pith. machine review for the scientific record. sign in

arxiv: 2604.24516 · v1 · submitted 2026-04-27 · 🌌 astro-ph.SR · astro-ph.IM

Recognition: unknown

StarCLR: Contrastive Learning Representation for Astronomical Light Curves

Junyao Ding , Xiaodian Chen , Xinyi Gao , Xiaoyu Tang , Shu Wang , Yang Huang , Xinyu Qi , Guirong Xue
show 2 more authors
Ali Luo Jifeng Liu
Authors on Pith no claims yet

Pith reviewed 2026-05-08 01:27 UTC · model grok-4.3

classification 🌌 astro-ph.SR astro-ph.IM
keywords contrastive learninglight curvesvariable starspretrainingTESSZTFGaiatime series classification
0
0 comments X

The pith

A contrastive pretraining method on TESS light curves learns temporal features from overlapping segments and improves variable star classification on TESS, ZTF, and Gaia surveys.

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

StarCLR applies contrastive learning to astronomical light curves by treating partially overlapping sub-sequences as positive pairs. This pretraining step, done on TESS data, produces representations that are then fine-tuned for classifying variable stars across three surveys with different sampling densities and class counts. The resulting model records macro-F1 scores of 84.35 percent on TESS, 87.82 percent on ZTF, and 92.73 percent on Gaia, together with higher micro-F1 scores, and beats LSTM and Transformer baselines trained from scratch, especially on the sparsely sampled ZTF curves. Systematic checks on embedding choices, pooling, and pretraining settings indicate that the gains come from better capture of temporal structure rather than survey-specific artifacts. The work therefore tests whether self-supervised temporal signals alone can transfer usefully between light-curve datasets that differ in cadence and label space.

Core claim

By constructing positive pairs from partially overlapping sub-sequences of light curves, StarCLR learns temporal representations during pretraining on TESS that, after fine-tuning, raise classification accuracy on TESS (18 classes), ZTF (11 classes), and Gaia (24 classes) relative to LSTM and Transformer models trained from scratch, with the clearest advantage appearing on the sparsely sampled ZTF data.

What carries the argument

StarCLR contrastive pretraining framework that builds positive pairs from partially overlapping light-curve sub-sequences to train temporal representations without labels.

If this is right

  • StarCLR outperforms LSTM and Transformer baselines on TESS and ZTF classification, with the largest margin on sparsely sampled ZTF light curves.
  • On Gaia, which has a broader class space, the pretrained backbone contributes less because performance depends more on astrophysical features.
  • Ablation studies confirm that embedding design, pooling strategy, and pretraining settings all affect how much temporal information the representations retain.
  • The approach demonstrates that contrastive signals from overlapping segments can transfer across surveys without hand-crafted features.

Where Pith is reading between the lines

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

  • The same overlapping-segment contrastive recipe could be applied to other unevenly sampled time series such as exoplanet transit data or solar flare records.
  • Standardized cross-survey benchmarks would make it easier to measure how much the pretrained backbone actually reduces the labeled-data requirement.
  • If the learned representations prove stable under changes in cadence and filter, they could serve as a reusable backbone for new surveys that lack large labeled sets.

Load-bearing premise

That positive pairs formed from partially overlapping sub-sequences of light curves supply enough training signal for temporal representations to generalize across surveys that differ in sampling rate, noise properties, and class definitions, without any astrophysical priors or explicit domain adaptation.

What would settle it

A head-to-head experiment on the ZTF variable-star task in which a from-scratch LSTM or Transformer reaches macro-F1 scores equal to or higher than the fine-tuned StarCLR model while using identical training and test splits.

Figures

Figures reproduced from arXiv: 2604.24516 by Ali Luo, Guirong Xue, Jifeng Liu, Junyao Ding, Shu Wang, Xiaodian Chen, Xiaoyu Tang, Xinyi Gao, Xinyu Qi, Yang Huang.

Figure 1
Figure 1. Figure 1: Examples of light curves from three surveys: (a) TESS, (b) ZTF, and (c) Gaia. The horizontal axis is the observation time with the minimum value subtracted, and the vertical axis is the flux. The TESS light curve spans a single sector (∼27.4 days), the shortest baseline among the three, whereas ZTF and Gaia extend to several thousand days. In terms of cadence, TESS has the shortest time interval between su… view at source ↗
Figure 2
Figure 2. Figure 2: Workflow of hierarchical contrastive loss computation. impact of noise and outliers. A detailed analysis of different pooling strategies is provided in Section 5.3. The resulting global representation is then concatenated with supplementary stellar attributes to construct a joint feature vector, thereby integrating light-curve features with static physical properties. Finally, this vector is passed through… view at source ↗
Figure 3
Figure 3. Figure 3: The architecture of StarCLR. The input consists of preprocessed time and flux values. The flux sequence is projected into a 512-dimensional feature space through a feed-forward layer, while the corresponding time are encoded into 512-dimensional positional embeddings using the sinusoidal positional encoding. The two representations are combined via element-wise addition to form an input embedding matrix of… view at source ↗
Figure 4
Figure 4. Figure 4: StarCLR fine-tuning architecture. The input light curve is represented as an L × 2 matrix, where L is the number of time steps and each step consists of time and flux. The sequence length is standardized according to the survey characteristics; for example, 8192 for TESS, 1024 for ZTF, and 256 for Gaia. The backbone is a Transformer encoder that produces an output of dimension L × 512. Subsequently, mean p… view at source ↗
Figure 5
Figure 5. Figure 5: Loss curves during the pretraining stage on the TESS dataset. The horizontal axis denotes the training epochs, while the vertical axis represents the loss value. Both training and validation losses are shown as a function of epochs, and the model achieves its optimal performance at around the third epoch. all reported classification results are averaged over at least three independent runs with different r… view at source ↗
Figure 6
Figure 6. Figure 6: Confusion matrix of the StarCLR model on the TESS fine-tuning task. Rows correspond to the true labels and columns to the predicted labels. and CEPII. In particular, CEPII, which is severely underrepresented, shows the weakest performance with an F1-score of 63.06% ± 3.12%. Overall, while rare and closely related subclasses remain difficult, performance on ZTF is robust. Importantly, these results demonstr… view at source ↗
Figure 7
Figure 7. Figure 7: Confusion matrix of the StarCLR model on the ZTF fine-tuning task. 5. DISCUSSION This section interprets the observed performance trends and identifies key factors governing the effectiveness of Star￾CLR. We examine model architectures and training strategies, analyze input embedding design and pooling operations, and study learned representation structure through UMAP visualizations. 5.1. Model Comparison… view at source ↗
Figure 8
Figure 8. Figure 8: Confusion matrix of the StarCLR model on the Gaia fine-tuning task. size of 512, which takes the same normalized flux inputs as StarCLR, combined with the corresponding time-aware positional encodings; (ii) a Transformer model trained from scratch, which shares the same architecture as StarCLR but is randomly initialized; and (iii) pretrained StarCLR, where the Transformer backbone was initialized from con… view at source ↗
Figure 9
Figure 9. Figure 9: Fine-tuning results of three models on the TESS, ZTF, and Gaia datasets. The horizontal axis denotes training epochs; the left vertical axis shows validation loss with solid lines, and the right vertical axis shows validation macro-F1 with dashed lines. From top to bottom are the results on TESS, ZTF, and Gaia validation sets, respectively. sector spans ∼ 27.4 days and often contains ∼ 103 cadence points a… view at source ↗
Figure 10
Figure 10. Figure 10: UMAP projection of the TESS dataset, showing the global distribution and a locally magnified region. UMAP-1 and UMAP-2 denote the two embedding dimensions obtained after dimensionality reduction. dimensional information is compressed into two dimensions, inevitably discarding some discriminative features while amplifying residual noise or unstable components, which appear more prominently as scattered poi… view at source ↗
Figure 11
Figure 11. Figure 11: UMAP projection of the ZTF dataset, showing the global distribution and a locally magnified region view at source ↗
Figure 12
Figure 12. Figure 12: UMAP projection of the Gaia dataset, showing the global distribution and a locally magnified region view at source ↗
read the original abstract

With the rapid development of time-domain surveys, the availability of massive light curve data offers new opportunities for studying stellar evolution and variable star classification, while simultaneously posing challenges for feature extraction and modeling. We present StarCLR, a contrastive pretraining framework for large-scale light curves. By constructing positive pairs from partially overlapping sub-sequences, StarCLR encourages the model to learn temporal representations. We pretrain StarCLR on the TESS dataset and fine-tune it for variable star classification on three surveys with distinct observational characteristics, namely TESS (18 types), ZTF (11 types), and Gaia (24 types). StarCLR achieves macro-F1 scores of 84.35%, 87.82%, and 92.73%, and micro-F1 scores of 94.46%, 92.83%, and 99.49%, respectively. Compared with LSTM and Transformer trained from scratch, StarCLR performs better on TESS and ZTF, with the largest gains on sparsely sampled ZTF light curves, demonstrating promising generalization. For Gaia, which involves a broader class space, the evaluation is not directly comparable, and performance is likely influenced by astrophysical features, resulting in a more limited contribution from the pretrained backbone. Systematic ablations on embedding design, pooling strategy, and pretraining settings further indicate that the pretrained representations provide performance gains by capturing informative temporal characteristics of light curves. Looking ahead, with standardized datasets and more diverse labeling schemes, the generalization ability of StarCLR can be further enhanced.

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 introduces StarCLR, a contrastive pretraining framework for astronomical light curves. Positive pairs are formed from partially overlapping sub-sequences of TESS data to learn temporal representations; the pretrained encoder is then fine-tuned for variable-star classification on TESS (18 classes), ZTF (11 classes), and Gaia (24 classes). It reports macro-F1 scores of 84.35%, 87.82%, and 92.73% and micro-F1 scores of 94.46%, 92.83%, and 99.49% respectively, outperforming LSTM and Transformer models trained from scratch, with the largest gains on sparsely sampled ZTF light curves. Systematic ablations on embedding design, pooling, and pretraining settings are provided to support that the gains arise from informative temporal features.

Significance. If the central empirical claims hold, the work provides concrete evidence that contrastive pretraining on overlapping light-curve windows can yield transferable representations across surveys with differing cadences and class spaces. The reported F1 margins (especially on ZTF) and the accompanying ablations on embedding/pooling/pretraining choices constitute a useful empirical demonstration that self-supervised methods can leverage large unlabeled time-domain datasets, which is of practical value for upcoming surveys.

major comments (2)
  1. [Section 3] Section 3 (pretraining framework): The headline claim that contrastive pretraining on TESS-derived overlapping sub-sequences produces survey-agnostic temporal features that transfer to ZTF rests on the assumption that positive-pair construction alone suffices to factor out cadence and gap differences. TESS provides near-continuous high-cadence segments while ZTF is irregular and gapped; without explicit domain-shift diagnostics (e.g., cadence-augmented pretraining or embedding-distance analysis between surveys), it remains possible that the encoder partially memorizes TESS-specific autocorrelation statistics. This assumption is load-bearing for interpreting the largest reported gains on ZTF.
  2. [Section 4] Section 4 (experiments and results): The performance tables do not report statistical error bars or significance tests on the F1 improvements over the scratch-trained baselines. Given that these margins are central to the generalization narrative, the absence of uncertainty quantification weakens the ability to judge whether the gains are robust rather than within run-to-run variability.
minor comments (2)
  1. [Abstract and Section 4] The abstract and results section should explicitly cross-reference the specific tables or figures that present the systematic ablations on embedding dimension, pooling strategy, and pretraining settings.
  2. [Section 4] Additional implementation details for the LSTM and Transformer baselines (exact hyperparameter search ranges, layer counts, and training schedules) would improve reproducibility of the reported comparisons.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed comments, which highlight important aspects of our pretraining framework and experimental evaluation. We address each major comment below and will revise the manuscript to incorporate additional analysis and statistical reporting as outlined.

read point-by-point responses

  1. Referee: [Section 3] Section 3 (pretraining framework): The headline claim that contrastive pretraining on TESS-derived overlapping sub-sequences produces survey-agnostic temporal features that transfer to ZTF rests on the assumption that positive-pair construction alone suffices to factor out cadence and gap differences. TESS provides near-continuous high-cadence segments while ZTF is irregular and gapped; without explicit domain-shift diagnostics (e.g., cadence-augmented pretraining or embedding-distance analysis between surveys), it remains possible that the encoder partially memorizes TESS-specific autocorrelation statistics. This assumption is load-bearing for interpreting the largest reported gains on ZTF.

    Authors: We acknowledge the importance of demonstrating that the learned representations are not merely capturing TESS-specific autocorrelation. The systematic ablations on embedding design, pooling strategy, and pretraining settings (Section 4) show that performance gains arise specifically from informative temporal features of light curves. The fact that the largest improvements occur on ZTF—which has markedly different cadence and gap patterns—further supports that the contrastive objective on overlapping subsequences promotes transferable temporal structure rather than survey-specific memorization. To directly address domain shift, we will add an embedding-distance analysis comparing TESS and ZTF representations in the revised manuscript. revision: yes

  2. Referee: [Section 4] Section 4 (experiments and results): The performance tables do not report statistical error bars or significance tests on the F1 improvements over the scratch-trained baselines. Given that these margins are central to the generalization narrative, the absence of uncertainty quantification weakens the ability to judge whether the gains are robust rather than within run-to-run variability.

    Authors: We agree that uncertainty quantification is necessary to establish the robustness of the reported F1 improvements. In the revised manuscript we will report mean and standard deviation of macro- and micro-F1 scores across multiple independent training runs (five random seeds) for both StarCLR and the scratch-trained baselines. We will also include results of paired statistical significance tests on the observed differences. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical claims rest on held-out cross-survey benchmarks, not self-referential definitions or fitted inputs.

full rationale

The paper defines its contrastive objective by construction (positive pairs from overlapping subsequences) as an explicit modeling choice, then evaluates via direct F1 comparisons to scratch-trained LSTM/Transformer baselines on separate TESS/ZTF/Gaia test sets. No equations, ablations, or results reduce by the paper's own definitions to quantities already fixed by the inputs or by self-citation chains. The method is self-contained against external benchmarks; the largest reported gains on ZTF are presented as empirical outcomes, not forced by the pretraining construction itself.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard contrastive-learning machinery plus one domain-specific assumption about light-curve subsequences; no new physical entities are introduced and hyperparameters are typical for the method.

free parameters (2)
  • contrastive temperature
    Standard hyperparameter in the contrastive loss; value chosen to optimize downstream performance.
  • embedding dimension and pooling strategy
    Architectural choices tuned during ablations.
axioms (1)
  • domain assumption Partially overlapping sub-sequences of the same light curve share the same underlying stellar variability process
    Invoked to construct positive pairs for contrastive pretraining.

pith-pipeline@v0.9.0 · 5600 in / 1561 out tokens · 72586 ms · 2026-05-08T01:27:10.189076+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

44 extracted references · 38 canonical work pages · 4 internal anchors

  1. [1]

    2024, in Proceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating

    Ansel, J., Yang, E., He, H., et al. 2024, in Proceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating

    2024

  2. [2]

    PyTorch 2: Faster Ma- chine Learning Through Dynamic Python Bytecode Transformation and Graph Compilation

    Systems, Volume 2, ASPLOS ’24 (New York, NY, USA: Association for Computing Machinery), 929–947, doi: 10.1145/3620665.3640366 Astropy Collaboration, Robitaille, T. P., Tollerud, E. J., et al. 2013, A&A, 558, A33, doi: 10.1051/0004-6361/201322068 28Ding et al. Astropy Collaboration, Price-Whelan, A. M., Sip˝ ocz, B. M., et al. 2018, AJ, 156, 123, doi: 10.3...

    work page doi:10.1145/3620665.3640366 2013

  3. [3]

    H., Bedding, T

    Barbara, N. H., Bedding, T. R., Fulcher, B. D., Murphy, S. J., & Van Reeth, T. 2022, MNRAS, 514, 2793, doi: 10.1093/mnras/stac1515

    work page doi:10.1093/mnras/stac1515 2022

  4. [4]

    M., Batalha, N., et al

    Basri, G., Walkowicz, L. M., Batalha, N., et al. 2011, AJ, 141, 20, doi: 10.1088/0004-6256/141/1/20

    work page doi:10.1088/0004-6256/141/1/20 2011

  5. [5]

    C., Kulkarni, S

    Bellm, E. C., Kulkarni, S. R., Graham, M. J., et al. 2019, PASP, 131, 018002, doi: 10.1088/1538-3873/aaecbe

    work page doi:10.1088/1538-3873/aaecbe 2019

  6. [6]

    Kepler Planet-Detection Mission: Introduction and First Results.Science2010,327, 977

    Borucki, W. J., Koch, D., Basri, G., et al. 2010, Science, 327, 977, doi: 10.1126/science.1185402

    work page doi:10.1126/science.1185402 2010

  7. [7]

    Machine Learning , author =

    Breiman, L. 2001, Machine Learning, 45, 5, doi: 10.1023/A:1010933404324

    work page doi:10.1023/a:1010933404324 2001

  8. [8]

    A Simple Framework for Contrastive Learning of Visual Representations

    Chen, T., Kornblith, S., Norouzi, M., & Hinton, G. 2020a, arXiv e-prints, arXiv:2002.05709, doi: 10.48550/arXiv.2002.05709

    work page internal anchor Pith review doi:10.48550/arxiv.2002.05709 2002

  9. [9]

    , keywords =

    Chen, X., Wang, S., Deng, L., et al. 2020b, ApJS, 249, 18, doi: 10.3847/1538-4365/ab9cae

    work page doi:10.3847/1538-4365/ab9cae

  10. [10]

    T., Jayasinghe, T., Stanek, K

    Christy, C. T., Jayasinghe, T., Stanek, K. Z., et al. 2023, MNRAS, 519, 5271, doi: 10.1093/mnras/stac3801

    work page doi:10.1093/mnras/stac3801 2023

  11. [11]

    J., & Feng, F

    Cui, K., Armstrong, D. J., & Feng, F. 2024, ApJS, 274, 29, doi: 10.3847/1538-4365/ad62fd

    work page doi:10.3847/1538-4365/ad62fd 2024

  12. [12]

    BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

    Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. 2018, arXiv e-prints, arXiv:1810.04805, doi: 10.48550/arXiv.1810.04805

    work page Pith review doi:10.48550/arxiv.1810.04805 2018

  13. [13]

    2025, arXiv e-prints, arXiv:2502.02717, doi: 10.48550/arXiv.2502.02717 —

    Donoso-Oliva, C., Becker, I., Protopapas, P., et al. 2025, arXiv e-prints, arXiv:2502.02717, doi: 10.48550/arXiv.2502.02717 —. 2023, A&A, 670, A54, doi: 10.1051/0004-6361/202243928

    work page doi:10.48550/arxiv.2502.02717 2025

  14. [14]

    1996, in Proceedings of the Second International Conference on Knowledge Discovery and Data Mining, KDD’96 (AAAI Press), 226–231

    Ester, M., Kriegel, H.-P., Sander, J., & Xu, X. 1996, in Proceedings of the Second International Conference on Knowledge Discovery and Data Mining, KDD’96 (AAAI Press), 226–231

    1996

  15. [15]

    2023, ApJS, 268, 4, doi: 10.3847/1538-4365/acdee5

    Fetherolf, T., Pepper, J., Simpson, E., et al. 2023, ApJS, 268, 4, doi: 10.3847/1538-4365/acdee5

    work page doi:10.3847/1538-4365/acdee5 2023

  16. [16]

    2023 , month = jun, journal =

    Fulcher, B. D., & Jones, N. S. 2017, Cell systems, 5, 527 Gaia Collaboration, Vallenari, A., Brown, A. G. A., et al. 2023, A&A, 674, A1, doi: 10.1051/0004-6361/202243940

    work page doi:10.1051/0004-6361/202243940 2017

  17. [17]

    Classification of Periodic Variable Stars from TESS

    Gao, X., Chen, X., Wang, S., & Liu, J. 2025, ApJS, 276, 57, doi: 10.3847/1538-4365/ad9dd6

    work page doi:10.3847/1538-4365/ad9dd6 2025

  18. [18]

    2022, Accelerate: Training and inference at scale made simple, efficient and adaptable., https://github.com/huggingface/accelerate

    Gugger, S., Debut, L., Wolf, T., et al. 2022, Accelerate: Training and inference at scale made simple, efficient and adaptable., https://github.com/huggingface/accelerate

    2022

  19. [19]

    R., Millman, K

    Harris, C. R., Millman, K. J., van der Walt, S. J., et al. 2020, Nature, 585, 357–362, doi: 10.1038/s41586-020-2649-2

    work page doi:10.1038/s41586-020-2649-2 2020

  20. [20]

    Huang, K.-W., & Koposov, S. E. 2022, MNRAS, 510, 3575, doi: 10.1093/mnras/stab3654

    work page doi:10.1093/mnras/stab3654 2022

  21. [21]

    Hunter, J. D. 2007, Computing in Science & Engineering, 9, 90, doi: 10.1109/MCSE.2007.55

    work page doi:10.1109/mcse.2007.55 2007

  22. [22]

    The ASAS-SN catalogue of variable stars – II. Uniform classification of 412 000 known variables

    Jayasinghe, T., Stanek, K. Z., Kochanek, C. S., et al. 2019, MNRAS, 486, 1907, doi: 10.1093/mnras/stz844

    work page doi:10.1093/mnras/stz844 2019

  23. [23]

    , keywords =

    Kochanek, C. S., Shappee, B. J., Stanek, K. Z., et al. 2017, PASP, 129, 104502, doi: 10.1088/1538-3873/aa80d9

    work page doi:10.1088/1538-3873/aa80d9 2017

  24. [24]

    Lhoest, Q., Villanova del Moral, A., Jernite, Y., et al. 2021, in Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing: System Demonstrations (Online and Punta Cana, Dominican Republic: Association for Computational Linguistics), 175–184. https://arxiv.org/abs/2109.02846

    work page arXiv 2021

  25. [25]

    Lomb, N. R. 1976, Ap&SS, 39, 447, doi: 10.1007/BF00648343

    work page doi:10.1007/bf00648343 1976

  26. [26]

    Publications of the Astronomical Society of the Pacific , author =

    Masci, F. J., Laher, R. R., Rusholme, B., et al. 2019, PASP, 131, 018003, doi: 10.1088/1538-3873/aae8ac

    work page internal anchor Pith review doi:10.1088/1538-3873/aae8ac 2019

  27. [27]

    UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction

    McInnes, L., Healy, J., & Melville, J. 2020, UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction. https://arxiv.org/abs/1802.03426

    work page internal anchor Pith review arXiv 2020

  28. [28]

    2010 , keywords =

    McKinney, W. 2010, SciPy 2010, doi: 10.25080/Majora-92bf1922-00a

    work page doi:10.25080/majora-92bf1922-00a 2010

  29. [29]

    J., & Peel, D

    McLachlan, G. J., & Peel, D. 2000, Finite mixture models (John Wiley & Sons)

    2000

  30. [30]

    2023, arXiv e-prints, arXiv:2308.06404, doi: 10.48550/arXiv.2308.06404

    Moreno-Cartagena, D., Cabrera-Vives, G., Protopapas, P., et al. 2023, arXiv e-prints, arXiv:2308.06404, doi: 10.48550/arXiv.2308.06404

    work page doi:10.48550/arxiv.2308.06404 2023

  31. [31]

    2019, OpenAI blog, 1, 9

    Radford, A., Wu, J., Child, R., et al. 2019, OpenAI blog, 1, 9

    2019

  32. [32]

    R., Winn, J

    Ricker, G. R., Winn, J. N., Vanderspek, R., et al. 2015, Journal of Astronomical Telescopes, Instruments, and Systems, 1, 014003, doi: 10.1117/1.JATIS.1.1.014003

    work page Pith review doi:10.1117/1.jatis.1.1.014003 2015

  33. [33]

    2023, A&A, 674, A14, doi: 10.1051/0004-6361/202245591

    Rimoldini, L., Holl, B., Gavras, P., et al. 2023, A&A, 674, A14, doi: 10.1051/0004-6361/202245591

    work page doi:10.1051/0004-6361/202245591 2023

  34. [34]

    Rizhko, M., & Bloom, J. S. 2024, arXiv e-prints, arXiv:2411.08842, doi: 10.48550/arXiv.2411.08842

    work page doi:10.48550/arxiv.2411.08842 2024

  35. [35]

    Scargle, J. D. 1982, ApJ, 263, 835, doi: 10.1086/160554

    work page doi:10.1086/160554 1982

  36. [36]

    2025, PASP, 137, 044503, doi: 10.1088/1538-3873/adc5a2

    Shan, Y., Chen, J., Zhang, Z., et al. 2025, PASP, 137, 044503, doi: 10.1088/1538-3873/adc5a2

    work page doi:10.1088/1538-3873/adc5a2 2025

  37. [37]

    , keywords =

    Stassun, K. G., Oelkers, R. J., Paegert, M., et al. 2019, AJ, 158, 138, doi: 10.3847/1538-3881/ab3467

    work page doi:10.3847/1538-3881/ab3467 2019

  38. [38]

    K., Soszynski, I., & Poleski, R

    Udalski, A., Szymanski, M. K., Soszynski, I., & Poleski, R. 2008, AcA, 58, 69, doi: 10.48550/arXiv.0807.3884

    work page doi:10.48550/arxiv.0807.3884 2008

  39. [39]

    OGLE-IV: Fourth Phase of the Optical Gravitational Lensing Experiment

    Udalski, A., Szyma´ nski, M. K., & Szyma´ nski, G. 2015, AcA, 65, 1, doi: 10.48550/arXiv.1504.05966

    work page Pith review doi:10.48550/arxiv.1504.05966 2015

  40. [40]

    Attention Is All You Need

    Vaswani, A., Shazeer, N., Parmar, N., et al. 2023, Attention Is All You Need. https://arxiv.org/abs/1706.03762

    work page internal anchor Pith review arXiv 2023

  41. [41]

    The Astrophysical Journal , author =

    Wang, S., & Chen, X. 2019, ApJ, 877, 116, doi: 10.3847/1538-4357/ab1c61 StarCLR: Contrastive Learning Representation for Astronomical Light Curves29

    work page doi:10.3847/1538-4357/ab1c61 2019

  42. [42]

    2020, in Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations (Online: Association for Computational Linguistics), 38–45

    Wolf, T., Debut, L., Sanh, V., et al. 2020, in Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations (Online: Association for Computational Linguistics), 38–45. https://www.aclweb.org/anthology/2020.emnlp-demos.6

    2020

  43. [43]

    2022, TS2Vec: Towards Universal Representation of Time Series

    Yue, Z., Wang, Y., Duan, J., et al. 2022, TS2Vec: Towards Universal Representation of Time Series. https://arxiv.org/abs/2106.10466

    work page arXiv 2022

  44. [44]

    2025, arXiv e-prints, arXiv:2504.20290, doi: 10.48550/arXiv.2504.20290 30Ding et al

    Zuo, X., Tao, Y., Huang, Y., et al. 2025, arXiv e-prints, arXiv:2504.20290, doi: 10.48550/arXiv.2504.20290 30Ding et al. APPENDIX A.SUPPORTING FIGURES AND TABLES Tables A1 present the detailed per-class classification performance of StarCLR on the TESS, ZTF, and Gaia test sets, including uncertainty estimates from multiple runs. T able A1. Per-class class...

    work page doi:10.48550/arxiv.2504.20290 2025