Pith. sign in

REVIEW 4 major objections 4 minor 49 references

TopoCL: Topological Contrastive Learning for Time Series

T0 review · 4 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read TopoCL claims that adding a cross-modal contrastive loss between temporal and topological embeddings of time series prevents augmentation from destroying semantic information, improving classification, anomaly detection, forecasting, and…

desk verdict A genuinely new combination of persistence diagrams and time-series contrastive learning with a broad but under-reported empirical study; the claim that topology specifically drives the gains is plausible but not fully pinned down. read the letter →

arxiv 2502.02924 v1 pith:LFMGYTUT submitted 2025-02-05 cs.LG cs.AI

Namwoo Kim , Hyungryul Baik , Yoonjin Yoon This is my paper · ORCID
classification cs.LGcs.AI
keywords timeseriesrepresentationlearningcontrastivepersistenthomologytopologicaldataanalysisclassificationanomalydetectionforecastingtransfer
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that data augmentation in time-series contrastive learning destroys seasonal patterns and temporal dependencies, and that this information loss can be compensated by aligning the learned temporal embeddings with topological features computed by persistent homology. The authors treat time and topology as two modalities: a temporal encoder learns from augmented series under the usual hierarchical contrastive loss, while a small permutation-invariant encoder reads persistence diagrams obtained from delay-embedded, per-channel series via the Vietoris–Rips filtration. A cross-modal contrastive loss pulls the two views of the same sample together, and the combined objective is added to existing CL backbones. On classification, anomaly detection, forecasting, and transfer learning benchmarks, the paper reports consistent gains over TS2Vec, CoST, and TS-TCC, including in low-data and heavy-augmentation settings. If correct, this establishes topological invariants as a useful universal signal for time-series representation learning.

What carries the argument

The load-bearing object is the cross-modal time–topology contrastive loss $\mathcal{L}_{\text{cross}}$ together with the topological feature extractor that feeds it. The extractor turns each time series into a multiset of persistence-diagram points, adds the persistence coordinate via $\delta(a,b)=(a,b,b-a)$, and encodes the multiset with shared MLPs and a max-pooling symmetric function, so the representation is invariant to the $M!$ orderings of the point cloud. $\mathcal{L}_{\text{cross}}$ is a temperature-scaled contrastive loss that treats the averaged temporal embedding of a sample's two augmentations as positive for that sample's topological embedding and all other samples in the batch as negatives. Its role is to make the temporal encoder preserve the topological structure that survives augmentation, compensating for the semantic loss that masking, cropping, permutation, and scaling would otherwise introduce.

What would settle it

A decisive check would be to retrain TopoCL with the topological branch receiving no genuine signal: set the delay-embedding dimension to $m=1$ (so no $H_1$ features exist) or randomly permute the assignment of diagrams to time series within each batch while keeping $\mathcal{L}_{\text{cross}}$ active. If the reported gains over the base contrastive models persist under either condition, the improvements do not come from the topological content of the diagrams.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that persistent homology provides a complementary modality that survives the transformations that ruin augmented time series, and that a neural network can learn to use it. Concretely, each channel of a series is mapped to a point cloud by Takens delay embedding; 0- and 1-dimensional persistence diagrams are computed with the Vietoris–Rips complex; diagram points are lifted by $(a,b)\mapsto(a,b,b-a)$; and a permutation-invariant MLP with max-pooling encodes the resulting point set into a vector $h_i$. The temporal encoder produces embeddings $z_i$ and $z_i'$ for two augmentations, and a cross-modal contrastive loss $\mathcal{L}_{\text{cross}}$ maximizes agreement between the averaged temporal embedding $z_i^a$ and the topological embedding $y_i$, with $\mathcal{L}=\mathcal{L}_{\text{time}}+\alpha\mathcal{L}_{\text{cross}}$ as the overall objective. The reported evidence: TS2Vec+TopoCL raises F1 by 1.9 and 2.2 percentage points on Yahoo and KPI anomaly detection (and 1.8 and 0.5 in cold-start), raises average classification accuracy to 0.853 on 125 UCR and 0.748 on 29 UEA datasets, cuts average MSE by 1.94% (univariate) and 1.76% (multivariate) for CoST+TopoCL and by 3.39% and 3.33% for TS2Vec+TopoCL on ETT forecasting, and improves TS-TCC on four transfer benchmarks. Ablations attribute the gain to the joint objective, to both $H_0$ and $H_1$ features, and to max-pooling over average pooling.

Load-bearing premise

The framework's usefulness rests on the assumption that topological summaries built from sliding-window delay embeddings of each channel faithfully capture task-relevant structure; since the main text reports neither the embedding dimension, delay, and truncation choices nor a sensitivity analysis, this premise is not yet established.

Editorial extensions

If this is right

  • Any existing time-series contrastive learner can be upgraded by adding a topological branch and one extra loss term, without changing the downstream classifier or regressor, making the improvement a plug-in rather than a new architecture.
  • The gains reported in low-data and cold-start settings imply that topological features carry information that is cheap to extract from few samples, which matters for domains where labels and data are scarce.
  • The robustness results under jittering, scaling, shifting, permuting, and flipping imply that time-topology alignment acts as a regularizer against augmentation-induced distribution shift.
  • The combination with three different backbones across four tasks supports the paper's claim that topology is a general invariant signal for universal time-series representation learning, a step toward foundation models for time series.
  • The authors acknowledge that computing persistent homology on large-scale data is costly, so the approach's scalability, not its accuracy, is the main obstacle to foundation-model pretraining.

Reading between the lines

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

  • A natural testable extension is to tune the delay-embedding dimension $m$, the delay $\gamma$, and the truncation $M$ per dataset; the main text reports none of these values, so the sensitivity of the reported gains to these choices remains open.
  • The same cross-modal alignment idea could be stacked with frequency-domain consistency to form a time-frequency-topology triple-modal objective, a combination the paper does not test.
  • If the topological features are truly invariant, the method should also transfer to masked-modeling pretraining approaches cited in the related work, which are not tested here; positive results there would strengthen the universality claim.
  • Because the topological branch is permutation-invariant and cheap relative to the temporal encoder, the method might work especially well on multivariate sensor data where channel-wise shapes repeat, but this regime is not isolated in the paper's experiments.
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 / 4 minor

Summary. The paper proposes TopoCL, a self-supervised representation-learning method for time series that adds a topological modality to a contrastive learning backbone. The method computes delay-embedded persistence diagrams for H0 and H1 homology, encodes them with a PointNet-style max-pooling network, and trains jointly with a time-domain contrastive loss (L_time) and a cross-modal time-topology contrastive loss (L_cross). The authors wrap TopoCL around TS2Vec, CoST, and TS-TCC and evaluate on classification (125 UCR and 29 UEA datasets), anomaly detection (Yahoo and KPI), forecasting (ETT datasets), and transfer learning (SleepEEG pre-training with four target datasets), reporting consistent improvements over baselines and an ablation study.

Significance. If the reported gains are reproducible, the paper makes a useful empirical contribution: it demonstrates that persistent-homology features can act as a complementary invariant signal for time-series contrastive learning, and it evaluates the idea across four downstream tasks with competitive baselines, a Nemenyi test for classification, and several ablations. A notable strength is that downstream metrics are computed independently of the training objective, so the reported improvements are not circular. However, the manuscript in its current form has substantial limitations that prevent acceptance: the appendix is missing, key hyperparameters for the topological representation are unreported, there are no error bars or seed variance for most results, and the main ablation does not isolate the topological content of the auxiliary branch.

major comments (4)
  1. [§VI-A (Table VI)] The ablation labeled 'w/o time-topology alignment' removes the entire topological branch together with L_cross, so the 2.4% accuracy drop cannot be attributed specifically to the topological features; an auxiliary branch with the same architecture and a generic contrastive loss could produce the same regularizing effect. Please add a control that keeps the auxiliary branch alive but replaces the persistence-diagram inputs with, for example, random point clouds or a fixed non-topological summary, to test whether persistent homology itself is what matters.
  2. [§IV-C and §VI (overall)] The method's usefulness depends on parameters m (embedding size), gamma (time delay), M (maximum number of topological features), alpha (cross-modal loss weight), and temperature tau, but the manuscript reports no values for any of these and no sensitivity analysis; the appendix, which is repeatedly referenced for implementation details, is absent from the v1. Please provide these values and a sensitivity study, because without them the reader cannot judge whether the reported gains are robust to the choice of topological representation.
  3. [Tables I–V] No standard deviations or seed variance are reported for any of the main experimental tables, and several headline improvements are very small (e.g., 0.5% F1 on KPI cold-start in Table I, 0.1% MSE on ETTh2 univariate forecasting at horizon 720 in Table III). Please report means over multiple seeds or provide significance tests for the central comparisons, since the claimed state-of-the-art performance rests on these differences.
  4. [§V-D (Table V)] The transfer-learning result for TS-TCC+TopoCL on EMG reports perfect accuracy and F1 of 1.0000, which is unusual and is not discussed in the text; this warrants verification and explanation, as a data or evaluation error in this result would directly affect the transfer-learning claim.
minor comments (4)
  1. [§V-B (Figure 3)] Figure 3 is referenced in the text but is not actually present in the manuscript text provided; please ensure the Critical Difference diagram is included and that the claimed Nemenyi significance is actually visible in the figure.
  2. [Table IV] Table IV contains what appear to be typographical errors: the ETTh2/336 row reports TS2Vec MAE 2.197 with MSE 2.136, and the ETTh1/720 row repeats the ETTh1/336 values for Informer and LogTrans; please correct these entries.
  3. [§VI-A] The description of the 'w/o time domain contrastive loss' variant is ambiguous: it says the instance and temporal contrast is removed from f_topo, but f_topo is the topological encoder and does not use that loss; presumably the removal is from the temporal encoder, and the text should say so.
  4. [Throughout] There are several typos and inconsistencies, including 'alining' in Section IV-F, 'peformance' in Section V-B, 'TS2vec' versus 'TS2Vec' in multiple places, and 'F ordA'/'F ordB' in Section V-A; please proofread the manuscript.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: TopoCL's losses and held-out evaluations are independent; the only flagged concern is an ablation control, not a derivation-level reduction.

full rationale

TopoCL's derivation chain is self-contained: the temporal encoder is trained with the TS2Vec-style contrastive loss (Eq. 8), the topological encoder consumes persistence-diagram point clouds (Sec. IV-C and IV-D), and the two are coupled only through the cross-modal contrastive loss Lcross (Eq. 14). Downstream metrics are computed independently on held-out portions of public benchmarks (UCR/UEA, Yahoo/KPI, ETT, SleepEEG transfer), so the reported gains are not equal by construction to any fitted parameter or to the training objective. The topological features are deterministic functions of the input time series, but this is the intended multi-modal setup, not a definitional reduction of the claimed result. The ablation removing Lcross also removes the topological branch, which weakens attribution to persistent homology specifically, but that is an experimental-control limitation rather than a circular step; no load-bearing premise is justified solely by a self-citation or by an imported uniqueness claim. The main unverifiable commitments (delay embedding size m, delay gamma, truncation M, and alpha) are unreported in the v1 with no appendix, but absence of hyperparameters is a reproducibility concern, not circularity.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The method rests on standard TDA tools plus unstated hyperparameters. The free-parameter list captures values that are named but never given; the axioms capture topological assumptions that are invoked but not established for these particular augmentations.

free parameters (6)
  • delay embedding dimension m
    Controls the point cloud from which persistence diagrams are computed; named in Section III-A but no value or selection procedure is reported.
  • time delay gamma
    Controls the lag in delay embedding; named in Section III-A but no value or selection procedure is reported.
  • maximum number of topological features M
    Persistence diagram points are represented as xp in R^(M times 3), so M truncates or pads the topological point set; no value is reported in the main text.
  • cross-modal loss weight alpha
    Balances Ltime and Lcross in the overall objective L = Ltime + alpha Lcross (Equation 15); no value or search range is reported.
  • temperature tau
    Used in the similarity function for cross-modal contrastive loss (Equation 13); no value is reported.
  • SVM RBF penalty C = grid search over {10^i | i in [-4,4]}
    Downstream classifier hyperparameter for the classification task; selected by validation grid search, not a theoretical constant.
assumptions (4)
  • standard math The algebraic definitions of simplicial complexes, homology, and persistence diagrams are correct as stated in Section III.
    The method inherits standard TDA results; the paper cites textbook references instead of proving them.
  • domain assumption Takens delay embedding produces a point cloud whose persistence diagram is informative for the time series.
    Section III-A invokes Takens's theorem; real noisy finite series may not satisfy its genericity conditions, and the parameters m and gamma are not reported.
  • domain assumption Persistent homology captures structure invariant under the data augmentations used in contrastive learning.
    The Introduction states topological features remain invariant under transformations; this is asserted for augmentations such as jittering, scaling, shifting, and permuting rather than demonstrated.
  • ad hoc to paper A union of per-channel persistence diagrams is a sufficient topological representation for multivariate time series.
    Section IV-C defines Dgm(x) as the union over channels of Dgmc(x); no result justifies concatenating per-channel topology for multivariate structure.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TopoCL: Topological Contrastive Learning for Time Series." pith.science (2026). https://pith.science/paper/LFMGYTUT

@misc{pith2026250202924,
  author       = {Pith},
  title        = {Pith review of: TopoCL: Topological Contrastive Learning for Time Series},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LFMGYTUT}},
  note         = {Machine review of arXiv:2502.02924}
}
read the original abstract

Universal time series representation learning is challenging but valuable in real-world applications such as classification, anomaly detection, and forecasting. Recently, contrastive learning (CL) has been actively explored to tackle time series representation. However, a key challenge is that the data augmentation process in CL can distort seasonal patterns or temporal dependencies, inevitably leading to a loss of semantic information. To address this challenge, we propose Topological Contrastive Learning for time series (TopoCL). TopoCL mitigates such information loss by incorporating persistent homology, which captures the topological characteristics of data that remain invariant under transformations. In this paper, we treat the temporal and topological properties of time series data as distinct modalities. Specifically, we compute persistent homology to construct topological features of time series data, representing them in persistence diagrams. We then design a neural network to encode these persistent diagrams. Our approach jointly optimizes CL within the time modality and time-topology correspondence, promoting a comprehensive understanding of both temporal semantics and topological properties of time series. We conduct extensive experiments on four downstream tasks-classification, anomaly detection, forecasting, and transfer learning. The results demonstrate that TopoCL achieves state-of-the-art performance.

Figures

Figures reproduced from arXiv: 2502.02924 by the authors.

Figure 1
Figure 1. Illustration of persistence diagram IV. PROPOSED TOPOCL FRAMEWORK In this section, we present TopoCL framework. A. Problem Formulation Our goal is to train a neural network fθ to transform each time series instance xi ∈ R T ×C in the time series set Xtime = {x1, x2, ..., xN } into its representation ri . Here, N denotes the number of samples, T denotes the number of timestamps, C represents the number of variables. … view at source ↗
Figure 2
Figure 2. (a) Overall framework of TopoCL. TopoCL consists of two distinct modules: a temporal module and a topological [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. A Critical Difference (CD) diagram of representation [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Classification results in a limited data sample scenario. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 4
Figure 4. Figure 4: Classification results on Crop data with various aug￾mentation techniques C. Limited data scenario Due to sensor malfunctions, data transmission errors, or man￾ual data entry issues, sufficient amount of data are not always guaranteed. To verify the effectiveness of th…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 39 canonical work pages

  1. [1]

    A simple framework for contrastive learning of visual representations,

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple framework for contrastive learning of visual representations,” in ICML, 2020

    2020

  2. [2]

    Momentum contrast for unsupervised visual representation learning,

    K. He, H. Fan, Y . Wu, S. Xie, and R. Girshick, “Momentum contrast for unsupervised visual representation learning,” in CVPR, 2020

    2020

  3. [3]

    SimCSE: Simple contrastive learning of sentence embeddings,

    T. Gao, X. Yao, and D. Chen, “SimCSE: Simple contrastive learning of sentence embeddings,” in EMNLP, Nov. 2021

    work page 2021

  4. [4]

    Ts2vec: Towards universal representation of time series,

    Z. Yue, Y . Wang, J. Duan, T. Yang, C. Huang, Y . Tong, and B. Xu, “Ts2vec: Towards universal representation of time series,” in AAAI, 2022

    work page 2022

  5. [5]

    Timesurl: Self-supervised contrastive learning for universal time series representation learning,

    J. Liu and S. Chen, “Timesurl: Self-supervised contrastive learning for universal time series representation learning,” in AAAI, 2024

    work page 2024

  6. [6]

    CoST: Contrastive learning of disentangled seasonal-trend representations for time series forecasting,

    G. Woo, C. Liu, D. Sahoo, A. Kumar, and S. Hoi, “CoST: Contrastive learning of disentangled seasonal-trend representations for time series forecasting,” in ICLR, 2022. [Online]. Available: https://openreview.net/forum?id=PilZY3omXV2

    work page 2022

  7. [7]

    Time-series representation learning via temporal and contextual con- trasting,

    E. Eldele, M. Ragab, Z. Chen, M. Wu, C. K. Kwoh, X. Li, and C. Guan, “Time-series representation learning via temporal and contextual con- trasting,” in IJCAI, 2021

    work page 2021

  8. [8]

    MHCCL: masked hierarchical cluster-wise contrastive learning for multivariate time series,

    Q. Meng, H. Qian, Y . Liu, L. Cui, Y . Xu, and Z. Shen, “MHCCL: masked hierarchical cluster-wise contrastive learning for multivariate time series,” in AAAI, 2023

    work page 2023

Show all 49 references
  1. [9]

    Topology and data,

    G. Carlsson, “Topology and data,” Bulletin of the American Mathemat- ical Society, vol. 46, no. 2, pp. 255–308, 2009

    2009

  2. [10]

    Stability of persis- tence diagrams,

    D. Cohen-Steiner, H. Edelsbrunner, and J. Harer, “Stability of persis- tence diagrams,” in Proceedings of the twenty-first annual symposium on Computational geometry , 2005, pp. 263–271

    2005

  3. [11]

    Statistical topological data analysis using persistence landscapes

    P. Bubenik et al., “Statistical topological data analysis using persistence landscapes.” J. Mach. Learn. Res. , vol. 16, no. 1, pp. 77–102, 2015

    2015

  4. [12]

    Persistence images: A stable vector representation of persistent homology,

    H. Adams, T. Emerson, M. Kirby, R. Neville, C. Peterson, P. Shipman, S. Chepushtanova, E. Hanson, F. Motta, and L. Ziegelmeier, “Persistence images: A stable vector representation of persistent homology,” Journal of Machine Learning Research , vol. 18, no. 8, pp. 1–35, 2017

    2017

  5. [13]

    Supervised learning with indefinite topolog- ical kernels,

    T. Padellini and P. Brutti, “Supervised learning with indefinite topolog- ical kernels,” Statistics, vol. 55, no. 4, pp. 765–786, 2021

    2021

  6. [14]

    Stochastic multiresolution persistent homology kernel

    X. Zhu, A. Vartanian, M. Bansal, D. Nguyen, and L. Brandl, “Stochastic multiresolution persistent homology kernel.” in IJCAI, 2016

    2016

  7. [15]

    The persistence landscape and some of its properties,

    P. Bubenik, “The persistence landscape and some of its properties,” in Topological Data Analysis: The Abel Symposium 2018. Springer, 2020, pp. 97–117

    2018

  8. [16]

    Deep learning with topological signatures,

    C. Hofer, R. Kwitt, M. Niethammer, and A. Uhl, “Deep learning with topological signatures,” in NeurIPS, 2017

    2017

  9. [17]

    Perslay: A neural network layer for persistence diagrams and new graph topological signatures,

    M. Carri `ere, F. Chazal, Y . Ike, T. Lacombe, M. Royer, and Y . Umeda, “Perslay: A neural network layer for persistence diagrams and new graph topological signatures,” in AISTATS, 2020

    2020

  10. [18]

    Topological autoen- coders,

    M. Moor, M. Horn, B. Rieck, and K. Borgwardt, “Topological autoen- coders,” in ICML, 2020

    2020

  11. [19]

    Pllay: Efficient topological layer based on persistent landscapes,

    K. Kim, J. Kim, M. Zaheer, J. Kim, F. Chazal, and L. Wasserman, “Pllay: Efficient topological layer based on persistent landscapes,” in NerIPS, 2020

    2020

  12. [20]

    Mixing up contrastive learning: Self-supervised representation learning for time series,

    K. Wickstrøm, M. Kampffmeyer, K. Ø. Mikalsen, and R. Jenssen, “Mixing up contrastive learning: Self-supervised representation learning for time series,” Pattern Recognition Letters, vol. 155, pp. 54–61, 2022

    2022

  13. [21]

    Simmtm: A simple pre-training framework for masked time-series modeling,

    J. Dong, H. Wu, H. Zhang, L. Zhang, J. Wang, and M. Long, “Simmtm: A simple pre-training framework for masked time-series modeling,” in NeurIPS, 2024

    2024

  14. [22]

    Time series contrastive learning with information-aware augmentations,

    D. Luo, W. Cheng, Y . Wang, D. Xu, J. Ni, W. Yu, X. Zhang, Y . Liu, Y . Chen, H. Chen, and X. Zhang, “Time series contrastive learning with information-aware augmentations,” in AAAI, 2023

    2023

  15. [23]

    Self-supervised contrastive pre-training for time series via time-frequency consistency,

    X. Zhang, Z. Zhao, T. Tsiligkaridis, and M. Zitnik, “Self-supervised contrastive pre-training for time series via time-frequency consistency,” in NeurIPS, 2022

    2022

  16. [24]

    Detecting strange attractors in turbulence,

    F. Takens, “Detecting strange attractors in turbulence,” in Dynamical Systems and Turbulence, Warwick 1980: proceedings of a symposium held at the University of Warwick 1979/80 . Springer, 2006

    1980

  17. [25]

    ¨Uber den h ¨oheren zusammenhang kompakter r ¨aume und eine klasse von zusammenhangstreuen abbildungen,

    L. Vietoris, “ ¨Uber den h ¨oheren zusammenhang kompakter r ¨aume und eine klasse von zusammenhangstreuen abbildungen,” Mathematische Annalen, 1927

    1927

  18. [26]

    Pointnet: Deep learning on point sets for 3d classification and segmentation,

    C. R. Qi, H. Su, K. Mo, and L. J. Guibas, “Pointnet: Deep learning on point sets for 3d classification and segmentation,” in CVPR, 2017

    2017

  19. [27]

    Deep sets,

    M. Zaheer, S. Kottur, S. Ravanbakhsh, B. Poczos, R. R. Salakhutdinov, and A. J. Smola, “Deep sets,” in NeurIPS, 2017

    2017

  20. [28]

    Time-series anomaly detection service at Microsoft,

    H. Ren, B. Xu, Y . Wang, C. Yi, C. Huang, X. Kou, T. Xing, M. Yang, J. Tong, and Q. Zhang, “Time-series anomaly detection service at Microsoft,” in SIGKDD, 2019

    2019

  21. [29]

    A benchmark dataset for time series anomaly detection,

    Y . B. Nikolay Laptev, Saeed Amizadeh, “A benchmark dataset for time series anomaly detection,” https://yahooresearch.tumblr.com/post/ 114590420346/a-benchmark-dataset-for-time-series-anomaly, 2015

    2015

  22. [30]

    Anomaly detec- tion in streams with extreme value theory,

    A. Siffer, P.-A. Fouque, A. Termier, and C. Largouet, “Anomaly detec- tion in streams with extreme value theory,” in SIGKDD, 2017

    2017

  23. [31]

    Unsupervised anomaly detection via variational auto- encoder for seasonal kpis in web applications,

    H. Xu, W. Chen, N. Zhao, Z. Li, J. Bu, Z. Li, Y . Liu, Y . Zhao, D. Pei, Y . Feng, et al. , “Unsupervised anomaly detection via variational auto- encoder for seasonal kpis in web applications,” in WWW, 2018

    2018

  24. [32]

    Fourier transform based spatial outlier mining,

    F. Rasheed, P. Peng, R. Alhajj, and J. Rokne, “Fourier transform based spatial outlier mining,” in International Conference on Intelligent Data Engineering and Automated Learning . Springer, 2009, pp. 317–324

    2009

  25. [33]

    A novel technique for long-term anomaly detection in the cloud,

    O. Vallis, J. Hochenbaum, and A. Kejariwal, “A novel technique for long-term anomaly detection in the cloud,” in 6th USENIX workshop on hot topics in cloud computing (HotCloud 14) , 2014

    2014

  26. [34]

    Luminol (github repository),

    V . Brennan and M. Ritesh, “Luminol (github repository),” https://github. com/linkedin/luminol, 2018

    2018

  27. [35]

    The ucr time series archive,

    H. A. Dau, A. Bagnall, K. Kamgar, C.-C. M. Yeh, Y . Zhu, S. Gharghabi, C. A. Ratanamahatana, and E. Keogh, “The ucr time series archive,” IEEE/CAA Journal of Automatica Sinica , vol. 6, no. 6, pp. 1293–1305, 2019

    2019

  28. [36]

    The UEA multivariate time series classification archive, 2018,

    A. J. Bagnall, H. A. Dau, J. Lines, M. Flynn, J. Large, A. Bostrom, P. Southam, and E. J. Keogh, “The UEA multivariate time series classification archive, 2018,” CoRR, vol. abs/1811.00075, 2018. [Online]. Available: http://arxiv.org/abs/1811.00075

    2018 arXiv

  29. [37]

    Unsupervised scalable representation learning for multivariate time series,

    J.-Y . Franceschi, A. Dieuleveut, and M. Jaggi, “Unsupervised scalable representation learning for multivariate time series,” in NeurIPS, 2019. IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS 10

    2019

  30. [38]

    A transformer-based framework for multivariate time series representation learning,

    G. Zerveas, S. Jayaraman, D. Patel, A. Bhamidipaty, and C. Eickhoff, “A transformer-based framework for multivariate time series representation learning,” in SIGKDD, New York, NY , USA, 2021

    2021

  31. [39]

    Unsupervised represen- tation learning for time series with temporal neighborhood coding,

    S. Tonekaboni, D. Eytan, and A. Goldenberg, “Unsupervised represen- tation learning for time series with temporal neighborhood coding,” in ICLR, 2021

    2021

  32. [40]

    Statistical comparisons of classifiers over multiple data sets,

    J. Dem ˇsar, “Statistical comparisons of classifiers over multiple data sets,” The Journal of Machine Learning Research , vol. 7, pp. 1–30, 2006

    2006

  33. [41]

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

    H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, and W. Zhang, “Informer: Beyond efficient transformer for long sequence time-series forecasting,” in AAAI, 2021

    2021

  34. [42]

    En- hancing the locality and breaking the memory bottleneck of transformer on time series forecasting,

    S. Li, X. Jin, Y . Xuan, X. Zhou, W. Chen, Y .-X. Wang, and X. Yan, “En- hancing the locality and breaking the memory bottleneck of transformer on time series forecasting,” in NeurIPS, 2019

    2019

  35. [43]

    An empirical evaluation of generic convolutional and recurrent networks for sequence modeling,

    S. Bai, J. Z. Kolter, and V . Koltun, “An empirical evaluation of generic convolutional and recurrent networks for sequence modeling,” CoRR, vol. abs/1803.01271, 2018. [Online]. Available: http://arxiv.org/abs/1803.01271

    2018 arXiv

  36. [44]

    Modeling long-and short- term temporal patterns with deep neural networks,

    G. Lai, W.-C. Chang, Y . Yang, and H. Liu, “Modeling long-and short- term temporal patterns with deep neural networks,” in SIGIR, 2018

    2018

  37. [45]

    Indications of nonlinear deterministic and finite-dimensional structures in time series of brain electrical activity: Dependence on recording region and brain state,

    R. G. Andrzejak, K. Lehnertz, F. Mormann, C. Rieke, P. David, and C. E. Elger, “Indications of nonlinear deterministic and finite-dimensional structures in time series of brain electrical activity: Dependence on recording region and brain state,” Physical Review E , 2001

    2001

  38. [46]

    uwave: Accelerometer-based personalized gesture recognition and its applica- tions,

    J. Liu, L. Zhong, J. Wickramasuriya, and V . Vasudevan, “uwave: Accelerometer-based personalized gesture recognition and its applica- tions,” in PMC, 2009

    2009

  39. [47]

    Condition monitoring of bearing damage in electromechanical drive systems by using motor current signals of electric motors: A benchmark data set for data-driven classification,

    C. Lessmeier, J. K. Kimotho, D. Zimmer, and W. Sextro, “Condition monitoring of bearing damage in electromechanical drive systems by using motor current signals of electric motors: A benchmark data set for data-driven classification,” in PHM, 2016

    2016

  40. [48]

    Physionet: components of a new research resource for complex physiologic signals,

    P. PhysioBank, “Physionet: components of a new research resource for complex physiologic signals,” Circulation, 2000

    2000

  41. [49]

    Ti-mae: Self-supervised masked time series autoencoders,

    Z. Li, Z. Rao, L. Pan, P. Wang, and Z. Xu, “Ti-mae: Self-supervised masked time series autoencoders,” arXiv preprint arXiv:2301.08871 , 2023

    2023 arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.