Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

EXCON: Extreme Instance-based Contrastive Representation Learning of Severely Imbalanced Multivariate Time Series for Solar Flare Prediction

T0 review · 3 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A single extreme training instance per class, selected by complete linkage in catch22 feature space and used as an LSTM reconstruction target, lets a downstream logistic-regression classifier predict rare major solar flares better than…

desk verdict A useful but incomplete recipe for solar flare representation learning; the headline TSS gain is plausible but the paper never shows that the embedding, rather than the catch22 features, is doing the work. read the letter →

arxiv 2411.11249 v1 pith:3A7HBLCT submitted 2024-11-18 cs.LG cs.AI

Onur Vural , Shah Muhammad Hamdi , Soukaina Filali Boubrahimi This is my paper · ORCID
classification cs.LGcs.AI
keywords solarflarepredictioncontrastiverepresentationlearningmultivariatetimeseriesclassificationclassimbalanceextremeinstanceselectionLSTMembeddingcatch22featuresSWAN-SFdataset
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

EXCON claims that a severe class imbalance in time series classification can be countered by choosing one extreme instance per class and training an LSTM embedding to collapse every instance of that class onto it. The extreme is picked in a catch22 feature space as the instance with maximal complete-linkage distance to all other classes, and the embedding is trained with a mean squared error loss against that prototype. On the SWAN-SF solar flare benchmark, using successive segments as train-test pairs, the learned embeddings give a mean TSS of 0.7098 and ROC AUC of 0.8549, outperforming the LPVV baseline by 7.2% in TSS and 3.6% in ROC AUC. The same recipe posts the best score on four of the eight UEA archive datasets tested, covering binary and multiclass, univariate and multivariate problems. The paper's case is that a single point of contrast per class suffices to reduce intra-class variation and inter-class confusion without negative sampling or augmentation.

What carries the argument

The load-bearing mechanism is the extreme reconstruction loss of Equation 2, which turns representation learning into supervised regression toward a single contrasting point per class. That point, the extreme instance, is fixed in advance by Equation 1: for each class $C_c$, it is the training instance whose multi-catch22 vector $\mathbf{V}^{(m)} \in \mathbb{R}^{22N}$ is farthest from every instance of the other classes under complete linkage. The loss $\mathcal{L}_{ER}$ averages, over all instances of a class, the squared error between the LSTM's projected embedding $e^{(m)}_{C_c}$ and the class extreme $E_{C_c}$. Pulling all embeddings of a class to the same remote point simultaneously shrinks intra-class spread and enlarges the gap between classes, which is what the downstream logistic regression then exploits.

What would settle it

Train the same LSTM-plus-logistic-regression pipeline on SWAN-SF segments but replace the extreme reconstruction loss with ordinary cross-entropy or a triplet loss; if TSS does not drop materially, the extreme target is not the deciding factor. Alternatively, compute the catch22 extreme for each of S1-S4 and measure the distance between consecutive extremes; if the prototype moves more than the typical within-class scatter, the regression target is nonstationary and the reported transfer is likely an artifact of particular segments.

Watch

Extended reading notes

Core claim

The central claim is that a supervised contrastive target -- one extreme instance per class rather than pairwise positive-negative contrasts -- is enough to learn discriminative embeddings for severely imbalanced time series. The paper shows that after compressing every multivariate instance into a 22N-dimensional catch22 vector, selecting the class extreme with complete linkage and then regressing LSTM final hidden states onto it via the extreme reconstruction loss produces a representation space in which major flares separate from quiet-Sun instances. This is evidenced by the mean results over three consecutive SWAN-SF segments (S1-S2, S2-S3, S3-S4), where EXCON reaches 0.7306 accuracy, 0.7098 TSS, and 0.8549 ROC AUC, and by competitive results on the UEA archive. The paper also traces a single X-class instance, the Valentine's Day Flare, through the pipeline to show it is correctly classified as a major flare.

Load-bearing premise

The method assumes that the single extreme instance chosen from the training segment is a stable, representative target for the embeddings of the held-out future segment, so if that extreme is a nonstationary outlier, the regression target is misspecified.

Editorial extensions

If this is right

  • On the SWAN-SF benchmark, the learned embeddings beat the latest-timestamp vector baseline (LPVV) by 5.1% in accuracy, 7.2% in TSS, and 3.6% in ROC AUC, indicating that temporal dynamics matter more than a single snapshot for flare prediction.
  • The method needs no negative-pair sampling or augmentation, so it applies to rare-event time series problems where such contrastive tricks are hard to define.
  • The same four-stage pipeline works across UEA datasets with binary and multiclass labels, meaning the extreme prototype idea is not specific to solar physics.
  • Because the loss operates per class, the framework extends to multiclass prediction directly, potentially allowing M-versus-X discrimination rather than a single major-flare category.
  • The reported mean scores come from three successive train-test segment pairs, suggesting the method transfers across time rather than fitting one solar cycle window.

Reading between the lines

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

  • The paper does not ablate the extreme reconstruction loss against the same LSTM trained without it; a direct comparison would reveal how much of the gain comes from the extreme target rather than from the sequence model itself.
  • Since extremes are computed on training segments and applied to future segments, a stability check across S1-S5 would tell whether the chosen prototype is a moving target; if it shifts substantially, a time-averaged or ensemble extreme would be a safer anchor.
  • The complete-linkage selection yields a single prototype per class, but classes with internal subclusters might benefit from multiple extremes with a soft assignment, a variant the paper does not explore.
  • A quantitative separation metric on held-out embeddings (for example, nearest-class mean distance) would complement the t-SNE plot and make the claimed inter-class separation directly measurable.
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

3 major / 6 minor

Summary. The paper presents EXCON, a four-stage contrastive representation learning framework for severely imbalanced multivariate time series classification, applied mainly to solar flare prediction on the SWAN-SF benchmark. The pipeline first extracts multi-catch22 feature vectors from each MVTS instance (Section IV.A), then selects a single 'extreme' instance per class via complete linkage in that feature space (Section IV.B, Eq. 1), trains an LSTM plus fully connected encoder to map each raw instance to its class extreme under a mean-squared-error loss (Section IV.C, Eq. 2), and finally trains a logistic regression (or other) classifier on the learned embeddings. Experiments on three successive SWAN-SF segment pairs report a mean TSS of 0.7098 and ROC AUC of 0.8549, which the paper claims surpasses baselines LPVV, MVTS2V, SEQ, and ROCKET, plus additional experiments on eight UEA archive datasets.

Significance. If the central claim is correct, EXCON would be a useful contribution to solar flare prediction and imbalanced time-series representation learning. Strengths include a temporally held-out evaluation protocol (training on segment Si, testing on Si+1), use of TSS as the primary metric as recommended in the flare-prediction literature, code release, and extension to multiple UEA datasets. However, the significance is conditional: the headline numbers are not yet established as attributable to the proposed embedding because of a missing feature-level control and because configuration choices appear to be selected on the same test segments used for reporting.

major comments (3)
  1. [Section V.G, Table II] The comparison lacks a baseline that trains the same logistic regression directly on the multi-catch22 feature vectors V(m) of Section IV.A, using the same preprocessing and segment splits. Since catch22 is a strong, established time-series feature set and the final classifier operates on embeddings in the same 528-dimensional space, the observed TSS gain over LPVV, MVTS2V, SEQ, and ROCKET does not identify whether the extreme-reconstruction embedding is responsible for the performance. Please add this catch22+LR control (and report its TSS and ROC AUC) before claiming that EXCON's embedding improves over the feature representation alone.
  2. [Section V.C, Tables I and II, Fig. 5] The downstream classifier and LSTM hyperparameters were chosen by comparing performance on the same three train-test segment pairs (S1-S2, S2-S3, S3-S4) that are later used to report the headline mean TSS in Table II. This selection-on-test procedure makes 0.7098 a best-of-many configuration rather than an unbiased out-of-sample estimate, and it can advantage EXCON over baselines whose hyperparameters are not tuned in the same way. Please either use nested validation (e.g., tune on S1-S2, evaluate on S3-S4) or report all configurations with the selection rule explicitly; the same applies to the LSTM vs. RNN vs. GRU hidden-dimension comparisons in Fig. 5.
  3. [Section IV.B, Eq. 1, and Section IV.C] The method assumes that the single complete-linkage extreme instance computed from training segment Si is a valid regression target for embeddings of held-out future segment Si+1. The manuscript provides no stability analysis across segments S1-S5, such as the distance between successive extremes or whether the selected extreme is a representative prototype rather than a nonstationary outlier. If the extreme changes substantially between segments, the loss target in Eq. 2 is misspecified. Please add a stability check or an ablation comparing the extreme target against the class mean or multiple prototypes to justify this load-bearing assumption.
minor comments (6)
  1. [Equation 3] The index range is stated as '1 < t < τ'; it should likely be '1 ≤ t ≤ τ' to cover all timestamps of the time series.
  2. [Throughout] The dataset name is rendered with an internal space as 'SW AN-SF'; the standard spelling is 'SWAN-SF' and should be used consistently.
  3. [Table III] The table header says 'UAE dataset'; this should be 'UEA' (University of East Anglia) as used elsewhere in the paper.
  4. [Fig. 6] The t-SNE visualization is qualitative. Consider adding a quantitative separation measure (e.g., silhouette score or nearest-class distance) to support the claim of improved class separability.
  5. [Section V.H] The text states that EXCON 'shows the best performance in four datasets,' but no numeric table accompanies Fig. 8; please provide per-dataset scores so the claim can be verified.
  6. [Equations 1 and 2] The symbol d is used both for the dimension of the multi-catch22 vector (Section IV.A) and as the distance function in Eq. 1. Please disambiguate, e.g., by using 'dim' for dimension and a named distance metric in Eq. 1.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity: the extreme-reconstruction pipeline is fully specified and evaluated on temporally held-out segments and external benchmarks; only minor self-citation and test-selection leakage keep the score above zero.

full rationale

The core derivation is self-contained. catch22 feature extraction (Sec. IV.A), complete-linkage extreme selection (Eq. 1), the extreme reconstruction loss (Eq. 2), and the downstream LR classifier are all explicitly defined in the paper, with no test labels entering the training or extreme-selection stage. The extreme targets are legitimate supervised signal derived from the training segment, and the temporal train/test splits (S1-S2, S2-S3, S3-S4) mean the reported TSS and ROC AUC are not fits of the test set. The only self-citation is [6], used to say the loss is 'inspired by' the authors' own prior work; because Eq. (2) fully defines the loss and the multiclass extension is described, the citation is not load-bearing. A real but non-circular limitation is that Sec. V.C selected the classifier and architecture using the same evaluation segments whose scores are then reported in Tables I-II, so the absolute numbers carry mild selection bias; this is an evaluation-quality concern, not a by-construction reduction. The absence of a direct catch22+LR ablation weakens attribution of the gain to the embedding but does not make the derivation circular. Overall, no step reduces to its input by construction, and the central comparison against LPVV, SEQ, and ROCKET has independent content.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on five premises: catch22 sufficiency, Euclidean geometry of the feature space, stability of the single-point extreme prototype, temporal stationarity between adjacent segments, and the exclusion of B/C classes. The method itself is a supervised prototype regression, and the only hand-tuned inputs are the six hyperparameters listed above. No new physical entities are postulated.

free parameters (6)
  • LSTM hidden state dimension = 128
    Chosen from {32, 64, 128, 256, 512} by mean TSS on the same segment pairs used for final reporting (Section V.C, Fig. 5).
  • Learning rate = 1e-2
    Selected from {1e-1, 1e-2, 1e-3, 1e-4} on the evaluation segments (Section V.C).
  • Dropout rate = 0.5
    Fixed regularization hyperparameter of the temporal embedding module (Section V.C).
  • Training epochs = 30
    Chosen training budget for the temporal feature embedding module (Section V.C).
  • Fully connected layer depth = 1 layer
    Selected from one, two, or three-layer configurations on the evaluation segments (Section V.C).
  • Downstream classifier choice = Logistic regression
    LR chosen as top classifier across the three train-test segments (Section V.C, Table I), making the classifier part of the tuned method.
assumptions (5)
  • domain assumption catch22 features (22 per channel) are sufficiently discriminative summaries of MVTS dynamics for class separation
    Section IV.A compresses every instance into a 22N-vector and builds all contrastive targets from these features, inheriting catch22's discriminative power without validation on SWAN-SF.
  • domain assumption Euclidean distance in catch22 space captures meaningful inter-class separation
    Equation 1 selects the extreme instance that maximizes Euclidean distance to all other-class instances; no alternative metric or normalization is tested.
  • ad hoc to paper The single complete-linkage extreme instance per class is a stable prototype that transfers across time segments
    Equation 2 regresses all training and test embeddings toward the extreme computed from the training segment only; the paper never checks extreme stability across segments S1-S5.
  • domain assumption Model trained on segment Si generalizes to adjacent future segment Si+1
    Section V.C uses successive segments as train-test pairs, assuming temporal stationarity of the flare-prediction relationship.
  • domain assumption B and C class magnetograms are too similar to M/X to be useful in training
    Section V.B excludes B and C from the training sets, citing reference [1]; this changes the effective task and is a premise for the high TSS.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EXCON: Extreme Instance-based Contrastive Representation Learning of Severely Imbalanced Multivariate Time Series for Solar Flare Prediction." pith.science (2026). https://pith.science/paper/3A7HBLCT

@misc{pith2026241111249,
  author       = {Pith},
  title        = {Pith review of: EXCON: Extreme Instance-based Contrastive Representation Learning of Severely Imbalanced Multivariate Time Series for Solar Flare Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3A7HBLCT}},
  note         = {Machine review of arXiv:2411.11249}
}
read the original abstract

In heliophysics research, predicting solar flares is crucial due to their potential to impact both space-based systems and Earth's infrastructure substantially. Magnetic field data from solar active regions, recorded by solar imaging observatories, are transformed into multivariate time series to enable solar flare prediction using temporal window-based analysis. In the realm of multivariate time series-driven solar flare prediction, addressing severe class imbalance with effective strategies for multivariate time series representation learning is key to developing robust predictive models. Traditional methods often struggle with overfitting to the majority class in prediction tasks where major solar flares are infrequent. This work presents EXCON, a contrastive representation learning framework designed to enhance classification performance amidst such imbalances. EXCON operates through four stages: obtaining core features from multivariate time series data; selecting distinctive contrastive representations for each class to maximize inter-class separation; training a temporal feature embedding module with a custom extreme reconstruction loss to minimize intra-class variation; and applying a classifier to the learned embeddings for robust classification. The proposed method leverages contrastive learning principles to map similar instances closer in the feature space while distancing dissimilar ones, a strategy not extensively explored in solar flare prediction tasks. This approach not only addresses class imbalance but also offers a versatile solution applicable to univariate and multivariate time series across binary and multiclass classification problems. Experimental results, including evaluations on the benchmark solar flare dataset and multiple time series archive datasets with binary and multiclass labels, demonstrate EXCON's efficacy in enhancing classification performance.

Figures

Figures reproduced from arXiv: 2411.11249 by the authors.

Figure 1
Figure 1. Segments of SWAN-SF benchmark dataset with the frequen [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. MVTS feature extraction process with catch22. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Process of deriving extreme instance of class [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: In the EXCON framework, at timestamp t, vector x <t> of MVTS instance is processed by the t th LSTM cell within temporal feature embedding module. In the last timestamp τ , the output h <τ> is projected into d-dimensional space by the fully connected layer. The downstr…
Figure 5
Figure 5. Figure 5: Comparison of mean TSS performance for selected sequence [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: t-SNE visualization of Segment 2 as test data: raw format and [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Bar chart of EXCON vs. baselines in SWAN-SF solar flare [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Bar chart of EXCON vs. baselines in ROC AUC performance [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. AVATAR: Adversarial Autoencoders with Autoregressive Refinement for Time Series Generation

    cs.LG 2025-01 conditional novelty 4.0 of 10

    AVATAR augments adversarial autoencoders with an autoregressive supervisor and a mean-std distribution loss, reporting improved time series generation over five baselines on three datasets.

Reference graph

Works this paper leans on

26 extracted references · 17 canonical work pages · cited by 1 Pith paper

  1. [6]

    Contrastive Representation Learning for Predicting Solar Flares from Extremely Imbalanced Multivariate Time Series Data

    O. Vural, S. M. Hamdi, and S. F. Boubrahimi, “Contrastive represen- tation learning for predicting solar flares from extremely imbalanced multivariate time series data,” arXiv preprint arXiv:2410.00312 , 2024

    work page Pith review arXiv 2024

  2. [1]

    How to train your flare prediction model: Revisiting robust sampling of rare events,

    A. Ahmadzadeh, B. Aydin, M. K. Georgoulis, D. J. Kempton, S. S. Mahajan, and R. A. Angryk, “How to train your flare prediction model: Revisiting robust sampling of rare events,” The Astrophysical Journal Supplement Series, vol. 254, no. 2, p. 23, 2021

    work page 2021

  3. [2]

    A time series classification-based approach for solar flare prediction,

    S. M. Hamdi, D. Kempton, R. Ma, S. F. Boubrahimi, and R. A. Angryk, “A time series classification-based approach for solar flare prediction,” in 2017 IEEE International Conference on Big Data (Big Data) . IEEE, 2017, pp. 2543–2551

    work page 2017

  4. [3]

    Solar flare prediction using sdo/hmi vector magnetic field data with a machine-learning algorithm,

    M. G. Bobra and S. Couvidat, “Solar flare prediction using sdo/hmi vector magnetic field data with a machine-learning algorithm,” The Astrophysical Journal, vol. 798, no. 2, p. 135, 2015

    work page 2015

  5. [4]

    Multivariate time series dataset for space weather data analytics,

    R. A. Angryk, P. C. Martens, B. Aydin, D. Kempton, S. S. Mahajan, S. Basodi, A. Ahmadzadeh, X. Cai, S. Filali Boubrahimi, S. M. Hamdi et al., “Multivariate time series dataset for space weather data analytics,” Scientific data, vol. 7, no. 1, p. 227, 2020

    2020

  6. [5]

    Multivariate time series-based solar flare prediction by functional network embedding and sequence modeling,

    S. M. Hamdi, A. F. Ahmad, and S. Filali Boubrahimi, “Multivariate time series-based solar flare prediction by functional network embedding and sequence modeling,” in CIKM workshop for Applied Machine Learning Methods for Time Series Forecasting (AMLTS) 2022. , 2022

    work page 2022

  7. [7]

    Contrastive representation learning: A framework and review,

    P. H. Le-Khac, G. Healy, and A. F. Smeaton, “Contrastive representation learning: A framework and review,” Ieee Access , vol. 8, pp. 193 907– 193 934, 2020

    2020

  8. [8]

    The classification of sunspot groups,

    P. S. McIntosh, “The classification of sunspot groups,” Solar Physics , vol. 125, pp. 251–267, 1990

    work page 1990

Show all 26 references
  1. [9]

    Spatio- temporal interpolation methods for solar events metadata,

    S. F. Boubrahimi, B. Aydin, D. Kempton, and R. Angryk, “Spatio- temporal interpolation methods for solar events metadata,” in 2016 IEEE International Conference on Big Data (Big Data) . IEEE, 2016, pp. 3149–3157

    2016

  2. [10]

    Testing automated solar flare forecast- ing with 13 years of michelson doppler imager magnetograms,

    J. P. Mason and J. Hoeksema, “Testing automated solar flare forecast- ing with 13 years of michelson doppler imager magnetograms,” The Astrophysical Journal, vol. 723, no. 1, p. 634, 2010

    2010

  3. [11]

    Statistical assessment of photospheric magnetic features in imminent solar flare predictions,

    H. Song, C. Tan, J. Jing, H. Wang, V . Yurchyshyn, and V . Abramenko, “Statistical assessment of photospheric magnetic features in imminent solar flare predictions,” Solar Physics, vol. 254, pp. 101–125, 2009

    2009

  4. [12]

    Short-term solar flare prediction using a sequential supervised learning method,

    D. Yu, X. Huang, H. Wang, and Y . Cui, “Short-term solar flare prediction using a sequential supervised learning method,” Solar Physics, vol. 255, pp. 91–105, 2009

    2009

  5. [13]

    Solar flare prediction using advanced feature extraction, machine learning, and feature selection,

    O. W. Ahmed, R. Qahwaji, T. Colak, P. A. Higgins, P. T. Gallagher, and D. S. Bloomfield, “Solar flare prediction using advanced feature extraction, machine learning, and feature selection,” Solar Physics, vol. 283, pp. 157–175, 2013

    2013

  6. [14]

    Solar flare prediction using multivariate time series decision trees,

    R. Ma, S. F. Boubrahimi, S. M. Hamdi, and R. A. Angryk, “Solar flare prediction using multivariate time series decision trees,” in 2017 IEEE International Conference on Big Data (Big Data) . IEEE, 2017, pp. 2569–2578

    2017

  7. [15]

    Sequence model-based end-to-end solar flare classification from multivariate time series data,

    A. A. M. Muzaheed, S. M. Hamdi, and S. F. Boubrahimi, “Sequence model-based end-to-end solar flare classification from multivariate time series data,” in 2021 20th IEEE International Conference on Machine Learning and Applications (ICMLA) . IEEE, 2021, pp. 435–440

    2021

  8. [16]

    Self-supervised time series repre- sentation learning by inter-intra relational reasoning,

    H. Fan, F. Zhang, and Y . Gao, “Self-supervised time series repre- sentation learning by inter-intra relational reasoning,” arXiv preprint arXiv:2011.13548, 2020

    2011 arXiv

  9. [17]

    Time-series representation learning via temporal and contextual con- trasting,

    E. Eldele, M. Ragab, Z. Chen, M. Wu, C. K. Kwoh, X. Li, and C. Guan, “Time-series representation learning via temporal and contextual con- trasting,” arXiv preprint arXiv:2106.14112 , 2021

    2021 arXiv

  10. [18]

    Time series contrastive learning with information-aware augmentations,

    D. Luo, W. Cheng, Y . Wang, D. Xu, J. Ni, W. Yu, X. Zhang, Y . Liu, Y . Chen, H. Chen et al. , “Time series contrastive learning with information-aware augmentations,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 37, no. 4, 2023, pp. 4534– 4542

    2023

  11. [19]

    Soft contrastive learning for time series,

    S. Lee, T. Park, and K. Lee, “Soft contrastive learning for time series,” arXiv preprint arXiv:2312.16424 , 2023

    2023

  12. [20]

    Simts: rethinking contrastive representation learning for time series forecasting,

    X. Zheng, X. Chen, M. Sch ¨urch, A. Mollaysa, A. Allam, and M. Krauthammer, “Simts: rethinking contrastive representation learning for time series forecasting,” arXiv preprint arXiv:2303.18205 , 2023

    2023 arXiv

  13. [21]

    catch22: Canonical time-series characteristics: Selected through highly comparative time-series analysis,

    C. H. Lubba, S. S. Sethi, P. Knaute, S. R. Schultz, B. D. Fulcher, and N. S. Jones, “catch22: Canonical time-series characteristics: Selected through highly comparative time-series analysis,” Data Mining and Knowledge Discovery, vol. 33, no. 6, pp. 1821–1852, 2019

    2019

  14. [22]

    The great multivariate time series classification bake off: a review and experimental evaluation of recent algorithmic advances,

    A. P. Ruiz, M. Flynn, J. Large, M. Middlehurst, and A. Bagnall, “The great multivariate time series classification bake off: a review and experimental evaluation of recent algorithmic advances,” Data Mining and Knowledge Discovery , vol. 35, no. 2, pp. 401–449, 2021

    2021

  15. [23]

    The uea multivariate time series classi- fication archive, 2018,

    A. Bagnall, H. A. Dau, J. Lines, M. Flynn, J. Large, A. Bostrom, P. Southam, and E. Keogh, “The uea multivariate time series classi- fication archive, 2018,” arXiv preprint arXiv:1811.00075 , 2018

    2018 arXiv

  16. [24]

    Impacts of data preprocessing and sampling techniques on solar flare prediction from multivariate time series data of photospheric magnetic field pa- rameters,

    M. EskandariNasab, S. M. Hamdi, and S. F. Boubrahimi, “Impacts of data preprocessing and sampling techniques on solar flare prediction from multivariate time series data of photospheric magnetic field pa- rameters,” The Astrophysical Journal Supplement Series, vol. 275, no. 1,...

    2024

  17. [25]

    Visualizing data using t-sne

    L. Van der Maaten and G. Hinton, “Visualizing data using t-sne.” Journal of machine learning research , vol. 9, no. 11, 2008

    2008

  18. [26]

    Rocket: exceptionally fast and accurate time series classification using random convolutional kernels,

    A. Dempster, F. Petitjean, and G. I. Webb, “Rocket: exceptionally fast and accurate time series classification using random convolutional kernels,” Data Mining and Knowledge Discovery , vol. 34, no. 5, pp. 1454–1495, 2020

    2020

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.