Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

Mamba Adaptive Anomaly Transformer with association discrepancy for time series

T0 review · 5 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read MAAT, an Anomaly Transformer augmented with sparse attention, a Mamba state-space block, and gated skip connections, reports higher anomaly-detection F1 than prior unsupervised methods on eight benchmarks.

desk verdict A plausible incremental architecture whose headline results are contradicted by its own Table 3; needs major revision before the empirical claims can be trusted. read the letter →

arxiv 2502.07858 v3 pith:OSVSMBXG submitted 2025-02-11 cs.LG

Abdellah Zakaria Sellam , Ilyes Benaissa , Abdelmalik Taleb-Ahmed , Luigi Patrono , Cosimo Distante This is my paper · ORCID
classification cs.LG
keywords MAATTransformerAssociationDiscrepancyGatedAttentionMamba-SSMSparseAnomalyDetectionUnsupervisedLearning
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 MAAT, a modification of the Anomaly Transformer for unsupervised time-series anomaly detection. MAAT replaces dense self-attention with block-wise sparse attention and adds a Mamba selective state-space model to the reconstruction module, fusing the two pathways with a learned gate and a skip connection. The claim is that these changes improve both halves of the anomaly score—association discrepancy (the gap between expected and observed temporal dependencies) and reconstruction error—so that anomalies separate more cleanly from normal, noisy, non-stationary patterns. If the claim is right, MAAT is a practical way to upgrade an anomaly detector without labels, reporting F1 gains on all eight benchmarks and precision/recall improvements in most comparisons. A reader should care because time-series anomaly detection underpins industrial monitoring, infrastructure reliability, and environmental sensing.

What carries the argument

The load-bearing mechanism is the MAAT block, which combines three components: Anomaly Sparse Attention, which restricts the series-association softmax to a local window of size block_size; the Mamba-Selective State Space Model, a linear-time sequence model that captures long-range dependencies; and Gated Attention, which computes a scalar gate from the concatenation of the main path and the skip connection and fuses them as $x_{\text{adapt}} = g \odot x_{\text{skip}} + (1-g) \odot x$. Sparse attention does the efficiency work, Mamba contributes global context and denoising, and the gate decides how much of each to trust at every time step. This fusion is what carries the argument: the final anomaly score weights each point by the softmax of the negative association discrepancy and the reconstruction error against $x_{\text{adapt}}$. The ablations confirm that the interaction, not either component alone, yields the reported F1 gains; for example, on SMD sparse attention alone drops F1 from 90.33 to 89.18 while the full MAAT model reaches 92.30.

What would settle it

Run the eight benchmark comparisons without using each dataset's known anomaly ratio—estimate the ratio from data or fix one threshold across all datasets—and check whether MAAT still beats Anomaly Transformer and DCdetector; if the ranking flips on SMD or the NIPS-TS sets, the reported gains are an artifact of oracle thresholding.

Watch

Extended reading notes

Core claim

MAAT's central claim is that anomaly detection improves when the Anomaly Transformer's association discrepancy is computed over sparse local attention and its reconstruction is augmented by a Mamba-SSM through a gated skip connection. The prior-association branch still models expected temporal dependencies with a learnable Gaussian kernel; the series-association branch now computes a block-wise sparse softmax over windows $\Omega_i = \{j : |j-i| \le \text{block\_size}/2\}$, which cuts the attention cost while keeping the dependencies that matter. The reconstruction module blends the sparse-attention output $x$ with a Mamba-transformed residual $x_{\text{mamba}}$ via a gate $g = \sigma(W[x; x_{\text{skip}}] + b)$, yielding $x_{\text{adapt}} = g \odot x_{\text{skip}} + (1-g) \odot x$, and the anomaly score becomes $\text{Softmax}(-\text{AssDis})$ element-wise multiplied by the squared error between the input and this adaptive reconstruction. On the reported benchmarks, MAAT beats both the Anomaly Transformer and DCdetector: SMD F1 rises from 90.33 to 92.30 and from 84.95 to 92.30, and gains appear on MSL, SMAP, SWaT, PSM, NIPS-TS-GECCO, and NIPS-TS-SWAN. The ablation study shows the gated combination matters—sparse attention alone can lower F1 relative to the baseline—while the full MAAT model recovers and exceeds it.

Load-bearing premise

The reported F1 improvements assume the detector knows the true anomaly ratio of each test set to set its detection threshold; if that ratio is unknown or misestimated in deployment, MAAT's advantage over the baselines may not hold.

Editorial extensions

If this is right

  • Unsupervised anomaly detection on multivariate sensor data can improve without labels by changing attention sparsity and reconstruction fusion.
  • Sparse softmax over local windows makes the attention cost scale linearly with sequence length for a fixed block size, making longer time series feasible for association-discrepancy models.
  • The gated Mamba skip path improves anomaly localization, as reflected in higher affiliation and range-based recall on SMD and PSM.
  • The architecture transfers across domains—spacecraft, water treatment, server machines, and solar magnetograms—with consistent F1 gains over both Anomaly Transformer and DCdetector.
  • Reconstruction-loss comparisons indicate the gains come with lower reconstruction error on MSL, SMAP, and SMD.

Reading between the lines

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

  • A threshold-free test would be revealing: the paper sets detection thresholds from each dataset's known anomaly ratio (0.5–1%), so estimating that ratio without labels is the most likely place the reported lead could shrink.
  • The paper describes the gate as switching between Mamba's global context and sparse attention's local focus, but it does not visualize learned gate values; inspecting the gate on sudden versus gradual anomalies would directly test this mechanism.
  • Mamba's linear-time processing suggests real-time or edge deployment, but the paper only discusses training efficiency qualitatively; an inference-latency and memory comparison would make that implication concrete.
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

5 major / 6 minor

Summary. The paper proposes MAAT, a modification of the Anomaly Transformer that replaces dense self-attention with block-wise sparse attention, adds a Mamba state-space block in a skip path, and fuses the two branches with a learned gating mechanism. The anomaly score is the Anomaly Transformer's association-discrepancy score computed with the adaptively reconstructed output. The authors report F1, precision, and recall on five standard benchmarks (SMD, MSL, SMAP, SWaT, PSM) and two NIPS-TS datasets, plus ablation studies, and claim that MAAT significantly outperforms prior methods, especially the Anomaly Transformer and DCdetector.

Significance. If the empirical claims are reproducible, MAAT would be a modest but useful incremental improvement: it combines two established ingredients (sparse attention and a Mamba SSM block) with the Anomaly Transformer's association-discrepancy framework, and the gated fusion of local and long-range pathways is a plausible way to improve reconstruction-based detection. The paper does not offer a new theory or a parameter-free derivation, and the contribution is primarily architectural. The claimed gains are, however, inconsistent across the paper's own result tables, and the evaluation protocol uses ground-truth anomaly ratios to set thresholds, so the significance of the headline result cannot currently be assessed. The paper also omits the Mamba equations, the training loss, and the sparse-attention block size, which prevents independent reimplementation.

major comments (5)
  1. [Tables 1 and 3] The reported F1 scores for the same model and dataset conflict between Table 1 and Table 3. For example, MAAT SMAP is 96.99 in Table 1 but 96.29 in Table 3; MAAT SWaT is 96.50 in Table 1 but 95.93 in Table 3; AnomalyTrans SMD is 90.33 in Table 1 but 87.18 in Table 3; and DCdetector SWaT is 96.42 in Table 1 but 96.33 in Table 3. Under Table 3, MAAT is not the best method on SMAP, SWaT, or MSL, so the abstract's claim that MAAT 'significantly outperforms previous methods' fails under one of the paper's own tables. The authors must reconcile these numbers under a single scoring protocol and clarify which table corresponds to the final evaluation.
  2. [Section 4.2 and Table A.7] The detection threshold is set using the ground-truth Anomaly Ratio for each dataset, as stated in Section 4.2 and listed in Table A.7. This is an oracle threshold that is not available in unsupervised deployment, and it can inflate F1 scores relative to a fully unsupervised threshold-selection procedure. The paper should either justify that this protocol follows the original Anomaly Transformer evaluation, report results with an unsupervised threshold, or explicitly discuss the sensitivity of the headline gains to the chosen ratio. Without this, the claimed improvements over baselines may reflect threshold tuning rather than model quality.
  3. [Section 3.3 and Eq. (9)] The method description is incomplete. Section 3.3 ends with 'The following equations define Mamba SSM:' but no equations are provided. The loss function used for training is never specified. In Eq. (9), the association discrepancy AssDis(P,S;X) is described as 'as defined in Eq. (9)', which is circular; it should be defined explicitly or by reference to the Anomaly Transformer with the sparse-attention modification. The block size in Eq. (3) is not listed in the hyperparameter table, so the actual sparsity pattern is unreproducible from the text. These omissions prevent independent implementation and verification.
  4. [Section 6.1 and Table 5] The ablation study inherits the same numerical inconsistency: Table 5 reports MAAT SMAP F1=96.99, while Table 3 reports 96.29 for the same model and dataset. In addition, the ablation text contains unresolved placeholders, such as 'a recall of Y%' for SMAP and 'Recall percentage points' for SWAN, and the discussion states that SWaT baseline AnomalyTrans has precision 93.59, but Table 5 and Table 1 report 93.59 while Table 3 reports different values. These inconsistencies make it impossible to determine whether the claimed component-wise contributions are real or artifacts of the conflicting evaluations.
  5. [Section 5.1] No error bars, standard deviations, or multiple-seed results are reported. Several headline improvements are very small, such as +0.08% F1 over DCdetector on SWaT in Table 1, and without variance information these gains cannot be distinguished from random seed noise. The paper should include at least three independent runs with means and standard deviations for the core comparison in Table 1 or Table 3.
minor comments (6)
  1. [Throughout] The manuscript contains numerous typos and incomplete sentences, including 'Accuracystness' in the Introduction, 'Implementation' in Section 4.2, 'outpuAtten-tion' in Section 3.4, 'Recallllinity' in Section 5.2, and 'precisionion' and 'Recallllcore' in Section 6.1. A thorough proofreading is needed.
  2. [Section 6.2] The text refers to Figure C.2(a) for both MSL and SMD, and to Figure C.2(f) for SMAP, but the figure panels are labeled (a)-(h); the references should be corrected to the actual panels.
  3. [References] Several references are incomplete: [47], [48], [49], and [50] list only 'A. et al.' or 'T. Z. et al.' without full author lists or paper titles, and [51] misattributes the MatrixProfile reference. The Mamba reference [33] duplicates [26] with an incorrect title.
  4. [Abstract and Introduction] The paper mentions AnomalyBERT and DDMT in the related work but does not compare against them in the experiments; if these are claimed as related state-of-the-art methods, their omission from Table 1 should be justified.
  5. [Eq. (1) vs. Eq. (9)] The anomaly score is defined twice, in Eq. (1) with X_hat and in Eq. (9) with X_adapt; the relationship between these two definitions should be clarified so that the reader knows which one is used in the experiments.
  6. [Appendix D] Appendix D states that fixed random seeds were used but does not report the seed values or the number of runs; reporting them would support reproducibility.

Circularity Check

1 steps flagged · score 1.0 of 10

No significant circularity; one non-load-bearing definitional self-reference in Eq. (9).

  1. self definitional [Section 3.4.2, Eq. (9) and its bullet list]
    "The anomaly score now leverages the adaptively fused reconstruction X adapt by balancing the association discrepancy from Eq. (9) ... AnomalyScore(X) = Softmax(−AssDis(P, S; X)) ⊙ ∥Xi,: − X adapt i,:∥2 2 ... • AssDis(P, S; X) is the association discrepancy, as defined in Eq. (9)."

    Eq. (9) defines AnomalyScore in terms of AssDis(P, S; X), and the bullet immediately below states that AssDis is 'defined in Eq. (9)'. As written, this makes AssDis defined by the very formula that uses it, a circular definition. In practice, AssDis is inherited unchanged from the external Anomaly Transformer [16] rather than newly defined here, so the circularity is a textual slip rather than a load-bearing derivation step; the empirical F1 comparisons rest on external benchmarks and independent scoring, not on this self-reference.

full rationale

The derivation chain is essentially architectural: sparse attention (Eqs. 2-5), the Mamba gated skip block (Eqs. 6-8), and the Anomaly Transformer association-discrepancy criterion (Eqs. 1 and 9) are assembled from prior external components, each with its own stated equations or citations. No parameter is fitted to a subset of the benchmark data and then renamed as a prediction; the reported F1, precision, and recall values are empirical evaluations against public datasets. The only genuinely self-referential passage is the Eq. (9) bullet that defines AssDis as 'defined in Eq. (9)' while Eq. (9) is the anomaly-score formula itself; since AssDis is actually carried over from Anomaly Transformer, this is a definitional oversight, not a load-bearing circularity. The use of the true anomaly ratio to set thresholds (Section 4.2) and the F1 discrepancy between Table 1 and Table 3 are evaluation-protocol and consistency concerns, not derivation-level circularity. Overall, the central claim does not reduce to its own inputs, so no significant circularity is found.

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

The method is an architectural modification of a known model; it introduces no new physical quantities or theoretical constants. All free parameters are training and evaluation hyperparameters, several unspecified.

free parameters (6)
  • Anomaly ratio threshold per dataset = 0.5, 0.85, 0.5, 1.0, 0.5, 0.5, 0.9 (percent) for SMD, MSL, SMAP, PSM, SWaT, GECCO, SWAN
    Used to set the detection threshold; derived from ground-truth anomaly proportions, which can inflate F1.
  • Local window block size for sparse attention = not reported
    Needed in Eq. (3) to define the local window, but Table A.7 omits it; performance may be sensitive to this choice.
  • Window size = 100 or 105 depending on dataset
    Sliding window length, chosen per dataset.
  • Batch size = 32 to 256
    Training hyperparameter, varies by dataset.
  • Model dimension d_model = 512
    Fixed across all datasets.
  • Mamba state and convolution dimensions = not reported
    The Mamba block parameters such as d_state and d_conv are mentioned in Algorithm 1 but never given.
assumptions (3)
  • domain assumption Association discrepancy (prior vs. series associations) is a valid criterion for separating anomalies from normal points.
    Adopted from the Anomaly Transformer (Ref [16]); MAAT does not re-derive or test this criterion.
  • domain assumption The minimax training loss of the Anomaly Transformer is unchanged and sufficient for the new architecture.
    The paper never states the loss used to train MAAT; it assumes the Anomaly Transformer's two-phase loss carries over.
  • ad hoc to paper The local-window sparse attention (Eq. 3) combined with a residual Mamba path preserves enough dependency structure for reconstruction-based detection.
    No analysis is given; the ablation shows sparse attention alone hurts on some datasets (e.g., SWaT), so the gating must compensate.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mamba Adaptive Anomaly Transformer with association discrepancy for time series." pith.science (2026). https://pith.science/paper/OSVSMBXG

@misc{pith2026250207858,
  author       = {Pith},
  title        = {Pith review of: Mamba Adaptive Anomaly Transformer with association discrepancy for time series},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OSVSMBXG}},
  note         = {Machine review of arXiv:2502.07858}
}
read the original abstract

Anomaly detection in time series is essential for industrial monitoring and environmental sensing, yet distinguishing anomalies from complex patterns remains challenging. Existing methods like the Anomaly Transformer and DCdetector have progressed, but they face limitations such as sensitivity to short-term contexts and inefficiency in noisy, non-stationary environments. To overcome these issues, we introduce MAAT, an improved architecture that enhances association discrepancy modeling and reconstruction quality. MAAT features Sparse Attention, efficiently capturing long-range dependencies by focusing on relevant time steps, thereby reducing computational redundancy. Additionally, a Mamba-Selective State Space Model is incorporated into the reconstruction module, utilizing a skip connection and Gated Attention to improve anomaly localization and detection performance. Extensive experiments show that MAAT significantly outperforms previous methods, achieving better anomaly distinguishability and generalization across various time series applications, setting a new standard for unsupervised time series anomaly detection in real-world scenarios.

Figures

Figures reproduced from arXiv: 2502.07858 by the authors.

Figure 1
Figure 1. The figure illustrates the MAAT framework for time series anomaly detection. [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. RAPT: Model-Predictive Out-of-Distribution Detection and Failure Diagnosis for Sim-to-Real Humanoid Deployment

    cs.RO 2026-02 conditional novelty 6.0 of 10

    A simulation-trained recurrent model detects out-of-distribution states on a real humanoid at 50 Hz and uses gradient saliency plus an LLM to diagnose failure causes.

Reference graph

Works this paper leans on

54 extracted references · 40 canonical work pages · cited by 1 Pith paper

  1. [1]

    G. E. P. Box, G. M. Jenkins, Time Series Analysis: Forecasting and Control, Holden-Day, San Francisco, 1970

    work page 1970

  2. [2]

    C. E. Rasmussen, C. K. I. Williams, Gaussian Processes for Machine Learning, MIT Press, Cambridge, MA, 2006

    work page 2006

  3. [3]

    G. E. Box, G. M. Jenkins, G. C. Reinsel, G. M. Ljung, Time series analysis: forecasting and control, John Wiley & Sons, 2015

    2015

  4. [4]

    Scholkopf, J

    B. Scholkopf, J. C. Platt, J. Shawe-Taylor, A. J. Smola, R. C. Williamson, Support vector method for novelty detection, in: Advances in neural information processing systems, 2000, pp. 582–588

    work page 2000

  5. [5]

    Rabiner, B.-H

    L. Rabiner, B.-H. Juang, Introduction to hidden markov models, IEEE ASSP magazine 3 (1) (1986) 4–16

    work page 1986

  6. [6]

    Hochreiter, J

    S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural com- putation 9 (8) (1997) 1735–1780

    1997

  7. [7]

    B. J. Radford, L. M. Apolonio, A. J. Trias, J. A. Simpson, Network traf- fic anomaly detection using recurrent neural networks, arXiv preprint arXiv:1803.10769 (2018). URL https://arxiv.org/abs/1803.10769

    arXiv 2018

  8. [8]

    Malhotra, A

    P. Malhotra, A. Ramakrishnan, G. Anand, L. Vig, P. Agarwal, G. Shroff, Lstm-based encoder-decoder for multi-sensor anomaly detection, in: Proceedings of the 2016 International Conference on Machine Learning and Applications (ICMLA), IEEE, 2016, pp. 409–414

    work page 2016

Show all 54 references
  1. [9]

    H. Xu, Y. Chen, W. Zhao, J. Bu, C. Li, D. Chen, W. Yu, Unsupervised anomaly detection via variational auto-encoder for seasonal kpis in web applications, arXiv preprint arXiv:1802.03903 (2018). URL https://arxiv.org/abs/1802.03903 28

    2018 arXiv

  2. [10]

    Somepalli, Y

    G. Somepalli, Y. Wu, Y. Balaji, B. Vinzamuri, S. Feizi, Unsupervised anomaly detection with adversarial mirrored autoencoders, in: Proceed- ings of the 37th Conference on Uncertainty in Artificial Intelligence, 2021, pp. 1610–1619. URL https://arxiv.org/abs/2003.10713

    2021 arXiv

  3. [11]

    D. P. Kingma, M. Welling, Auto-encoding variational bayes, arXiv preprint arXiv:1312.6114 (2013)

    2013 arXiv

  4. [12]

    G. E. Hinton, R. R. Salakhutdinov, Reducing the dimensionality of data with neural networks, in: Science, Vol. 313, American Association for the Advancement of Science, 2006, pp. 504–507

    2006

  5. [13]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, . Kaiser, I. Polosukhin, Attention is all you need, Advances in neural information processing systems 30 (2017)

    2017

  6. [14]

    Zhang, Z

    K. Zhang, Z. Wang, J. Zhou, C. Wang, Gated attention mechanisms for time series forecasting, IEEE Transactions on Neural Networks and Learning Systems 34 (6) (2023) 1234–1245. doi:10.1109/TNNLS.2023. 10637425

    2023 doi

  7. [15]

    Y. Guo, H. Zhang, X. Liu, Efficient sparse attention for long sequence time-series forecasting, Scientific Reports 14 (2024) 1–10. doi:10.1038/ s41598-024-66886-1

    2024

  8. [16]

    J. Xu, H. Wu, J. Wang, M. Long, Anomaly transformer: Time series anomaly detection with association discrepancy, in: International Con- ference on Learning Representations, 2022. URL https://openreview.net/forum?id=LzQQ89U1qm_

    2022

  9. [17]

    Y. Yang, C. Zhang, T. Zhou, Q. Wen, L. Sun, Dcdetector: Dual atten- tion contrastive representation learning for time series anomaly detec- tion, arXiv preprint arXiv:2306.10347 (2023)

    2023 arXiv

  10. [18]

    F. T. Liu, K. M. Ting, Z.-H. Zhou, Isolation forest, 2008 Eighth IEEE International Conference on Data Mining (2008) 413–422

    2008

  11. [19]

    Distante, L

    C. Distante, L. Fineo, L. Mainetti, L. Manco, B. Taccardi, R. Vergallo, Hf-sca: Hands-free strong customer authentication based on a memory- guided attention mechanisms, Journal of Risk and Financial Manage- ment 15 (8) (2022) 342. 29

    2022

  12. [20]

    Schlegl, P

    T. Schlegl, P. Seeb¨ ock, S. M. Waldstein, G. Langs, U. Schmidt-Erfurth, f-anogan: Fast unsupervised anomaly detection with generative adver- sarial networks, Medical Image Analysis 54 (2019) 30–44. URL https://doi.org/10.1016/j.media.2019.01.010

    2019 doi

  13. [21]

    D. Li, D. Chen, B. Jin, L. Shi, J. Goh, S.-K. Ng, Mad-gan: Multivari- ate anomaly detection for time series data with generative adversarial networks, in: International Conference on Artificial Neural Networks, Springer, 2019, pp. 703–716. URL https://arxiv.org/abs/1901.04997

    2019 arXiv

  14. [22]

    Jeong, E

    Y. Jeong, E. Yang, J. H. Ryu, I. Park, M. Kang, Anomalybert: Self- supervised transformer for time series anomaly detection using data degradation scheme, arXiv preprint arXiv:2305.04468 (2023)

    2023 arXiv

  15. [23]

    C. Yang, T. Wang, X. Yan, Ddmt: Denoising diffusion mask transformer models for multivariate time series anomaly detection, arXiv preprint arXiv:2310.08800 (2023)

    2023 arXiv

  16. [24]

    H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, W. Zhang, In- former: Beyond efficient transformer for long sequence time-series fore- casting, in: Proceedings of the AAAI Conference on Artificial Intelli- gence, Vol. 35, 2021, pp. 11106–11115

    2021

  17. [25]

    Zhang, S

    Y. Zhang, S. S. Rangapuram, Y. Wang, C. Chen, A. Smola, Multi- task time series forecasting with shared attention, arXiv preprint arXiv:2101.09645 (2021)

    2021 arXiv

  18. [27]

    Gupta, G

    A. Gupta, G. Dar, S. Goodman, D. Ciprut, J. Berant, Memory-efficient transformers via top-k attention, in: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 2021, pp. 5796– 5809

    2021

  19. [28]

    G. Lai, W. Chang, Y. Yang, H. Liu, Revisiting deep learning for time series forecasting, in: Proceedings of the AAAI Conference on Artificial Intelligence, 2021. 30

    2021

  20. [29]

    Child, S

    R. Child, S. Gray, A. Radford, I. Sutskever, Generating long sequences with sparse transformers, arXiv preprint arXiv:1904.10509 (2019). URL https://arxiv.org/abs/1904.10509

    2019 arXiv

  21. [30]

    Zaheer, G

    M. Zaheer, G. Guruganesh, A. Dubey, J. Ainslie, C. Alberti, S. On- tanon, P. Pham, A. Ravula, Q. Wang, L. Yang, A. Ahmed, Big bird: Transformers for longer sequences, Advances in Neural Information Pro- cessing Systems (NeurIPS) (2020). URL https://arxiv.org/abs/2007.14062

    2020 arXiv

  22. [31]

    Kitaev, L

    N. Kitaev, L. Kaiser, A. Levskaya, Reformer: The efficient transformer, International Conference on Learning Representations (ICLR) (2020). URL https://arxiv.org/abs/2001.04451

    2020 arXiv

  23. [32]

    A. Gu, T. Dao, A. T. Suresh, C. R´ e, Efficiently modeling long sequences with structured state spaces, in: Advances in Neural Information Pro- cessing Systems (NeurIPS), 2022. URL https://arxiv.org/abs/2111.00396

    2022 arXiv

  24. [33]

    Y. Tay, M. Dehghani, V. Aribandi, H. W. Chung, W. Fedus, C. Raffel, D. Metzler, Mamba: Linear-time sequence modeling with selective state spaces, arXiv preprint arXiv:2312.00752 (2023). URL https://arxiv.org/abs/2312.00752

    2023 arXiv

  25. [34]

    NASA, Mars science laboratory (msl) dataset, available at: https:// github.com/nasa/telemanom

  26. [35]

    NASA, Soil moisture active passive (smap) dataset, available at: https: //github.com/nasa/telemanom

  27. [36]

    Huang, X

    S. Huang, X. Li, W. Hu, S. Liu, W. Peng, X. He, Practical approach to anomaly detection in multivariate time series with missing values, in: Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, ACM, 2021, pp. 2162–2170

    2021

  28. [37]

    Y. Su, Y. Zhao, C. Niu, R. Liu, W. Sun, D. Pei, Robust anomaly de- tection for multivariate time series through stochastic recurrent neu- ral network, in: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, ACM, 2019, pp. 2828–2837. 31

    2019

  29. [38]

    Mathur, N

    A. Mathur, N. O. Tippenhauer, Swat: A water treatment testbed for research and training on ics security, in: 2016 International Workshop on Cyber-physical Systems for Smart Water Networks (CySWater), IEEE, 2016, pp. 31–36

    2016

  30. [39]

    NASA, Space weather hmi active region patches (sharp) dataset, avail- able at: https://ntrs.nasa.gov/citations/20150003032

  31. [40]

    D. J. Miller, A. Nagaraj, R. Gerdes, C. Rieger, Anomaly detection in drinking water quality data from a real-world water distribution system, in: Proceedings of the Genetic and Evolutionary Computation Confer- ence Companion, ACM, 2018, pp. 157–158

    2018

  32. [41]

    Chandola, A

    V. Chandola, A. Banerjee, V. Kumar, Anomaly detection: A survey, ACM Computing Surveys (CSUR) 41 (3) (2009) 15:1–15:58

    2009

  33. [42]

    Hundman, V

    K. Hundman, V. Constantinou, C. Laporte, I. Colwell, T. Soderstrom, Detecting spacecraft anomalies using lstms and nonparametric dynamic thresholding, in: Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2018, pp. 387–395

    2018

  34. [43]

    Blazquez-Garcia, J

    A. Blazquez-Garcia, J. Ruiz, I. Pazos, J. A. Lozano, Multivariate anomaly detection in time series data using causal convolutional net- works, IEEE Access 7 (2019) 130463–130473

    2019

  35. [44]

    S. D. Mathias Perslev, Michael Hejselbak Jensen, P. J. Jennum, U-Time: A Fully Convolutional Network for Time Series Segmentation Applied to Sleep Staging, Advances in Neural Information Processing Systems (NeurIPS) (2019)

    2019

  36. [45]

    B. Z. et al., Deep Autoencoding Gaussian Mixture Model for Unsuper- vised Anomaly Detection, in: Proceedings of the International Confer- ence on Learning Representations (ICLR), 2018

    2018

  37. [46]

    Yairi, N

    T. Yairi, N. Takeishi, T. Oda, Y. Nakajima, N. Nishimura, N. Takata, A data-driven health monitoring method for satellite housekeeping data based on probabilistic clustering and dimensionality reduction, IEEE Transactions on Aerospace and Electronic Systems 53 (6) (2017) 2794–

    2017

  38. [47]

    et al., Contrastive Learning with Mixtures of Probabilistic PCA, Journal Name (Year)

    A. et al., Contrastive Learning with Mixtures of Probabilistic PCA, Journal Name (Year)

  39. [48]

    et al., Time-Series Contrastive Predictive Coding, Journal Name (Year)

    A. et al., Time-Series Contrastive Predictive Coding, Journal Name (Year)

  40. [49]

    L. R. et al., Deep One-Class Classification, in: Proceedings of the Inter- national Conference on Machine Learning (ICML), 2018

    2018

  41. [50]

    T. Z. et al., BeatGAN: Anomalous Rhythm Detection using GANs for Time Series Anomaly Detection, in: Proceedings of the AAAI Confer- ence on Artificial Intelligence (AAAI), 2022

    2022

  42. [51]

    A. V. Benschoten, A. Ouyang, F. Bischoff, T. Marrs, Mpa: a novel cross- language api for time series analysis, Journal of Open Source Software 5 (49) (2020) 2179. doi:10.21105/joss.02179. URL https://doi.org/10.21105/joss.02179

    2020 doi

  43. [52]

    S. B. Taieb, G. Bontempi, A. F. Atiya, A. Sorjamaa, Machine learning strategies for time series forecasting, Lecture Notes in Business Infor- mation Processing 138 (2012) 62–77

    2012

  44. [53]

    A. Huet, J. M. Navarro, D. Rossi, Local evaluation of time series anomaly detection algorithms, in: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2022, pp. 635–645

    2022

  45. [54]

    Paparrizos, P

    J. Paparrizos, P. Boniol, T. Palpanas, R. S. Tsay, A. Elmore, M. J. Franklin, Volume under the surface: a new accuracy evaluation measure for time-series anomaly detection, Proceedings of the VLDB Endowment 15 (11) (2022) 2774–2787. Appendix A. Appendix: Hyperparameters for Mo...

    2022

  46. [2809]

    URL https://ieeexplore.ieee.org/document/7998252 32

    doi:10.1109/TAES.2017.2734500. URL https://ieeexplore.ieee.org/document/7998252 32

    2017

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.