The DOI matches its Crossref record; 223 of the 232 references it cites were found in a registry or an index; 86 of the 189 sources its claims cite were checked for support.
Self-Supervised Learning for Time Series Analysis: Taxonomy, Progress, and Prospects
Open accessAbstract
A comprehensive survey of current state-of-the-art SSL methods for time series data, proposing a taxonomy categorizing methods into generative-based, contrastive-based, and adversarial-based across ten subcategories, along with benchmark datasets for forecasting, classification, anomaly detection, and clustering.
Overview
CoChat Summary
TL;DR
This survey reviews self-supervised learning (SSL) for time series, proposing a taxonomy of generative-based, contrastive-based, and adversarial-based methods divided into ten subcategories. It summarizes applications/datasets and future directions, and compiles quantitative comparisons showing no single SSL method dominates across tasks.
Key findings
- The paper proposes a taxonomy with three broad SSL paradigms and ten subcategories: autoregressive-based forecasting, autoencoder-based reconstruction, diffusion-based generation; sampling contrast, prediction contrast, augmentation contrast, prototype contrast, expert knowledge contrast; generation/imputation and auxiliary representation enhancement.
- For anomaly detection, contrastive methods such as Skip-CPC and Dcdetector are reported as achieving the best performance on four datasets; Dcdetector reaches F1 scores of 97.94 on PSM, 96.60 on MSL, 97.02 on SMAP, and 96.33 on SWaT, while Skip-CPC reaches F1 98.55 on PSM, 94.46 on SMD, and 92.75 on MSL.
- For forecasting, LaST and CoST show significant advantages; LaST reports ETTh H=24 MSE/MAE of 0.324/0.368, ETTm 0.218/0.289, Electricity 0.125/0.222, and Weather 0.105/0.134.
- For classification and clustering, results on 30 datasets from HAR, UCR, and UEA show that no single method achieves the best performance on all datasets, and most methods are based on self-supervised contrastive learning.
Methods
- Design: a comprehensive literature survey that reviews existing SSL and time series surveys, then organizes time series SSL methods by learning paradigm and by how positive/negative samples or pretext tasks are constructed.
- Data/resources: summarizes commonly used datasets for anomaly detection, forecasting, classification, and clustering, including PSM, SMD, MSL, SMAP, SWaT, WADI, ETTh, ETTm, Electricity, Weather, HAR, UCR 128, and UEA 30.
- Analysis: qualitative review of frameworks, mathematical formulations, advantages/disadvantages, plus compiled quantitative comparisons from prior work on anomaly detection (five datasets), forecasting (four datasets), and classification/clustering (30 datasets); no new experiments are reported.
Limitations
- The paper states that prior time series SSL surveys only discuss a small part of self-supervised contrastive learning, lack benchmark dataset summaries, and have scarce future directions; it positions itself as filling that gap.
- It also summarizes method-family limitations: generative-based methods are sensitive to rare instances and lack high-level semantic information; contrastive-based methods face sampling bias and augmentation uncertainty; adversarial-based methods suffer unstable training and are rarely used for feature extraction.
- Evident from the text: as a survey, it does not run a unified benchmark; the quantitative tables are collected from prior papers and contain missing entries (“-”), so comparisons are incomplete, and the paper itself calls for reliable benchmark evaluation as a future direction.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Figures






Verification
Claims the paper makes
66 Valid · 26 Not valid · 63 Not checked · 80 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Methods for constructing multiple contrastive views based on multi-granularity or multi-scale augmentations also include MRLF [131], CMLF [115], and SSLAPP [132].”
What the sources say
- Not checked — no full text of the sourcecites [131] source
- Not checked — no full text of the sourcecites [115] source
- Partly supportedcites [132] sourceExperimental — accuracy not yet measured
Sections 3.2 and 3.3 describe SSLAPP constructing two augmented views as a positive pair and learning from global and local features. But Section 3.2 explicitly says SSLAPP uses pairwise representation learning instead of contrastive learning; the global and local distinction applies to its loss terms, not to multi-scale augmentations. The citation identifies SSLAPP correctly but mischaracterizes its method. (LLM, conf=0.88)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
Every source the paper cites for this claim was checked and supports it.
In the paper
“However, introducing reasonable inductive bias or prior is helpful for many deep neural networks-based modeling tasks [140], [196], [197].”
What the sources say
- Supportedcites [140] source
The Introduction (§1.1; page not provided) says integrating prior knowledge into a learning framework is effective for sparse data and describes its use across several applications. This supports the broad claim that suitable priors can help deep-learning modeling tasks. (LLM, conf=0.85)
- Supportedcites [196] source
The Network Architecture section says that adding domain-specific components to neural networks can boost performance and that task-targeted architectures can generalize better. It gives examples for different data and tasks, supporting the broad claim that appropriate inductive biases are helpful in neural-network modeling. (LLM, conf=0.91)
- Supportedcites [197] Chen, 2022, Integration of knowledge and data in machine learning source
The Introduction (page not available in our copy) says that embedding domain knowledge can improve model accuracy and robustness, and that it can improve machine-learning performance. Section 3.1 describes several ways to incorporate prior knowledge into neural networks across different applications, supporting the citing sentence’s broad, qualified claim. (LLM, conf=0.91)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Rests on 2 assumptions not shown here.
- Supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“Unlike previous frameworks, SSGAN's classification network makes full use of label information, which helps the model achieve more accurate imputations.”
What the sources say
- Supportedcites [162] source
The Classifier Network passage says SSGAN’s classifier uses self-training to make full use of label information in partially labeled time series and guides the generator during imputation. The Ablation Study reports that removing the classifier reduces imputation accuracy by an average of 1.36%, supporting its contribution to more accurate imputations. (LLM, conf=0.94)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“It is beneficial to introduce an adversarial-based SSL, which can further amplify the difference between normal and anomalous data, such as USAD [66] and DAEMON [184].”
What the sources say
- Partly supportedcites [66] sourceExperimental — accuracy not yet measured
Section 5.4 (Ablation Study; page unavailable) reports that adversarial training improves USAD’s F1 score and helps detect anomalies close to normal data. Section 3.2 describes USAD as an unsupervised autoencoder method with adversarial training, but the copy does not identify it as self-supervised learning (SSL). (LLM, conf=0.82)
- Partly supportedcites [184] sourceExperimental — accuracy not yet measured
The Abstract and §2.2 identify DAEMON as an unsupervised anomaly detector that uses adversarial training to learn normal time-series patterns; the model-variant results in §4.5.2 report a benefit from adversarial regularization. The paper does not describe DAEMON as self-supervised learning (SSL) or establish that adversarial training amplifies the difference between normal and anomalous data. The cited paper is DAEMON, so this verdict does not assess USAD. (LLM, conf=0.84)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“In recent years, many scholars have also explored the potential of generative representation learning in time series generation and imputation, such as C-RNN-GAN [154], TimeGAN [155], TTS-GAN [156], and E 2 GAN[157].”
What the sources say
- Partly supportedcites [154] Mogren, 2016, C-RNN-GAN: Continuous recurrent neural networks with adversarial training sourceExperimental — accuracy not yet measured
The Introduction (p. 1) identifies C-RNN-GAN as an adversarial model that generates continuous sequences, demonstrated with classical music. This supports citing it as an example of sequence generation, but the paper does not describe time-series imputation or establish it as an imputation method. (LLM, conf=0.84)
- Not matchedcites [155] Yoon, Time-Series Generative Adversarial Networks
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Supportedcites [156] sourceExperimental — accuracy not yet measured
The TTS-GAN paper identifies itself as a transformer-based GAN for generating synthetic time-series data (title and abstract, p. 1). Citation [156] is attached to TTS-GAN as an example in the list; the paper supports its role as a time-series generation method. (LLM, conf=0.93)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
44 more claims · 15 with their cited sources checked
Create a free account to see all 235 claims and to verify the 10 references still open.
- Probably not a substantial claim. 158
- It may state two separate claims as one. 79
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ I.2 Correlation analysis between methods and tasks · p. 26
“| Method | 15 UCR Datasets | 15 UCR Datasets | 15 UCR Datasets | 15 UCR Datasets |”
§ I.2 Correlation analysis between methods and tasks · p. 26
“| | Beef | ChCo | Coffee | CricketX | DiPOA | DiPOC | ECG200 | ECGF | Car | Ham | Herring | Wine | Meat | MiPOA | Wafer | Yoga |”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ I.1 Quantitative performance comparisons · p. 25
“| RANSynCoder [63] | 99 | 96 | 93 | 90 | 85 | 83 |”
§ H.1 Anomaly detection · p. 23
“F1-score is the harmonic mean of precision and recall, i.e.,”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ I.1 Quantitative performance comparisons · p. 25
“| ImDiffusion [101] | 98.11 | 97.53 | 97.81 | 95.2 | 95.09 | 94.88 |”
§ H.1 Anomaly detection · p. 23
“F1-score is the harmonic mean of precision and recall, i.e.,”
Citations
Cites
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2018) · arXiv · doi:10.48550/arxiv.1810.04805 · Cited by 73,173 · cited 2×
“With the help of well-designed pretext tasks, SSL has recently achieved great success in the domains of Computer Vision (CV) [7], [8], [9], [10] and Natural Language Processing (NLP) [11], [12].”
BackgroundMethodUnpaired image-to-image translation using cycle-consistent adversarial networks
Zhu, J.-Y., Park, T., Isola, P., & Efros, A. A. (2017) · 2017 IEEE International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2017.244 · Cited by 23,602
Image-to-Image Translation with Conditional Adversarial Networks
Isola, P., Zhu, J.-Y., Zhou, T., & Efros, A. A. (2017) · 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2017.632 · Cited by 22,930 · cited 1×
“The generator in the GAN can generate synthetic data close to the real data, so adversarial representation learning has a wide range of applications in the data generation field [141], especially in image generation [142]–[145].”
BackgroundAuto-Encoding Variational Bayes
Kingma, D. P., & Welling, M. (2013) · arXiv · doi:10.48550/arxiv.1312.6114 · Cited by 20,067
Momentum Contrast for Unsupervised Visual Representation Learning
He, K., Fan, H., Wu, Y., Xie, S., & Girshick, R. (2020) · 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr42600.2020.00975 · Cited by 13,347 · cited 1×
“The major categories include end-to-end [16], memory bank [34], momentum encoder [35], and clustering [36].”
BackgroundRepresentation Learning: A Review and New Perspectives
Bengio, Y., Courville, A., & Vincent, P. (2013) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2013.50 · Cited by 13,001 · cited 1×
“Using the learning paradigm as a taxonomy is arguably the most popular among the existing SSL surveys, including [20], [22], [23], [24], [25], [26], [27].”
MethodKirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A. C., Lo, W.-Y., et al. (2023) · 2023 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv51070.2023.00371 · Cited by 11,953 · cited 1×
“A similar phenomenon also appears in computational vision [203].”
BackgroundA Style-Based Generator Architecture for Generative Adversarial Networks
Karras, T., Laine, S., & Aila, T. (2019) · 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2019.00453 · Cited by 11,714 · cited 1×
“The generator in the GAN can generate synthetic data close to the real data, so adversarial representation learning has a wide range of applications in the data generation field [149], especially in image generation [150], [151], [152], [153].”
BackgroundA simple framework for contrastive learning of visual representations
Chen, T., Kornblith, S., Norouzi, M., & Hinton, G. (2020) · arXiv · doi:10.48550/arxiv.2002.05709 · Cited by 10,857
Explaining and Harnessing Adversarial Examples
Goodfellow, I. J., Shlens, J., & Szegedy, C. (2014) · arXiv · doi:10.48550/arxiv.1412.6572 · Cited by 10,381 · cited 1×
“…various recent and typical adversarial attack and defense methods for fault classifiers in data-driven fault detection and classificationsystems,includingwhite-boxattack(FGSM[209], IGSM[210],C & W attack [211], DeepFool [212]) and gray-box and black-box attack (UAP [213], SPSA [214], Random noise).”
Background
Research works printed but not linked: 12
- Cleveland, 1990, STL: A seasonal-trend decomposition, J. Official Stat., 6, 3
- 10.5555/3524938.3525087 · doi:10.5555/3524938.3525087
- Dai, Graph-augmented normalizing flows for anomaly detection of multiple time series, 1
- Chen, Deep variational graph convolutional recurrent network for multivariate time series anomaly detection, 3621
- 10.1109/tkde.2024.3361474/mm1 · doi:10.1109/tkde.2024.3361474/mm1
- Zhang, Time series anomaly detection using skip-step contrastive predictive coding, 1
- 10.1371/journal.pone.0254841 · DOI · Paper page
- 10.5555/2969033.2969125 · doi:10.5555/2969033.2969125
- 10.1609/aaai.v35i10.17070 · DOI · Paper page
- Zhang, Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting, 1
- Wang, MICN: Multi-scale local and global context modeling for long-term series forecasting, 1
- Li, Moco SFL: Enabling cross-client collaborative self-supervised learning
Cited by
Jin, M., Koh, H. Y., Wen, Q., Zambon, D., Alippi, C., Webb, G. I., King, I., & Pan, S. (2024) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2024.3443141 · Cited by 567
Self-citationKim, J., Kim, H., Kim, H., Lee, D., & Yoon, S. (2025) · Artificial Intelligence Review · doi:10.1007/s10462-025-11223-9 · Cited by 263
Data-driven stock forecasting models based on neural networks: A review
Bao, W., Cao, Y., Yang, Y., Che, H., Huang, J., & Wen, S. (2024) · Information Fusion · doi:10.1016/j.inffus.2024.102616 · Cited by 211
Correlation-Aware Spatial–Temporal Graph Learning for Multivariate Time-Series Anomaly Detection
Zheng, Y., Koh, H. Y., Jin, M., Chi, L., Phan, K. T., Pan, S., Chen, Y. P., & Xiang, W. (2023) · IEEE Transactions on Neural Networks and Learning Systems · doi:10.1109/tnnls.2023.3325667 · Cited by 158
“Due to the capacity of modeling interrelationship of different entities in various domains, GNNs have been widely used in domains and applications including traffic [35], [42], recommender systems [43], drug discovery [44], [45], and anomaly detection [46], [47].”
BackgroundMethodSelf-citationA Survey on Time-Series Pre-Trained Models
Ma, Q., Liu, Z., Zheng, Z., Huang, Z., Zhu, S., Yu, Z., & Kwok, J. T.-Y. (2024) · IEEE Transactions on Knowledge and Data Engineering · doi:10.1109/tkde.2024.3475809 · Cited by 182
Metadata
- Authors
- Zhang, K., Wen, Q., Zhang, C., Cai, R., Jin, M., Liu, Y., Zhang, J. Y., Liang, Y., Pang, G., Song, D., & Pan, S.
- Affiliations
- Zhejiang University (CN)
- Alibaba Group (China) (CN)
- Zhejiang Normal University (CN)
- Monash University (AU)
- Ant Group (China) (CN)
- Hong Kong University of Science and Technology (HK)
- Singapore Management University (SG)
- University of Connecticut (US)
- Griffith University (AU)
- ORCID
- Venue
- IEEE Transactions on Pattern Analysis and Machine Intelligence
- Published
- 2024-10
- Volume
- 46
- Issue
- 10
- Pages
- 6775-6794
- ISSN
- 0162-8828 (ISSN-L)
- 2160-9292
- 1939-3539
- OpenAlex
- W4394699135
- Cited by
- 295
- References
- 232
- Citations by year
- 2026: 124 · 2025: 115 · 2024: 51
- FWCI
- 51.74
- Type
- journal article
- Language
- en
- Open access
- green
- arxiv.org
- Licence
- Locations
- IEEE Transactions on Pattern Analysis and Machine Intelligence (publishedVersion)
- PubMed (publishedVersion)
- arXiv (Cornell University) (submittedVersion, public-domain)
- Singapore Management University Institutional Knowledge (InK) (Singapore Management University) (submittedVersion, cc-by-nc-nd)
- Rare & Special e-Zone (The Hong Kong University of Science and Technology) (submittedVersion)
- Griffith Research Online (submittedVersion)
- arXiv (Cornell University)
- Funders
- Key R&D Project of Zhejiang Province (2024C01172)
- National Key R&D Program of China (2021YFB2012300)
- Topics
- Physical Sciences › Computer Science › Signal Processing › Time Series Analysis and Forecasting
- Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
- Social Sciences › Decision Sciences › Management Science and Operations Research › Stock Market Forecasting Methods
- Keywords
- self-supervised learning, time series analysis, contrastive learning, adversarial learning, pre-training and fine-tuning, time series forecasting, time series classification, time series anomaly detection, time series clustering, benchmark dataset
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Self-Supervised Learning for Time Series Analysis: Taxonomy, Progress, and Prospects. https://app.cochat.ai/papers/10.1109/tpami.2024.3387317
- The paper
- Zhang, K., Wen, Q., Zhang, C., Cai, R., Jin, M., Liu, Y., Zhang, J. Y., Liang, Y., Pang, G., Song, D., & Pan, S. (2024). Self-Supervised Learning for Time Series Analysis: Taxonomy, Progress, and Prospects. IEEE Transactions on Pattern Analysis and Machine Intelligence. https://doi.org/10.1109/tpami.2024.3387317
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.