Pith. sign in

REVIEW 2 major objections 7 minor 37 references

Multivariate time series anomalies that break cross-channel coordination can be caught by checking whether attention query vectors stay predictable from recent history.

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 · grok-4.5

2026-07-14 22:00 UTC pith:3QO3L6Q4

load-bearing objection Solid dual-score detector: query prediction plus reconstruction beats strong baselines on TSB-AD and vehicle telemetry; premise is plausible and ablated, not fully proven as coordination-specific. the 2 major comments →

arxiv 2603.12916 v3 pith:3QO3L6Q4 submitted 2026-03-13 cs.LG cs.AI

Surprised by Attention: Predictable Query Dynamics for Time Series Anomaly Detection

Kadir-Kaan \"Ozer , Ren\'e Ebeling , Markus Enzweiler This is my paper
classification cs.LG cs.AI
keywords time series anomaly detectionmulti-head attentionquery predictabilitymultivariate telemetryunsupervised detectionreconstruction residualEMA target encodercoordination breaks
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.

Many real faults in multi-sensor systems do not push any single channel out of range; they break the usual coupling between channels, for example a steering command that no longer matches lateral acceleration. Flexible residual detectors can still reconstruct each channel and therefore miss those coordination breaks. This paper introduces AxonAD, which treats multi-head attention queries as a short-horizon predictable process: a history-only predictor forecasts future query vectors against an exponential-moving-average target encoder, and the inference score adds a tail-aggregated cosine query mismatch to ordinary reconstruction error. On proprietary vehicle telemetry and the TSB-AD multivariate suite of 180 series, the dual score improves ranking quality and temporal localization over strong residual and transformer baselines. Ablations show that predicting queries (not keys, values, or attention maps) and combining both scores are the main sources of the gains.

Core claim

Under stationary nominal dynamics, multi-head attention query vectors evolve predictably over short horizons. A history-only predictor trained with a masked cosine objective against an EMA target encoder therefore yields a query-mismatch signal that is complementary to reconstruction residual. Robustly standardizing and summing the two scores detects both amplitude excursions and structural dependency shifts, improving unsupervised ranking and range-aware localization on in-vehicle telemetry and the TSB-AD multivariate benchmark.

What carries the argument

Tail-aggregated query mismatch: the average cosine distance, over recent valid timesteps and heads, between multi-head queries forecast by a causal history-only temporal network and the corresponding EMA target queries; this is the same cosine used in the masked training loss and is added after robust standardization to reconstruction MSE.

Load-bearing premise

The method rests on the premise that, under normal operation, the attention query vectors that control routing evolve in a short-horizon predictable way, so a large cosine gap between predicted and target queries reliably flags a coordination break rather than noise or model mismatch.

What would settle it

On a held-out multivariate series whose annotated anomalies are confirmed coordination breaks (not mere amplitude spikes), if the query-mismatch component alone fails to raise AUC-PR or Range-F1 above a pure reconstruction baseline after identical robust standardization, the claim that query predictability diagnoses structural dependency shifts would be falsified.

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

If this is right

  • Residual-only detectors systematically under-detect coordination-type faults in multi-channel systems such as vehicle telemetry.
  • Scoring the predictability of attention queries, rather than attention weights or value residuals, is a usable unsupervised anomaly signal.
  • Dual scoring with robust standardization transfers across heterogeneous multivariate datasets without label-based threshold calibration.
  • Short forecast horizons and moderate EMA momentum are preferred for stable query prediction; longer horizons or extreme target lag degrade ranking.
  • Per-window scoring latency low enough for streaming supports integration into fleet monitoring pipelines.

Where Pith is reading between the lines

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

  • Query-mismatch heatmaps over heads and recent timesteps could localize which subsystem decoupled without extra labels.
  • Nonstationary mode switches (load, weather, operating regime) may require adaptive EMA or multi-mode predictors before continuous vehicle use is safe.
  • The same short-horizon query-predictability idea could be tried on other attention-based fusion models outside pure time-series detectors.
  • Pairing query mismatch with explicit sensor-graph relation models may further separate structural breaks from pure amplitude spikes.

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

2 major / 7 minor

Summary. AxonAD is an unsupervised multivariate time-series anomaly detector that couples a bidirectional self-attention reconstruction pathway with a history-only causal predictor of multi-head query vectors, trained by masked cosine loss against an EMA target encoder (JEPA-style). At inference, robustly standardized reconstruction MSE (d_rec) is added to a tail-aggregated cosine query mismatch (d_q) to form the anomaly score S. The central claim is that short-horizon query predictability under nominal dynamics yields a signal complementary to residuals, improving sensitivity to cross-channel coordination breaks while retaining amplitude detection. Empirically, the method reports large gains on proprietary in-vehicle telemetry (interval annotations) and leads on threshold-free ranking and range-aware localization on the TSB-AD multivariate suite (17 datasets, 180 series), with ablations attributing gains primarily to query prediction and combined scoring; code is released.

Significance. If the results hold, the paper offers a practically useful and conceptually clean extension of residual detectors: scoring predictability of attention queries rather than attention weights or value residuals, with training–inference metric consistency and label-free robust score fusion. Strengths include chronological unsupervised splits, hyperparameters fixed from the official TSB-AD tuning set and transferred without telemetry labels, multi-seed means±std, paired Wilcoxon tests, range-aware metrics (VUS, Event/Range/Affiliation-F1), extensive ablations (scoring, EMA, masking, capacity, horizon, prediction target), mechanistic diagnostics, runtime profiling, and public code. The dual-score geometry and the finding that predicting queries outperforms keys/values/maps/hidden states are of interest beyond the specific application. The proprietary telemetry setting (coordination-type faults) is a relevant industrial use case for fleet monitoring.

major comments (2)
  1. [§1, §3.2–3.4, §5–6, Tables 1–3] The load-bearing premise—that elevated d_q selectively flags structural dependency/coordination shifts when d_rec does not (Introduction; Fig. 1; §3.2–3.4; Eq. 4)—is only weakly established as the driver of gains. Ablations (Table 3) show that removing d_q, replacing cosine with MSE, or predicting non-query targets hurts aggregate ranking/localization, and §6 reports moderate Spearman ρ≈0.677 between ||ΔQ|| and tail KL, small median ρ≈0.211 between d_rec and d_q, and ~19% high-d_q/low-d_rec anomalous windows. These support complementarity but not that d_q is caused by coordination breaks rather than generic nonstationarity, representation drift, or EMA lag. TSB-AD results (Table 2, Fig. 5) are not stratified by anomaly type; the proprietary stream (Table 1), where coordination breaks are described as dominant, is inaccessible. Please add either (i) type-stratified or synthetic coordinati
  2. [§3.3–3.4, Eq. (2), Eq. (4), Fig. 2] Inference-time definition of the EMA target pathway is underspecified relative to the training description. §3.3 defines θ_tgt via online EMA updates during training; §3.4 and Eq. (4) use q_tgt at inference without stating whether θ_tgt (and W_q^tgt) are frozen after training or continue to track, and whether the target encoder is applied to the full current window (as written) while the predictor remains history-only. If the target is a frozen nominal encoder applied to anomalous windows, that should be stated explicitly; if it continues to adapt, the mismatch signal’s interpretation changes. Please specify the exact inference graph (frozen vs updating, stop-grad, which parameters) and confirm that no test labels or future windows enter θ_tgt.
minor comments (7)
  1. [§4.1] Clarify that offline TSB-AD score assignment to the window center (lookahead ⌊(T−1)/2⌋) is evaluation-only and that real-time deployment uses endpoint assignment (§4.1); a short note in the caption of any timeline figure would help readers avoid conflating the two.
  2. [Fig. 2, §3.4, Table 3] Fig. 2 legend mentions “Attention divergence (KL tail) is not included in the default scoring pipeline,” which is good; ensure the main text never implies KL is part of S (Eq. 6), and that Table 3’s “Score MSE+JEPA KL” row is clearly diagnostic-only.
  3. [§3.2] Notation: eH vs Ĥ / bQpred vs Q̂pred is inconsistent in the text vs equations (e.g., §3.2); unify hats and tildes for predicted quantities.
  4. [Table 1, §5] Table 1: AxonAD ties LSTMAD on PA-F1 and is not best on Affiliation-F1; a one-sentence discussion of when point-adjusted vs range/affiliation metrics diverge would strengthen interpretation of “temporal localization.”
  5. [§2.3] Related work could briefly contrast with association-discrepancy scoring in Anomaly Transformer [34] beyond the one-sentence distinction in §2.3, since both use attention-derived signals.
  6. [§3.4, Eq. (5)] Report whether robust median/IQR statistics for rz(·) (Eq. 5) are computed once on the training set and frozen, or recomputed per series; this affects reproducibility of S across the 180 TSB-AD series.
  7. [Header / arXiv line] Minor: arXiv version string and “v3 28 May 2026” in the header look like a future date typo; fix for archival consistency.

Circularity Check

0 steps flagged

No circularity: training-inference metric consistency is intentional design, not a tautological reduction; gains are empirical on held-out labeled benchmarks.

full rationale

AxonAD's derivation chain is architectural and empirical, not definitional. The reconstruction pathway (Eq. 1) and history-only query predictor trained by masked cosine against EMA targets (Eqs. 2-3) are standard self-supervised components (JEPA-style + EMA). At inference the same cosine is reused as tail-aggregated dq (Eq. 4) and additively combined after robust standardization (Eqs. 5-6). This is explicitly framed as 'training and inference consistency' so the predictor is optimized for the deployed score; it is not a self-definitional loop because parameters and rz statistics are fit exclusively on nominal training windows, while ranking/localization metrics are computed on held-out series with external interval labels (proprietary telemetry + TSB-AD 180 series). Hyperparameters are selected once on the official TSB-AD tuning subset and frozen; no labels enter fitting, thresholding or early stopping. Ablations (Table 3) and paired Wilcoxon tests demonstrate that removing dq, switching distance, or predicting alternative targets degrades metrics, confirming the components contribute non-trivially rather than by construction. No self-citation supplies a uniqueness theorem, no fitted parameter is renamed a prediction of a closely related quantity, and no known result is merely re-labeled. The paper is therefore self-contained against external benchmarks; the reader's noted consistency is good engineering, not circularity.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 2 invented entities

The central claim is empirical: a dual reconstruction + query-predictability score improves unsupervised multivariate AD. It rests on standard attention/EMA/JEPA machinery plus the domain premise that short-horizon query dynamics track coordination under nominal stationarity. Free parameters are ordinary architecture and training choices fixed once on a tuning set; no new physical constants. Invented entities are the detector components themselves, with independent evidence only via public-benchmark metrics and ablations, not external physical measurements.

free parameters (7)
  • window length T = 100
    Fixed at 100 for all datasets; controls temporal context and latency.
  • embedding dim D and number of heads Nh = D=128, Nh=8
    Capacity knobs (D=128, Nh=8 default); ablated at D=64 and Nh=4.
  • forecast horizon s = 1
    History-only shift depth; default s=1, sensitivity shows larger s hurts ranking.
  • tail aggregation length k = 10
    Number of recent timesteps for dq; default 10, Range-F1 peaks there.
  • EMA momentum m = 0.9
    Target encoder update rate; default 0.9, extremes 0/0.99/0.999 degrade AUC-PR.
  • mask ratio and block fraction for JEPA patches = 0.5 / 0.5
    Training supervision density (0.5 / 0.5); mask 0.8 hurts Event-F1.
  • learning rate, batch size, weight decay, patience = 5e-4, 128, 1e-5, patience 3
    Optimization hyperparameters fixed after TSB-AD tuning (e.g. 5e-4, 128, 1e-5, patience 3).
axioms (5)
  • domain assumption Standard multi-head self-attention and residual reconstruction can model nominal multivariate windows.
    Used throughout §3.1 as the reconstruction pathway; inherited from transformer AD literature.
  • ad hoc to paper Under stationary nominal dynamics, attention query vectors evolve predictably over short horizons.
    Core modeling premise in Introduction and Fig. 1; not proven, motivated by coordination-break examples.
  • domain assumption EMA target encoders provide stable stop-gradient targets for predictive self-supervision (BYOL/JEPA style).
    Invoked in §3.3 with citations to Grill et al. and Assran et al.; standard SSL practice.
  • domain assumption Robust median/IQR standardization of drec and dq on nominal train windows yields comparable additive scores across datasets without labels.
    Eqs. (5)–(6); common robust scoring practice, assumed sufficient for combination.
  • domain assumption Training only on anomaly-free prefixes and evaluating with official TSB-AD metrics is a valid unsupervised protocol.
    §4 protocol; standard for the benchmark but depends on label quality and chronological purity.
invented entities (2)
  • AxonAD dual pathway (reconstruction + history-only query predictor with EMA target) independent evidence
    purpose: Produce complementary amplitude and structural anomaly scores from one model.
    Named architecture and training scheme introduced in this paper; not a prior named detector.
  • Tail-aggregated query mismatch score dq (cosine) independent evidence
    purpose: Inference signal for coordination shifts via predicted vs target queries on recent timesteps.
    Defined in Eq. (4); evidence is empirical correlation with attention KL and complementary coverage of anomalies.

pith-pipeline@v1.1.0-grok45 · 29005 in / 3541 out tokens · 30828 ms · 2026-07-14T22:00:43.015039+00:00 · methodology

0 comments
read the original abstract

Multivariate time series anomalies often manifest as shifts in cross-channel dependencies rather than simple amplitude excursions. In autonomous driving, for instance, a steering command might be internally consistent but decouple from the resulting lateral acceleration. Residual-based detectors can miss such anomalies when flexible sequence models still reconstruct signals plausibly despite altered coordination. We introduce AxonAD, an unsupervised detector that treats multi-head attention query evolution as a short horizon predictable process. A gradient-updated reconstruction pathway is coupled with a history-only predictor that forecasts future query vectors from past context. This is trained via a masked predictor-target objective against an exponential moving average (EMA) target encoder. At inference, reconstruction error is combined with a tail-aggregated query mismatch score, which measures cosine deviation between predicted and target queries on recent timesteps. This dual approach provides sensitivity to structural dependency shifts while retaining amplitude-level detection. On proprietary in-vehicle telemetry with interval annotations and on the TSB-AD multi-variate suite (17 datasets, 180 series) with threshold-free and range-aware metrics, AxonAD improves ranking quality and temporal localization over strong baselines. Ablations confirm that query prediction and combined scoring are the primary drivers of the observed gains. Code is available at the URL https://github.com/iis-esslingen/AxonAD.

Figures

Figures reproduced from arXiv: 2603.12916 by Kadir-Kaan \"Ozer, Markus Enzweiler, Ren\'e Ebeling.

Figure 1
Figure 1. Figure 1: Query predictability in 2D query space (schematic, single head). Gray: past query trajectory. Blue dashed: predicted query. Red: EMA target query. (a) Nominal: predictor and target agree. (b) Coordination anomaly: the target query diverges from the predicted trajectory, producing large dq even when per-channel amplitudes are within normal bounds. such anomalies matters directly for fleet monitoring, warran… view at source ↗
Figure 2
Figure 2. Figure 2: AxonAD overview. The online reconstruction encoder computes self attention using queries Qrec. In parallel, a history-only predictor forecasts Qb pred and is trained to match EMA target queries Qtgt (stop-gradient). Query mismatch on the last valid timesteps yields dq, and reconstruction yields drec. Attention divergence (KL tail) is not included in the default scoring pipeline. 2.4 Self-Supervised Predict… view at source ↗
Figure 3
Figure 3. Figure 3: Score complementarity (schematic). Nominal windows spread along both axes but cluster near the origin in both scores simultaneously. Near-boundary amplitude and coordination anomalies are moderate on both axes, falling inside both single-axis thresholds (dotted lines) but separated by the additive S (dashed diagonal). where τ0 enforces both validity under the s step history constraint and tail focus of nom… view at source ↗
Figure 4
Figure 4. Figure 4: Chronological split and anomaly onset for the proprietary telemetry stream. 180 series) under the official pipeline [22,25]. Training is strictly unsupervised. All parameters and robust scoring statistics are fit on nominal training windows only, with labels reserved for evaluation. Hyperparameters for all methods are selected on the official TSB-AD tuning component (20 multivariate series) and then fixed.… view at source ↗
Figure 5
Figure 5. Figure 5: Paired AUC-PR comparison on TSB-AD multivariate (n = 180). Left: win/loss counts. Right: median paired difference with lollipop connectors from zero. All paired Wilcoxon tests yield p < 10−4 with entirely negative 95% bootstrap CIs (full statistics in the Appendix). Scoring components. The base configuration (Base) achieves the stron￾gest balanced profile across ranking and localization metrics. Removing t… view at source ↗
Figure 6
Figure 6. Figure 6: Sensitivity to forecast horizon s (left) and tail aggregation length k (right) on the tuning subset. Capacity and horizon. Reducing the number of attention heads from 8 to 4 (Heads=4 ) lowers AUC-PR by 0.033 with a smaller effect on localization metrics. Halving the model dimension from 128 to 64 (D=64) reduces AUC-PR by 0.042 and AUC-ROC by 0.025. Increasing the forecast horizon to s = 25 (Horizon 25 ) re… view at source ↗
Figure 7
Figure 7. Figure 7: Runtime on the telemetry stream (80,000 samples, stride 1). Left: wall-clock time split into fitting and scoring. Right: per window scoring latency. Classical baselines have no iterative training, so their time is entirely attributed to scoring. 0.553) and is generally lower for larger horizons, as a harder prediction task increases score variance. For tail length, threshold-free ranking is stable across k… 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

37 extracted references · 5 canonical work pages · 1 internal anchor

  1. [1]

    Competition and Attraction Improve Model Fusion

    Abrantes, J., Lange, R.T., Tang, Y.: Competition and Attraction Improve Model Fu- sion (Aug 2025).https://doi.org/10.48550/arXiv.2508.16204, http://arxiv. org/abs/2508.16204, arXiv:2508.16204

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2508.16204 2025

  2. [2]

    https://doi.org/10.48550/arXiv.2301.08243, http://arxiv.org/ abs/2301.08243

    Assran, M., Duval, Q., Misra, I., Bojanowski, P., Vincent, P., Rabbat, M., LeCun, Y., Ballas, N.: Self-supervised learning from images with a joint-embedding predictive architecture. https://doi.org/10.48550/arXiv.2301.08243, http://arxiv.org/ abs/2301.08243

    Pith/arXiv arXiv doi:10.48550/arxiv.2301.08243

  3. [3]

    In: Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Min- ing

    Audibert, J., Michiardi, P., Guyard, F., Marti, S., Zuluaga, M.A.: USAD: UnSuper- vised anomaly detection on multivariate time series. In: Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Min- ing. pp. 3395–3404. ACM (2020).https://doi.org/10.1145/3394486.3403392, https://dl.acm.org/doi/10.1145/3394486.3403392

    doi:10.1145/3394486.3403392 2020

  4. [4]

    org/abs/1607.06450

    Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization (2016),https://arxiv. org/abs/1607.06450

    Pith/arXiv arXiv 2016

  5. [5]

    https://doi.org/10.48550/arXiv.2511.08544, http:// arxiv.org/abs/2511.08544

    Balestriero, R., LeCun, Y.: LeJEPA: Provable and scalable self-supervised learning without the heuristics. https://doi.org/10.48550/arXiv.2511.08544, http:// arxiv.org/abs/2511.08544

    Pith/arXiv arXiv doi:10.48550/arxiv.2511.08544

  6. [6]

    Boniol, P., Krishna, A.K., Bruel, M., Liu, Q., Huang, M., Palpanas, T., Tsay, R.S., Elmore, A., Franklin, M.J., Paparrizos, J.: Vus: Effective and efficient accuracy measures for time-series anomaly detection (2025),https://arxiv.org/abs/2502. 13318

    2025

  7. [7]

    https://doi.org/10.1145/335191.335388, https://dl.acm.org/doi/10.1145/335191.335388 Predictable Query Dynamics for Time Series Anomaly Detection 15

    Breunig, M.M., Kriegel, H.P., Ng, R.T., Sander, J.: LOF: identifying density-based local outliers29(2), 93–104 (2000). https://doi.org/10.1145/335191.335388, https://dl.acm.org/doi/10.1145/335191.335388 Predictable Query Dynamics for Time Series Anomaly Detection 15

    doi:10.1145/335191.335388 2000

  8. [8]

    Candès, E.J., Li, X., Ma, Y., Wright, J.: Robust principal component analysis?58(3), 1–37 (2011).https://doi.org/10.1145/1970392.1970395, https://dl.acm.org/ doi/10.1145/1970392.1970395

    doi:10.1145/1970392.1970395 2011

  9. [9]

    In: 2018 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition

    Cipolla, R., Gal, Y., Kendall, A.: Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In: 2018 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. pp. 7482–7491 (2018).https://doi.org/ 10.1109/CVPR.2018.00781

    doi:10.1109/cvpr.2018.00781 2018

  10. [10]

    In: Proceedings of the 15th International Joint Conference on Computational Intelligence

    Correia, L., Goos, J.C., Klein, P., Bäck, T., Kononova, A.: MA-VAE: Multi-Head Attention-Based Variational Autoencoder Approach for Anomaly Detection in Multivariate Time-Series Applied to Automotive Endurance Powertrain Testing:. In: Proceedings of the 15th International Joint Conference on Computational Intelligence. pp. 407–418. SCITEPRESS - Science an...

    work page doi:10.5220/0012163100003595 2023

  11. [11]

    IEEE Transactions on Information Theory13(1), 21–27 (January 1967).https://doi.org/10.1109/TIT

    Cover, T., Hart, P.: Nearest neighbor pattern classification. IEEE Transactions on Information Theory13(1), 21–27 (January 1967).https://doi.org/10.1109/TIT. 1967.1053964

    doi:10.1109/tit 1967

  12. [12]

    Deng,A.,Hooi,B.:GraphNeuralNetwork-BasedAnomalyDetectioninMultivariate Time Series (Jun 2021).https://doi.org/10.48550/arXiv.2106.06947, http:// arxiv.org/abs/2106.06947, arXiv:2106.06947

    Pith/arXiv arXiv doi:10.48550/arxiv.2106.06947 2021

  13. [13]

    Goldstein, M., Dengel, A.: Histogram-based outlier score (hbos): A fast unsupervised anomaly detection algorithm (09 2012)

    2012

  14. [14]

    Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P.H., Buchatskaya, E., Doersch, C., Pires, B.A., Guo, Z.D., Azar, M.G., Piot, B., Kavukcuoglu, K., Munos, R., Valko, M.: Bootstrap your own latent: A new approach to self-supervised learning (2020),https://arxiv.org/abs/2006.07733

    Pith/arXiv arXiv 2020

  15. [15]

    IEEE Trans- actions on Knowledge and Data Engineering33(4), 1479–1489 (Apr 2021)

    Hariri, S., Kind, M.C., Brunner, R.J.: Extended Isolation Forest. IEEE Trans- actions on Knowledge and Data Engineering33(4), 1479–1489 (Apr 2021). https://doi.org/10.1109/TKDE.2019.2947676, https://ieeexplore.ieee.org/ document/8888179/

    doi:10.1109/tkde.2019.2947676 2021

  16. [16]

    Pat- tern Recognition Letters24(9), 1641–1650 (2003)

    He, Z., Xu, X., Deng, S.: Discovering cluster-based local outliers. Pat- tern Recognition Letters24(9), 1641–1650 (2003). https://doi.org/https: //doi.org/10.1016/S0167-8655(03)00003-5, https://www.sciencedirect.com/ science/article/pii/S0167865503000035

    doi:10.1016/s0167-8655(03)00003-5 2003

  17. [17]

    CoRRabs/1611.05267(2016), http://arxiv.org/abs/1611.05267

    Lea, C., Flynn, M.D., Vidal, R., Reiter, A., Hager, G.D.: Temporal convolutional networks for action segmentation and detection. CoRRabs/1611.05267(2016), http://arxiv.org/abs/1611.05267

    Pith/arXiv arXiv 2016

  18. [18]

    IEEE Trans- actions on Neural Networks and Learning Systems32(3), 1177–1191 (Mar 2021)

    Li, L., Yan, J., Wang, H., Jin, Y.: Anomaly Detection of Time Series With Smoothness-Inducing Sequential Variational Auto-Encoder. IEEE Trans- actions on Neural Networks and Learning Systems32(3), 1177–1191 (Mar 2021). https://doi.org/10.1109/TNNLS.2020.2980749, https://ieeexplore. ieee.org/document/9064715/

    work page doi:10.1109/tnnls.2020.2980749 2021

  19. [19]

    https://doi.org/10.48550/arXiv.2009.09463, http://arxiv.org/abs/ 2009.09463

    Li, Z., Zhao, Y., Botta, N., Ionescu, C., Hu, X.: COPOD: Copula-based outlier de- tection. https://doi.org/10.48550/arXiv.2009.09463, http://arxiv.org/abs/ 2009.09463

    Pith/arXiv arXiv doi:10.48550/arxiv.2009.09463 2009

  20. [20]

    International Jour- nal of Forecasting37(4), 1748–1764 (Oct 2021)

    Lim, B., Arık, S.Ö., Loeff, N., Pfister, T.: Temporal Fusion Transform- ers for interpretable multi-horizon time series forecasting. International Jour- nal of Forecasting37(4), 1748–1764 (Oct 2021). https://doi.org/10.1016/j. ijforecast.2021.03.012, https://www.sciencedirect.com/science/article/ pii/S0169207021000637 16 K.-K. Özer et al

    doi:10.1016/j 2021

  21. [21]

    In: 2008 Eighth IEEE In- ternational Conference on Data Mining

    Liu, F.T., Ting, K.M., Zhou, Z.H.: Isolation Forest. In: 2008 Eighth IEEE In- ternational Conference on Data Mining. pp. 413–422 (Dec 2008).https://doi. org/10.1109/ICDM.2008.17, https://ieeexplore.ieee.org/document/4781136, iSSN: 2374-8486

    doi:10.1109/icdm.2008.17 2008

  22. [22]

    In: Advances in Neural Information Processing Systems 37

    Liu, Q., Paparrizos, J.: The Elephant in the Room: Towards A Reliable Time-Series Anomaly Detection Benchmark. In: Advances in Neural Information Processing Systems 37. pp. 108231–108261. Neural Information Processing Systems Founda- tion, Inc. (NeurIPS), Vancouver, BC, Canada (2024).https://doi.org/10.52202/ 079017-3437,http://www.proceedings.com/079017-...

    2024

  23. [23]

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization (2019),https: //arxiv.org/abs/1711.05101

    Pith/arXiv arXiv 2019

  24. [24]

    Malhotra, P., Vig, L., Shroff, G., Agarwal, P.: Long short term memory networks for anomaly detection in time series (04 2015)

    2015

  25. [25]

    Proceedings of the VLDB Endowment15(11), 2774–2787 (Jul 2022)

    Paparrizos, J., Boniol, P., Palpanas, T., Tsay, R.S., Elmore, A., Franklin, M.J.: Volume under the surface: a new accuracy evaluation measure for time-series anomaly detection. Proceedings of the VLDB Endowment15(11), 2774–2787 (Jul 2022). https://doi.org/10.14778/3551793.3551830, https://dl.acm.org/doi/ 10.14778/3551793.3551830

    doi:10.14778/3551793.3551830 2022

  26. [26]

    In: 2018 17th IEEE International Conference on Machine Learning and Applications (ICMLA)

    Pereira, J., Silveira, M.: Unsupervised Anomaly Detection in Energy Time Series Data Using Variational Recurrent Autoencoders with Attention. In: 2018 17th IEEE International Conference on Machine Learning and Applications (ICMLA). pp. 1275–1282. IEEE, Orlando, FL (Dec 2018).https://doi.org/10.1109/ICMLA. 2018.00207,https://ieeexplore.ieee.org/document/8614232/

    doi:10.1109/icmla 2018

  27. [27]

    International Journal of Data Mining and Bioinfor- matics22(4), 389 (2019)

    Pereira, J., Silveira, M.: Unsupervised representation learning and anomaly de- tection in ECG sequences. International Journal of Data Mining and Bioinfor- matics22(4), 389 (2019). https://doi.org/10.1504/IJDMB.2019.101395, http: //www.inderscience.com/link.php?id=101395

    work page doi:10.1504/ijdmb.2019.101395 2019

  28. [28]

    In: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining

    Su, Y., Zhao, Y., Niu, C., Liu, R., Sun, W., Pei, D.: Robust Anomaly De- tection for Multivariate Time Series through Stochastic Recurrent Neural Net- work. In: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. pp. 2828–2837. ACM, Anchorage AK USA (Jul 2019). https://doi.org/10.1145/3292500.3330672, https://...

    doi:10.1145/3292500.3330672 2019

  29. [29]

    https://doi.org/10.48550/arXiv.2201.07284, http://arxiv.org/abs/ 2201.07284, arXiv:2201.07284

    Tuli, S., Casale, G., Jennings, N.R.: TranAD: Deep Transformer Net- works for Anomaly Detection in Multivariate Time Series Data (May 2022). https://doi.org/10.48550/arXiv.2201.07284, http://arxiv.org/abs/ 2201.07284, arXiv:2201.07284

    Pith/arXiv arXiv doi:10.48550/arxiv.2201.07284 2022

  30. [30]

    In: Proceedings of the 31st International Conference on Neural Information Processing Systems

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. In: Proceedings of the 31st International Conference on Neural Information Processing Systems. p. 6000–6010 (2017)

    2017

  31. [31]

    Engineering Applications of Artificial Intelligence 104, 104354 (Sep 2021)

    Von Schleinitz, J., Graf, M., Trutschnig, W., Schröder, A.: VASP: An autoencoder- based approach for multivariate anomaly detection and robust time series prediction with application in motorsport. Engineering Applications of Artificial Intelligence 104, 104354 (Sep 2021). https://doi.org/10.1016/j.engappai.2021.104354, https://linkinghub.elsevier.com/ret...

    work page doi:10.1016/j.engappai.2021.104354 2021

  32. [32]

    https://doi.org/10.48550/arXiv.2210.02186, http://arxiv.org/abs/ 2210.02186, arXiv:2210.02186 Predictable Query Dynamics for Time Series Anomaly Detection 17

    Wu, H., Hu, T., Liu, Y., Zhou, H., Wang, J., Long, M.: TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis (Apr 2023). https://doi.org/10.48550/arXiv.2210.02186, http://arxiv.org/abs/ 2210.02186, arXiv:2210.02186 Predictable Query Dynamics for Time Series Anomaly Detection 17

    Pith/arXiv arXiv doi:10.48550/arxiv.2210.02186 2023

  33. [33]

    48550/arXiv.1802.03903, http://arxiv.org/abs/1802.03903, arXiv:1802.03903

    Xu, H., Chen, W., Zhao, N., Li, Z., Bu, J., Li, Z., Liu, Y., Zhao, Y., Pei, D., Feng, Y., Chen, J., Wang, Z., Qiao, H.: Unsupervised Anomaly Detection via Variational Auto- Encoder for Seasonal KPIs in Web Applications (Feb 2018).https://doi.org/10. 48550/arXiv.1802.03903, http://arxiv.org/abs/1802.03903, arXiv:1802.03903

    Pith/arXiv arXiv 2018

  34. [34]

    Xu, J., Wu, H., Wang, J., Long, M.: Anomaly Transformer: Time Series Anomaly Detection with Association Discrepancy (Jun 2022).https://doi.org/10.48550/ arXiv.2110.02642,http://arxiv.org/abs/2110.02642, arXiv:2110.02642

    Pith/arXiv arXiv 2022

  35. [35]

    Xu, Z., Zeng, A., Xu, Q.: FITS: Modeling time series with $10k$ parameters.https: //doi.org/10.48550/arXiv.2307.03756,http://arxiv.org/abs/2307.03756

    Pith/arXiv arXiv doi:10.48550/arxiv.2307.03756

  36. [36]

    Yairi, T., Kato, Y., Hori, K.: Fault detection by mining association rules from housekeeping data (06 2001)

    2001

  37. [37]

    Mean± standard deviation over all evaluated series

    Zhou, T., Niu, P., Wang, X., Sun, L., Jin, R.: One fits all:power general time series analysis by pretrained lm (2023),https://arxiv.org/abs/2302.11939 Predictable Query Dynamics for Time-Series Anomaly Detection Supplementary Material S1 Additional Tables Table S1.TSB-AD multivariate benchmark (17 datasets, 180 time series). Mean± standard deviation over...

    Pith/arXiv arXiv 2023