Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

Causality-Aware Contrastive Learning for Robust Multivariate Time-Series Anomaly Detection

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that injecting an estimated causal graph into contrastive learning separates normal time-series variations from anomalies better than reconstruction-based and value-based contrastive baselines, and validates the claim on…

desk verdict A genuinely new causal-contrastive recipe for MTSAD, but the paper's own ablations undercut the claim that causality is what drives the gains. read the letter →

arxiv 2506.03964 v1 pith:2ZAH25VN submitted 2025-06-04 cs.LG cs.AI

HyunGi Kim , Jisoo Mok , Dongjun Lee , Jaihyun Lew , Sungjae Kim , Sungroh Yoon This is my paper · ORCID
classification cs.LGcs.AI
keywords multivariatetime-seriesanomalydetectioncausality-awarecontrastivelearningcausaldiscoverycausality-preservingaugmentationcausality-disturbingsimilarity-filteredone-classlossunsupervised
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that multivariate time-series anomaly detection is improved when the model is made aware of the causal graph among variables, rather than only of statistical regularities in the data. It proposes CAROTS, a pipeline that first learns the normal causal structure with a forecasting-based causal discovery model, then generates two kinds of augmented samples: causality-preserving ones that extend the normal class, and causality-disturbing ones that simulate anomalies. These samples act as positives and negatives in a contrastive learning objective, and a similarity-filtered one-class loss keeps semantically diverse normal patterns from collapsing into a single cluster. At test time, a window is scored by combining the distance from its embedding to the normal centroid with the forecasting error of the causal model. The paper reports that this causality-aware training consistently outperforms reconstruction and contrastive baselines, with the largest margins on synthetic anomalies that were designed to break causal relationships.

What carries the argument

The central object is the learned causality matrix $A$ together with the causal forecasting functions $f_i$, produced by the forecasting-based causal discoverer (CUTS+). This pair carries the entire argument: $A$ and $f_i$ define what counts as a causality-preserving variation (noise on causes, effects repredicted through $f_i$) versus a causality-disturbing anomaly (a DFS-selected subgraph with random biases), and the same $f_i$ provides the forecasting-error anomaly score $A_{CD}$. The contrastive encoder is trained with a similarity-filtered one-class loss that only includes positive samples above a cosine-similarity threshold $\alpha$, scheduled from 0.5 to 0.9, so the latent space preserves multiple normal semantic clusters rather than collapsing them; the final score then sums the z-normalized embedding-distance score $A_{CL}$ and $A_{CD}$.

What would settle it

Train CAROTS on a synthetic dataset with a known ground-truth causal graph, once with the true graph and once with a randomly permuted graph; if the random-graph version matches the true-graph version in AUROC, the learned causality is not what drives detection. A second check is the similarity filter: if fixing the threshold at no filtering matches or beats the scheduled version across all datasets, the progressive-diversity mechanism is not doing its stated work.

Watch

Extended reading notes

Core claim

CAROTS treats anomalies as violations of time-invariant causal relationships among variables. A forecasting-based causal discovery model first estimates a causality matrix $A$ and per-variable functions $f_i$ from clean training data. The causality-preserving augmentor adds noise to selected cause variables and uses the learned functions to repredict the affected effect variables, producing diverse samples that still obey the normal causal graph. The causality-disturbing augmentor extracts random connected subgraphs from $A$ via depth-first search and injects random biases, producing synthetic samples where causal links are broken. Contrastive learning then trains an encoder that pulls causality-preserving samples together and pushes causality-disturbing samples away, while the similarity-filtered one-class loss starts with only highly similar positives and gradually admits more diverse ones. The final anomaly score sums the z-normalized distance to the positive centroid and the z-normalized forecasting error. The paper's central claim is that this causal formulation yields consistently higher AUROC, AUPRC, and F1 scores than reconstruction- and contrastive-based methods, including on hard synthetic anomalies that all baselines miss.

Load-bearing premise

The load-bearing premise is that the causal discovery step faithfully recovers the true normal causal graph and functional relationships from the training data, and that these relationships stay fixed over time, so the generated positives really are normal and the generated negatives really violate causality.

Editorial extensions

If this is right

  • The encoder's latent space separates causality-preserving samples from causality-disturbing ones, so test windows can be scored without labels by their distance to the normal centroid.
  • Ensembling the contrastive score with the causal forecaster's error improves robustness, since removing either component lowers detection scores in the ablations.
  • On hard synthetic anomalies where reconstruction and contrastive baselines score near chance, CAROTS reports high AUROC, most notably 0.998 on point-global anomalies in the Lorenz96 dataset.
  • The framework remains competitive across different causal discovery models (NGC, CUTS, CUTS+), different encoders (LSTM, GRU, iTransformer), and varying temperature and batch sizes.
  • The similarity-filtered one-class loss contributes directly to performance: replacing it with a naive one-class contrastive loss drops SWaT AUROC from 0.852 to 0.819.

Reading between the lines

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

  • If CAROTS's gains replicate, anomaly detection systems could shift from deep reconstruction towers toward lightweight encoders guided by a learned causal prior, since the paper shows a one-layer LSTM suffices when causality informs the training signal.
  • The same preserve-versus-disturb augmentation pair could transfer to other self-supervised tasks such as out-of-distribution detection or domain-shift detection, where causality-preserving transformations define the in-distribution manifold and causality-disturbing ones define novel inputs.
  • The paper assumes time-invariant causal structure, so a natural untested extension is sliding-window causal discovery that refreshes $A$ and $f_i$ over time, turning CAROTS into an online detector able to track slow concept drift.
  • The similarity-filtering threshold is annealed on a fixed schedule; an adaptive schedule driven by loss or cluster structure could be tested against the fixed schedule to see whether the progressive-diversity mechanism is being exploited optimally.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes CAROTS, a multivariate time-series anomaly detection (MTSAD) pipeline that first learns an adjacency-based causal graph and causal forecasting functions from normal training data, then uses two augmentors—a causality-preserving augmentor (CPA) and a causality-disturbing augmentor (CDA)—to construct positive and negative samples for contrastive learning with a similarity-filtered one-class (SOC) loss. Anomaly scores are formed by ensembling a contrastive distance (ACL) with the forecasting error of the causal model (ACD). The authors claim that injecting learned causal structure into contrastive learning yields consistent and significant improvements over reconstruction and contrastive baselines on five real-world and two synthetic datasets.

Significance. If the central claim were established, the paper would make a useful contribution by connecting causal discovery with contrastive MTSAD: it proposes a coherent pipeline, provides code, and includes extensive experiments, ablations, hyperparameter robustness checks, and a t-SNE qualitative analysis. The synthetic Lorenz96 results, where CAROTS outperforms all baselines by a large margin, are genuinely interesting. However, the paper's own tables contradict several of its headline claims: the full model underperforms its own ACD-free ablation on most real datasets, the WADI performance is below several baselines, and the VAR average performance is not the best. Because the causal-attribution claim is load-bearing and is not supported by the currently reported experiments, the significance is not yet established.

major comments (4)
  1. [§5.2, Table 1] The statement in §5.2 that "CAROTS consistently achieves the highest detection scores in nearly all metrics and datasets" is not supported by the reported numbers. On WADI, the full model's AUROC is 0.502, which is below CTAD (0.599), CARLA (0.533), and even the ablated CAROTS† (0.622), and its AUPRC (0.056) is far below CTAD (0.327). Moreover, CAROTS† (without ACD) has higher AUROC than CAROTS on 5 of the 7 real-world datasets: SWaT 0.861 vs 0.852, WADI 0.622 vs 0.502, SMD 2-1 0.726 vs 0.703, SMD 3-7 0.779 vs 0.769, and MSL P-14 0.782 vs 0.764. This directly contradicts the claim in the same section that "the slight superiority of CAROTS to CAROTS† implies that ACD makes CAROTS even more robust." The authors should either revise the claim, re-tune the ensemble, or report CAROTS† as the primary model.
  2. [§5.3, Table 2] The synthetic-data claim that CAROTS "significantly outperforms methods that overlook causal relationships" is not uniformly true. On the VAR dataset, CAROTS's average AUROC is 0.805, below USAD (0.810) and essentially tied with TimesNet (0.803). The Lorenz96 results are strong, but the blanket statement in the abstract and §5.3 should be qualified to the datasets and anomaly types where the improvement is actually observed. In addition, the claimed "remarkable improvement in AUROC up to 50%" is not explicitly backed by a numerical comparison in the text; Figure 3 lacks clear axis labels and numeric annotations, making it difficult to verify this specific claim.
  3. [§5.4, Table 5] The paper's central attribution—that the reported gains come from causal structure—is not established because no non-causal control is reported. Table 5 shows that swapping CUTS+ for NGC or CUTS changes AUROC by only about 0.01–0.04 on some datasets (e.g., SWaT 0.85 vs 0.85/0.86) and by up to 0.10 on MSL P-15 (0.76 vs 0.66). Without a control in which the adjacency matrix is replaced by a random or empty matrix, the observed improvements could plausibly be produced by the generic contrastive recipe (noise injection, bias perturbation, one-class loss) rather than by causal relationships. The authors should add such a control and, depending on the outcome, temper the causal-attribution claims in the abstract and conclusion.
  4. [§4.3, §5.2] ACD is computed from the same forecaster F that generates the CPA positive augmentations, so the ensemble score in Eq. (7) combines two signals derived from a single model. The empirical result that adding ACD reduces AUROC on most real-world datasets (Table 1) suggests that the two scores are not complementary as claimed. The paper should either remove the ensemble from the main comparison, use a learned or better-calibrated combination, or provide an analysis of why ACD helps only on a minority of datasets before claiming that it makes CAROTS "even more robust."
minor comments (5)
  1. [Table 1 caption] The caption contains a typo: "TimsNet" should be "TimesNet."
  2. [§5.3] The text contains a typo: "dtection capability" should be "detection capability."
  3. [§5.4] The sentence beginning "The contrastive learning-based anomaly score (ACL) is shown to play an essential role" is repeated nearly verbatim in the same paragraph; one of the two instances should be removed.
  4. [Figure 3] The figure is difficult to interpret: the x-axis values (1.0, 2.0, 3.0, 4.0) are not labeled, and the legend is ambiguous. Please add axis labels, explain what the values represent, and consider adding error bars or numeric annotations for the headline claims.
  5. [Appendix B] The statement that SWaT and WADI were downsampled by a factor of 5 for computational efficiency is reported only in the appendix; please state this in the main experimental setup, since it may affect comparability with baseline results reported elsewhere.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the causal graph and forecaster are learned from normal training data only, and both anomaly scores are evaluated on held-out test windows; only minor non-load-bearing self-citations and a shared CPA/ACD forecaster entanglement are present.

full rationale

The derivation chain is self-contained against external benchmarks. The causal matrix A and functions f_i are obtained by minimizing forecasting MSE on normal training data (Eq. 1); no test labels or anomaly annotations enter training, and the anomaly scores of Eq. 7 are computed on held-out test windows. ACL is the distance to the centroid of causality-preserving embeddings produced by a contrastively trained encoder, and ACD is the out-of-sample forecasting error of F_theta,A, a standard reconstruction/forecasting-based anomaly signal rather than a parameter fitted to the target result. The only entanglement is that the same forecaster F generates the CPA positive samples (Eq. 3) and supplies the ACD score (Eq. 7), so the two ensemble components share one fitted module and are not statistically independent; however, this does not reduce to fitting the labels or to a definitional equivalence, and the paper's own ablation (Table 10) shows ACD often lowers AUROC, so the ensemble is not forced by construction. Self-citations (CTAD augmentation strategies and LSTM choice in Section 5.1 and Appendix B; the point-adjusted F1 exclusion citing Kim et al. 2022; Choi et al. 2024 in related work) are methodological and not load-bearing; no uniqueness theorem is imported from prior author work. The stated limitation in Section D that CAROTS assumes time-invariant causal structure is an assumption, and Table 12 provides empirical support for it on three datasets. The skeptical concern that full CAROTS underperforms CAROTS without ACD on 5 of 7 real datasets (Table 10) and that no random-graph control is run bears on the attributability of the gains to causal structure, which is a correctness and evidence concern, not a circularity of derivation. No step exhibits a by-construction reduction of a claimed prediction to its inputs, so the circularity score is low.

Assumptions & free parameters 8 free parameters · 6 assumptions · 0 invented entities

The central claim rests primarily on the accuracy of the learned causal graph and on the augmentation and contrastive hyperparameters listed above. No parameters are fitted to test labels; all are fixed or selected by validation. The most fragile input is the causal discovery model's output, which determines the meaning of both positive and negative augmented samples.

free parameters (8)
  • Similarity filter threshold α = Scheduled linear increase from 0.5 to 0.9 over epochs
    Controls which positives participate in the SOC loss; selected based on SWaT validation experiments (Table 9).
  • Contrastive temperature τ = 0.1
    Scales cosine similarities in the contrastive loss; robustness shown in Table 3.
  • CPA noise σ = 0.2
    Standard deviation of Gaussian perturbation added to causing variables in Eq. (3); robustness shown in Table 11.
  • CDA DFS cut-off probability p = 0.1
    Termination probability for subgraph extraction in the causality-disturbing augmentor (Appendix B).
  • CDA bias set = Values in {-0.5,-0.4,...,0.5}
    Random biases injected into causality-disturbing samples; chosen to create anomalies of varying difficulty.
  • CPA number of causing variables M = 1
    One causing variable is perturbed per causality-preserving augmentation (Appendix B).
  • Window size w = 10 (2 for Lorenz96, 4 for VAR)
    Sliding-window length used to segment time series; follows prior work for synthetic data.
  • Anomaly score distance metric = L2 default; cosine for WADI and VAR
    Per-dataset choice reported in Appendix B; selection is not justified a priori.
assumptions (6)
  • domain assumption Time-invariant normal causal structure
    Section 3.2 assumes A and f_i are time-invariant; Appendix C checks this empirically on three datasets but it remains an assumption.
  • domain assumption Causal discovery faithfulness
    Section 4.1 builds all augmentations and the ACD score on the A and F learned by CUTS+; if these are wrong, positives and negatives are mislabeled.
  • domain assumption Anomaly-as-causal-violation definition
    Equation (2) defines anomalies as deviations from the learned causal model; real anomalies that do not correspond to such deviations are out of scope.
  • domain assumption CDA represents real anomalies
    Random bias injection into causal subgraphs is assumed to be a good proxy for real anomalies in training; the paper does not validate this on real attack data.
  • domain assumption CPA validity under distribution shift
    The forecaster F is used on inputs with added noise; the resulting predictions are treated as valid normal samples, which presumes F generalizes beyond its training inputs.
  • domain assumption Window-level labeling validity
    Section 5.1 labels a window as anomalous if it contains at least one anomaly point; this coarse protocol can inflate metric values.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Causality-Aware Contrastive Learning for Robust Multivariate Time-Series Anomaly Detection." pith.science (2026). https://pith.science/paper/2ZAH25VN

@misc{pith2026250603964,
  author       = {Pith},
  title        = {Pith review of: Causality-Aware Contrastive Learning for Robust Multivariate Time-Series Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2ZAH25VN}},
  note         = {Machine review of arXiv:2506.03964}
}
read the original abstract

Utilizing the complex inter-variable causal relationships within multivariate time-series provides a promising avenue toward more robust and reliable multivariate time-series anomaly detection (MTSAD) but remains an underexplored area of research. This paper proposes Causality-Aware contrastive learning for RObust multivariate Time-Series (CAROTS), a novel MTSAD pipeline that incorporates the notion of causality into contrastive learning. CAROTS employs two data augmentors to obtain causality-preserving and -disturbing samples that serve as a wide range of normal variations and synthetic anomalies, respectively. With causality-preserving and -disturbing samples as positives and negatives, CAROTS performs contrastive learning to train an encoder whose latent space separates normal and abnormal samples based on causality. Moreover, CAROTS introduces a similarity-filtered one-class contrastive loss that encourages the contrastive learning process to gradually incorporate more semantically diverse samples with common causal relationships. Extensive experiments on five real-world and two synthetic datasets validate that the integration of causal relationships endows CAROTS with improved MTSAD capabilities. The code is available at https://github.com/kimanki/CAROTS.

Figures

Figures reproduced from arXiv: 2506.03964 by the authors.

Figure 1
Figure 1. An overall pipeline of CAROTS. (a) First, a causal discoverer is trained to learn the normal causal relationships in the training data. (b) Based on the causal discoverer, the causality-preserving and -disturbing augmentors construct a mini-batch of four groups of samples: original (G1), causality-preserving augmentations (G2), causality-disturbing augmentations of G1 (G3), and causality￾disturbing augmentations of … view at source ↗
Figure 2
Figure 2. (a-1) CPA adds Gaussian noise to randomly selected variables (Red) and uses the causality-integrated forecaster to predict and replace the values of affected variables (Blue). (a-2) CDA selects a random variable (Yellow) and extracts a directed subgraph of a causality matrix through DFS. Random perturbations that are unrelated to causal relationships are injected into the selected variables.(b) [Left] In early itera… view at source ↗
Figure 3
Figure 3. Anomaly detection performance measured by AUROC across different anomaly types in the Lorenz96 dataset, evaluated based on anomaly difficulty levels controlled by the factor λ [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: T-SNE visualizations of embeddings during training and validation after training. thus, they show a significant overlap and appear inseparable from a causal viewpoint. As training progresses, positive and negative samples begin to separate, showing that the proposed ca…
Figure 5
Figure 5. Figure 5: Learned causality matrix from the causal discovery model for each dataset. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. VAN-AD: Visual Masked Autoencoder with Normalizing Flow For Time Series Anomaly Detection

    cs.LG 2026-03 unverdicted novelty 6.0 of 10

    VAN-AD adapts a pretrained visual MAE with distribution mapping and normalizing flow modules to detect anomalies in time series data more effectively across different datasets.

  2. Cluster-Aware Causal Mixer for Online Anomaly Detection in Multivariate Time Series

    cs.LG 2025-05 conditional novelty 6.0 of 10

    A cluster-aware causal MLP-mixer with sequential evidence-accumulation scoring achieves state-of-the-art best-F1 on six multivariate time-series anomaly detection benchmarks.

Reference graph

Works this paper leans on

55 extracted references · 42 canonical work pages · cited by 2 Pith papers

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Practical approach to asynchronous multivariate time series anomaly detection and localization

    Abdulaal, A., Liu, Z., and Lancewicki, T. Practical approach to asynchronous multivariate time series anomaly detection and localization. In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, pp.\ 2485--2494, 2021

    work page 2021

  3. [3]

    and Zhang, L

    Absar, S. and Zhang, L. Discovering time-invariant causal structure from temporal data. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pp.\ 2807--2811, 2021

    work page 2021

  4. [4]

    Neural time-invariant causal discovery from time series data

    Absar, S., Wu, Y., and Zhang, L. Neural time-invariant causal discovery from time series data. In 2023 International Joint Conference on Neural Networks (IJCNN), pp.\ 1--8. IEEE, 2023

    work page 2023

  5. [5]

    M., Palleti, V

    Ahmed, C. M., Palleti, V. R., and Mathur, A. P. Wadi: A water distribution testbed for research in the design of secure cyber physical systems. In Proceedings of the 3rd International Workshop on Cyber-Physical Systems for Smart Water Networks, CySWATER '17, pp.\ 25–28, New York, NY, USA, 2017. Association for Computing Machinery. ISBN 9781450349758. doi:...

    arXiv 2017

  6. [6]

    K., Devijver, E., and Gaussier, E

    Assaad, C. K., Devijver, E., and Gaussier, E. Survey and evaluation of causal discovery methods for time series. Journal of Artificial Intelligence Research, 73: 0 767--819, 2022

    work page 2022

  7. [7]

    Audibert, J., Michiardi, P., Guyard, F., Marti, S., and Zuluaga, M. A. Usad: Unsupervised anomaly detection on multivariate time series. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pp.\ 3395--3404, 2020

    work page 2020

  8. [8]

    Ben-Moshe, N., Tsutsui, K., Biton, S., Zvuloni, E., S \"o rnmo, L., and Behar, J. A. Rawecgnet: Deep learning generalization for atrial fibrillation detection from the raw ecg. IEEE Journal of Biomedical and Health Informatics, 2024

    work page 2024

Show all 55 references
  1. [9]

    A simple framework for contrastive learning of visual representations

    Chen, T., Kornblith, S., Norouzi, M., and Hinton, G. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pp.\ 1597--1607. PMLR, 2020

    2020

  2. [10]

    CUTS : Neural causal discovery from irregular time-series data

    Cheng, Y., Yang, R., Xiao, T., Li, Z., Suo, J., He, K., and Dai, Q. CUTS : Neural causal discovery from irregular time-series data. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=UG8bQcD3Emv

    2023

  3. [11]

    Cuts+: High-dimensional causal discovery from irregular time-series

    Cheng, Y., Li, L., Xiao, T., Li, Z., Suo, J., He, K., and Dai, Q. Cuts+: High-dimensional causal discovery from irregular time-series. Proceedings of the AAAI Conference on Artificial Intelligence, 38 0 (10): 0 11525--11533, Mar. 2024 a . doi:10.1609/aaai.v38i10.29034. URL htt...

    2024 doi

  4. [12]

    Causaltime: Realistically generated time-series for benchmarking of causal discovery

    Cheng, Y., Wang, Z., Xiao, T., Zhong, Q., Suo, J., and He, K. Causaltime: Realistically generated time-series for benchmarking of causal discovery. In The Twelfth International Conference on Learning Representations, 2024 b . URL https://openreview.net/forum?id=iad1yyyGme

    2024

  5. [13]

    Learning phrase representations using RNN encoder -- decoder for statistical machine translation

    Cho, K., van Merri \"e nboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., and Bengio, Y. Learning phrase representations using RNN encoder -- decoder for statistical machine translation. In Moschitti, A., Pang, B., and Daelemans, W. (eds.), Proceedings of the 20...

    2014 doi

  6. [14]

    Deep learning for anomaly detection in time-series data: Review, analysis, and guidelines

    Choi, K., Yi, J., Park, C., and Yoon, S. Deep learning for anomaly detection in time-series data: Review, analysis, and guidelines. IEEE access, 9: 0 120043--120065, 2021

    2021

  7. [15]

    Self-supervised time-series anomaly detection using learnable data augmentation

    Choi, K., Yi, J., Mok, J., and Yoon, S. Self-supervised time-series anomaly detection using learnable data augmentation. arXiv preprint arXiv:2406.12260, 2024

    2024 arXiv

  8. [16]

    D., Liu, C., Moody, B., Li-wei, H

    Clifford, G. D., Liu, C., Moody, B., Li-wei, H. L., Silva, I., Li, Q., Johnson, A., and Mark, R. G. Af classification from a short single lead ecg recording: The physionet/computing in cardiology challenge 2017. In 2017 Computing in Cardiology (CinC), pp.\ 1--4. IEEE, 2017

    2017

  9. [17]

    SARAD : Spatial association-aware anomaly detection and diagnosis for multivariate time series

    Dai, Z., He, L., Yang, S., and Leeke, M. SARAD : Spatial association-aware anomaly detection and diagnosis for multivariate time series. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview.net/forum?id=gmf5Aj01Hz

    2024

  10. [18]

    Z., Webb, G

    Darban, Z. Z., Webb, G. I., Pan, S., Aggarwal, C. C., and Salehi, M. Carla: Self-supervised contrastive representation learning for time series anomaly detection. Pattern Recognition, 157: 0 110874, 2025

    2025

  11. [19]

    Dickey, D. A. and Fuller, W. A. Distribution of the estimators for autoregressive time series with a unit root. Journal of the American statistical association, 74 0 (366a): 0 427--431, 1979

    1979

  12. [20]

    G., Moore, K., Thapa, C., Liu, M., Saikrishna, V., Ma, J., and Xia, F

    Febrinanto, F. G., Moore, K., Thapa, C., Liu, M., Saikrishna, V., Ma, J., and Xia, F. Entropy causal graphs for multivariate time series anomaly detection. arXiv preprint arXiv:2312.09478, 2023

    2023 arXiv

  13. [21]

    N., and Mathur, A

    Goh, J., Adepu, S., Junejo, K. N., and Mathur, A. A dataset to support research in the design of secure water treatment systems. In International conference on critical information infrastructures security, pp.\ 88--99. Springer, 2016

    2016

  14. [22]

    Granger, C. W. Investigating causal relations by econometric models and cross-spectral methods. Econometrica: journal of the Econometric Society, pp.\ 424--438, 1969

    1969

  15. [23]

    Momentum contrast for unsupervised visual representation learning

    He, K., Fan, H., Wu, Y., Xie, S., and Girshick, R. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 9729--9738, 2020

    2020

  16. [24]

    and Schmidhuber, J

    Hochreiter, S. and Schmidhuber, J. Long short-term memory. Neural computation, 9 0 (8): 0 1735--1780, 1997

    1997

  17. [25]

    Detecting spacecraft anomalies using lstms and nonparametric dynamic thresholding

    Hundman, K., Constantinou, V., Laporte, C., Colwell, I., and Soderstrom, T. Detecting spacecraft anomalies using lstms and nonparametric dynamic thresholding. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pp.\ 387--395, 2018

    2018

  18. [26]

    and Kang, P

    Kang, H. and Kang, P. Transformer-based multivariate time series anomaly detection using inter-variable attention mechanism. Knowledge-Based Systems, 290: 0 111507, 2024

    2024

  19. [27]

    and Paul, M

    Karimi, A. and Paul, M. R. Extensive chaos in the lorenz-96 model. Chaos: An interdisciplinary journal of nonlinear science, 20 0 (4), 2010

    2010

  20. [28]

    Contrastive time-series anomaly detection

    Kim, H., Kim, S., Min, S., and Lee, B. Contrastive time-series anomaly detection. IEEE Transactions on Knowledge and Data Engineering, 2023

    2023

  21. [29]

    Towards a rigorous evaluation of time-series anomaly detection

    Kim, S., Choi, K., Choi, H.-S., Lee, B., and Yoon, S. Towards a rigorous evaluation of time-series anomaly detection. Proceedings of the AAAI Conference on Artificial Intelligence, 36 0 (7): 0 7194--7201, Jun. 2022. doi:10.1609/aaai.v36i7.20680. URL https://ojs.aaai.org/index....

    2022 doi

  22. [30]

    Kingma, D. P. and Ba, J. Adam: A method for stochastic optimization. In Bengio, Y. and LeCun, Y. (eds.), 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings , 2015. URL http://arxiv.org/abs/1412.6980

    2015 arXiv

  23. [31]

    Causalformer: An interpretable transformer for temporal causal discovery

    Kong, L., Li, W., Yang, H., Zhang, Y., Guan, J., and Zhou, S. Causalformer: An interpretable transformer for temporal causal discovery. IEEE Transactions on Knowledge and Data Engineering, 2024

    2024

  24. [32]

    R., Mathieu, B., Lahmadi, A., and Boutaba, R

    Ky, J. R., Mathieu, B., Lahmadi, A., and Boutaba, R. Cats: Contrastive learning for anomaly detection in time series. In 2024 IEEE International Conference on Big Data (BigData), pp.\ 1352--1359. IEEE, 2024

    2024

  25. [33]

    Revisiting time series outlier detection: Definitions and benchmarks

    Lai, K.-H., Zha, D., Xu, J., Zhao, Y., Wang, G., and Hu, X. Revisiting time series outlier detection: Definitions and benchmarks. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. URL https://openreview.net/forum...

    2021

  26. [34]

    Soft contrastive learning for time series

    Lee, S., Park, T., and Lee, K. Soft contrastive learning for time series. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=pAsQSWlDUf

    2024

  27. [35]

    and Jung, J

    Li, G. and Jung, J. J. Deep learning for anomaly detection in multivariate time series: Approaches, applications, and challenges. Information Fusion, 91: 0 93--102, 2023

    2023

  28. [36]

    itransformer: Inverted transformers are effective for time series forecasting

    Liu, Y., Hu, T., Zhang, H., Wu, H., Wang, S., Ma, L., and Long, M. itransformer: Inverted transformers are effective for time series forecasting. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=JePfAI8fah

    2024

  29. [37]

    Gcad: Anomaly detection in multivariate time series from the perspective of granger causality

    Liu, Z., Gao, M., and Jiao, P. Gcad: Anomaly detection in multivariate time series from the perspective of granger causality. Proceedings of the AAAI Conference on Artificial Intelligence, 39 0 (18): 0 19041--19049, Apr. 2025. doi:10.1609/aaai.v39i18.34096. URL https://ojs.aaa...

    2025 doi

  30. [38]

    and Hutter, F

    Loshchilov, I. and Hutter, F. SGDR: stochastic gradient descent with warm restarts. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings . OpenReview.net, 2017. URL https://openreview.net/forum?...

    2017

  31. [39]

    and Katarya, R

    Nandanwar, H. and Katarya, R. Deep learning enabled intrusion detection system for industrial iot environment. Expert Systems with Applications, 249: 0 123808, 2024

    2024

  32. [40]

    Ngu, H. C. V. and Lee, K. M. Cl-tad: A contrastive-learning-based method for time series anomaly detection. Applied Sciences, 13 0 (21): 0 11938, 2023

    2023

  33. [41]

    On the difficulty of training recurrent neural networks

    Pascanu, R., Mikolov, T., and Bengio, Y. On the difficulty of training recurrent neural networks. In Dasgupta, S. and McAllester, D. (eds.), Proceedings of the 30th International Conference on Machine Learning, volume 28 of Proceedings of Machine Learning Research, pp.\ 1310--...

    2013

  34. [42]

    A., and Li, W

    Qiu, H., Liu, Y., Subrahmanya, N. A., and Li, W. Granger causality for time-series anomaly detection. In 2012 IEEE 12th international conference on data mining, pp.\ 1074--1079. IEEE, 2012

    2012

  35. [43]

    Detecting and quantifying causal associations in large nonlinear time series datasets

    Runge, J., Nowack, P., Kretschmer, M., Flaxman, S., and Sejdinovic, D. Detecting and quantifying causal associations in large nonlinear time series datasets. Science advances, 5 0 (11): 0 eaau4996, 2019

    2019

  36. [44]

    \ SSD \ : A unified framework for self-supervised outlier detection

    Sehwag, V., Chiang, M., and Mittal, P. \ SSD \ : A unified framework for self-supervised outlier detection. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=v5gjXpmR8J

    2021

  37. [45]

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

    Su, Y., Zhao, Y., Niu, C., Liu, R., Sun, W., and Pei, D. Robust anomaly detection for multivariate time series through stochastic recurrent neural network. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pp.\ 2828--2837, 2019

    2019

  38. [46]

    Csi: Novelty detection via contrastive learning on distributionally shifted instances

    Tack, J., Mo, S., Jeong, J., and Shin, J. Csi: Novelty detection via contrastive learning on distributionally shifted instances. Advances in neural information processing systems, 33: 0 11839--11852, 2020

    2020

  39. [47]

    Tank, A., Covert, I., Foti, N., Shojaie, A., and Fox, E. B. Neural granger causality. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44 0 (8): 0 4267--4279, 2021

    2021

  40. [48]

    and Hinton, G

    Van der Maaten, L. and Hinton, G. Visualizing data using t-sne. Journal of machine learning research, 9 0 (11), 2008

    2008

  41. [49]

    A survey of deep anomaly detection in multivariate time series: Taxonomy, applications, and directions

    Wang, F., Jiang, Y., Zhang, R., Wei, A., Xie, J., and Pang, X. A survey of deep anomaly detection in multivariate time series: Taxonomy, applications, and directions. Sensors (Basel, Switzerland), 25 0 (1): 0 190, 2025

    2025

  42. [50]

    Timesnet: Temporal 2d-variation modeling for general time series analysis

    Wu, H., Hu, T., Liu, Y., Zhou, H., Wang, J., and Long, M. Timesnet: Temporal 2d-variation modeling for general time series analysis. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=ju_Uqw384Oq

    2023

  43. [51]

    Gcformer: Granger causality based attention mechanism for multivariate time series anomaly detection

    Xing, S., Niu, J., and Ren, T. Gcformer: Granger causality based attention mechanism for multivariate time series anomaly detection. In 2023 IEEE International Conference on Data Mining (ICDM), pp.\ 1433--1438. IEEE, 2023

    2023

  44. [52]

    Unsupervised anomaly detection via variational auto-encoder for seasonal kpis in web applications

    Xu, H., Chen, W., Zhao, N., Li, Z., Bu, J., Li, Z., Liu, Y., Zhao, Y., Pei, D., Feng, Y., et al. Unsupervised anomaly detection via variational auto-encoder for seasonal kpis in web applications. In Proceedings of the 2018 World Wide Web Conference, pp.\ 187--196, 2018

    2018

  45. [53]

    Calibrated one-class classification for unsupervised time series anomaly detection

    Xu, H., Wang, Y., Jian, S., Liao, Q., Wang, Y., and Pang, G. Calibrated one-class classification for unsupervised time series anomaly detection. IEEE Transactions on Knowledge and Data Engineering, 2024

    2024

  46. [54]

    Anomaly transformer: Time series anomaly detection with association discrepancy

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

    2022

  47. [55]

    Dcdetector: Dual attention contrastive representation learning for time series anomaly detection

    Yang, Y., Zhang, C., Zhou, T., Wen, Q., and Sun, L. Dcdetector: Dual attention contrastive representation learning for time series anomaly detection. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp.\ 3033--3045, 2023

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.