pith. sign in

arxiv: 2604.13924 · v2 · submitted 2026-04-15 · 💻 cs.LG · cs.AI· cs.CV

ASTER: Latent Pseudo-Anomaly Generation for Unsupervised Time-Series Anomaly Detection

Romain Hermary , Samet Hicsonmez , Dan Pineau , Abd El Rahman Shabayek , Djamila Aouada This is my paper

Pith reviewed 2026-05-10 13:09 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.CV
keywords time-series anomaly detectionpseudo-anomaly generationlatent spaceunsupervised learningtransformer classifierlarge language modelsanomaly classification
0
0 comments X

The pith

A latent-space decoder generates pseudo-anomalies to train a Transformer classifier for unsupervised time-series anomaly detection.

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

The paper presents ASTER as a way to perform unsupervised time-series anomaly detection by creating synthetic anomalies directly inside a learned latent representation instead of using reconstruction error or manually designed injections. A decoder tailors these pseudo-anomalies to train a Transformer classifier, while a pre-trained LLM supplies richer temporal context to the same space. This removes the need for domain expertise in anomaly synthesis and addresses the scarcity of real labels. On three standard benchmark datasets the resulting detector reaches state-of-the-art accuracy.

Core claim

ASTER generates tailored pseudo-anomalies in latent space with a decoder and uses them to train a Transformer-based anomaly classifier whose input representations are enriched by a pre-trained LLM, thereby achieving state-of-the-art performance on benchmark time-series datasets without handcrafted anomaly injection or domain-specific knowledge.

What carries the argument

The latent-space decoder that produces task-specific pseudo-anomalies used to supervise the Transformer classifier.

If this is right

  • The method works without any labeled anomalies or domain-tuned distance metrics.
  • LLM-derived contextual features become directly usable inside the latent space for temporal anomaly tasks.
  • Detection performance improves on heterogeneous anomaly patterns that defeat pure reconstruction approaches.
  • The same latent-generation step can be applied to new datasets without redesigning anomaly rules.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The approach may generalize to multivariate or irregularly sampled series if the latent decoder is retrained on the new sampling structure.
  • Replacing the fixed pre-trained LLM with a domain-adapted language model could further lift performance on specialized monitoring tasks.
  • Because pseudo-anomalies are produced on demand, the framework could support continual learning by regenerating anomalies as data distributions drift.

Load-bearing premise

The decoder-generated pseudo-anomalies are representative enough of genuine anomalies that training on them does not introduce systematic bias.

What would settle it

A controlled experiment in which real anomalies drawn from a held-out test set are replaced by the model's generated pseudo-anomalies and the classifier's detection F1 score falls below that of a simple reconstruction baseline would falsify the central claim.

Figures

Figures reproduced from arXiv: 2604.13924 by Abd El Rahman Shabayek, Dan Pineau, Djamila Aouada, Romain Hermary, Samet Hicsonmez.

Figure 1
Figure 1. Figure 1: Overview of the proposed method. It consists of three components: Con [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Detailed description of the Perturbator architecture. It includes a com [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Main differences between previous methods and ASTER; (a) the pseudo [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: PCA visualisation of normal (N), real anomaly (A), and generated pseudo￾anomaly (PA) data. Features are extracted before the clas￾sifier, resulting in an unclear boundary between normal and anomalous data at this stage [PITH_FULL_IMAGE:figures/full_fig_p022_5.png] view at source ↗
read the original abstract

Time-series anomaly detection (TSAD) is critical in domains such as industrial monitoring, healthcare, and cybersecurity, but it remains challenging due to rare and heterogeneous anomalies and the scarcity of labelled data. This scarcity makes unsupervised approaches predominant, yet existing methods often rely on reconstruction or forecasting, which struggle with complex data, or on embedding-based approaches that require domain-specific anomaly synthesis and fixed distance metrics. We propose ASTER, a framework that generates pseudo-anomalies directly in the latent space, avoiding handcrafted anomaly injections and the need for domain expertise. A latent-space decoder produces tailored pseudo-anomalies to train a Transformer-based anomaly classifier, while a pre-trained LLM enriches the temporal and contextual representations of this space. Experiments on three benchmark datasets show that ASTER achieves state-of-the-art performance and sets a new standard for LLM-based TSAD.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes ASTER, a framework for unsupervised time-series anomaly detection that generates pseudo-anomalies directly in the latent space via a decoder to train a Transformer-based classifier, while using a pre-trained LLM to enrich temporal and contextual representations. It avoids handcrafted anomaly injections and domain expertise, claiming state-of-the-art performance on three benchmark datasets and establishing a new standard for LLM-based TSAD.

Significance. If the central assumption holds, ASTER could meaningfully advance unsupervised TSAD by reducing reliance on domain-specific synthesis and fixed metrics, offering a more scalable integration of LLMs for applications like industrial monitoring and healthcare. The latent-space generation approach is a promising direction that, if properly validated, might improve generalization over reconstruction-based or embedding methods.

major comments (2)
  1. [Abstract] The abstract states that the latent-space decoder 'produces tailored pseudo-anomalies' to train the classifier, but provides no mechanism (distribution matching loss, adversarial alignment, or post-hoc statistical test) ensuring these points lie within the support of real anomaly embeddings rather than extrapolating from normal modes. This assumption is load-bearing for the unsupervised training pipeline and the SOTA claim.
  2. [Experiments] The experimental claims of SOTA performance on three benchmarks rest on the pseudo-anomalies being sufficiently representative, yet the abstract (and by extension the evaluation) gives no specifics on baselines, metrics, statistical significance tests, or ablation on the decoder's output distribution. Without this, it is impossible to verify whether improvements reflect genuine anomaly detection or artifacts of the generation process.
minor comments (2)
  1. [Abstract] The abstract lacks any mention of the specific benchmark datasets, evaluation metrics (e.g., F1, AUC), or comparison methods, which hinders immediate assessment of the SOTA claim.
  2. [Method] Notation for the latent space, decoder, and LLM enrichment components could be introduced more formally with equations to improve clarity for readers.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their insightful comments and the opportunity to improve our manuscript. We address each major comment in detail below, providing clarifications and indicating the revisions we will make.

read point-by-point responses

  1. Referee: [Abstract] The abstract states that the latent-space decoder 'produces tailored pseudo-anomalies' to train the classifier, but provides no mechanism (distribution matching loss, adversarial alignment, or post-hoc statistical test) ensuring these points lie within the support of real anomaly embeddings rather than extrapolating from normal modes. This assumption is load-bearing for the unsupervised training pipeline and the SOTA claim.

    Authors: We agree that the abstract does not elaborate on the mechanism. The full paper explains in Section 3 that the decoder generates pseudo-anomalies by applying perturbations in the latent space guided by the LLM representations, trained to increase the anomaly classification loss. This creates points that are representative of anomalies by design, as they are optimized to be classified as anomalous. While we do not employ an explicit distribution matching loss or adversarial alignment, the empirical results and latent space visualizations support that they do not simply extrapolate from normal modes. To address the concern, we will revise the abstract to mention the self-supervised generation objective and include a dedicated paragraph in the methods discussing the assumption, along with additional statistical analysis of the generated embeddings. revision: partial

  2. Referee: [Experiments] The experimental claims of SOTA performance on three benchmarks rest on the pseudo-anomalies being sufficiently representative, yet the abstract (and by extension the evaluation) gives no specifics on baselines, metrics, statistical significance tests, or ablation on the decoder's output distribution. Without this, it is impossible to verify whether improvements reflect genuine anomaly detection or artifacts of the generation process.

    Authors: The experimental section of the manuscript provides details on the three benchmark datasets, the baselines used (reconstruction-based, forecasting-based, and other embedding methods), the evaluation metrics (including F1, AUC, and AUPRC), and reports mean performance with standard deviations. Ablations on the decoder are included, varying the latent perturbation strength and LLM enrichment. However, we acknowledge that the abstract lacks these specifics and that statistical tests could be more prominently featured. We will update the abstract to include a summary of the evaluation protocol and ensure the main text highlights the statistical significance of the results (using t-tests with p-values reported). This revision will make it easier to verify the claims. revision: yes

Circularity Check

0 steps flagged

No circularity: framework is empirically validated without tautological reductions

full rationale

The provided abstract and manuscript description introduce ASTER as a latent-space pseudo-anomaly generation framework using a decoder and pre-trained LLM to train a Transformer classifier, with SOTA claims resting on experiments across three benchmark datasets. No equations, derivations, or mathematical steps are exhibited that could reduce by construction to fitted inputs, self-definitions, or self-citation chains. The central premise (pseudo-anomalies being representative) is presented as an empirical assumption tested via benchmarks rather than derived tautologically from the method itself. No load-bearing uniqueness theorems, ansatzes smuggled via citation, or renaming of known results appear. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no details on parameters, axioms or entities; review limited to summary.

pith-pipeline@v0.9.0 · 5467 in / 1148 out tokens · 34351 ms · 2026-05-10T13:09:39.833425+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

76 extracted references · 76 canonical work pages · 1 internal anchor

  1. [1]

    In: SIGKDD (2021)

    Abdulaal, A., Liu, Z., Lancewicki, T.: Practical approach to asynchronous multi- variate time series anomaly detection and localization. In: SIGKDD (2021)

    work page 2021

  2. [2]

    Alnegheimish, S., Nguyen, L., et al.: Can large language models be anomaly de- tectors for time series? In: DSAA (2024)

    work page 2024

  3. [3]

    TMLR (2024)

    Ansari, A.F., et al.: Chronos: Learning the language of time series. TMLR (2024)

    work page 2024

  4. [4]

    In: SIGKDD (2020)

    Audibert, J., Michiardi, P., et al.: Usad: Unsupervised anomaly detection on mul- tivariate time series. In: SIGKDD (2020)

    work page 2020

  5. [5]

    TPAMI (2013)

    Bengio, Y., Courville, A., Vincent, P.: Representation learning: A review and new perspectives. TPAMI (2013)

    work page 2013

  6. [6]

    CoRRabs/2109.09265(2021)

    Bhatnagar, A., Kassianik, P., et al.: Merlion: A machine learning library for time series. CoRRabs/2109.09265(2021)

    work page arXiv 2021

  7. [7]

    In: VLDB (2022)

    Boniol, P., Paparrizos, J., et al.: Theseus: navigating the labyrinth of time-series anomaly detection. In: VLDB (2022)

    work page 2022

  8. [8]

    In: SIGMOD (2000)

    Breunig, M.M., Kriegel, H.P., et al.: LOF: identifying density-based local outliers. In: SIGMOD (2000)

    work page 2000

  9. [9]

    TIST (2025)

    Chang, C., Wang, W., et al.: LLM4TS: aligning pre-trained llms as data-efficient time-series forecasters. TIST (2025)

    work page 2025

  10. [10]

    IEEE-TSMC-A (2007)

    Chaovalitwongse, W.A., Fan, Y.J., et al.: On the time seriesk-nearest neighbor classification of abnormal brain activity. IEEE-TSMC-A (2007)

    work page 2007

  11. [11]

    In: ICML (2020)

    Chen, T., Kornblith, S., et al.: A simple framework for contrastive learning of visual representations. In: ICML (2020)

    work page 2020

  12. [12]

    In: VLDB (2023)

    Chen, Y., Zhang, C., et al.: Imdiffusion: Imputed diffusion models for multivariate time series anomaly detection. In: VLDB (2023)

    work page 2023

  13. [13]

    In: IEEE-SSCI (2015)

    Dal Pozzolo, A., Caelen, O., et al.: Calibrating probability with undersampling for unbalanced classification. In: IEEE-SSCI (2015)

    work page 2015

  14. [14]

    CoRRabs/2502.08262 (2025)

    Darban, Z.Z., Wang, Q., Webb, G.I., Pan, S., Aggarwal, C.C., Salehi, M.: Ge- nias: Generator for instantiating anomalies in time series. CoRRabs/2502.08262 (2025)

    work page arXiv 2025

  15. [15]

    Pattern Recognit

    Darban, Z.Z., Webb, G.I., et al.: CARLA: Self-supervised contrastive representa- tion learning for time series anomaly detection. Pattern Recognit. (2025)

    work page 2025

  16. [16]

    In: ICML (2024)

    Das, A., Kong, W., et al.: A decoder-only foundation model for time-series fore- casting. In: ICML (2024)

    work page 2024

  17. [17]

    In: NAACL-HLT (2019)

    Devlin, J., Chang, M., et al.: BERT: pre-training of deep bidirectional transformers for language understanding. In: NAACL-HLT (2019)

    work page 2019

  18. [18]

    In: SIGKDD (2021) 14 R

    Feng, C., Tian, P.: Time series anomaly detection for cyber-physical systems via neural system identification and bayesian filtering. In: SIGKDD (2021) 14 R. Hermary et al

    work page 2021

  19. [19]

    Multivariate Industrial Time Series with Cyber-Attack Simulation: Fault Detection Using an LSTM-based Predictive Data Model

    Filonov, P., Lavrentyev, A., Vorontsov, A.: Multivariate industrial time series with cyber-attacksimulation:Faultdetectionusinganlstm-basedpredictivedatamodel. CoRRabs/1612.06676(2016)

    work page Pith review arXiv 2016

  20. [20]

    Technical report, Solenix Engineering GmbH, Darmstadt, Germany (2023)

    Fleith, P.: Controlled anomalies time series (CATS) dataset. Technical report, Solenix Engineering GmbH, Darmstadt, Germany (2023)

    work page 2023

  21. [21]

    In: NIPS (2014)

    Goodfellow, I.J., et al.: Generative adversarial nets. In: NIPS (2014)

    work page 2014

  22. [22]

    In: ICML (2024)

    Goswami,M.,Szafer,K.,etal.:MOMENT:Afamilyofopentime-seriesfoundation models. In: ICML (2024)

    work page 2024

  23. [23]

    Greenwald, S.D., Patil, R.S., Mark, R.G.: Improved detection and classification of arrhythmias in noise-corrupted electrocardiograms using contextual information. CinC pp. 461–464 (1990)

    work page 1990

  24. [24]

    NIPS (2023)

    Gruver, N., Finzi, M., et al.: Large language models are zero-shot time series fore- casters. NIPS (2023)

    work page 2023

  25. [25]

    In: AAAI (2024)

    Gu, Z., Zhang, J., et al.: Rethinking reverse distillation for multi-modal anomaly detection. In: AAAI (2024)

    work page 2024

  26. [26]

    In: WACV

    Hermary, R., et al.: Removing geometric bias in one-class anomaly detection with adaptive feature perturbation. In: WACV. pp. 6612–6622 (2025)

    work page 2025

  27. [27]

    In: ICLR (2022)

    Hu, E.J., Shen, Y., et al.: LoRA: Low-rank adaptation of large language models. In: ICLR (2022)

    work page 2022

  28. [28]

    In: KDD (2018)

    Hundman, K., Constantinou, V., et al.: Detecting spacecraft anomalies using lstms and nonparametric dynamic thresholding. In: KDD (2018)

    work page 2018

  29. [29]

    In: ICASSP (2020)

    Jiang, J., Xia, G., et al.: Transformer VAE: A hierarchical model for structure- aware and interpretable music representation learning. In: ICASSP (2020)

    work page 2020

  30. [30]

    In: ICLR (2024)

    Jin, M., Wang, S., et al.: Time-llm: Time series forecasting by reprogramming large language models. In: ICLR (2024)

    work page 2024

  31. [31]

    TKDE (2023)

    Kim, H., et al.: Contrastive time-series anomaly detection. TKDE (2023)

    work page 2023

  32. [32]

    In: AAAI (2022)

    Kim, S., Choi, K., et al.: Towards a rigorous evaluation of time-series anomaly detection. In: AAAI (2022)

    work page 2022

  33. [33]

    Kingma, D.P., Welling, M.: Auto-encoding variational bayes (2014)

    work page 2014

  34. [34]

    In: TLRW (2025)

    Kowsher, M., Sobuj, M.S.I., et al.: LLM-mixer: Multiscale mixing in LLMs for time series forecasting. In: TLRW (2025)

    work page 2025

  35. [35]

    In: IEEE Big Data (2024)

    Ky, J.R., Mathieu, B., et al.: CATS: contrastive learning for anomaly detection in time series. In: IEEE Big Data (2024)

    work page 2024

  36. [36]

    LeCun, Y., Bengio, Y., Hinton, G.E.: Deep learning. Nat. (2015)

    work page 2015

  37. [37]

    Pattern Recognit

    Lee, D., Malacarne, S., Aune, E.: Explainable time series anomaly detection using masked latent generative modeling. Pattern Recognit. (2024)

    work page 2024

  38. [38]

    In: IJCAI (2024)

    Liu, C., He, S., et al.: Large language model guided knowledge distillation for time series anomaly detection. In: IJCAI (2024)

    work page 2024

  39. [39]

    In: SDM (2008)

    Liu, F.T., Ting, K.M., et al.: Isolation forest. In: SDM (2008)

    work page 2008

  40. [40]

    In: SIGKDD (2025)

    Liu, J., Zhang, C., et al.: Large language models can deliver accurate and inter- pretable time series anomaly detection. In: SIGKDD (2025)

    work page 2025

  41. [41]

    In: AAAI (2025)

    Liu, P., Guo, H., et al.: Calf: Aligning llms for time series forecasting via cross- modal fine-tuning. In: AAAI (2025)

    work page 2025

  42. [42]

    In: ACM Web Conference (2024)

    Liu, X., Hu, J., et al.: Unitime: A language-empowered unified model for cross- domain time series forecasting. In: ACM Web Conference (2024)

    work page 2024

  43. [43]

    In: ICML (2024)

    Liu, Y., Zhang, H., et al.: Timer: generative pre-trained transformers are large time series models. In: ICML (2024)

    work page 2024

  44. [44]

    In: ICLR (2025)

    Liu, Z., Wang, Y., et al.: KAN: kolmogorov-arnold networks. In: ICLR (2025)

    work page 2025

  45. [45]

    In: AAAI (2022) ASTER: Time-Series Anomaly Detection 15

    Lu, K., Grover, A., et al.: Frozen pretrained transformers as universal computation engines. In: AAAI (2022) ASTER: Time-Series Anomaly Detection 15

    work page 2022

  46. [46]

    In: CySWater (2016)

    Mathur, A.P., Tippenhauer, N.O.: Swat: A water treatment testbed for research and training on ics security. In: CySWater (2016)

    work page 2016

  47. [47]

    In: ICLR

    Nie, Y., Nguyen, N.H., Sinthong, P., Kalagnanam, J.: A time series is worth 64 words: Long-term forecasting with transformers. In: ICLR. OpenReview.net (2023)

    work page 2023

  48. [48]

    Representation Learning with Contrastive Predictive Coding

    van den Oord, A., Li, Y., Vinyals, O.: Representation learning with contrastive predictive coding. CoRRabs/1807.03748(2018)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  49. [49]

    In: VLDB (2022)

    Paparrizos, J., Boniol, P., et al.: Volume under the surface: a new accuracy evalu- ation measure for time-series anomaly detection. In: VLDB (2022)

    work page 2022

  50. [50]

    Paszke, A., Gross, S., et al.: Automatic differentiation in pytorch (2017)

    work page 2017

  51. [51]

    In: VLDB (2025)

    Qiu, X., Li, Z., et al.: TAB: Unified benchmarking of time series anomaly detection methods. In: VLDB (2025)

    work page 2025

  52. [52]

    Radford, A., Wu, J., et al.: Language models are unsupervised multitask learners. Tech. rep. (2019)

    work page 2019

  53. [53]

    CoRRabs/2507.23615(2025)

    Roque, L., Soares, C., et al.: L-GTA: latent generative modeling for time series augmentation. CoRRabs/2507.23615(2025)

    work page arXiv 2025

  54. [54]

    In: ICML (2018)

    Ruff, L., Vandermeulen, R., et al.: Deep one-class classification. In: ICML (2018)

    work page 2018

  55. [55]

    In: ICCV (2025)

    Rukhovich, D., Dupont, E., et al.: Cad-recode: Reverse engineering cad code from point clouds. In: ICCV (2025)

    work page 2025

  56. [56]

    In: WSDM (2024)

    Sui, Y., Zhou, M., et al.: Table meets LLM: can large language models understand structured table data? A benchmark and empirical study. In: WSDM (2024)

    work page 2024

  57. [57]

    In: ICLR (2024)

    Sun, C., Li, H., Li, Y., Hong, S.: TEST: Text prototype aligned embedding to activate LLM’s ability for time series. In: ICLR (2024)

    work page 2024

  58. [58]

    Tan, M., Merrill, M., et al.: Are language models actually useful for time series forecasting? NeurIPS (2024)

    work page 2024

  59. [59]

    In: ICME (2025)

    Tao, W., Qu, X., et al.: Madllm: Multivariate anomaly detection via pre-trained llms. In: ICME (2025)

    work page 2025

  60. [60]

    In: VLDB (2022)

    Tuli,S.,Casale,G.,Jennings,N.R.:Tranad:deeptransformernetworksforanomaly detection in multivariate time series data. In: VLDB (2022)

    work page 2022

  61. [61]

    NIPS (2017)

    Vaswani, A., Shazeer, N., et al.: Attention is all you need. NIPS (2017)

    work page 2017

  62. [62]

    Wen, Q., Zhou, T., et al.: Transformers in time series: A survey (2023)

    work page 2023

  63. [63]

    In: ICLR (2023)

    Wu, H., Hu, T., et al.: Timesnet: Temporal 2d-variation modeling for general time series analysis. In: ICLR (2023)

    work page 2023

  64. [64]

    IEEE Trans

    Wu, R., Keogh, E.J.: Current time series anomaly detection benchmarks are flawed and are creating the illusion of progress. IEEE Trans. Knowl. Data Eng.35(3), 2421–2429 (2023)

    work page 2023

  65. [65]

    In: SIGKDD (2023)

    Xiao, C., Gou, Z., et al.: Imputation-based time-series anomaly detection with conditional weight-incremental diffusion models. In: SIGKDD (2023)

    work page 2023

  66. [66]

    In: VLDB (2025)

    Xie, Z., Li, Z., et al.: Chatts: Aligning time series with llms via synthetic data for enhanced understanding and reasoning. In: VLDB (2025)

    work page 2025

  67. [67]

    Pattern Recognit

    Yao, Y., Ma, J., et al.: Regularizing autoencoders with wavelet transform for se- quence anomaly detection. Pattern Recognit. (2023)

    work page 2023

  68. [68]

    In: AAAI (2022)

    Yue, Z., Jin, H., et al.: TS2Vec: Towards universal representation of time series. In: AAAI (2022)

    work page 2022

  69. [69]

    In: ISCIT

    Zare Moayedi, H., Masnadi-Shirazi, M.: Arima model for network traffic prediction and anomaly detection. In: ISCIT. vol. 4, pp. 1–6 (2008)

    work page 2008

  70. [70]

    In: ICDM (2018)

    Zenati, H., et al.: Adversarially learned anomaly detection. In: ICDM (2018)

    work page 2018

  71. [71]

    In: VLDB (2023)

    Zhang, A., Deng, S., et al.: An experimental evaluation of anomaly detection in time series. In: VLDB (2023)

    work page 2023

  72. [72]

    In: IJCAI (2019) 16 R

    Zhou, B., Liu, S., et al.: BeatGAN: Anomalous rhythm detection using adversari- ally generated time series. In: IJCAI (2019) 16 R. Hermary et al

    work page 2019

  73. [73]

    NeurIPS (2023)

    Zhou, T., Niu, P., et al.: One fits all: Power general time series analysis by pre- trained lm. NeurIPS (2023)

    work page 2023

  74. [74]

    Zhou, Z., Yu, R.: Can LLMs understand time series anomalies? In: ICLR (2025)

    work page 2025

  75. [75]

    In: ICLR (2018)

    Zong, B., Song, Q., Min, M.R., Cheng, W., et al.: Deep autoencoding gaussian mixture model for unsupervised anomaly detection. In: ICLR (2018)

    work page 2018

  76. [76]

    d666a9796412c17619e599308cdff 9edfa619e6a

    Zou, K., Yu, B., Seo, H.: 3d facial expression generator based on transformer VAE. In: ICIP (2023) ASTER: Latent Pseudo-Anomaly Generation for Unsupervised Time-Series Anomaly Detection –Supplementary– A Complementary Figure: Positioning Multivariate Time-Series Learned Transformations in Latent Space Trained Binary Classifier −1 0 1 Original Amplify Bias...

    work page 2023