is a load-bearing gap because the semantic-preservation guarantee is the paper's central theoretical contribution.
  3. [Section 3.2.1 and Proposition 3] Even if independence of xcrit_f and xdist_f were granted, the proof relies on the assumption that the noisy components xdist_f are independent of the label y. The paper calls this assumption 'reliable' but provides neither theoretical justification nor empirical verification. Because the mask wcrit is learned without labels and is a function of the data, there is no mechanism ensuring that the discarded components carry zero label information. The semantic-preservation result is thus close to being an assumption restated through entropy manipulations rather than a proven property. A concrete test would be to estimate MI(xdist_f; y) on several datasets; without such evidence, the claim is not supported.
  4. [Section 3.3, 'How does the learning objective benefit view generation?'] The text states that optimizing Eq. (9) provably yields a view generator satisfying Definition 1, but no such proof is provided. Definition 1 requires minimizing MI(A(x);x) subject to the equality constraint MI(A(x);y)=MI(x;y). The informal argument that InfoNCE maximizes a lower bound of MI(x,xcrit) and that the regularization term reduces MI(A_s(x),x) does not demonstrate that the constrained minimum is attained. The paper should either present a precise theorem and proof or weaken the claim to say the objective 'encourages' satisfaction of the optimality condition rather than 'proves' it.
minor comments (6)
  1. [Section 3.2.1, 'Independent'] The phrase 'the orthogonal nature of the Fourier basis allow easier isolation and independent modifications' conflates orthogonality of the decomposition basis with statistical independence of the random coefficients; the manuscript should distinguish between deterministic orthogonal decomposition and probabilistic independence.
  2. [Appendix A.3, Proposition 3 proof] The notation 'H(x_f ;y)' appears to be a typo for conditional entropy H(x_f | y), and the expression 'H((wcrit⊙x_f + wdist⊙x_f|y)' is missing a closing parenthesis; these should be corrected for readability.
  3. [Figure 1 and Appendix A.1] The caption of Figure 1 labels the curve as 'MI(s(x); y)' but the figure actually plots timestamp-wise mutual information, as described in Appendix A.1; the caption should state this directly to avoid implying an estimate of the full-sequence mutual information.
  4. [Table 1 footnotes] The footnote marker for the SoftCLT entry in the UEA row is missing or ambiguous; the text should clarify which entries were adopted from the original paper and for which no rank is available.
  5. [Conclusion] The conclusion contains a typo, 'Fr-eRA', which should read 'FreRA'; please correct this in the final version.
  6. [Appendix A.4, InfoMin baseline] The description of the InfoMin baseline states that the original flow-based view generator is replaced with the authors' frequency-domain refinement; this adaptation should be described in the main text when the baseline is first introduced so that readers understand that 'InfoMin+' is a modified version.

Circularity Check

1 steps flagged · score 6.0 of 10

Proposition 3's semantic-preservation 'proof' restates its assumption that the discarded frequency components are label-independent; the central theoretical claim is conditional on the unverified property it claims to establish.

  1. self definitional [Appendix A.3, Proposition 3; also Section 3.2.1 ('Lastly, we demonstrate...')]
    "With the reliable assumption that the noisy frequency components are independent to the label, FreRA is a semantic preserving transformation, i.e., MI(A_s(x);y)=MI(x;y). ... Knowing xcrit_f and xdist_f are independent, we have H(x_f)=H(xcrit_f)+H(xdist_f). ... (xdist_f, as irrelevant components, is independent to y) ... = MI(xcrit_f;y)."

    The proposition's hypothesis is that the noisy frequency components carry no label information; the proof uses exactly this assumption to cancel H(xdist_f) and obtain MI(x_f;y)=MI(xcrit_f;y), and similarly MI(A_s(x);y)=MI(xcrit_f;y). The conclusion MI(A_s(x);y)=MI(x;y) is therefore the assumption 'the modified/discarded part is label-independent' restated through entropy algebra. Since the separator s is learned without labels and the paper never verifies that the learned s satisfies the hypothesis, the 'proof' of semantic preservation is conditional on the very property it claims to establish. This is a self-definitional reduction: 'unimportant' is defined as label-independent, and then modifying unimportant components is shown to preserve label mutual information.

full rationale

FreRA's central theoretical claim is Proposition 3, MI(A_s(x);y)=MI(x;y). The proof in Appendix A.3 assumes (i) xcrit_f and xdist_f are statistically independent and (ii) xdist_f is independent of y. Under these assumptions, the entropy algebra cancels H(xdist_f) and shows both MI(x_f;y) and MI(A_s(x);y) equal MI(xcrit_f;y). Assumption (ii) is exactly the semantic-preservation property for the modified part: 'noisy frequency components are independent to the label.' Thus the proposition is a conditional restatement of its hypothesis rather than an unconditional derivation. The paper also justifies (i) by the orthogonality of the Fourier basis, conflating orthogonality of fixed basis vectors with statistical independence of the random Fourier coefficients; for finite-length non-Gaussian series this factorization is not established. I treat the orthogonality/independence conflation as a correctness risk, not as an additional circularity. The extensive benchmark evaluations (UCR/UEA archives, transfer learning, anomaly detection) are external and independent, and no load-bearing self-citation chain is present. The score reflects partial circularity: the key theoretical 'prediction' reduces by construction to its own assumption, while the empirical contribution remains independently informative.

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

The central theoretical claim relies on unproven independence assumptions about frequency components, an implicit entropy-conservation claim that is false for differential entropy, and the assumption that unsupervised contrastive training discovers the true semantic split. These are not standard math axioms; they are domain assumptions specific to this paper.

free parameters (3)
  • s (frequency importance vector) = learned per dataset, not reported numerically
    Core trainable parameter; scores each frequency component's semantic importance and controls both w_crit and w_dist.
  • lambda (L1 regularization weight) = dataset-specific, e.g., 1 for UCIHAR, 10 for MS, 3 for WISDM
    Balances the InfoNCE loss and critical-component sparsity; searched from {0.1, 0.3, 1, 3, 10, 30}.
  • tau_w (Gumbel-Softmax temperature) = selected from {0.1, 0.2}
    Controls the discretization of w_crit; a minor tuned hyperparameter.
assumptions (4)
  • domain assumption Noisy frequency components are independent of the label.
    Assumed before Proposition 3; without labels, the learned separation is not guaranteed to isolate label-independent components.
  • domain assumption Selected critical and unimportant frequency components are independent random variables.
    Used in the proofs of Propositions 3 and 4; orthogonality of the Fourier basis does not imply statistical independence for arbitrary random vectors.
  • ad hoc to paper DFT preserves entropy.
    Proposition 1 states H(x)=H(x_f); this is false for differential entropy under a non-unitary DFT unless a discrete bijection is assumed, and the paper does not state such a discretization.
  • domain assumption The contrastive objective recovers the true semantic distribution without labels.
    The method relies on the training signal to identify critical components; no proof is given that this recovers the MI-maximizing split.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FreRA: A Frequency-Refined Augmentation for Contrastive Learning on Time Series Classification." pith.science (2026). https://pith.science/paper/ICZGHEQK

@misc{pith2026250523181,
  author       = {Pith},
  title        = {Pith review of: FreRA: A Frequency-Refined Augmentation for Contrastive Learning on Time Series Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ICZGHEQK}},
  note         = {Machine review of arXiv:2505.23181}
}
read the original abstract

Contrastive learning has emerged as a competent approach for unsupervised representation learning. However, the design of an optimal augmentation strategy, although crucial for contrastive learning, is less explored for time series classification tasks. Existing predefined time-domain augmentation methods are primarily adopted from vision and are not specific to time series data. Consequently, this cross-modality incompatibility may distort the semantically relevant information of time series by introducing mismatched patterns into the data. To address this limitation, we present a novel perspective from the frequency domain and identify three advantages for downstream classification: global, independent, and compact. To fully utilize the three properties, we propose the lightweight yet effective Frequency Refined Augmentation (FreRA) tailored for time series contrastive learning on classification tasks, which can be seamlessly integrated with contrastive learning frameworks in a plug-and-play manner. Specifically, FreRA automatically separates critical and unimportant frequency components. Accordingly, we propose semantic-aware Identity Modification and semantic-agnostic Self-adaptive Modification to protect semantically relevant information in the critical frequency components and infuse variance into the unimportant ones respectively. Theoretically, we prove that FreRA generates semantic-preserving views. Empirically, we conduct extensive experiments on two benchmark datasets, including UCR and UEA archives, as well as five large-scale datasets on diverse applications. FreRA consistently outperforms ten leading baselines on time series classification, anomaly detection, and transfer learning tasks, demonstrating superior capabilities in contrastive representation learning and generalization in transfer learning scenarios across diverse datasets.

Figures

Figures reproduced from arXiv: 2505.23181 by the authors.

Figure 1
Figure 1. Our method (blue curve) achieves the highest MI between [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. An overview of the proposed FreRA. The left-hand side presents the detailed design of FreRA: [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Performance of FreRA on the 3 HAR datasets under varying [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Take the UCIHAR dataset as an example, the energy in [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: We aim to achieve the intersection point pointed by the [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 2
Figure 2. Figure 2: For time series classification tasks, we record the best [PITH_FULL_IMAGE:figures/full_fig_p012_2.png]
Figure 6
Figure 6. Figure 6: Despite the diverse distributions of global semantics across three datasets (Libras, ArticularyWordRecognition, and Epilepsy), as [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Critical difference diagrams on UEA and UCR archives. [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 39 canonical work pages

  1. [1]

    Davide Anguita, Alessandro Ghio, Luca Oneto, Xavier Parra, and Jorge L Reyes- Ortiz. 2012. Human activity recognition on smartphones using a multiclass hardware-friendly support vector machine. InAmbient Assisted Living and Home Care: 4th International Workshop, IW AAL 2012, Vitoria-Gasteiz, Spain. Proceedings

    work page 2012

  2. [2]

    Anthony Bagnall, Hoang Anh Dau, Jason Lines, Michael Flynn, James Large, Aaron Bostrom, Paul Southam, and Eamonn Keogh. 2018. The UEA multivariate time series classification archive, 2018.arXiv preprint arXiv:1811.00075(2018)

    arXiv 2018

  3. [3]

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. 2020. A Simple Framework for Contrastive Learning of Visual Representations. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, Virtual Event (Proceedings of Machine Learning Research, Vol. 119). PMLR

    work page 2020

  4. [4]

    Xinlei Chen and Kaiming He. 2021. Exploring Simple Siamese Representation Learning. InIEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual. Computer Vision Foundation / IEEE

    work page 2021

  5. [5]

    Hoang Anh Dau, Anthony Bagnall, Kaveh Kamgar, Chin-Chia Michael Yeh, Yan Zhu, Shaghayegh Gharghabi, Chotirat Ann Ratanamahatana, and Eamonn Keogh

  6. [6]

    Berken Utku Demirel and Christian Holz. 2024. Finding order in chaos: A novel data augmentation method for time series in contrastive learning.Advances in Neural Information Processing Systems36 (2024)

    work page 2024

  7. [7]

    Emadeldeen Eldele, Mohamed Ragab, Zhenghua Chen, Min Wu, Chee-Keong Kwoh, Xiaoli Li, and Cuntai Guan. 2023. Self-supervised contrastive representa- tion learning for semi-supervised time-series classification.IEEE Transactions on Pattern Analysis and Machine Intelligence45, 12 (2023)

    work page 2023

  8. [8]

    Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. 2020. Bootstrap your own latent-a new approach to self-supervised learning.Advances in neural information processing systems33 (2020)

    work page 2020

Show all 46 references
  1. [9]

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. 2020. Mo- mentum contrast for unsupervised visual representation learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition

    2020

  2. [10]

    Zhizhong Huang, Junping Zhang, and Hongming Shan. 2023. Twin Contrastive Learning with Noisy Labels. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada. IEEE

    2023

  3. [11]

    Iris AM Huijben, Arthur Andreas Nijdam, Sebastiaan Overeem, Merel M Van Gilst, and Ruud Van Sloun. 2023. Som-cpc: Unsupervised contrastive learning with self-organizing maps for structured representations of high-rate time series. In International Conference on Machine Learning. PMLR

    2023

  4. [12]

    Ashish Jaiswal, Ashwin Ramesh Babu, Mohammad Zaki Zadeh, Debapriya Baner- jee, and Fillia Makedon. 2020. A survey on contrastive self-supervised learning. Technologies9, 1 (2020)

    2020

  5. [13]

    Eric Jang, Shixiang Gu, and Ben Poole. 2017. Categorical Reparameterization with Gumbel-Softmax. In5th International Conference on Learning Representations, ICLR 2017, Toulon, France, Conference Track Proceedings

    2017

  6. [14]

    Kwapisz, Gary M

    Jennifer R. Kwapisz, Gary M. Weiss, and Samuel Moore. 2010. Activity recognition using cell phone accelerometers.SIGKDD Explor.12, 2 (2010)

    2010

  7. [15]

    Seunghan Lee, Taeyoung Park, and Kibok Lee. 2024. Soft Contrastive Learning for Time Series. InThe Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria

    2024

  8. [16]

    Christian Lessmeier, James Kuria Kimotho, Detmar Zimmer, and Walter Sextro

  9. [17]

    Dongsheng Luo, Wei Cheng, Yingheng Wang, Dongkuan Xu, Jingchao Ni, Wen- chao Yu, Xuchao Zhang, Yanchi Liu, Yuncong Chen, Haifeng Chen, et al. 2023. Time series contrastive learning with information-aware augmentations. InPro- ceedings of the AAAI Conference on Artificial Intel...

    2023

  10. [18]

    Clegg, Andrea Cavallaro, and Hamed Had- dadi

    Mohammad Malekzadeh, Richard G. Clegg, Andrea Cavallaro, and Hamed Had- dadi. 2019. Mobile sensor data anonymization. InProceedings of the International Conference on Internet of Things Design and Implementation, IoTDI 2019, Montreal, QC, Canada. ACM

    2019

  11. [19]

    Qianwen Meng, Hangwei Qian, Yong Liu, Lizhen Cui, Yonghui Xu, and Zhiqi Shen. 2023. MHCCL: masked hierarchical cluster-wise contrastive learning for multivariate time series. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 37

    2023

  12. [20]

    Qianwen Meng, Hangwei Qian, Yong Liu, Yonghui Xu, Zhiqi Shen, and Lizhen Cui. 2023. Unsupervised representation learning for time series: A review.arXiv preprint arXiv:2308.01578(2023)

    2023 arXiv

  13. [21]

    Daniela Micucci, Marco Mobilio, and Paolo Napoletano. 2016. UniMiB SHAR: a new dataset for human activity recognition using acceleration data from smart- phones. abs/1611.07688 (2016). arXiv:1611.07688

    2016 arXiv

  14. [22]

    Marc-Antoine Parseval. 1806. Mémoire sur les séries et sur l’intégration complète d’une équation aux différences partielles linéaires du second ordre, à coefficients constants.Mém. prés. par divers savants, Acad. des Sciences, Paris,(1)1 (1806)

  15. [23]

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. 2019. Pytorch: An imperative style, high-performance deep learning library.Advances in neural information processing systems...

    2019

  16. [24]

    Hangwei Qian. 2019. A Novel Distribution-Embedded Neural Network for Sensor- Based Activity Recognition (IJCAI-19). InProceedings of the 28th International Joint Conference on Artificial Intelligence (IJCAI), Macao, China

    2019

  17. [25]

    Hangwei Qian, Tian Tian, and Chunyan Miao. 2022. What Makes Good Con- trastive Learning on Small-Scale Wearable-based Tasks?. InKDD ’22: The 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Washington, DC, USA. ACM

    2022

  18. [26]

    2001.The discrete Fourier transform: theory, algorithms and applications

    Duraisamy Sundararajan. 2001.The discrete Fourier transform: theory, algorithms and applications. World Scientific

    2001

  19. [27]

    Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. 2020. What Makes for Good Views for Contrastive Learning?. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS ...

    2020

  20. [28]

    Sana Tonekaboni, Danny Eytan, and Anna Goldenberg. 2021. Unsupervised Representation Learning for Time Series with Temporal Neighborhood Coding. In9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria. OpenReview.net

    2021

  21. [29]

    Aäron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation Learning with Contrastive Predictive Coding. abs/1807.03748 (2018). arXiv:1807.03748

    2018 arXiv

  22. [30]

    Tuomas Virtanen, Jort Florent Gemmeke, Bhiksha Raj, and Paris Smaragdis. 2015. Compositional models for audio processing: Uncovering the structure of sound mixtures.IEEE Signal Processing Magazine32, 2 (2015)

    2015

  23. [31]

    Zhiguang Wang, Weizhong Yan, and Tim Oates. 2017. Time series classification from scratch with deep neural networks: A strong baseline. In2017 International Joint Conference on Neural Networks, IJCNN 2017, Anchorage, AK, USA. IEEE

    2017

  24. [32]

    Junkang Wu, Jiawei Chen, Jiancan Wu, Wentao Shi, Xiang Wang, and Xiang- nan He. 2024. Understanding contrastive learning via distributionally robust optimization.Advances in Neural Information Processing Systems36 (2024)

    2024

  25. [33]

    Yu, and Dahua Lin

    Zhirong Wu, Yuanjun Xiong, Stella X. Yu, and Dahua Lin. 2018. Unsupervised Feature Learning via Non-Parametric Instance Discrimination. In2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA. Computer Vision Foundation / IEEE Comp...

    2018

  26. [34]

    Jie Xu, Shuo Chen, Yazhou Ren, Xiaoshuang Shi, Hengtao Shen, Gang Niu, and Xiaofeng Zhu. 2024. Self-Weighted Contrastive Learning among Multiple Views for Mitigating Representation Degeneration.Advances in Neural Information Processing Systems36 (2024)

    2024

  27. [35]

    Ling Yang and Shenda Hong. 2022. Unsupervised time-series representation learn- ing with iterative bilinear temporal-spectral fusion. InInternational Conference on Machine Learning. PMLR

    2022

  28. [36]

    Chun-Hsiao Yeh, Cheng-Yao Hong, Yen-Chi Hsu, Tyng-Luh Liu, Yubei Chen, and Yann LeCun. 2022. Decoupled contrastive learning. InEuropean conference on computer vision. Springer

    2022

  29. [37]

    Kun Yi, Qi Zhang, Wei Fan, Shoujin Wang, Pengyang Wang, Hui He, Ning An, Defu Lian, Longbing Cao, and Zhendong Niu. 2023. Frequency-domain mlps are more effective learners in time series forecasting.Advances in Neural Information Processing Systems36 (2023)

    2023

  30. [38]

    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

    2022

  31. [39]

    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

    2022

  32. [40]

    Xiang Zhang, Ziyuan Zhao, Theodoros Tsiligkaridis, and Marinka Zitnik. 2022. Self-supervised contrastive pre-training for time series via time-frequency con- sistency.Advances in Neural Information Processing Systems35 (2022)

    2022

  33. [41]

    Xu Zheng, Tianchun Wang, Wei Cheng, Aitian Ma, Haifeng Chen, Mo Sha, and Dongsheng Luo. 2024. Parametric Augmentation for Time Series Contrastive Learning. InThe Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria

    2024

  34. [42]

    Tian Zhou, Ziqing Ma, Xue Wang, Qingsong Wen, Liang Sun, Tao Yao, Wotao Yin, and Rong Jin. 2022. FiLM: Frequency improved Legendre Memory Model for Long-term Time Series Forecasting. InNeurIPS

    2022

  35. [43]

    Tian Zhou, Ziqing Ma, Qingsong Wen, Xue Wang, Liang Sun, and Rong Jin

  36. [2016]

    InPHM Society European Conference, Vol

    Condition monitoring of bearing damage in electromechanical drive sys- tems by using motor current signals of electric motors: A benchmark data set for data-driven classification. InPHM Society European Conference, Vol. 3

  37. [2019]

    The UCR time series archive.IEEE/CAA Journal of Automatica Sinica6, 6 (2019)

    2019

  38. [2022]

    InInternational Conference on Machine Learning, ICML 2022, Baltimore, Maryland, USA (Proceedings of Machine Learning Research, Vol

    FEDformer: Frequency Enhanced Decomposed Transformer for Long-term Series Forecasting. InInternational Conference on Machine Learning, ICML 2022, Baltimore, Maryland, USA (Proceedings of Machine Learning Research, Vol. 162). PMLR. FreRA: A Frequency-Refined Augmentation for Co...

    2022

Pith tools

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