Pith. sign in

REVIEW 3 major objections 3 minor 51 references

LEFT treats time series anomalies as broken agreement among time, frequency, and multi-scale views, and detects them by enforcing cycle consistency.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 03:12 UTC pith:A2BCHRII

load-bearing objection A solid tri-view consistency detector for unsupervised TSAD with strong empirical gains, but the central cross-view premise needs a direct stress test before the SOTA claim is fully convincing. the 3 major comments →

arxiv 2602.08638 v2 pith:A2BCHRII submitted 2026-02-09 cs.LG cs.AI

LEFT: Learnable Fusion of Tri-view Tokens for Unsupervised Time Series Anomaly Detection

Dezheng Wang , Tong Chen , Guansong Pang , Congyan Chen , Shihua Li , Hongzhi Yin This is my paper
classification cs.LG cs.AI
keywords time series anomaly detectionunsupervised learningcross-view consistencycycle consistencymulti-scale representationlearnable filterbanktime-frequency analysisanomaly scoring
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that many time series anomalies are too subtle to appear in any single representation — time, frequency, or a single resolution — and instead show up as inconsistencies between views. To exploit this, it proposes LEFT, a model that builds three token streams from time, frequency, and multi-scale views, and trains them with a bidirectional time-frequency cycle consistency constraint plus a coarse-to-fine multi-scale reconstruction objective. If the central claim is right, cross-view agreement is a usable unsupervised training signal, and enforcing it yields both better detection accuracy and a much lighter model: about 3.5% higher VUS-ROC and 6.9% higher VUS-PR than the strongest baseline, with over 80% fewer FLOPs and about 8x faster training. The framework is one of the first to combine time-frequency and explicit multi-scale structure in a single unsupervised detector.

Core claim

LEFT's core claim is that anomalies can be modeled as violations of agreement among three complementary views of the same series: time-domain tokens capturing local dynamics, frequency-domain tokens from a differentiable short-time Fourier transform capturing periodicity, and multi-scale tokens produced by a learnable Nyquist-constrained filterbank that band-limits and downsamples the signal at several resolutions. The model is trained so that normal windows satisfy (i) bidirectional analysis-synthesis consistency — the predicted spectrum must reconstruct the time signal and the reconstructed signal must re-produce the spectrum — and (ii) cross-path consistency, where the multi-scale reconst

What carries the argument

The central object is the tri-view tokenization plus two consistency objectives. The learnable filterbank parameterizes band edges monotonically with sigmoids and normalizes masks to partition the spectrum, with a proven bound (Lemma 3.1) on aliasing energy after downsampling. The bidirectional cycle operates through a differentiable STFT; Lemma 3.2 shows time and time-frequency reconstruction errors bound each other when the transform has frame bounds, and Lemma 3.3 gives a lower-bound guarantee that if prototype assignments of time and frequency views stay separated on an interval, the JS divergence anomaly evidence stays large. These lemmas turn 'agreement' into a measurable, trainable qu

Load-bearing premise

The load-bearing premise is that normal data maintain agreement across time, frequency, and resolution views and anomalies break that agreement; if normal-but-nonstationary or noisy data also break cross-view agreement, the training signal produces false alarms and the guarantees in Lemma 3.3 no longer separate normal from anomalous windows.

What would settle it

A concrete test: construct a synthetic benchmark where normal series are piecewise nonstationary with occasional legitimate frequency shifts (no anomaly), and measure whether LEFT's cycle and cross-path losses spike on those normal windows. If the anomaly score on normal nonstationary segments is comparable to true anomalies, the consistency premise is falsified. Alternatively, run LEFT on full-channel MSL/SMAP (without discarding discrete channels) and compare VUS-PR to the reported single-channel results; a large drop would falsify the claim that results transfer to full multivariate telemet

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Unsupervised detectors can be trained without annotations by enforcing cross-view and cross-resolution agreement; the consistency signal is self-supervised.
  • Because agreement is checked in both directions (analysis and synthesis), the frequency branch cannot cheat by matching coarse energy statistics; it must be physically consistent, making the model's internal representations more interpretable.
  • The Nyquist-constrained learnable filterbank removes the need for hand-tuned downsampling factors; the model learns dataset-specific band allocations while controlling aliasing, which should transfer to other signal-processing tasks.
  • The efficiency results (lightweight encoders, selected fusion) suggest that cross-view consistency can substitute for large model capacity, which matters for deployment in industrial monitoring.
  • On SMAP the gains are largest (VUS-ROC +13.55%, VUS-PR +19.61%), indicating the benefit scales with the degree to which anomalies are cross-view rather than single-view.

Where Pith is reading between the lines

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

  • If the consistency premise holds generally, LEFT's framework could be adapted to other unsupervised representation-learning tasks where cross-domain agreement (e.g., audio-visual, sensor-modal) is a natural training signal, not just time-frequency.
  • The dependency on the consistency premise suggests a testable boundary: on datasets with heavy nonstationarity or frequent normal regime shifts, the model may over-flag normal changes; one could measure false alarm rate as a function of normal-data spectral change.
  • The theory in Lemma 3.3 provides a lower bound on anomaly evidence given separation of prototype assignments; a practical next step is to use this bound to set detection thresholds rather than relying on SPOT.
  • Since MSL and SMAP are evaluated after discarding discrete channels, the claimed transfer to full multivariate telemetry is untested; a direct test on full-channel MSL/SMAP would clarify scope.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 3 minor

Summary. The paper proposes LEFT, an unsupervised time series anomaly detection method that models anomalies as inconsistencies across three complementary views: time-domain tokens, frequency-domain tokens from a differentiable STFT, and multi-scale structural tokens produced by a Nyquist-constrained learnable filterbank. Training combines multi-scale reconstruction, bidirectional time-frequency cycle consistency, and a cross-path consistency objective; inference uses a weighted combination of cycle-based discrepancy and multi-scale reconstruction residuals. Experiments on seven benchmarks against 19 baselines report consistent VUS-ROC/VUS-PR improvements, with particularly large gains on SMAP, together with a substantial reduction in FLOPs and training time. Ablations identify the cycle-consistency objective as the main contributor. The paper also provides Lemmas 3.1-3.3 with proofs in Appendix A.10.

Significance. If the empirical claims hold, this is a useful contribution: it integrates time, frequency, and multi-scale evidence in a single lightweight unsupervised detector, reports state-of-the-art results on several widely used benchmarks, and includes code. The ablation study is thorough and the efficiency analysis is informative. However, the theoretical framing overstates what the lemmas establish, and the central consistency premise — that normal data maintain cross-view agreement while anomalies violate it — is not directly tested. The paper would be strengthened considerably by evidence that the score separates normal nonstationary windows from anomalous ones, and by a clear hyperparameter-selection protocol.

major comments (3)
  1. [Section 3.1 / Lemma 3.3 (Eqs. 32-35)] The consistency premise — normal data agree across domains/scales, anomalies violate this — is stated but not tested. Lemma 3.3 gives a conditional upper bound when training losses are small and conditional lower bounds when residuals are large; it does not establish a separating threshold for normal nonstationary windows. The large ablation drop for cycle consistency (Table A3, Row 13) could also arise from regularization, not from anomaly-specific disagreement. Add direct evidence: score distributions on normal vs anomalous windows, false-positive rates on regime-shift/trend segments, and/or synthetic injections. The Future Work sentence in Sec. 5 admits the boundary condition is open; this should be closed before claiming the premise.
  2. [Section 4.1 / Table 1] No standard deviations, number of runs, or t-test design are reported; markers ‡/† appear only on LEFT values, and SMD has no marker despite the text claiming improvement on every benchmark (SMD margins are 0.0058 VUS-ROC, 0.0045 VUS-PR). State the pairing scheme, number of seeds/splits, and degrees of freedom, and report mean±std. Without this, the significance claims and the 'every benchmark' statement are not verifiable.
  3. [Section 3.7/3.8 and Fig. 2/A1-A3] The loss weights (λ_ms, λ_cyc, λ_cons) and score weights (α_cyc, α_ms, ...) show strong sensitivity, but the paper never states how these are chosen. If the grid search maximizes VUS on the labeled test set, the unsupervised evaluation is compromised. Specify a fixed hyperparameter policy (e.g., validation-based unsupervised criterion or fixed values) and report the per-dataset values. This matters because the reported gains are not large enough to be insensitive to arbitrary choices.
minor comments (3)
  1. [Appendix A.1 / Table 1] MSL and SMAP are evaluated with only the first continuous channel, so they are effectively univariate in this study. Please disclose this in the main text and, if possible, report results on the full multivariate versions or explicitly discuss the limitation.
  2. [Section 3.2.3] The text mentions 'residual coverage' to justify structural completeness, but Eqs. (5)-(9) do not formalize how unmodeled high-frequency content is covered. Clarify whether a residual term is included in the filterbank or whether the multi-scale pathway is intentionally band-limited.
  3. [Abstract / Introduction / Fig. 2] Minor typos and presentation issues: the abstract says 'multi-scale tokens that learn abnormal patterns' (should be 'learns' or rephrase); the introduction contains 'that can are subtle or supressed'; and Fig. 2's caption repeats '(c)(d)'.

Circularity Check

0 steps flagged

No significant circularity: LEFT's consistency premise and reconstruction/cycle residuals define the method, but no fitted target or self-citation chain drives the reported empirical claims.

full rationale

LEFT's derivation is self-contained in the sense that its central premise—'normal data maintain agreement across domains and resolutions, while anomalies violate it' (Sec. 3.1)—is an assumption, not a conclusion derived from the target labels. The training objective (Eq. 29) and the inference score (Eq. 35/36/37) intentionally share residual terms: Acyc and Ams are moving-averaged reconstruction and cycle discrepancies of exactly the kind minimized by Lcyc, Lms, and Lcons. This is the standard reconstruction/consistency paradigm for unsupervised TSAD and is not circular: no anomaly labels, oracle scores, or external benchmark values enter the training or model-selection loop, and the reported VUS numbers are computed against held-out labeled test windows. Lemma 3.3 is (at most) a consistency/tautological check: it says that if residuals are small the score is small, and if residuals are large on a segment the score is large—which is true by definition of A(t). Its weakness is that the 'training achieves ε' upper bound (Eq. 32) is not automatically valid for arbitrary test windows, and the lower bound assumes the very residual largeness it seeks to detect; that is a rigor/scope deficiency in the theoretical contribution, not circular reasoning. No load-bearing self-citations were found: all cited prior work (e.g., CrossAD [10], [17], [36], prior MSL/SMAP channel-selection work [10,28,33]) is by other authors and is used for motivation, comparison, or data preprocessing. The disclosed MSL/SMAP 'keep only the first continuous channel' choice (Appendix A.1) is a transferability limitation, and the Future Work admission that 'cross-view inconsistency is most informative' needs study is an honest statement of an untested premise; neither constitutes a derivation that reduces to its inputs. Consequently the score is 0.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central empirical claim rests on a named consistency premise about normal versus anomalous data, on frame-bound and normalization conditions used by the lemmas, and on a set of per-dataset hyperparameters. No new physical entities are introduced; the prototype banks and filterbank are model components, not independent postulates.

free parameters (5)
  • lambda_cyc, lambda_cons, lambda_ms = not reported explicitly; grid-tuned via sensitivity analysis
    Chosen per dataset, and sensitivity plots (Figs. 2a-b, A1, A2) show performance depends on the joint choice; no selection protocol is stated.
  • alpha_cyc, alpha_ms, alpha_f, alpha_t, alpha_g, alpha_c = not reported explicitly; grid-tuned
    Scoring-function weights are selected by grid search (Figs. 2e-f, A1, A2); no held-out validation protocol is given.
  • multi-scale kernel sizes / downsampling factors r_k = e.g., SMAP (16, 8, 4)
    Per-dataset hand choice; Appendix A.8 admits the reported settings may not be optimal, so results may be configuration-dependent.
  • moving-average smoothing window kappa = not reported
    Window size for final anomaly-score smoothing; chosen without ablation or justification.
  • number of scales K, STFT window/hop, prototype bank size M, temperature gamma = not reported
    Architecture hyperparameters not ablated (only fusion depth/width are); their values affect the tri-view tokenization and scoring.
axioms (5)
  • domain assumption Normal data maintain agreement across time, frequency, and multi-scale views; anomalies violate this agreement.
    Stated in Sec. 3.1: 'It is based on a consistency premise: normal data maintain agreement across domains and resolutions, while anomalies violate it.' All training and scoring depend on this premise.
  • domain assumption The differentiable STFT W_theta satisfies uniform frame bounds 0 < A <= B < infinity (Eq. 20 in Lemma 3.2).
    Needed for the cycle-consistency error bound; not proven during training for the learned window/basis parameters.
  • domain assumption Learned filter masks are nonnegative and normalized so that 0 <= m_tilde_k <= 1 and they partition the spectrum (Eqs. 7-9).
    Assumed in the proof of Lemma 3.1 and enforced by normalization; if masks became negative or exceeded 1, the aliasing bound would not hold.
  • domain assumption The moving-average smoother condition in Eq. (30) holds for the chosen boundary rule, i.e., a finite rho_kappa exists.
    Used in Lemma 3.3 to pass from raw cycle magnitude to the smoothed score; plausible but not verified for the exact boundary handling.
  • standard math Standard inequalities: Pinsker's inequality and the SmoothL1 residual inequality |r| <= 2*l(r,0)+1.
    Used in the proof of Lemma 3.3 and are standard, correct mathematical facts.

pith-pipeline@v1.3.0-alltime-deepseek · 22068 in / 13915 out tokens · 159213 ms · 2026-08-03T03:12:18.150253+00:00 · methodology

0 comments
read the original abstract

As a fundamental data mining task, unsupervised time series anomaly detection (TSAD) aims to build a model for identifying abnormal timestamps without assuming the availability of annotations. A key challenge in unsupervised TSAD is that many anomalies are too subtle to exhibit detectable deviation in any single view (e.g., time domain), and instead manifest as inconsistencies across multiple views like time, frequency, and a mixture of resolutions. However, most cross-view methods rely on feature or score fusion and do not enforce analysis-synthesis consistency, meaning the frequency branch is not required to reconstruct the time signal through an inverse transform, and vice versa. In this paper, we present Learnable Fusion of Tri-view Tokens (LEFT), a unified unsupervised TSAD framework that models anomalies as inconsistencies across complementary representations. LEFT learns feature tokens from three views of the same input time series: frequency domain tokens that embed periodicity information, time domain tokens that capture local dynamics, and multi-scale tokens that learn abnormal patterns at varying time series granularities. By learning a set of adaptive Nyquist-constrained spectral filters, the original time series is rescaled into multiple resolutions and then encoded, allowing these multi-scale tokens to complement the extracted frequency and time domain information. When generating the fused representation, we introduce a novel objective that reconstructs fine-grained targets from coarser multi-scale structure, and put forward an innovative time-frequency cycle consistency constraint to explicitly regularize cross-view agreement. As cross-view agreement is explicitly regularized during training, LEFT can adopt lightweight tri-view encoders while maintaining effective coordination among the three views.

Figures

Figures reproduced from arXiv: 2602.08638 by Congyan Chen, Dezheng Wang, Guansong Pang, Hongzhi Yin, Shihua Li, Tong Chen.

Figure 1
Figure 1. Figure 1: The architecture of LEFT. Transformer detects anomalies via association inconsistency [41], while DCdetector uses discrepancy-aware contrastive learning [42]. Time–frequency modeling for unsupervised TSAD Be￾cause anomaly evidence can be weak in a single domain, many works incorporate time–frequency information to capture spectral changes such as periodic shifts and band-wise energy redistribution [9]. Com… view at source ↗
Figure 2
Figure 2. Figure 2: Fusion hyper-parameters sensitivity analysis on [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

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

51 extracted references · 1 linked inside Pith

  1. [1]

    Ahmed Abdulaal, Zhuanghua Liu, and Tomer Lancewicki. 2021. Practical Ap- proach to Asynchronous Multivariate Time Series Anomaly Detection and Lo- calization. InProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. 2485–2494

    2021

  2. [2]

    Ali Behrouz, Michele Santacatterina, and Ramin Zabih. 2024. Chimera: Effec- tively modeling multivariate time series with 2-dimensional state space models. Advances in Neural Information Processing Systems37 (2024), 119886–119918

    2024

  3. [3]

    Breunig, Hans-Peter Kriegel, Raymond T

    Markus M. Breunig, Hans-Peter Kriegel, Raymond T. Ng, and Jörg Sander. 2000. LOF: identifying density-based local outliers. InProceedings of the 2000 ACM SIGMOD International Conference on Management of Data. 93–104

    2000

  4. [4]

    David Campos, Tung Kieu, Chenjuan Guo, Feiteng Huang, Kai Zheng, Bin Yang, and Christian S. Jensen. 2021. Unsupervised time series outlier detection with diversity-driven convolutional ensembles.Proceedings of the VLDB15, 3 (Nov. 2021), 611–623

    2021

  5. [5]

    Cheung, Hong-Ning Dai, Byron Choi, and Jiming Liu

    Zhaoliang Chen, Zhihao Wu, William K. Cheung, Hong-Ning Dai, Byron Choi, and Jiming Liu. 2025. MSHTrans: Multi-Scale Hypergraph Transformer with Time-Series Decomposition for Temporal Anomaly Detection. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2. 274–285

    2025

  6. [6]

    Yuchen Fang, Jiandong Xie, Yan Zhao, Lu Chen, Yunjun Gao, and Kai Zheng. 2024. Temporal-frequency masked autoencoders for time series anomaly detection. In 2024 IEEE 40th International Conference on Data Engineering (ICDE). 1228–1241

    2024

  7. [7]

    Markus Goldstein and Andreas Dengel. 2012. Histogram-based outlier score (hbos): A fast unsupervised anomaly detection algorithm.KI-2012: Poster and Demo Track1 (2012), 59–63

    2012

  8. [8]

    Kyle Hundman, Valentino Constantinou, Christopher Laporte, Ian Colwell, and Tom Soderstrom. 2018. Detecting Spacecraft Anomalies Using LSTMs and Non- parametric Dynamic Thresholding. InProceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 387–395

    2018

  9. [9]

    Webb, Irwin King, and Shirui Pan

    Ming Jin, Huan Yee Koh, Qingsong Wen, Daniele Zambon, Cesare Alippi, Ge- offrey I. Webb, Irwin King, and Shirui Pan. 2024. A survey on graph neural networks for time series: Forecasting, classification, imputation, and anomaly detection.IEEE Transactions on Pattern Analysis and Machine Intelligence(2024)

    2024

  10. [10]

    Beibu Li, Qichao Shentu, Yang Shu, Hui Zhang, Ming Li, Ning Jin, Bin Yang, and Chenjuan Guo. 2025. CrossAD: Time Series Anomaly Detection with Cross-scale Associations and Cross-window Modeling. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems

    2025

  11. [11]

    Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. 2008. Isolation Forest. InThe Eighth IEEE International Conference on Data Mining

    2008

  12. [12]

    Qinghua Liu and John Paparrizos. 2024. The elephant in the room: Towards a re- liable time-series anomaly detection benchmark.Advances in Neural Information Processing Systems37 (2024), 108231–108261

    2024

  13. [13]

    Donghao Luo and Xue Wang. 2024. Moderntcn: A modern pure convolution structure for general time series analysis. InThe twelfth International Conference on Learning Representations. 1–43

    2024

  14. [14]

    Jiaming Ma, Binwu Wang, Qihe Huang, Guanjun Wang, Pengkun Wang, Zhengyang Zhou, and Yang Wang. 2025. Mofo: Empowering long-term time series forecasting with periodic pattern modeling. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems

    2025

  15. [15]

    Mathur and Nils Ole Tippenhauer

    Aditya P. Mathur and Nils Ole Tippenhauer. 2016. SWaT: A water treatment testbed for research and training on ICS security. In2016 International Workshop on Cyber-physical Systems for Smart Water Networks (CySWater). 31–36

    2016

  16. [16]

    SHYU Mei-Ling. 2003. A Novel Anomaly Detection Scheme Based on Principal Component Classifier. InProc. of ICDM Foundation and New Direction of Data Mining

    2003

  17. [17]

    Youngeun Nam, Susik Yoon, Yooju Shin, Minyoung Bae, Hwanjun Song, Jae- Gil Lee, and Byung Suk Lee. 2024. Breaking the Time-Frequency Granularity Discrepancy in Time-Series Anomaly Detection. InProceedings of the ACM Web Conference 2024. 4204–4215

    2024

  18. [18]

    Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam

    Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. 2022. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. In The Eleventh International Conference on Learning Representations

    2022

  19. [19]

    John Paparrizos, Paul Boniol, Qinghua Liu, and Themis Palpanas. 2025. Advances in Time-Series Anomaly Detection: Algorithms, Benchmarks, and Evaluation Measures. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2. 6151–6161

    2025

  20. [20]

    Tsay, Aaron Elmore, and Michael J

    John Paparrizos, Paul Boniol, Themis Palpanas, Ruey S. Tsay, Aaron Elmore, and Michael J. Franklin. 2022. Volume under the surface: a new accuracy evaluation measure for time-series anomaly detection.Proceedings of the VLDB15, 11 (July 2022), 2774–2787

    2022

  21. [21]

    Tomáš Pevný. 2016. Loda: Lightweight on-line detector of anomalies.Machine Learning102, 2 (Feb. 2016), 275–304

    2016

  22. [22]

    Jensen, Zhenli Sheng, and Bin Yang

    Xiangfei Qiu, Jilin Hu, Lekui Zhou, Xingjian Wu, Junyang Du, Buang Zhang, Chenjuan Guo, Aoying Zhou, Christian S. Jensen, Zhenli Sheng, and Bin Yang

  23. [23]

    Jensen, and Bin Yang

    Xiangfei Qiu, Zhe Li, Wanghui Qiu, Shiyan Hu, Lekui Zhou, Xingjian Wu, Zhengyu Li, Chenjuan Guo, Aoying Zhou, Zhenli Sheng, Jilin Hu, Christian S. Jensen, and Bin Yang. 2025. TAB: Unified Benchmarking of Time Series Anomaly Detection Methods. http://arxiv.org/abs/2506.18046 arXiv:2506.18046 [cs]

    Pith/arXiv arXiv 2025

  24. [24]

    Mayu Sakurada and Takehisa Yairi. 2014. Anomaly Detection Using Autoencoders with Nonlinear Dimensionality Reduction. InProceedings of the MLSDA 2014 2nd Workshop on Machine Learning for Sensory Data Analysis. 4–11

    2014

  25. [25]

    Sebastian Schmidl, Felix Naumann, and Thorsten Papenbrock. 2024. AutoTSAD: Unsupervised Holistic Anomaly Detection for Time Series Data.Proceedings of the VLDB Endowment17, 11 (July 2024), 2987–3002

    2024

  26. [26]

    Williamson, Alex Smola, John Shawe-Taylor, and John Platt

    Bernhard Schölkopf, Robert C. Williamson, Alex Smola, John Shawe-Taylor, and John Platt. 1999. Support vector method for novelty detection.Advances in Neural Information Processing Systems12 (1999)

    1999

  27. [27]

    Lifeng Shen, Zhongzhong Yu, Qianli Ma, and James T. Kwok. 2021. Time series anomaly detection with multiresolution ensemble decoding. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 35. 9567–9575

    2021

  28. [28]

    Qichao Shentu, Beibu Li, Kai Zhao, Yang Shu, Zhongwen Rao, Lujia Pan, Bin Yang, and Chenjuan Guo. 2024. Towards a General Time Series Anomaly Detector with Adaptive Bottlenecks and Dual Adversarial Decoders. InThe Thirteenth International Conference on Learning Representations

    2024

  29. [29]

    Alban Siffer, Pierre-Alain Fouque, Alexandre Termier, and Christine Largouet

  30. [30]

    Junho Song, Keonwoo Kim, Jeonglyul Oh, and Sungzoon Cho. 2023. Memto: Memory-guided transformer for multivariate time series anomaly detection. Advances in Neural Information Processing Systems36 (2023), 57947–57963

    2023

  31. [31]

    Ya Su, Youjian Zhao, Chenhao Niu, Rong Liu, Wei Sun, and Dan Pei. 2019. Robust Anomaly Detection for Multivariate Time Series through Stochastic Recurrent Neural Network. InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2828–2837

    2019

  32. [32]

    Tax and Robert P.W

    David M.J. Tax and Robert P.W. Duin. 2004. Support Vector Data Description. Machine Learning54, 1 (Jan. 2004), 45–66

    2004

  33. [33]

    Chengsen Wang, Zirui Zhuang, Qi Qi, Jingyu Wang, Xingyu Wang, Haifeng Sun, and Jianxin Liao. 2023. Drift doesn’t matter: Dynamic decomposition with diffusion reconstruction for unstable multivariate time series anomaly detection. Advances in Neural Information Processing Systems36 (2023), 10758–10774

    2023

  34. [34]

    Zhang, and JUN ZHOU

    Shiyu Wang, Haixu Wu, Xiaoming Shi, Tengge Hu, Huakun Luo, Lintao Ma, James Y. Zhang, and JUN ZHOU. 2024. TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting. InThe Twelfth International Conference on Learning Representations

    2024

  35. [35]

    Gaogang Xie

    Zexin Wang, Changhua Pei, Minghua Ma, Xin Wang, Zhihan Li, Dan Pei, Sara- van Rajmohan, Dongmei Zhang, Qingwei Lin, Haiming Zhang, Jianhui Li, and Conference’17, July 2017, Washington, DC, USA Dezheng et al. Gaogang Xie. 2024. Revisiting VAE for Unsupervised Time Series Anomaly Detection: A Frequency Perspective. InProceedings of the ACM Web Conference

    2017

  36. [36]

    Hershey, Kevin Wilson, Jeremy Thorpe, Michael Chinen, Brian Patton, and Rif A

    Scott Wisdom, John R. Hershey, Kevin Wilson, Jeremy Thorpe, Michael Chinen, Brian Patton, and Rif A. Saurous. 2019. Differentiable consistency constraints for improved deep speech enhancement. InICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). 900–904

    2019

  37. [37]

    Haixu Wu, Tengge Hu, Yong Liu, Hang Zhou, Jianmin Wang, and Mingsheng Long. 2022. TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis. InThe Eleventh International Conference on Learning Representations

    2022

  38. [38]

    Xingjian Wu, Xiangfei Qiu, Zhengyu Li, Yihang Wang, Jilin Hu, Chenjuan Guo, Hui Xiong, and Bin Yang. 2024. CATCH: Channel-Aware Multivariate Time Series Anomaly Detection via Frequency Patching. InThe Thirteenth International Conference on Learning Representations

    2024

  39. [39]

    Mingyuan Xia, Chunxu Zhang, Zijian Zhang, Hao Miao, Qidong Liu, Yuanshao Zhu, and Bo Yang. [n. d.]. TimeEmb: A Lightweight Static-Dynamic Disentan- glement Framework for Time Series Forecasting. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems

  40. [40]

    Yongzheng Xie, Hongyu Zhang, and Muhammad Ali Babar. 2025. Multivariate Time Series Anomaly Detection by Capturing Coarse-Grained Intra- and Inter- Variate Dependencies. InProceedings of the ACM on Web Conference 2025. 697– 705

    2025

  41. [41]

    Jiehui Xu, Haixu Wu, Jianmin Wang, and Mingsheng Long. 2021. Anomaly Transformer: Time Series Anomaly Detection with Association Discrepancy. In International Conference on Learning Representations

    2021

  42. [42]

    Yiyuan Yang, Chaoli Zhang, Tian Zhou, Qingsong Wen, and Liang Sun. 2023. DCdetector: Dual Attention Contrastive Representation Learning for Time Se- ries Anomaly Detection. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 3033–3045

    2023

  43. [43]

    Tianyi Yin, Jingwei Wang, Yunlong Ma, Han Wang, Chenze Wang, Yukai Zhao, Min Liu, and Weiming Shen. 2025. Apollo-Forecast: Overcoming Aliasing and Inference Speed Challenges in Language Models for Time Series Forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 22173–22181. Issue: 21

    2025

  44. [44]

    Chuxu Zhang, Dongjin Song, Yuncong Chen, Xinyang Feng, Cristian Lumezanu, Wei Cheng, Jingchao Ni, Bo Zong, Haifeng Chen, and Nitesh V. Chawla. 2019. A deep neural network for unsupervised anomaly detection and diagnosis in multivariate time series data. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 33. 1409–1416

    2019

  45. [45]

    Yahao Zhang, Xiaofeng Zhou, Yichi Zhang, Shuai Li, and Shurui Liu. 2025. Im- proving time series forecasting in frequency domain using a multi resolution dual branch mixer with noise insensitive ArcTanLoss.Scientific Reports15, 1 (2025), 12557

    2025

  46. [46]

    Guojin Zhong, Jin Yuan, Zhiyong Li, and Long Chen. 2025. Multi-resolution decomposable diffusion model for non-stationary time series anomaly detection. InThe Thirteenth International Conference on Learning Representations

    2025

  47. [47]

    Tian Zhou, Ziqing Ma, Qingsong Wen, Xue Wang, Liang Sun, and Rong Jin. 2022. Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting. InInternational Conference on Machine Learning. 27268–27286

    2022

  48. [48]

    Tian Zhou, Peisong Niu, Liang Sun, and Rong Jin. 2023. One fits all: Power general time series analysis by pretrained lm.Advances in Neural Information Processing Systems36 (2023), 43322–43355

    2023

  49. [49]

    Bo Zong, Qi Song, Martin Renqiang Min, Wei Cheng, Cristian Lumezanu, Daeki Cho, and Haifeng Chen. 2018. Deep autoencoding gaussian mixture model for unsupervised anomaly detection. InInternational Conference on Learning Representations. Appendix A.1 Datasets To evaluate LEFT, we conduct experiments on multiple public TSAD benchmarks. We evaluate LEFT on b...

    arXiv 2018

  50. [2017]

    InProceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining

    Anomaly Detection in Streams with Extreme Value Theory. InProceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 1067–1075

  51. [2024]

    TFB: Towards Comprehensive and Fair Benchmarking of Time Series Forecasting Methods.Proceedings of the VLDB Endowment17, 9 (May 2024), 2363–2377

    2024