pith. sign in

arxiv: 2606.13277 · v1 · pith:TMX52YNEnew · submitted 2026-06-11 · 📊 stat.ML · cs.LG

ProtoX-AD: Self-Explainable Time Series Anomaly Detection and Characterization

Aitor S\'anchez-Ferrera , Elisabeth Wetzer , Kristoffer Wickstr{\o}m , Michael Kampffmeyer , Robert Jenssen This is my paper

Pith reviewed 2026-06-27 05:43 UTC · model grok-4.3

classification 📊 stat.ML cs.LG
keywords time series anomaly detectionprototype learningself-supervised learningexplainable anomaly detectionlatent representationsanomaly characterization
0
0 comments X

The pith

ProtoX-AD detects time series anomalies at black-box levels while explaining them via learned prototypes.

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

The paper presents ProtoX-AD as a prototype-based extension to self-supervised time series anomaly detection. It trains on transformed normal data to build a classifier but adds a set of interpretable prototypes in the latent space that link classification errors to specific anomaly profiles. This setup is intended to keep detection accuracy high while generating explanations that are more consistent and semantically meaningful than those from prior explainable methods. A reader would care if they need to understand the type of anomaly, not just its presence, in monitoring applications.

Core claim

ProtoX-AD learns transformation-aware latent representations alongside interpretable prototypes, enabling both accurate anomaly detection and the identification of distinct anomalous profiles through prototype-based explanations. Experimental results on synthetic and real-world datasets demonstrate that ProtoX-AD achieves detection performance comparable to its black-box counterparts while offering more consistent and semantically meaningful explanations than existing explainable baselines.

What carries the argument

Transformation-aware prototypes in the latent space that match anomalous samples to distinct profiles and drive the explanations.

If this is right

  • ProtoX-AD maintains detection performance comparable to black-box self-supervised methods on both synthetic and real-world data.
  • The method produces more consistent and semantically meaningful explanations than existing explainable baselines.
  • The framework supports systematic study of how transformation design choices affect both detection accuracy and explanation quality.

Where Pith is reading between the lines

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

  • The prototype matches could be used to group similar anomalies automatically for prioritized response in operational systems.
  • The same prototype mechanism might transfer to other self-supervised tasks where characterizing deviations matters as much as detecting them.

Load-bearing premise

The learned prototypes will reliably produce semantically meaningful and consistent explanations for anomalies.

What would settle it

Quantitative or expert-rated comparisons on real-world datasets showing that ProtoX-AD explanations are no more consistent or meaningful than those from existing explainable baselines would falsify the central claim.

Figures

Figures reproduced from arXiv: 2606.13277 by Aitor S\'anchez-Ferrera, Elisabeth Wetzer, Kristoffer Wickstr{\o}m, Michael Kampffmeyer, Robert Jenssen.

Figure 1
Figure 1. Figure 1: Pipeline of ProtoX-AD. The input time series [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Representative normal (blue) and anomalous (red) time series from the UMD dataset. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Representative normal (blue) and anomalous (red) yearly temperature anomaly time series from [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Representative normal (blue) and anomalous (red) water flow sequences from the Yorkshire Water [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Learned prototypes for the UMD dataset. Columns represent transformation-induced classes. [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Prototype-based explanations for representative test samples from the UMD dataset. Columns [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Alignment between latent assignment and input-space similarity in the UMD dataset. For ProtoX [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
read the original abstract

Recent advances in time series anomaly detection (TSAD) have highlighted the effectiveness of self-supervised classification-based approaches. These methods apply transformations to normal training samples, training a classifier to recognize transformation-specific patterns that help identify anomalies through increased classification errors. Despite their strong performance, a significant challenge is their lack of explainability, as they provide limited insight into the characteristics of flagged anomalies. To address this limitation, we propose ProtoX-AD, a prototype-based self-explainable framework for self-supervised TSAD. ProtoX-AD learns transformation-aware latent representations alongside interpretable prototypes, enabling both accurate anomaly detection and the identification of distinct anomalous profiles through prototype-based explanations. Additionally, it allows for systematic analysis of how transformation design impacts detection performance and explainability. Experimental results on synthetic and real-world datasets demonstrate that ProtoX-AD achieves detection performance comparable to its black-box counterparts while offering more consistent and semantically meaningful explanations than existing explainable baselines. Our code is publicly available at https://github.com/Aitorzan3/ProtoX-AD.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 2 minor

Summary. The paper proposes ProtoX-AD, a prototype-based self-explainable framework extending self-supervised classification-based time series anomaly detection (TSAD). It learns transformation-aware latent representations and interpretable prototypes to enable both anomaly detection via classification errors and characterization of anomalous profiles. The method also supports analysis of transformation design effects. Experiments on synthetic and real-world datasets claim detection performance comparable to black-box counterparts and superior consistency/semantics in explanations versus explainable baselines; code is released publicly.

Significance. If the empirical claims hold under the reported evaluation protocol, the work addresses a clear gap in TSAD by adding prototype-based explanations without sacrificing detection accuracy. Public code enables direct verification of prototype construction and any quantitative explainability metrics, which strengthens the contribution.

major comments (2)
  1. [§4] §4 (Method), prototype assignment step: the claim that prototypes yield 'semantically meaningful' explanations requires explicit validation that the latent space preserves transformation semantics; without a quantitative metric (e.g., prototype purity or human evaluation scores) tied to the transformation set, the superiority claim over baselines rests on qualitative inspection alone.
  2. [§5] §5 (Experiments), Table 2/3 comparison rows: the reported F1 scores for ProtoX-AD are within 1-2% of black-box baselines, but the explainability metrics (consistency, semantic alignment) lack statistical significance testing across the 22 runs; this weakens the 'more consistent' claim if variance overlaps with baselines.
minor comments (2)
  1. [§3] Notation: the transformation set T is introduced in §3 but the exact cardinality and composition (e.g., which augmentations) are only listed in the appendix; move the list to the main text for readability.
  2. [Figure 3] Figure 3 caption: the prototype visualization labels are not defined in the caption; add a legend or reference to the transformation indices used.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thoughtful review and constructive feedback. We address the major comments point by point below, proposing revisions where appropriate to enhance the clarity and rigor of our claims.

read point-by-point responses

  1. Referee: [§4] §4 (Method), prototype assignment step: the claim that prototypes yield 'semantically meaningful' explanations requires explicit validation that the latent space preserves transformation semantics; without a quantitative metric (e.g., prototype purity or human evaluation scores) tied to the transformation set, the superiority claim over baselines rests on qualitative inspection alone.

    Authors: We agree that a quantitative metric would strengthen the claim regarding semantically meaningful explanations. In the revised version, we will add a prototype purity metric defined as the fraction of samples assigned to a prototype that share the same transformation label in the latent space. This metric will be reported for ProtoX-AD and compared to baselines to provide quantitative evidence of semantic preservation. Additionally, we will include a brief analysis showing the correlation between prototype assignments and transformation types. revision: yes

  2. Referee: [§5] §5 (Experiments), Table 2/3 comparison rows: the reported F1 scores for ProtoX-AD are within 1-2% of black-box baselines, but the explainability metrics (consistency, semantic alignment) lack statistical significance testing across the 22 runs; this weakens the 'more consistent' claim if variance overlaps with baselines.

    Authors: We acknowledge the importance of statistical testing for the explainability metrics. We will revise the experimental section to include statistical significance tests (e.g., paired t-tests) on the consistency and semantic alignment scores across the 22 independent runs. Updated tables will report mean values with standard deviations and p-values to substantiate the 'more consistent' claim. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The abstract and description present ProtoX-AD as an extension of existing self-supervised classification-based TSAD methods, adding prototype learning for explanations. No equations, fitted parameters renamed as predictions, or self-citation load-bearing steps are visible in the provided text. The central claims rest on experimental results and prototype interpretability rather than reducing by construction to inputs or prior self-citations. This is the normal case of a self-contained proposal with public code for verification.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review prevents identification of specific free parameters, axioms, or invented entities; none are extractable from the provided text.

pith-pipeline@v0.9.1-grok · 5728 in / 915 out tokens · 23967 ms · 2026-06-27T05:43:31.627664+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

35 extracted references · 3 canonical work pages

  1. [1]

    Carreño, I

    A. Carreño, I. Inza, J. A. Lozano, Analyzing rare event, anomaly, novelty and outlier detection terms under the supervised classification framework, Artificial Intelligence Review 53 (2020) 3575–3594

    2020

  2. [2]

    Hilal, S

    W. Hilal, S. A. Gadsden, J. Yawney, Financial fraud: A review of anomaly de- tection techniques and recent advances, Expert Systems with Applications 193 (2022) 116429

    2022

  3. [3]

    A. A. Cook, G. Misirli, Z. Fan, Anomaly detection for IoT time-series data: A survey, IEEE Internet of Things Journal 7 (7) (2019) 6481–6494

    2019

  4. [4]

    X. Yang, X. Qi, X. Zhou, Deep learning technologies for time series anomaly detection in healthcare: A review, IEEE Access 11 (2023) 117788–117799

    2023

  5. [5]

    Blázquez-García, A

    A. Blázquez-García, A. Conde, U. Mori, J. A. Lozano, A review on out- lier/anomaly detection in time series data, ACM Computing Surveys 54 (3) (2021) 1–33. 20

    2021

  6. [6]

    Zamanzadeh Darban, G

    Z. Zamanzadeh Darban, G. I. Webb, S. Pan, C. Aggarwal, M. Salehi, Deep learn- ing for time series anomaly detection: A survey, ACM Computing Surveys 57 (1) (2024) 1–42

    2024

  7. [7]

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

    2021

  8. [8]

    X. Liu, F. Zhang, Z. Hou, L. Mian, Z. Wang, J. Zhang, J. Tang, Self-supervised learning: Generative or contrastive, IEEE Transactions on Knowledge and Data Engineering 35 (1) (2021) 857–876

    2021

  9. [9]

    Sánchez-Ferrera, B

    A. Sánchez-Ferrera, B. Calvo, J. A. Lozano, A review on self-supervised learning in time series anomaly detection: Recent advances and open challenges, ACM Computing Surveys 58 (5) (2025) 1–35

    2025

  10. [10]

    Hojjati, T

    H. Hojjati, T. K. K. Ho, N. Armanfard, Self-supervised anomaly detection in computer vision and beyond: A survey and outlook, Neural Networks 172 (2024) 106106

    2024

  11. [11]

    J. Yoo, T. Zhao, L. Akoglu, Data augmentation is a hyperparameter: Cherry- picked self-supervision for unsupervised anomaly detection is creating the illu- sion of success, Transactions on Machine Learning Research (2023)

    2023

  12. [12]

    D. Lee, S. Malacarne, E. Aune, Explainable time series anomaly detection using masked latent generative modeling, Pattern Recognition 156 (2024) 110826

    2024

  13. [13]

    Foorthuis, On the nature and types of anomalies: A review of deviations in data, International Journal of Data Science and Analytics 12 (4) (2021) 297–331

    R. Foorthuis, On the nature and types of anomalies: A review of deviations in data, International Journal of Data Science and Analytics 12 (4) (2021) 297–331

    2021

  14. [14]

    X. Bai, X. Wang, X. Liu, Q. Liu, J. Song, N. Sebe, B. Kim, Explainable deep learning for efficient and robust pattern recognition: A survey of recent develop- ments, Pattern Recognition 120 (2021) 108102

    2021

  15. [15]

    C. Chen, O. Li, D. Tao, A. Barnett, C. Rudin, J. K. Su, This looks like that: Deep learning for interpretable image recognition, Advances in Neural Informa- tion Processing Systems 32 (2019)

    2019

  16. [16]

    Gautam, M

    S. Gautam, M. M.-C. Höhne, S. Hansen, R. Jenssen, M. Kampffmeyer, This looks more like that: Enhancing self-explaining models by prototypical relevance prop- agation, Pattern Recognition 136 (2023) 109172

    2023

  17. [17]

    J. Gui, T. Chen, J. Zhang, Q. Cao, Z. Sun, H. Luo, D. Tao, A survey on self- supervised learning: Algorithms, applications, and future trends, IEEE Transac- tions on Pattern Analysis and Machine Intelligence 46 (12) (2024) 9052–9071

    2024

  18. [18]

    Blázquez-García, A

    A. Blázquez-García, A. Conde, U. Mori, J. A. Lozano, Water leak detection using self-supervised time series classification, Information Sciences 574 (2021) 528– 541. 21

    2021

  19. [19]

    Zheng, Z

    Y . Zheng, Z. Liu, R. Mo, Z. Chen, W.-s. Zheng, R. Wang, Task-oriented self- supervised learning for anomaly detection in electroencephalography, in: Inter- national Conference on Medical Image Computing and Computer-Assisted Inter- vention, Springer, 2022, pp. 193–203

    2022

  20. [20]

    J. Xu, Y . Zheng, Y . Mao, R. Wang, W.-S. Zheng, Anomaly detection on elec- troencephalography with self-supervised learning, in: 2020 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), IEEE, 2020, pp. 363– 368

    2020

  21. [21]

    C. Qiu, T. Pfrommer, M. Kloft, S. Mandt, M. Rudolph, Neural transformation learning for deep anomaly detection beyond images, in: International Conference on Machine Learning, PMLR, 2021, pp. 8703–8714

    2021

  22. [22]

    Sánchez-Ferrera, U

    A. Sánchez-Ferrera, U. Mori, B. Calvo, J. A. Lozano, NeuCoReClass AD: Redefining self-supervised time series anomaly detection, arXiv preprint arXiv:2508.00909 (2025)

    work page arXiv 2025

  23. [23]

    H. Hu, X. Wang, Y . Zhang, Q. Chen, Q. Guan, A comprehensive survey on con- trastive learning, Neurocomputing 610 (2024) 128645

    2024

  24. [24]

    Gautam, A

    S. Gautam, A. Boubekki, S. Hansen, S. Salahuddin, R. Jenssen, M. Höhne, M. Kampffmeyer, ProtoV AE: A trustworthy self-explainable prototypical vari- ational model, Advances in Neural Information Processing Systems 35 (2022) 17940–17952

    2022

  25. [25]

    B. Li, C. Jentsch, E. Müller, Prototypes as explanation for time series anomaly detection, arXiv preprint arXiv:2307.01601 (2023)

    work page arXiv 2023

  26. [26]

    Gautam, A

    S. Gautam, A. Boubekki, M. M. Höhne, M. Kampffmeyer, Prototypical self- explainable models without re-training, Transactions on Machine Learning Re- search (2024)

    2024

  27. [27]

    H. A. Dau, A. Bagnall, K. Kamgar, C.-C. M. Yeh, Y . Zhu, S. Gharghabi, C. A. Ratanamahatana, E. Keogh, The UCR time series archive, IEEE/CAA Journal of Automatica Sinica 6 (6) (2019) 1293–1305

    2019

  28. [28]

    Lenssen, G

    N. Lenssen, G. A. Schmidt, M. Hendrickson, P. Jacobs, M. J. Menne, R. Ruedy, A NASA GISTEMPv4 observational uncertainty ensemble, Journal of Geophysical Research: Atmospheres 129 (17) (2024) e2023JD040179

    2024

  29. [29]

    NOAA National Centers for Environmental Information, Climate at a glance: Global time series,https://www.ncei.noaa.gov/access/monitoring/ climate-at-a-glance/global/time-series, published May 2026, retrieved on May 18 2026 (2026)

    2026

  30. [30]

    F. T. Liu, K. M. Ting, Z.-H. Zhou, Isolation forest, in: 2008 eighth IEEE Interna- tional Conference on Data Mining, IEEE, 2008, pp. 413–422. 22 Table A.7: Stochastic parameters and sampling ranges used by the manual transformation modules for each dataset. All parameters are resampled at each forward pass of the transformation module. Transformation Stoc...

    2008

  31. [31]

    Schölkopf, R

    B. Schölkopf, R. C. Williamson, A. Smola, J. Shawe-Taylor, J. Platt, Support vector method for novelty detection, Advances in neural information processing systems 12 (1999)

    1999

  32. [32]

    M. M. Breunig, H.-P. Kriegel, R. T. Ng, J. Sander, LOF: Identifying density- based local outliers, in: Proceedings of the 2000 ACM SIGMOD international conference on Management of data, 2000, pp. 93–104

    2000

  33. [33]

    Hoffmann, C

    A. Hoffmann, C. Fanconi, R. Rade, J. Kohler, This looks like that... does it? shortcomings of latent space prototype interpretability in deep networks, arXiv preprint arXiv:2105.02968 (2021)

    work page arXiv 2021

  34. [34]

    Pedregosa, G

    F. Pedregosa, G. Varoquaux, A. Gramfort, V . Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V . Dubourg, et al., Scikit-learn: Machine learning in python, Journal of Machine Learning Research 12 (2011) 2825–2830

    2011

  35. [35]

    Z. Yue, Y . Wang, J. Duan, T. Yang, C. Huang, Y . Tong, B. Xu, TS2Vec: Towards universal representation of time series, in: Proceedings of the AAAI Conference on Artificial Intelligence, V ol. 36, 2022, pp. 8980–8987. Appendix A. Manual Transformations per dataset Following prior work [11], we design dataset-specific transformations that generate augmente...

    2022