Pith. sign in

REVIEW 1 major objections 6 minor 70 references

Benchmarking Unsupervised Strategies for Anomaly Detection in Multivariate Time Series

T0 review · 1 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The reconstruction-based iTransformer, which applies self-attention across variates instead of time, is the top unsupervised transformer detector in this benchmark.

desk verdict The pooling comparison is a genuinely useful contribution; the headline iTransformer claim is not established by the reported protocol because the winner is test-tuned while comparators get fixed configurations. read the letter →

arxiv 2506.20574 v1 pith:TQFY3PTR submitted 2025-06-25 cs.LG stat.ME

Laura Boggia , Rafael Teixeira de Lima , Bogdan Malaescu This is my paper
classification cs.LGstat.ME
keywords anomalydetectionmultivariatetimeseriesiTransformerinvertedembeddingreconstruction-basedlabelpoolingMatthewscorrelationcoefficientunsupervisedlearning
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 a simple change to how a transformer reads multivariate time series—inverting the input so self-attention runs across variates instead of across time—carries over from forecasting to unsupervised anomaly detection. The reconstruction-based version of this inverted transformer (iTransformer-reco) attains the highest Matthews correlation coefficient on the most datasets of any model tested, including 0.967 on SMD, 0.964 on SWaT, and 0.935 on MSL. The paper also establishes that the way multidimensional anomaly scores are pooled into a single label must be matched to the anomaly type: short, point-like anomalies are only caught by a per-variate inclusive OR, while longer collective anomalies favor majority voting or a global average. Together the results suggest that no further architectural modification is needed once the embedding is inverted, and that evaluation choices such as pooling rule and window size materially change the ranking.

What carries the argument

The load-bearing mechanism is the inverted embedding (the iTransformer): instead of treating each time step as a token with all variates as features, the input window is transposed so each variate becomes a token and self-attention is computed across variates, with a feed-forward layer over the time dimension. For reconstruction-based anomaly detection, the training loss is the mean-squared error between input and reconstructed window, and the per-variate reconstruction error at test time serves as the anomaly score. The paper's other central object is score-to-label pooling: global averaging of variates, local inclusive OR, or local majority voting, applied after thresholding each variate's anomaly score with the Peak-over-Threshold method, with MCC used to evaluate the resulting binary labels.

What would settle it

Re-run the same ten-dataset benchmark giving TranAD and USAD the same per-dataset search over window size, step size, internal size, and pooling rule, with no cap on variates, and compare MCC scores selected under identical test-label-based rules; if the ranking changes or the margins close, the claimed dominance is an artifact of asymmetric tuning. A second check is to replace the test-derived pooling choice with one fixed a priori by anomaly length and see whether iTransformer-reco still wins.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the inverted embedding introduced for long-horizon forecasting is also the key ingredient for transformer-based anomaly detection. When each input window is transposed so that the attention mechanism operates over the N variates rather than over time, a plain reconstruction model learns multivariate dependencies well enough to separate anomalies from normal behaviour: the reconstruction-based iTransformer scores best on eight of ten datasets, with MCC values of 0.967 (SMD), 0.964 (SWaT), 0.935 (MSL), 0.825 (WADI) and 0.629 (IEEECIS), and it matches the best model on the remaining two within one standard deviation. The paper further finds that the pooling rule is dataset-dependent and linked to anomaly length—local inclusive OR is required to catch point anomalies, while majority voting and global pooling suit longer collective anomalies—and that even tiny fractions of anomalous training data degrade performance, with Huber or Soft-DTW losses offering partial mitigation.

Load-bearing premise

The ranking depends on a protocol in which each model's window size, step size, internal size, and label-pooling rule are picked using the test set's anomaly labels, and in which iTransformer gets per-dataset tuning while TranAD and USAD run with fixed published settings; if that asymmetry is unfair, the iTransformer dominance could be a tuning artefact.

Editorial extensions

If this is right

  • Reconstruction-based iTransformer-reco is the recommended default among the tested models: it wins on the most datasets, handles full-size high-dimensional inputs that other transformers cannot, and smaller internal model sizes often suffice.
  • The pooling rule should be chosen from the anomaly type: use local inclusive OR when point anomalies or single-variate deviations matter, and majority voting or global pooling for long collective anomalies.
  • The inverted embedding, rather than deeper architectural changes, explains most of the performance difference: the vanilla Transformer with the same configuration scores lower on multivariate sets and nearly identical on univariate ones.
  • Forecasting-based anomaly detection can outperform reconstruction on point-anomaly data such as Credit Card, so the two paradigms are complementary rather than redundant.
  • When anomalous instances contaminate training data, performance degrades even at rates below 1.5%, and switching from MSE to Huber or Soft-DTW training loss can partly restore it.

Reading between the lines

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

  • If the comparison were rerun with equally permissive tuning for TranAD and USAD—same per-dataset search over window, step, internal size, and pooling rule, and no variate cap—the headline MCC margins would likely shrink; the paper's protocol gives iTransformer per-dataset configurations while fixing published defaults for the other models.
  • A testable extension is to fix the pooling rule by anomaly length rather than per dataset: the paper's Table 2 suggests a rule such as 'inclusive OR below 100 average anomaly length, majority or global above' that could be validated on new labelled data without test-set selection.
  • Because configuration and pooling selection is made on test labels, the reported absolute MCC values are optimistic upper bounds; the relative ordering across models is the more reliable quantity.
  • The finding that inverted attention transfers to reconstruction suggests trying the same embedding in hybrid detectors, for example using the inverted representation to weight or gate per-variate reconstruction errors, which the paper does not explore.
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

1 major / 6 minor

Summary. The paper investigates transformer-based approaches for unsupervised anomaly detection in multivariate time series, with a focus on the iTransformer architecture applied in both reconstruction-based and forecasting-based modes. It contributes (i) a parameter study of window size, step size, and model dimension for iTransformer; (ii) an analysis of three anomaly-label pooling strategies (global, local inclusive OR, local majority voting) and of evaluation metrics, recommending MCC; (iii) a study of the impact of anomalous training data and alternative loss functions (Huber, Soft-DTW); and (iv) a comparison of iTransformer, a vanilla Transformer, USAD, and TranAD across ten datasets. The central claims are that reconstruction-based iTransformer achieves the best MCC scores on the most datasets and that the inverted embedding transfers from forecasting to anomaly detection, and that pooling must be matched to anomaly type, with inclusive OR needed for point anomalies.

Significance. If the claims hold, the paper would provide useful evidence that the iTransformer's inverted embedding is effective for reconstruction-based anomaly detection, and it would offer practical guidance on label-pooling choices and on loss functions robust to anomalous training data. The paper is commendably transparent in its appendices, reports multiple random seeds with standard deviations, and makes code publicly available, which are strengths that facilitate verification and reproduction. However, as detailed in the major comments, the evaluation protocol as written selects model configurations and pooling rules using test-set anomaly labels, and it compares models under asymmetric tuning budgets; these issues currently prevent the headline performance claims from being accepted as established.

major comments (1)
  1. [Section 5.4, Appendix B, Tables 5 and 6] The model comparison in Section 6.4, Table 4, is asymmetric. iTransformer-reco and iTransformer-fc receive per-dataset tuned configurations (Tables 5 and 6), and the vanilla Transformer inherits the iTransformer-selected configuration, while USAD and TranAD are assigned fixed published configurations; TranAD is further restricted to the first 30 variates, and SWaT is downsampled to one tenth for models other than iTransformer-reco. These differences mean that the ranking in Table 4 could reflect tuning budget rather than architectural merit. The paper should either give all compared models a comparable tuning budget (with a validation-based selection), or explicitly frame the comparison as a deployment-oriented study with per-model practical constraints and soften the architecture-level conclusions accordingly.
minor comments (6)
  1. [Table 4] In the Credit Card block, TranAD's anomaly label method is listed as 'local (incl. voting)', which appears to be a typo for 'local (incl. OR)'.
  2. [Section 4.3, Section 6] The thresholding procedure used to convert continuous anomaly scores into binary labels is not fully specified for the reported experiments. Section 4.3 describes validation-threshold, percentile, and POT methods, and Section 6.4 mentions POT, but it is not stated explicitly which method is applied, per variate or globally, for each pooling strategy and dataset. Please specify this for reproducibility.
  3. [References] Reference [4] (iTransformer) lacks publication venue and year; several other references are also missing venues or years (e.g., [17], [18], [30]). The bibliography would benefit from a consistent style.
  4. [Section 5.4] The tie-breaking rule 'if the highest MCC score across five repetitions, along with its standard deviation, overlaps with the scores from other configurations' is ambiguous; please clarify whether 'overlaps' means interval overlap of mean±std.
  5. [Section 6.1] The statement that 'a smaller internal model size is preferred' is not fully supported by Table 5, where MSL and WADI select M equal to W (96 and 44, respectively). Consider qualifying the claim as a tendency for most datasets.
  6. [Appendix A, Table 1] The description of SWaT states that the updated version is used and that a downsampled tenth-size version is available for models that struggle; Table 1 should clarify which version is used for each model, since Section 6.4 indicates different versions for different models.

Circularity Check

1 steps flagged · score 2.0 of 10

No derivation cycle; the only circularity-adjacent element is that the best iTransformer MCC and the preferred label-pooling rule are selected on the test anomaly labels themselves, making part of the headline result a restatement of the selection criterion.

  1. fitted input called prediction [Appendix B, Tables 5-6; Section 6.4, Table 4]
    "For each parameter configuration and each of the three anomaly labelling methods, the MCC score is computed, and the configuration with the highest MCC score is selected. ... For our case, we use the values provided in Table 1. ... The iTransformer-reco is the model achieving the best scores across the most datasets"

    The 'Best MCC' reported for iTransformer-reco in Table 4 is, by construction, the maximum over up to 12 configurations and 3 pooling rules evaluated on the test anomaly labels (Appendix B), and the window-size grid is derived from test-set anomaly lengths in Table 1. Section 6.4 then compares this selected maximum against TranAD and USAD with one fixed published configuration each. The headline superiority is therefore partly a property of the per-dataset test-label search rather than an out-of-sample architectural prediction. This is transparent, but the announced 'best' number is fitted to the labels it is used to benchmark.

full rationale

The paper contains no formal derivation chain to be circular; it is an empirical benchmark. The only circularity-adjacent element is the use of test anomaly labels to select the iTransformer configuration and label-pooling rule (Appendix B), after which the selected maximum MCC is presented as the model's score and the chosen pooling rule is used to infer that inclusive OR is required for point anomalies (Section 6.2). That makes part of the headline result a restatement of the selection criterion rather than an independent prediction. The comparison against vanilla Transformer, USAD, and TranAD retains independent empirical content, and the protocol is described transparently enough to permit a corrected validation-based rerun. Under the strict definition of circularity (derivation equivalent to inputs by construction), this is a minor/partial issue rather than a full circularity; hence score 2.

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

This is an empirical benchmark, so the ledger records choices rather than postulates. The free-parameter column shows that the headline per-dataset numbers depend on four fitted choices: the (W, S, M) grid point, the forecasting window, the pooling rule, and the unstated thresholding method. The axioms are domain assumptions inherited from the reconstruction/forecasting paradigm plus the ad hoc protocol assumption that test labels may steer configuration selection. No new entities are introduced.

free parameters (4)
  • Per-dataset iTransformer-reco configuration (W, S, M) = e.g., Credit Card W=10 S=1 M=2; SMD W=350 S=35 M=70; see Table 5
    Selected per dataset as the grid point with the highest MCC computed on the test set (Appendix B); W+ and W- ranges use test-set anomaly lengths from Table 1. These choices are fitted to the evaluation data and prop up the headline best MCCs.
  • Per-dataset iTransformer-fc window size W = e.g., Credit Card W=96; GECCO W=12; see Table 6; S=1, M=2 fixed
    Chosen per dataset by the same test-MCC selection rule; a free choice, with S and M fixed for efficiency.
  • Anomaly-label pooling rule = per dataset in Table 2, e.g., local inclusive OR for Credit Card, global for WADI
    The pooling method assigned to each dataset is the one giving the best test MCC; this is a data-fitted choice that materially changes scores, as the paper itself shows.
  • Score thresholding rule for binary labels = not reported per dataset
    Section 4.3.1 presents three threshold strategies (validation-set static, anomaly-fraction percentile, POT); the tables never state which one produced the reported MCCs, leaving a decision-relevant free parameter undocumented.
assumptions (4)
  • domain assumption Anomalous instances are harder to reconstruct or forecast than normal ones, so reconstruction or prediction error is a valid anomaly score.
    Stated in Section 4.2 as the fundamental assumption for reconstruction- and forecasting-based detectors; the paper's own Section 6.3 shows it can fail when anomalies appear in training data.
  • ad hoc to paper Test-set anomaly statistics can be used to guide model configuration.
    Appendix B derives window-size choices W+ and W- from test-set average/minimal anomaly lengths and selects the final configuration by test MCC; this bakes the evaluation labels into the protocol.
  • domain assumption MCC is the appropriate metric for comparing detectors on imbalanced anomaly labels.
    Section 5.2 justifies MCC over accuracy, AUC, and F1 for imbalanced problems; reasonable and cited, but it is a choice that shapes all reported numbers.
  • domain assumption Extreme-value-theory-based POT thresholding (or a validation-labels threshold) is usable without test labels.
    Section 4.3.1 relies on POT and EVT from prior literature; whichever variant was used in the final experiments is not reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Benchmarking Unsupervised Strategies for Anomaly Detection in Multivariate Time Series." pith.science (2026). https://pith.science/paper/TQFY3PTR

@misc{pith2026250620574,
  author       = {Pith},
  title        = {Pith review of: Benchmarking Unsupervised Strategies for Anomaly Detection in Multivariate Time Series},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TQFY3PTR}},
  note         = {Machine review of arXiv:2506.20574}
}
read the original abstract

Anomaly detection in multivariate time series is an important problem across various fields such as healthcare, financial services, manufacturing or physics detector monitoring. Accurately identifying when unexpected errors or faults occur is essential, yet challenging, due to the unknown nature of anomalies and the complex interdependencies between time series dimensions. In this paper, we investigate transformer-based approaches for time series anomaly detection, focusing on the recently proposed iTransformer architecture. Our contributions are fourfold: (i) we explore the application of the iTransformer to time series anomaly detection, and analyse the influence of key parameters such as window size, step size, and model dimensions on performance; (ii) we examine methods for extracting anomaly labels from multidimensional anomaly scores and discuss appropriate evaluation metrics for such labels; (iii) we study the impact of anomalous data present during training and assess the effectiveness of alternative loss functions in mitigating their influence; and (iv) we present a comprehensive comparison of several transformer-based models across a diverse set of datasets for time series anomaly detection.

Figures

Figures reproduced from arXiv: 2506.20574 by the authors.

Figure 1
Figure 1. Taxonomy for anomalies in time series. First, point anomalies describe anomalies that are highly-localised and correspond to one anomalous time stamp. We further distinguish anomalies corresponding to an extreme point of the time series, which are called global outliers, from contextual anomalies, that describe time stamps that only appear anomalous given the context of their neighbouring time stamps. Second, collec… view at source ↗
Figure 2
Figure 2. Sliding window protocol for training and testing. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

70 extracted references · 28 canonical work pages

  1. [1]

    Aggarwal

    Charu C. Aggarwal. Outlier Analysis. Springer International Publishing, 2 edition. ISBN 978-3-319-47578-3. doi: 10.1007/978-3-319-47578-3. URL http://link.springer.com/ 10.1007/978-3-319-47578-3

    doi:10.1007/978-3-319-47578-3

  2. [2]

    Webb, Shirui Pan, Charu Aggarwal, and Mahsa Salehi

    Zahra Zamanzadeh Darban, Geoffrey I. Webb, Shirui Pan, Charu Aggarwal, and Mahsa Salehi. Deep learning for time series anomaly detection: A survey. page 3691338. ISSN 0360-0300, 1557-7341. doi: 10.1145/3691338. URL https://dl.acm.org/doi/10.1145/3691338

    doi:10.1145/3691338

  3. [3]

    ADBench: Anomaly detection benchmark

    Songqiao Han, Xiyang Hu, Hailiang Huang, Minqi Jiang, and Yue Zhao. ADBench: Anomaly detection benchmark. In Neural Information Processing Systems (NeurIPS) . doi: 10.48550/ arXiv.2206.09426

    arXiv doi:10.48550/arxiv.2206.09426

  4. [4]

    iTransformer: Inverted transformers are effective for time series forecasting,

    Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. iTransformer: Inverted transformers are effective for time series forecasting, . URL http://arxiv.org/abs/2310.06625

    arXiv

  5. [5]

    Aggarwal, and Jiawei Han

    Manish Gupta, Jing Gao, Charu C. Aggarwal, and Jiawei Han. Outlier detection for temporal data: A survey. 26(9):2250–2267. ISSN 1041-4347. doi: 10.1109/TKDE.2013.184. URL http://ieeexplore.ieee.org/document/6684530/

    work page arXiv 2013

  6. [6]

    Deep learning for anomaly detection: A survey

    Raghavendra Chalapathy and Sanjay Chawla. Deep learning for anomaly detection: A survey. URL http://arxiv.org/abs/1901.03407

    arXiv 1901

  7. [7]

    Ane Blázquez-García, Angel Conde, Usue Mori, and Jose A. Lozano. A review on out- lier/anomaly detection in time series data. 54(3). ISSN 0360-0300. doi: 10.1145/3444690. URL https://doi.org/10.1145/3444690. Place: New York, NY, USA Publisher: Association for Computing Machinery

    doi:10.1145/3444690

  8. [8]

    Anomaly detection in time series: a comprehensive evaluation

    Sebastian Schmidl, Phillip Wenig, and Thorsten Papenbrock. Anomaly detection in time series: a comprehensive evaluation. 15(9):1779–1797. ISSN 2150-8097. doi: 10.14778/3538598. 3538602. URL https://dl.acm.org/doi/10.14778/3538598.3538602

    arXiv

Show all 70 references
  1. [9]

    Anomaly detection in univariate time-series: A survey on the state-of-the-art

    Mohammad Braei and Sebastian Wagner. Anomaly detection in univariate time-series: A survey on the state-of-the-art. URL http://arxiv.org/abs/2004.00433

    2004 arXiv

  2. [10]

    Tsay, Themis Palpanas, and Michael J

    John Paparrizos, Yuhao Kang, Paul Boniol, Ruey S. Tsay, Themis Palpanas, and Michael J. Franklin. TSB-UAD: an end-to-end benchmark suite for univariate time-series anomaly detection. 15(8):1697–1711. ISSN 2150-8097. doi: 10.14778/3529337.3529354. URL https://dl.acm.org/doi/10....

  3. [11]

    Isolation forest

    Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. Isolation forest. In 2008 Eighth IEEE International Conference on Data Mining , pages 413–422. IEEE, . ISBN 978-0-7695-3502-9. doi: 10.1109/ICDM.2008.17. URL http://ieeexplore.ieee.org/document/4781136/

    2008

  4. [12]

    Support vector method for novelty detection

    Bernhard Schölkopf, Robert C Williamson, Alex Smola, John Shawe-Taylor, and John Platt. Support vector method for novelty detection. In Advances in Neural Information Processing Systems, volume 12. MIT Press. URL https://proceedings.neurips.cc/paper_files/ paper/1999/hash/8725...

    1999

  5. [13]

    Cover and P

    T. Cover and P. Hart. Nearest neighbor pattern classification. 13(1):21–27. doi: 10.1109/TIT. 1967.1053964

    1967

  6. [15]

    Revisiting time series outlier detection: Definitions and benchmarks

    Kwei-Herng Lai, Daochen Zha, Junjie Xu, Yue Zhao, Guanchu Wang, and Xia Hu. Revisiting time series outlier detection: Definitions and benchmarks. URL https://openreview.net/ forum?id=r8IvOsnHchr. 18

  7. [16]

    Anomaly detection: A survey

    Varun Chandola, Arindam Banerjee, and Vipin Kumar. Anomaly detection: A survey. 41(3): 1–58. ISSN 0360-0300, 1557-7341. doi: 10.1145/1541880.1541882. URL https://dl.acm. org/doi/10.1145/1541880.1541882

  8. [17]

    Julien Audibert, Pietro Michiardi, Frédéric Guyard, Sébastien Marti, and Maria A. Zuluaga. USAD: UnSupervised anomaly detection on multivariate time series. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages 3395–3404. A...

  9. [18]

    Jennings

    Shreshth Tuli, Giuliano Casale, and Nicholas R. Jennings. TranAD: deep transformer networks for anomaly detection in multivariate time series data. 15(6):1201–1214. ISSN 2150-8097. doi: 10.14778/3514061.3514067. URL https://dl.acm.org/doi/10.14778/3514061. 3514067

  10. [19]

    Graph neural network-based anomaly detection in multivariate time series

    Ailin Deng and Bryan Hooi. Graph neural network-based anomaly detection in multivariate time series. 35(5):4027–4035. ISSN 2374-3468, 2159-5399. doi: 10.1609/aaai.v35i5.16523. URL https://ojs.aaai.org/index.php/AAAI/article/view/16523

    doi

  11. [20]

    Multivariate time-series anomaly detection via graph attention network

    Hang Zhao, Yujing Wang, Juanyong Duan, Congrui Huang, Defu Cao, Yunhai Tong, Bixiong Xu, Jing Bai, Jie Tong, and Qi Zhang. Multivariate time-series anomaly detection via graph attention network. In 2020 IEEE International Conference on Data Mining (ICDM) , pages 841–850. doi: ...

    2020

  12. [21]

    Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio

    Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. URL http: //arxiv.org/abs/1406.2661

    arXiv

  13. [22]

    Modeles connexionnistes de l’apprentissage

    Yann LeCun and Francoise Soulie Fogelman. Modeles connexionnistes de l’apprentissage. 2. doi: 10.3406/intel.1987.1804

    1987

  14. [23]

    Bourlard and Y

    H. Bourlard and Y . Kamp. Auto-association by multilayer perceptrons and singular value decomposition. 59(4):291–294. ISSN 1432-0770. doi: 10.1007/BF00332918. URL https: //doi.org/10.1007/BF00332918

    doi

  15. [24]

    Developing population codes by minimizing description length

    Richard Zemel and Geoffrey E Hinton. Developing population codes by minimizing description length. In Advances in Neural Information Processing Systems , volume 6. Morgan-Kaufmann. URL https://papers.neurips.cc/paper_files/paper/1993/ hash/b86e8d03fe992d1b0e19656875ee557c-Abst...

    1993

  16. [25]

    Auto-encoding variational bayes

    Diederik P Kingma and Max Welling. Auto-encoding variational bayes. page arXiv:1312.6114. doi: 10.48550/arXiv.1312.6114. _eprint: 1312.6114

    doi arXiv

  17. [26]

    The graph neural network model

    Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. 20(1):61–80. doi: 10.1109/TNN.2008.2005605

    2008

  18. [27]

    McCulloch and Walter Pitts

    Warren S. McCulloch and Walter Pitts. A logical calculus of the ideas immanent in nervous activity. 5(4):115–133. ISSN 1522-9602. doi: 10.1007/BF02478259. URL https://doi. org/10.1007/BF02478259

    doi

  19. [28]

    Rumelhart, Geoffrey E

    David E. Rumelhart, Geoffrey E. Hinton, and Ronald J. Williams. Learning representations by back-propagating errors. 323(6088):533–536. ISSN 1476-4687. doi: 10.1038/323533a0. URL https://www.nature.com/articles/323533a0. Publisher: Nature Publishing Group

    doi

  20. [29]

    Long short-term memory

    Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. 9(8):1735–1780. ISSN 0899-7667. doi: 10.1162/neco.1997.9.8.1735. URL https://doi.org/10.1162/neco. 1997.9.8.1735. Place: Cambridge, MA, USA Publisher: MIT Press

    1997 doi

  21. [30]

    Deep autoencoding gaussian mixture model for unsupervised anomaly detection

    Bo Zong, Qi Song, Martin Renqiang Min, Wei Cheng, Cristian Lumezanu, Daeki Cho, and Haifeng Chen. Deep autoencoding gaussian mixture model for unsupervised anomaly detection. In International Conference on Learning Representations . URL https://openreview.net/ forum?id=BJJLHbb0-. 19

  22. [31]

    Robust anomaly detection for multivariate time series through stochastic recurrent neural network

    Ya Su, Youjian Zhao, Chenhao Niu, Rong Liu, Wei Sun, and Dan Pei. Robust anomaly detection for multivariate time series through stochastic recurrent neural network. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages 2828...

  23. [32]

    Unsupervised online anomaly detection on multivariate sensing time series data for smart manufacturing

    Ruei-Jie Hsieh, Jerry Chou, and Chih-Hsiang Ho. Unsupervised online anomaly detection on multivariate sensing time series data for smart manufacturing. In 2019 IEEE 12th Conference on Service-Oriented Computing and Applications (SOCA) , pages 90–97. IEEE. ISBN 978- 1-7281-5411...

    2019

  24. [33]

    Transformers in time series: A survey

    Qingsong Wen, Tian Zhou, Chaoli Zhang, Weiqi Chen, Ziqing Ma, Junchi Yan, and Liang Sun. Transformers in time series: A survey. URL http://arxiv.org/abs/2202.07125

    arXiv

  25. [34]

    Nielsen, Aakash Tripathi, Shamoon Siddiqui, Ravi P

    Sabeen Ahmed, Ian E. Nielsen, Aakash Tripathi, Shamoon Siddiqui, Ravi P. Ramachandran, and Ghulam Rasool. Transformers in time-series analysis: A tutorial. 42(12):7433–7466, . ISSN 0278-081X, 1531-5878. doi: 10.1007/s00034-023-02454-8. URL https://link.springer. com/10.1007/s0...

    doi

  26. [35]

    Variational transformer- based anomaly detection approach for multivariate time series

    Xixuan Wang, Dechang Pi, Xiangyan Zhang, Hao Liu, and Chang Guo. Variational transformer- based anomaly detection approach for multivariate time series. 191:110791. ISSN 0263-

  27. [36]

    Unsupervised anomaly detection in multivariate time series through transformer-based variational autoencoder

    Hongwei Zhang, Yuanqing Xia, Tijin Yan, and Guiyang Liu. Unsupervised anomaly detection in multivariate time series through transformer-based variational autoencoder. In 2021 33rd Chinese Control and Decision Conference (CCDC) , pages 281–286. doi: 10.1109/CCDC52312. 2021.9601669

    2021

  28. [37]

    Learning graph structures with transformer for multivariate time series anomaly detection in IoT

    Zekai Chen, Dingshuo Chen, Xiao Zhang, Zixuan Yuan, and Xiuzhen Cheng. Learning graph structures with transformer for multivariate time series anomaly detection in IoT. 9 (12):9179–9189. ISSN 2327-4662, 2372-2541. doi: 10.1109/JIOT.2021.3100509. URL http://arxiv.org/abs/2104.03466

  29. [38]

    Anomaly transformer: Time series anomaly detection with association discrepancy

    Jiehui Xu, Haixu Wu, Jianmin Wang, and Mingsheng Long. Anomaly transformer: Time series anomaly detection with association discrepancy. URLhttp://arxiv.org/abs/2110.02642

    arXiv

  30. [39]

    Transformer-based multivariate time series anomaly detection using inter-variable attention mechanism

    Hyeongwon Kang and Pilsung Kang. Transformer-based multivariate time series anomaly detection using inter-variable attention mechanism. 290:111507. ISSN 09507051. doi: 10. 1016/j.knosys.2024.111507. URL https://linkinghub.elsevier.com/retrieve/pii/ S0950705124001424

    2024

  31. [40]

    Laptev, S

    N. Laptev, S. Amizadeh and Y . Billawala. S5 - a labeled anomaly detection dataset, version 1.0 (16m). URL https://webscope.sandbox.yahoo.com/catalog.php?datatype=s& did=70&guccounter=1

  32. [41]

    Unsupervised real-time anomaly detection for streaming data

    Subutai Ahmad, Alexander Lavin, Scott Purdy, and Zuha Agha. Unsupervised real-time anomaly detection for streaming data. 262. doi: 10.1016/j.neucom.2017.04.070

    2017 doi

  33. [42]

    Detecting spacecraft anomalies using LSTMs and nonparametric dynamic thresholding

    Kyle Hundman, Valentino Constantinou, Christopher Laporte, Ian Colwell, and Tom Soderstrom. Detecting spacecraft anomalies using LSTMs and nonparametric dynamic thresholding. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, p...

  34. [43]

    Current time series anomaly detection benchmarks are flawed and are creating the illusion of progress

    Renjie Wu and Eamonn Keogh. Current time series anomaly detection benchmarks are flawed and are creating the illusion of progress. pages 1–1. ISSN 1041-4347, 1558-2191, 2326-3865. doi: 10.1109/TKDE.2021.3112126. URL https://ieeexplore.ieee.org/ document/9537291/. 20

    2021

  35. [44]

    Mathur and Nils Ole Tippenhauer

    Aditya P. Mathur and Nils Ole 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), pages 31–36. IEEE. ISBN 978-1-5090-1161-2. doi: 10.1109/ CySWater...

    2016

  36. [45]

    Chuadhry Mujeeb Ahmed, Venkata Reddy Palleti, and Aditya P. Mathur. WADI: a water distribution testbed for research in the design of secure cyber physical systems. In Proceedings of the 3rd International Workshop on Cyber-Physical Systems for Smart Water Networks , CySW ATER ’...

  37. [46]

    Anomaly detection tutorial

    JulienAu. Anomaly detection tutorial. URL https://github.com/JulienAu/Anomaly_ Detection_Tuto/blob/master/Data/serie2.json

  38. [47]

    URL https://dl.acm.org/doi/10.1145/3055366

    doi: 10.1145/3055366.3055375. URL https://dl.acm.org/doi/10.1145/3055366. 3055375

  39. [48]

    Internet of things: Online anomaly detection for drinking water quality

    Thomas Bartz-Beielstein. Internet of things: Online anomaly detection for drinking water quality. URL http://www.spotseven.de/gecco-challenge/gecco-challenge-2018

    2018

  40. [49]

    Johnson, and Gianluca Bontempi

    Andrea Dal Pozzolo, Olivier Caelen, Reid A. Johnson, and Gianluca Bontempi. Cali- brating probability with undersampling for unbalanced classification. In 2015 IEEE Sym- posium Series on Computational Intelligence , pages 159–166. doi: 10.1109/SSCI.2015

    2015 doi

  41. [50]

    https://www.openml.org/search?type=data&sort=runs&id=1597&status=active

    URL https://ieeexplore.ieee.org/document/7376606/?arnumber=7376606. https://www.openml.org/search?type=data&sort=runs&id=1597&status=active

  42. [51]

    Deep learning for anomaly detection: A review

    Guansong Pang, Chunhua Shen, Longbing Cao, and Anton van den Hengel. Deep learning for anomaly detection: A review. 54(2):1–38. ISSN 0360-0300, 1557-7341. doi: 10.1145/3439950. URL http://arxiv.org/abs/2007.02500

    2007 arXiv

  43. [52]

    IEEE-CIS fraud detection

    Addison Howard, Bernadette Bouchon-Meunier, IEEE CIS, inversion, John Lei, Lynn@Vesta, Marcus2010, and Prof Hussein Abbass. IEEE-CIS fraud detection. URL https://kaggle. com/competitions/ieee-fraud-detection

  44. [53]

    Fraud detection with a single-qubit quantum neural network

    Elena Peña Tapia, Giannicola Scarpa, and Alejandro Pozas-Kerstjens. Fraud detection with a single-qubit quantum neural network. URL http://arxiv.org/abs/2211.13191

    arXiv

  45. [54]

    Do we really need deep learning models for time series forecasting? URL http: //arxiv.org/abs/2101.02118

    Shereen Elsayed, Daniela Thyssens, Ahmed Rashed, Hadi Samer Jomaa, and Lars Schmidt- Thieme. Do we really need deep learning models for time series forecasting? URL http: //arxiv.org/abs/2101.02118

    arXiv

  46. [55]

    Towards a rigorous evaluation of time-series anomaly detection

    Siwon Kim, Kukjin Choi, Hyun-Soo Choi, Byunghan Lee, and Sungroh Yoon. Towards a rigorous evaluation of time-series anomaly detection. URL http://arxiv.org/abs/2109. 05257

  47. [56]

    Are we really making much progress? a worrying analysis of recent neural recommendation approaches

    Maurizio Ferrari Dacrema, Paolo Cremonesi, and Dietmar Jannach. Are we really making much progress? a worrying analysis of recent neural recommendation approaches. In Proceedings of the 13th ACM Conference on Recommender Systems , pages 101–109. ACM. ISBN 978-1- 4503-6243-6. d...

  48. [57]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems , volume 30. Curran Associates, Inc. URL https://proceedings.neurips.cc...

    2017

  49. [58]

    Are transformers effective for time series forecasting? URL http://arxiv.org/abs/2205.13504

    Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. Are transformers effective for time series forecasting? URL http://arxiv.org/abs/2205.13504

    arXiv

  50. [59]

    Long- term forecasting with TiDE: Time-series dense encoder

    Abhimanyu Das, Weihao Kong, Andrew Leach, Shaan Mathur, Rajat Sen, and Rose Yu. Long- term forecasting with TiDE: Time-series dense encoder. URLhttp://arxiv.org/abs/2304. 08424

  51. [60]

    Robust and unsupervised KPI anomaly detection based on conditional variational autoencoder

    Zeyan Li, Wenxiao Chen, and Dan Pei. Robust and unsupervised KPI anomaly detection based on conditional variational autoencoder. In 2018 IEEE 37th International Performance Computing and Communications Conference (IPCCC) , pages 1–9. doi: 10.1109/PCCC.2018. 8710885. URL https:...

    2018

  52. [61]

    Anomaly detection in streams with extreme value theory

    Alban Siffer, Pierre-Alain Fouque, Alexandre Termier, and Christine Largouet. Anomaly detection in streams with extreme value theory. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pages 1067–1075. ACM. ISBN 978-1-4503-4...

  53. [62]

    Jan Beirlant, Yuri Goegebeur, Johan Segers, and Jozef L. Teugels. Statistics of Extremes: Theory and Applications. John Wiley & Sons. ISBN 978-0-471-97647-9. Google-Books-ID: GtIYLAlTcKEC

  54. [63]

    Soft-DTW: a differentiable loss function for time-series

    Marco Cuturi and Mathieu Blondel. Soft-DTW: a differentiable loss function for time-series. URL http://arxiv.org/abs/1703.01541

    arXiv

  55. [64]

    David M. W. Powers. Evaluation: from precision, recall and f-measure to ROC, informedness, markedness and correlation. URL http://arxiv.org/abs/2010.16061

    2010 arXiv

  56. [65]

    The advantages of the matthews correlation coefficient (MCC) over f1 score and accuracy in binary classification evaluation

    Davide Chicco and Giuseppe Jurman. The advantages of the matthews correlation coefficient (MCC) over f1 score and accuracy in binary classification evaluation. 21(1):1–13. doi: 10. 1186/S12864-019-6413-7/TABLES/5. URL https://bmcgenomics.biomedcentral.com/ articles/10.1186/s12...

    doi

  57. [66]

    Dynamic programming algorithm optimization for spoken word recognition

    Hiroaki Sakoe. Dynamic programming algorithm optimization for spoken word recognition. 26:159–165. URL https://api.semanticscholar.org/CorpusID:17900407

  58. [67]

    Peter J. Huber. Robust estimation of a location parameter. 35(1):73–101. ISSN 0003- 4851, 2168-8990. doi: 10.1214/aoms/1177703732. URL https://projecteuclid. org/journals/annals-of-mathematical-statistics/volume-35/issue-1/ Robust-Estimation-of-a-Location-Parameter/10.1214/aom...

  59. [68]

    A dynamic programming approach to continuous speech recognition

    Hiroaki Sakoe and Seibi Chiba. A dynamic programming approach to continuous speech recognition. URL https://api.semanticscholar.org/CorpusID:107516844

  60. [70]

    Metric learning for temporal sequence alignment

    Damien Garreau, Rémi Lajugie, Sylvain Arlot, and Francis Bach. Metric learning for temporal sequence alignment. In Advances in Neural Information Processing Systems , volume 27. Cur- ran Associates, Inc. URL https://papers.nips.cc/paper_files/paper/2014/hash/ 75fa245a86d8a3580...

    2014

  61. [71]

    Fraud dataset benchmark and applications

    Prince Grover, Julia Xu, Justin Tittelfitz, Anqi Cheng, Zheng Li, Jakub Zablocki, Jianbo Liu, and Hao Zhou. Fraud dataset benchmark and applications. URL http://arxiv.org/abs/ 2208.14417. 22

    arXiv

  62. [2241]

    URL https://www.sciencedirect.com/ science/article/pii/S0263224122000914

    doi: 10.1016/j.measurement.2022.110791. URL https://www.sciencedirect.com/ science/article/pii/S0263224122000914

    2022

Pith tools

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