TS2TabPFN: Time Series Classification and Extrinsic Regression through Feature Extraction and a Tabular Foundation Model

Diego Furtado Silva; Gabriel da Costa Merlin

REVIEW 3 major objections 5 minor 23 references

TS2TabPFN: Time Series Classification and Extrinsic Regression through Feature Extraction and a Tabular Foundation Model

T0 review · 3 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read The paper claims that TS2TabPFN, combining explicit time-series feature extraction with the frozen tabular foundation model TabPFN 2.5, sets a new state of the art in time series extrinsic regression and matches HIVE-COTE 2.0 in…

desk verdict Sensible pipeline paper with a real novelty, but the TSER SOTA claim rests on comparison statistics that the paper never shows are computed on a common dataset subset, and duplicate win/rank numbers need explanation. read the letter →

arxiv 2608.04174 v1 pith:JWW6VWCN submitted 2026-08-04 cs.LG

Gabriel da Costa Merlin , Diego Furtado Silva This is my paper · ORCID

classification cs.LG

keywords timeseriesclassificationextrinsicregressionfeatureextractiontabularfoundationmodelin-contextlearningTabPFNtsfreshMultiROCKET

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

TS2TabPFN is a two-stage framework that turns time series into tables and lets a tabular foundation model do the prediction. The paper argues that this decoupling beats both feature-based and end-to-end approaches: its tsfresh-based variant achieves the best average rank in extrinsic regression, significantly ahead of DrCIF and FreshPRINCE, and its MultiROCKET variant matches HIVE-COTE 2.0 in classification accuracy while winning more datasets and running up to two orders of magnitude faster. If correct, this establishes that in-context learning on explicit features is a viable, cheaper route to time series state of the art.

What carries the argument

The load-bearing mechanism is the separation of representation from inference. A feature extractor ψ (tsfresh, catch22, or MultiROCKET) maps each time series to a vector; the featurized training set D and test queries X are concatenated into one table and passed in a single forward pass to TabPFN 2.5, a transformer pre-trained on synthetic data that approximates the Bayesian posterior predictive distribution. Because TabPFN's weights stay frozen, the pipeline is training-free and relies on in-context learning.

What would settle it

Run TS2TabPFN and its baselines on the 24 TSC and 8 TSER datasets that were excluded for exceeding 12 GB VRAM, then recompute average ranks; if the excluded datasets favor DrCIF or FreshPRINCE, the claimed state-of-the-art rank on the full archives would not reproduce.

Watch

Extended reading notes

Core claim

The central discovery is that a frozen tabular foundation model, TabPFN 2.5, becomes a highly effective time-series predictor once the raw series are reduced to a fixed-dimensional feature table. On the TSER benchmark, tsfresh-TabPFN has average rank 1.4364 and beats DrCIF and FreshPRINCE with p ≤ 1e-4 and a 51/0/4 win-tie-loss record. On TSC, MultiROCKET-TabPFN reaches 0.8627 mean accuracy versus HC2's 0.8674 (p = 0.6690), winning 69 head-to-head comparisons to HC2's 60, and dominating standard ROCKET with p ≤ 1e-4. These results are obtained with no per-dataset training or hyperparameter tuning.

Load-bearing premise

The reported comparisons assume that baseline accuracies and RMSEs published on the full UCR/UEA and TSML archives can be directly compared with TS2TabPFN scores computed only on the subset of datasets that fit in 12 GB of GPU memory; if the excluded datasets differ in difficulty, the claimed ranks and win records may not hold.

Editorial extensions

If this is right

tsfresh-TabPFN becomes the top-performing TSER model in the comparison, with a statistically significant edge over DrCIF and FreshPRINCE (p ≤ 1e-4, identical 51/0/4 records).
MultiROCKET-TabPFN matches HC2 accuracy (0.8627 vs 0.8674, p=0.6690) while winning more datasets, and beats ROCKET significantly.
End-to-end speedups reach 98.39× over HC2 in classification and 101.81× over DrCIF in regression; inference-only speedups reach 769.20×.
Total energy consumption is lower than the baselines despite higher instantaneous power draw, because runtime drops sharply.

Reading between the lines

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

The task-dependent best extractor (MultiROCKET for classification, tsfresh for regression) suggests the framework's performance is governed by how well the feature table matches TabPFN's prior, not by a single universal representation.
Because TabPFN's quadratic memory limits forced evaluation on 134/55 of the full archives, the headline ranks are conditional; re-running on the excluded datasets could shift them.
The same pipeline could be probed with other tabular foundation models or with latent features from time-series foundation models to test whether the gains come from the prior or from the feature extraction.

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.

Desk Editor's Note

read the letter

Colleague,

This is a straightforward empirical paper: take tsfresh, catch22, or MultiROCKET features, feed them into TabPFN 2.5, and report results on UCR/UEA and TSML. The new bit is applying a tabular foundation model as an in-context learning head over established time series feature extractors. That combination is not in the literature, and the scale of the evaluation (189 datasets, though a subset) is real work. The speedup and energy numbers are plausible and locally measured, at least for the small set of datasets with local baseline runs. The paper is also honest about the VRAM exclusions and the no-tuning protocol.

The soft spots are in the comparison statistics, and they are load-bearing for the headline claims. The paper evaluates TS2TabPFN on 134/55 datasets because of memory limits, but takes baseline accuracies and RMSE from the literature, which usually cover the full 158/63 archives. Nowhere does it say that the Wilcoxon tests, average ranks, and win-tie-loss records were recomputed on the common dataset intersection. Without that, paired signed-rank tests are not actually paired. That is not a minor detail; the TSER state-of-the-art claim rests on those p-values.

There are also two numbers that made me stop. In the TSER section, the internal CD diagram among three feature extractors gives tsfresh-TabPFN an average rank of 1.4364, and the external CD diagram among seven baselines gives the same 1.4364. The pairwise matrix reports identical 51/0/4 win-tie-loss records against both DrCIF and FreshPRINCE. These coincidences are possible but unlikely; they look like the sort of artifact you get from copy-pasting or from a script that overwrote values. The paper needs to show per-dataset tables to rule that out.

I am not saying the idea is wrong. The exclusion of 24 TSC and 8 TSER datasets could change ranks if those datasets differ in difficulty, and the duplicate statistics are a warning sign, not proof of fabrication. But as written, the central empirical claim is not verifiable from the text. The paper itself points to a GitHub repo, so this is fixable.

If this lands in a journal, I would send it to review rather than desk reject. The relevance is high for practitioners who want a cheap TSER alternative; the contribution is modest but real. The referee should ask for a clear statement of common-subset statistics and per-dataset results. If the numbers hold up, it is a useful contribution. If they do not, the paper becomes an efficiency study with a much weaker accuracy claim.

Bring it to reading group? Maybe, mainly to discuss how not to do benchmark comparisons.

Referee Report

3 major / 5 minor

Summary. The paper proposes TS2TabPFN, a two-stage framework for time series classification (TSC) and extrinsic regression (TSER). It first extracts features from raw series using one of three extractors (tsfresh, catch22, MultiROCKET) and then feeds the resulting tabular representation to the TabPFN 2.5 foundation model for in-context inference. The authors evaluate on 134 of 158 UCR/UEA datasets and 55 of 63 TSML regression datasets, constrained by 12 GB VRAM, and compare against literature-reported baselines. They report that MultiROCKET-TabPFN matches HC2 accuracy in TSC while being up to two orders of magnitude faster, and that tsfresh-TabPFN significantly outperforms DrCIF and FreshPRINCE in TSER (p <= 1e-4), with speedups up to 769x at inference.

Significance. The idea of decoupling temporal feature extraction from a tabular foundation model is simple, original in this combination, and potentially very useful. The paper follows accepted benchmarking protocols (fixed resamples/seeds, Wilcoxon with Holm correction), reports energy and runtime measurements, and provides a public repository. If the statistical claims are correct, TS2TabPFN would be a strong and computationally attractive alternative to state-of-the-art ensembles. However, the significance of the TSER result depends entirely on the correctness of the pairwise comparisons, which are currently not verifiable from the manuscript.

major comments (3)

[Sections 4.1 and 4.3] The central claim of significant TSER outperformance is not verifiable because the manuscript does not state the dataset intersection used for the paired Wilcoxon tests. TS2TabPFN results are reported on 55 of 63 TSML datasets (Section 4.1), while baseline results are taken from the literature (Section 4.3) and likely cover the full archive. If the 8 excluded datasets are included for baselines but missing for TS2TabPFN, the signed-rank tests, average ranks, and win/tie/loss records in Figures 5-7 are invalid. Please state explicitly that all statistics were computed on the common subset (and list the datasets) or recompute them.
[Section 5.2, Figures 5 and 6] The identical average rank 1.4364 for tsfresh-TabPFN in the internal 3-model comparison (Figure 5) and the external 7-model comparison (Figure 6) is highly implausible and suggests the statistic was not recomputed for Figure 6. Likewise, the identical 51/0/4 win-tie-loss records against DrCIF and FreshPRINCE in Figure 7, despite different mean RMSE differences, require explanation. Please provide per-dataset ranks and pairwise outcomes, and verify these numbers.
[Sections 4.1 and 6] The unqualified claim of 'state-of-the-art' performance is not supported if the evaluated subset is not representative of the full archives. With 24 TSC and 8 TSER datasets excluded for hardware reasons, the authors should either explicitly limit the claims to the evaluated subset or provide a sensitivity analysis showing that baseline rankings are similar on the excluded datasets.

minor comments (5)

[Section 5.2] The text mentions ResNet as one of the baselines, but ResNet is not listed in Section 4.3 nor shown in Figure 6; please reconcile.
[Section 5.2] The text says 'seven established TSER baselines' but only six are listed in Section 4.3; either correct the count or add the missing baseline.
[Figure 2] The CD diagram for internal TSC validation has no labels on the horizontal axis; please add a caption explaining the axis and the meaning of the rank numbers.
[Section 7] The conclusion calls the pipeline 'training-free', but tsfresh feature selection and the MultiROCKET fit are per-dataset procedures; please qualify this claim.
[Figure 4] The win/tie/loss convention (r>c / r=c / r<c) is not defined in the caption; please define how ties are determined.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: TS2TabPFN's claims are empirical benchmark comparisons against external baselines and standard feature extractors, with no derivation that reduces to its own inputs.

full rationale

The paper contains no mathematical derivation chain in which a claimed prediction is defined in terms of its own output. TS2TabPFN combines externally maintained feature extractors (tsfresh, catch22, MultiROCKET) with the externally pre-trained TabPFN 2.5, and its state-of-the-art claims are evaluated against literature-reported baselines on the UCR/UEA and TSML archives. There is no fitted parameter that is later renamed as a prediction, no uniqueness theorem imported from the authors' prior work, and no ansatz smuggled in via citation. The only self-citation (Guijo-Rubio et al. 2024, which includes co-author Silva) is used for background statements about FreshPRINCE and DrCIF being strong TSER baselines; it is not load-bearing for the reported outcome. The internal validation that selects the best feature extractor using the same benchmark test sets later used for the final comparison is a legitimate experimental-validity concern (model selection on test data), and the identical 51/0/4 records and identical average rank in Figures 5-7 warrant verification, but these are correctness or reporting risks, not circularity by construction. Accordingly, no circular step can be exhibited with a specific reduction, and the appropriate score is 0.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The central empirical claim rests on comparability of literature baselines, representativeness of the benchmarks, and uninformative dataset exclusion. The paper introduces no fitted numeric parameters and no new theoretical entities; its contribution is entirely empirical.

assumptions (5)

domain assumption The UCR/UEA and TSML benchmark archives are representative and their train/test splits are trustworthy.
Section 4.1 relies on these archives as the basis for all rank and significance claims.
ad hoc to paper Literature-reported baseline results are directly comparable to locally computed TS2TabPFN results despite different hardware, software versions, and dataset subsets.
Section 4.3 states baseline metrics were compiled from literature and not re-executed; this comparability is essential for the claimed significance.
ad hoc to paper Excluding datasets that exceed 12 GB VRAM does not bias the comparison.
Section 4.1 and Section 6 discuss memory-based exclusion of 24 TSC and 8 TSER datasets; the assumption that this does not affect rankings is implicit and load-bearing.
domain assumption TabPFN 2.5's pretrained prior sufficiently approximates the posterior predictive distribution for featurized time series data.
Section 2.3 describes TabPFN as approximating Bayesian inference; the paper relies on this for zero-shot predictive quality.
standard math Wilcoxon signed-rank tests with Holm correction are appropriate for paired benchmark comparisons.
Section 4.2 specifies the statistical protocol; standard assumptions of paired, non-parametric testing are used.

how reviews work

0 comments

Cite this review

Pith. "Pith review of TS2TabPFN: Time Series Classification and Extrinsic Regression through Feature Extraction and a Tabular Foundation Model." pith.science (2026). https://pith.science/paper/JWW6VWCN

@misc{pith2026260804174,
  author       = {Pith},
  title        = {Pith review of: TS2TabPFN: Time Series Classification and Extrinsic Regression through Feature Extraction and a Tabular Foundation Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JWW6VWCN}},
  note         = {Machine review of arXiv:2608.04174}
}

read the original abstract

Time series data are ubiquitous in practical applications, where classification (TSC) and extrinsic regression (TSER) have emerged as essential tasks for obtaining value from temporal sequences. While the literature has seen significant progress through feature-based and deep learning models, existing methods often focus either on the quality of feature extraction or on the intrinsic predictive power of complex architectures applied to raw data. This division creates a gap between the control offered by feature engineering and the automated performance of end-to-end models. This paper proposes TS2TabPFN, a framework that bridges this gap by integrating explicit feature extraction with TabPFN 2.5, a cutting-edge foundation model for tabular data, to leverage its predictive capabilities. Our extensive experimental evaluation demonstrates that TS2TabPFN significantly outperforms state-of-the-art models in TSER tasks with statistical significance, providing a robust and efficient alternative for TSC and surpassing most of the currently best-performing algorithms. These results suggest that combining foundation models with structured features overcomes single-paradigm limitations, establishing a new time series state-of-the-art.

Figures

Figures reproduced from arXiv: 2608.04174 by the authors.

**Figure 1.** TS2TabPFN framework pipeline. Raw dataset S is processed by a modular function ψ, transforming sequences into fixed-dimensional vectors (d for training, x for test). These features form a joint table (D ∪ X) by concatenating labeled context D and unlabeled queries X. Finally, the pre-trained TabPFN model processes this input to generate predictions for Classification or Extrinsic Regression tasks [PITH_FULL_IMAGE:f… view at source ↗

**Figure 2.** The MultiROCKET-based feature extraction yielded the best perfor [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗

**Figure 3.** They indicate that our approach achieves performance parity with [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Multiple comparison matrix showing mean differences, win/tie/loss records and Wilcoxon p-values. Bold values indicate statistical significance (p < 0.05). demonstrates clear statistical dominance (p ≤ 1e − 04), securing a 99/3/32 wintie-loss record. While previous exp…

**Figure 5.** Figure 5: Critical Difference diagram for TSER (Internal Validation) based on pairwise Wilcoxon signed-rank tests of regression RMSE. Closer to 1, the better [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]

**Figure 6.** Figure 6: Critical Difference for TSER (Comparative Analysis) based on pairwise Wilcoxon signed-rank tests of regression RMSE. Horizontal black bar indicates groups of models with no statistically significant difference. Closer to 1, the better. The pairwise comparison matrix, i…

**Figure 7.** Figure 7: Multiple comparison matrix showing mean differences, win/tie/loss records and Wilcoxon p-values. Bold values indicate statistical significance (p < 0.05). ditioned on extracted features. This is most pronounced in the BIDMC32SpO2 dataset, where DrCIF required over 48 h…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 17 canonical work pages

[1]

Bagnall, A., Dau, H.A., Lines, J., Flynn, M., Large, J., Bostrom, A., Southam, P., Keogh, E.: The uea multivariate time series classification archive, 2018 (2018)

work page 2018
[2]

Bagnall, A., Flynn, M., Large, J., Line, J., Bostrom, A., Cawley, G.: Is rotation forest the best classifier for problems with continuous features? arXiv preprint arXiv:1809.06705 (2018)

arXiv 2018
[3]

Neurocom- puting307, 72–77 (2018)

Christ, M., Braun, N., Neuffer, J., Kempa-Liehr, A.W.: Time series feature extrac- tion on basis of scalable hypothesis tests (tsfresh–a python package). Neurocom- puting307, 72–77 (2018)

work page 2018
[4]

IEEE/CAA Jour- nal of Automatica Sinica6(6), 1293–1305 (2019)

Dau, H.A., Bagnall, A., Kamgar, K., Yeh, C.C.M., Zhu, Y., Gharghabi, S., Ratanamahatana, C.A., Keogh, E.: The ucr time series archive. IEEE/CAA Jour- nal of Automatica Sinica6(6), 1293–1305 (2019)

work page 2019
[5]

Dau, H.A., Keogh, E., Kamgar, K., Yeh, C.C.M., Zhu, Y., Gharghabi, S., Ratanamahatana, C.A., Yanping, Hu, B., Begum, N., Bagnall, A., Mueen, b., Batista, G., Hexagon-ML: The UCR time series classification archive (October 2018)

work page 2018
[6]

Data Mining and Knowledge Discovery34(5), 1454–1495 (2020) 18 G

Dempster, A., Petitjean, F., Webb, G.I.: Rocket: exceptionally fast and accurate time series classification using random convolutional kernels. Data Mining and Knowledge Discovery34(5), 1454–1495 (2020) 18 G. C. Merlin and D. F. Silva

work page 2020
[7]

Journal of Machine learning research7(Jan), 1–30 (2006)

Demšar, J.: Statistical comparisons of classifiers over multiple data sets. Journal of Machine learning research7(Jan), 1–30 (2006)

2006
[8]

arXiv preprint arXiv:2502.15637 (2025)

Feofanov, V., Wen, S., Alonso, M., Ilbert, R., Guo, H., Tiomoko, M., Pan, L., Zhang, J., Redko, I.: Mantis: Lightweight calibrated foundation model for user- friendly time series classification. arXiv preprint arXiv:2502.15637 (2025)

arXiv 2025

Show all 23 references

[9]

arXiv preprint arXiv:2511.08667 (2025)

Grinsztajn, L., Flöge, K., Key, O., Birkel, F., Jund, P., Roof, B., Jäger, B., Safaric, D., Alessi, S., Hayler, A., et al.: Tabpfn-2.5: Advancing the state of the art in tabular foundation models. arXiv preprint arXiv:2511.08667 (2025)

2025 arXiv
[10]

Data Mining and Knowledge Discovery38(4), 2141–2185 (2024)

Guijo-Rubio, D., Middlehurst, M., Arcencio, G., Silva, D.F., Bagnall, A.: Unsu- pervised feature based algorithms for time series extrinsic regression. Data Mining and Knowledge Discovery38(4), 2141–2185 (2024)

2024
[11]

In: The Eleventh International Conference on Learning Representations (2023)

Hollmann, N., Müller, S., Eggensperger, K., Hutter, F.: Tabpfn: A transformer that solves small tabular classification problems in a second. In: The Eleventh International Conference on Learning Representations (2023)

2023
[12]

Data mining and knowledge discovery34(6), 1936–1962 (2020)

Ismail Fawaz, H., Lucas, B., Forestier, G., Pelletier, C., Schmidt, D.F., Weber, J., Webb, G.I., Idoumghar, L., Muller, P.A., Petitjean, F.: Inceptiontime: Finding alexnet for time series classification. Data mining and knowledge discovery34(6), 1936–1962 (2020)

2020
[13]

arXiv preprint arXiv:1910.09700 (2019)

Lacoste, A., Luccioni, A., Schmidt, V., Dandres, T.: Quantifying the carbon emis- sions of machine learning. arXiv preprint arXiv:1910.09700 (2019)

2019 arXiv
[14]

Lubba, C.H., Sethi, S.S., Knaute, P., Schultz, S.R., Fulcher, B.D., Jones, N.S.: catch22:Canonicaltime-seriescharacteristics:Selectedthroughhighlycomparative time-seriesanalysis.Dataminingandknowledgediscovery33(6),1821–1852(2019)

2019
[15]

In: International Conference on Pattern Recognition and Artificial Intelligence

Middlehurst, M., Bagnall, A.: The freshprince: A simple transformation based pipeline time series classifier. In: International Conference on Pattern Recognition and Artificial Intelligence. pp. 150–161. Springer (2022)

2022
[16]

Journal of Machine Learning Research25(289), 1–10 (2024)

Middlehurst, M., Ismail-Fawaz, A., Guillaume, A., Holder, C., Guijo-Rubio, D., Bulatova, G., Tsaprounis, L., Mentel, L., Walter, M., Schäfer, P., Bagnall, A.: aeon: a python toolkit for learning from time series. Journal of Machine Learning Research25(289), 1–10 (2024)

2024
[17]

Machine Learning 110(11), 3211–3243 (2021)

Middlehurst, M., Large, J., Flynn, M., Lines, J., Bostrom, A., Bagnall, A.: Hive- cote 2.0: a new meta ensemble for time series classification. Machine Learning 110(11), 3211–3243 (2021)

2021
[18]

middlehurst et al

Middlehurst, M., Schäfer, P., Bagnall, A.: Bake off redux: a review and experimen- tal evaluation of recent time series classification algorithms: M. middlehurst et al. Data Mining and Knowledge Discovery38(4), 1958–2031 (2024)

2024
[19]

ACM Computing Surveys56(9), 1–45 (2024)

Mohammadi Foumani, N., Miller, L., Tan, C.W., Webb, G.I., Forestier, G., Salehi, M.: Deep learning for time series classification and extrinsic regression: A current survey. ACM Computing Surveys56(9), 1–45 (2024)

2024
[20]

Data Mining and Knowledge Discovery35(3), 1032–1060 (2021)

Tan,C.W.,Bergmeir,C.,Petitjean,F.,Webb,G.I.:Timeseriesextrinsicregression: Predicting numeric values from time series data. Data Mining and Knowledge Discovery35(3), 1032–1060 (2021)

2021
[21]

Data Mining and Knowledge Discovery36(5), 1623–1646 (2022)

Tan, C.W., Dempster, A., Bergmeir, C., Webb, G.I.: Multirocket: multiple pooling operators and transformations for fast and effective time series classification: Cw tan. Data Mining and Knowledge Discovery36(5), 1623–1646 (2022)

2022
[22]

https://doi.org/10.5281/zenodo.11236865

TSML Archive Maintainers: Tsmlƒextended time series extrinsic regression archive 2024 (May 2024). https://doi.org/10.5281/zenodo.11236865

2024 doi
[23]

Wang, Z., Yan, W., Oates, T.: Time series classification from scratch with deep neuralnetworks:Astrongbaseline.In:2017Internationaljointconferenceonneural networks (IJCNN). pp. 1578–1585. IEEE (2017)

2017

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

Bagnall, A., Dau, H.A., Lines, J., Flynn, M., Large, J., Bostrom, A., Southam, P., Keogh, E.: The uea multivariate time series classification archive, 2018 (2018)

work page 2018

[2] [2]

Bagnall, A., Flynn, M., Large, J., Line, J., Bostrom, A., Cawley, G.: Is rotation forest the best classifier for problems with continuous features? arXiv preprint arXiv:1809.06705 (2018)

arXiv 2018

[3] [3]

Neurocom- puting307, 72–77 (2018)

Christ, M., Braun, N., Neuffer, J., Kempa-Liehr, A.W.: Time series feature extrac- tion on basis of scalable hypothesis tests (tsfresh–a python package). Neurocom- puting307, 72–77 (2018)

work page 2018

[4] [4]

IEEE/CAA Jour- nal of Automatica Sinica6(6), 1293–1305 (2019)

Dau, H.A., Bagnall, A., Kamgar, K., Yeh, C.C.M., Zhu, Y., Gharghabi, S., Ratanamahatana, C.A., Keogh, E.: The ucr time series archive. IEEE/CAA Jour- nal of Automatica Sinica6(6), 1293–1305 (2019)

work page 2019

[5] [5]

Dau, H.A., Keogh, E., Kamgar, K., Yeh, C.C.M., Zhu, Y., Gharghabi, S., Ratanamahatana, C.A., Yanping, Hu, B., Begum, N., Bagnall, A., Mueen, b., Batista, G., Hexagon-ML: The UCR time series classification archive (October 2018)

work page 2018

[6] [6]

Data Mining and Knowledge Discovery34(5), 1454–1495 (2020) 18 G

Dempster, A., Petitjean, F., Webb, G.I.: Rocket: exceptionally fast and accurate time series classification using random convolutional kernels. Data Mining and Knowledge Discovery34(5), 1454–1495 (2020) 18 G. C. Merlin and D. F. Silva

work page 2020

[7] [7]

Journal of Machine learning research7(Jan), 1–30 (2006)

Demšar, J.: Statistical comparisons of classifiers over multiple data sets. Journal of Machine learning research7(Jan), 1–30 (2006)

2006

[8] [8]

arXiv preprint arXiv:2502.15637 (2025)

Feofanov, V., Wen, S., Alonso, M., Ilbert, R., Guo, H., Tiomoko, M., Pan, L., Zhang, J., Redko, I.: Mantis: Lightweight calibrated foundation model for user- friendly time series classification. arXiv preprint arXiv:2502.15637 (2025)

arXiv 2025