Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

PromptTSS: A Prompting-Based Approach for Interactive Multi-Granularity Time Series Segmentation

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read PromptTSS claims that a single prompt-conditioned Transformer model can segment multivariate time series at multiple levels of state granularity, using sparse label and boundary prompts to guide inference and adapt to unseen patterns at…

desk verdict Good prompt-based TSS idea, but the transfer numbers rely on an undefined label space that the architecture as written cannot support. read the letter →

arxiv 2506.11170 v2 pith:WPOGTTPR submitted 2025-06-12 cs.LG cs.AI

Ching Chang , Ming-Chih Lo , Wen-Chih Peng , Tien-Fu Chen This is my paper · ORCID
classification cs.LGcs.AI
keywords timeseriessegmentationpromptingmulti-granularityinteractivestatedetectiontransferlearningTransformerpatching
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

This paper claims that one neural model, PromptTSS, can segment multivariate time series at several levels of state granularity at once, instead of requiring a separate model per granularity. The model takes sparse user prompts — a few timesteps labeled with their state and a few marked as state boundaries — and uses them to guide its predictions at test time, so it can adapt to new patterns or new granularities without retraining. On benchmark and industrial datasets the authors report 24.49% average accuracy improvement over the best baseline in multi-granularity segmentation, 17.88% in single-granularity, and up to 599.24% in transfer settings. If correct, the framework would make interactive time series labeling practical for predictive maintenance and activity monitoring, where users can correct or direct a segmenter with a handful of annotations.

What carries the argument

The load-bearing object is the two-way Transformer decoder (called TWTL), which performs self-attention on prompt embeddings, cross-attention from prompts to time series embeddings, and cross-attention in the reverse direction so that prompt guidance and observed signal mutually influence each other over stacked layers. Label prompts are encoded as a 2K-dimensional vector (one-hot positive plus multi-hot negative) and boundary prompts as a binary lookup-table embedding, both projected to the same dimension and added; the decoder then produces per-timestep state probabilities through a shared softmax over K states. Iterative training progressively adds sampled prompts across iterations so the model learns to integrate increasing amounts of user guidance without relying on it.

What would settle it

Take a dataset with fine states A and B merged into coarse state C; train PromptTSS on both granularities, then at test time prompt a few timesteps with label C and check whether the model assigns every constituent fine-state timestep to C (or to a token representing C). If the model instead predicts A or B on those timesteps, its multi-granularity mechanism does not actually implement hierarchical state merging, and the reported gains may stem from prompt memorization rather than granularity control.

Watch

Extended reading notes

Core claim

PromptTSS's central claim is that a single model with a shared label space and a prompting mechanism can represent and predict states at multiple granularities, and that prompt-guided inference can adapt it to unseen datasets and unseen granularity levels. The label and boundary prompts are encoded into embeddings that are added to the time series representation inside a two-way Transformer decoder, so the same weights produce coarse or fine segmentations depending on the prompt. The paper validates this on eight datasets, including a proprietary industrial dataset with naturally annotated coarse and fine states, and reports large accuracy gains over seven baselines in both multi-granularity and single-granularity settings.

Load-bearing premise

The model assumes all granularity levels fit into one shared label space of K states with a single output layer, yet the paper does not define how a coarse state maps to the fine states it contains, so the mechanism behind multi-granularity prediction is underspecified.

Editorial extensions

If this is right

  • One PromptTSS model can replace a family of granularity-specific segmenters, since granularity is selected by prompt at inference time.
  • Sparse supervision suffices: at 5% prompted timesteps the model exceeds 95% accuracy on several datasets, and 1% already gives 89.93% on USC-HAD.
  • Because prompts guide inference directly, the model can be applied to new, related datasets (e.g., different pump types) with no retraining and only sparse labels.
  • The same model transfers across granularity levels, letting users annotate coarse states and still get fine-grained segmentations, albeit at lower accuracy (38.59% in the hardest direction).
  • Training is efficient enough for long windows because patching cuts the token count.

Reading between the lines

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

  • The reported transfer gains suggest prompt-conditioning may act as a fast adaptation mechanism comparable to in-context learning for time series; a natural test is whether a model trained on many datasets could segment a brand-new domain from scratch with only prompt labels.
  • The paper leaves the coarse-to-fine label mapping implicit; if the shared K-dimensional label space stores each state as an independent token, then prompting with a coarse label may simply retrieve that coarse token rather than composing fine states, so exact behavior on union-of-states prompts is an open empirical question.
  • The two-way decoder idea could be reused for other sequential tagging tasks, such as event detection or audio segmentation, where users supply sparse corrections at inference time.
  • A testable extension is to quantify how prompt positions matter: the paper samples uniformly, but prompting only at state boundaries may be far more efficient.
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

4 major / 4 minor

Summary. PromptTSS proposes a single Transformer-based model for multivariate time series segmentation across multiple granularities, conditioned on two prompt types: label prompts (positive one-hot and negative multi-hot over a K-state vocabulary) and binary boundary prompts. The model uses a patched time series encoder, a prompt encoder, an additive combination of prompt embeddings, a two-way Transformer decoder, and an iterative training schedule that progressively adds randomly sampled prompts. Experiments cover multi-granularity segmentation (Table 2), single-granularity segmentation (Table 3), transfer to unseen datasets and unseen granularity levels (Table 4), and ablation studies on prompt percentage, training iterations, window length, and encoder architecture. The paper reports large average accuracy improvements and releases code.

Significance. If the results hold, the paper would contribute a practical interactive mechanism for TSS: a single model whose output granularity and test-time behavior can be steered by sparse user-provided labels and boundaries, which is a genuinely underexplored direction compared with static TSS baselines. The iterative training strategy that mimics incremental user prompting is a reasonable design, and the release of code plus the breadth of ablations are strengths. However, the evidence as presented is incomplete: the label-space construction is underspecified, the transfer results in Table 4 are not obviously realizable by the architecture in Eq. (7), and the baseline adaptation gives the baselines a fundamentally weaker form of prompting. The central idea is plausible but the empirical claims currently outrun the described method.

major comments (4)
  1. [§4.2.1, Eq. (7), Table 2] The label vocabulary K is never defined for multi-granularity data. The text says only that K is 'the total number of possible states' and that p_l,t ∈ {0,1}^{2K}, while Eq. (7) applies a single softmax over K states. It is not stated whether coarse labels are the same labels as fine labels (as unions), disjoint new labels, or a dataset-specific set. If coarse and fine labels are disjoint entries in one K-dimensional space, the model is simply doing multiclass classification over a union and no coarse-to-fine relationship is learned; if they are unions, no mechanism in Section 4 encodes that hierarchy. Table 2 therefore does not support the claim of 'multi-granularity' modeling as described; the authors need to define the label hierarchy and show how Eq. (7) can represent it.
  2. [§5.3.2, Table 4 bottom half, Eqs. (7) and (9)] The transfer-to-unseen-granularity results are not realizable by the described architecture. In the setting 'Pump V35 (Original, 2x Coarser) → Pump V35 (4x Coarser)', the training loss in Eq. (9) never contains examples labeled with 4x coarser states. If K is fixed to the training labels, the linear layer in Eq. (7) has no output unit for 4x states; if K includes 4x states from the start, the corresponding columns of Linear_s and Linear_l receive no gradient and remain at their random initialization. There is no described mechanism (e.g., an output embedding shared between coarse and fine states, or a dynamic classifier) by which a one-hot prompt for an unseen state can produce correct predictions. The reported 61.60% accuracy therefore needs either a concrete architectural explanation or a revised experiment with K and training coverage stated explicitly. The same concern applies to unseen-dataset transfer (V35→V36/V38) unless all target labels coincide with V35 labels, which is not stated.
  3. [§5, Baselines paragraph, Tables 2–4] The baseline adaptation is a post-hoc model-selection procedure, not prompt-conditioned inference. Each baseline trains separate sub-models at each granularity and then, at test time, the sub-model whose predictions best match the user-provided prompts is selected. This gives baselines no per-timestep conditioning on label and boundary information and no mechanism for using prompts to refine a prediction. The reported improvements of 24.49% and 17.88% therefore cannot be attributed solely to the prompting mechanism; they also reflect the fact that PromptTSS receives sparse supervision during inference while baselines do not. A fairer comparison would use a single baseline model trained jointly on all granularities with a granularity indicator as an additional input, or would fine-tune baselines on the prompted timesteps.
  4. [Tables 2–4 and Figures 4–6] No standard deviations, confidence intervals, or significance tests are reported for any result. All numbers are single-point estimates, and the headline transfer improvements (up to 599.24%) are relative gains over baselines near random accuracy, so the reader cannot assess the stability of these differences. The ablation curves also show point values without variance. The authors should report mean and standard deviation over multiple random seeds, and where feasible paired significance tests, for at least the main tables.
minor comments (4)
  1. [§5.3.2] The text states that the second-best baseline achieves 8.81% accuracy in the unseen-granularity setting, but Table 4 bottom half shows second-best values of 9.68%, 9.42%, and 7.33% for the three rows; the number should be reconciled.
  2. [§5.4.1] The text refers to 'a prompt encoder and a mask decoder,' while Section 4 describes a state decoder; the terminology should be made consistent.
  3. [§4.2.1, Eq. (5)] The indicator I(p_l,t) is not formally defined for the case where no label prompt is provided; since a missing prompt is represented by a zero vector, the reader cannot tell whether I is based on any nonzero entry or on an explicit mask. Please clarify the definition.
  4. [§4.1] Patching uses patch length 16 and stride 8, which implies overlapping patches, but the paper does not mention overlap in the method description; please state whether patches overlap and how the prompt timestep alignment is handled.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the prompting inputs are explicitly part of the interactive task, and the reported gains are empirical benchmark comparisons rather than derivations from the inputs.

full rationale

PromptTSS is an empirical ML system, not a claimed derivation chain with theorems. Its core mechanism is that label and boundary prompts are provided as inputs at inference, which the paper states explicitly: 'two types of ground truth information are used as prompts' and users provide 'sparse corrections or partial annotations.' The experiments evaluate prediction on all time steps while supplying prompts for only 5% of time steps, so the reported accuracy is not forced by construction: a model that simply copied the prompted labels would receive at most 5% accuracy over the full window. Equation (7) is a standard softmax over the state vocabulary, and Eq. (9) is ordinary cross-entropy; no output quantity is defined as a function of the quantity it is said to predict. The only overlapping-author citation, [29] COKE, appears in related-work discussion of hierarchical classification and is descriptive context, not load-bearing support for PromptTSS's design or results. The reviewer-identified K / unseen-granularity mapping issue is a genuine specification and architectural-support gap: Section 4.2.1 defines K as 'the total number of possible states' without specifying how K is fixed across granularities, so the transfer experiments in Table 4 may not be implementable as described. But that is a correctness and reproducibility concern, not circularity: no equation in the paper reduces the transfer-learning result to the training labels or to a fitted parameter renamed as a prediction. Under the hard rules, that does not qualify as a circular step.

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

No new physical entities or forces are introduced. The main under-specification is in label space construction and prompt conditioning, not in invented quantities.

free parameters (5)
  • Prompt coverage = 5% of time steps
    The paper chooses 5% prompting for all main results; accuracy depends strongly on prompt fraction (Fig. 4), so this is a hand-picked operating point.
  • Iterative training rounds N_r = 8
    Accuracy peaks at 8 and falls at 16 (Fig. 5); selected to maximize validation performance.
  • Window length T and stride S = T=256/S=64, USC-HAD T=512/S=128
    Set based on prior work; Fig. 6 shows accuracy varies with T, so this choice is consequential.
  • Patch length and stride = 16 and 8
    Hyperparameters chosen without sensitivity analysis; they affect temporal resolution of the encoder.
  • Prompt sample range (n_min, n_max) = 1 to 3 prompts per iteration
    Controls how quickly prompts accumulate during iterative training.
assumptions (5)
  • ad hoc to paper All granularity levels share one K-dimensional label vocabulary with a well-defined mapping between coarse and fine states
    Section 4.2.1 defines K as total number of possible states but never states whether K is the union of coarse and fine labels or how merged states map to fine labels; Eq. (7) uses a single softmax.
  • domain assumption Merging adjacent states yields valid coarse labels that a shared model can learn
    Section 5, datasets paragraph; coarse versions are generated synthetically and may not reflect natural label hierarchies.
  • domain assumption Randomly sampled ground-truth prompts simulate realistic interactive user input
    Section 4.4 and 5.1; no noisy, spatially biased, or erroneous prompts are tested, so real-time user interactions may be harder.
  • ad hoc to paper Baseline post-hoc submodel selection is a fair adaptation for prompt comparison
    Section 5, Baselines; this mechanism gives baselines access to prompt labels only through output selection, not through conditioning, so the comparison is favorable to PromptTSS.
  • domain assumption Cross-entropy trained two-way attention can propagate sparse prompt constraints to all timesteps
    Core architectural assumption; no formalization or convergence analysis is given in Section 4.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PromptTSS: A Prompting-Based Approach for Interactive Multi-Granularity Time Series Segmentation." pith.science (2026). https://pith.science/paper/WPOGTTPR

@misc{pith2026250611170,
  author       = {Pith},
  title        = {Pith review of: PromptTSS: A Prompting-Based Approach for Interactive Multi-Granularity Time Series Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WPOGTTPR}},
  note         = {Machine review of arXiv:2506.11170}
}
read the original abstract

Multivariate time series data, collected across various fields such as manufacturing and wearable technology, exhibit states at multiple levels of granularity, from coarse-grained system behaviors to fine-grained, detailed events. Effectively segmenting and integrating states across these different granularities is crucial for tasks like predictive maintenance and performance optimization. However, existing time series segmentation methods face two key challenges: (1) the inability to handle multiple levels of granularity within a unified model, and (2) limited adaptability to new, evolving patterns in dynamic environments. To address these challenges, we propose PromptTSS, a novel framework for time series segmentation with multi-granularity states. PromptTSS uses a unified model with a prompting mechanism that leverages label and boundary information to guide segmentation, capturing both coarse- and fine-grained patterns while adapting dynamically to unseen patterns. Experiments show PromptTSS improves accuracy by 24.49% in multi-granularity segmentation, 17.88% in single-granularity segmentation, and up to 599.24% in transfer learning, demonstrating its adaptability to hierarchical states and evolving time series dynamics. Our code is available at https://github.com/blacksnail789521/PromptTSS.

Figures

Figures reproduced from arXiv: 2506.11170 by the authors.

Figure 1
Figure 1. Multiple-Granularity States and Prompts. The top [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. PromptTSS framework. Consists of three modules: a time series encoder that generates time series embeddings, a [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Training and inference time on USC-HAD dataset. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Ablation Study: Number of Iterations in Iterative [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 7
Figure 7. Figure 7: Ablation Study: Comparison of different encoder [PITH_FULL_IMAGE:figures/full_fig_p009_7.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. A Survey of Reasoning and Agentic Systems in Time Series with Large Language Models

    cs.AI 2025-09 conditional novelty 5.0 of 10

    The authors organize LLM-based time series reasoning into three exclusive topologies (direct, chain, branch) crossed with four objectives, and use them to label 125 papers, benchmarks, and resources.

Reference graph

Works this paper leans on

48 extracted references · 35 canonical work pages · cited by 1 Pith paper

  1. [1]

    Andreas Bulling, Ulf Blanke, and Bernt Schiele. 2014. A Tutorial on Human Activity Recognition Using Body-worn Inertial Sensors. Comput. Surveys 46, 3 (2014), 33:1–33:33. doi:10.1145/2499621

    doi:10.1145/2499621 2014

  2. [2]

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexan- der Kirillov, and Sergey Zagoruyko. 2020. End-to-end object detection with transformers. In European conference on computer vision . Springer, 213–229

    2020

  3. [3]

    Mouxiang Chen, Lefei Shen, Han Fu, Zhuo Li, Jianling Sun, and Chenghao Liu

  4. [4]

    Bowen Cheng, Alex Schwing, and Alexander Kirillov. 2021. Per-pixel classifica- tion is not all you need for semantic segmentation.Advances in neural information processing systems 34 (2021), 17864–17875

    2021

  5. [5]

    Shohreh Deldari, Daniel V Smith, Amin Sadri, and Flora Salim. 2020. Espresso: Entropy and shape aware time-series segmentation for processing heteroge- neous sensor data. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 4, 3 (2020), 1–24

    work page 2020

  6. [6]

    Emadeldeen Eldele, Mohamed Ragab, Zhenghua Chen, Min Wu, Chee-Keong Kwoh, and Xiaoli Li. 2024. Label-efficient time series representation learning: A review. IEEE Transactions on Artificial Intelligence (2024)

    work page 2024

  7. [7]

    Stefan Gaugel and Manfred Reichert. 2023. PrecTime: A deep learning architec- ture for precise time series segmentation in industrial manufacturing operations. Engineering Applications of Artificial Intelligence 122 (2023), 106078

    work page 2023

  8. [8]

    Shaghayegh Gharghabi, Chin-Chia Michael Yeh, Yifei Ding, Wei Ding, Paul Hibbing, Samuel LaMunion, Andrew Kaplan, Scott E Crouter, and Eamonn Keogh

    work page

Show all 48 references
  1. [9]

    Dimitra Gkatzia, Helen Hastie, and Oliver Lemon. 2014. Comparing multi-label classification with reinforcement learning for summarisation of time-series data. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) . 1231–1240

    2014

  2. [10]

    Yuan Gong, Yu-An Chung, and James Glass. 2021. Ast: Audio spectrogram transformer. arXiv preprint arXiv:2104.01778 (2021)

    2021 arXiv

  3. [11]

    David Hallac, Sagar Vare, Stephen Boyd, and Jure Leskovec. 2017. Toeplitz inverse covariance-based clustering of multivariate time series data. In Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining. 215–223

    2017

  4. [12]

    Zeineb Hammami, Moamar Sayed-Mouchaweh, Wiem Mouelhi, and Lamjed Ben Said. 2020. Neural networks for online learning of non-stationary data streams: a review and application for smart grids flexibility improvement. Artifi- cial Intelligence Review 53, 8 (2020), 6111–6154

    2020

  5. [13]

    Huan He, Owen Queen, Teddy Koker, Consuelo Cuevas, Theodoros Tsiligkaridis, and Marinka Zitnik. 2023. Domain adaptation for time series under feature and label shifts. In International conference on machine learning . PMLR, 12746–12774

    2023

  6. [14]

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al

  7. [15]

    Jennifer R Kwapisz, Gary M Weiss, and Samuel A Moore. 2011. Activity recogni- tion using cell phone accelerometers. ACM SigKDD Explorations Newsletter 12, 2 (2011), 74–82

    2011

  8. [16]

    Zhichen Lai, Huan Li, Dalin Zhang, Yan Zhao, Weizhu Qian, and Christian S Jensen. 2024. E2Usd: Efficient-yet-effective Unsupervised State Detection for Multivariate Time Series. In Proceedings of the ACM on Web Conference 2024 . 3010–3021

    2024

  9. [17]

    Wei-Han Lee, Jorge Ortiz, Bongjun Ko, and Ruby Lee. 2018. Time series seg- mentation through automatic feature learning. arXiv preprint arXiv:1801.05394 (2018)

    2018 arXiv

  10. [18]

    Shijie Li, Yazan Abu Farha, Yun Liu, Ming-Ming Cheng, and Juergen Gall. 2020. Ms-tcn++: Multi-stage temporal convolutional network for action segmentation. IEEE transactions on pattern analysis and machine intelligence 45, 6 (2020), 6647– 6658

    2020

  11. [19]

    Qin Liu, Jaemin Cho, Mohit Bansal, and Marc Niethammer. 2024. Rethinking Interactive Image Segmentation with Low Latency High Quality and Diverse Prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 3773–3782

    2024

  12. [20]

    Weiwei Liu, Haobo Wang, Xiaobo Shen, and Ivor W Tsang. 2021. The emerging trends of multi-label learning. IEEE transactions on pattern analysis and machine intelligence 44, 11 (2021), 7955–7974

    2021

  13. [21]

    Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. 2024. iTransformer: Inverted Transformers Are Effective for Time Series Forecasting. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-1...

    2024

  14. [22]

    Sabarinath Mahadevan, Paul Voigtlaender, and Bastian Leibe. 2018. Iteratively trained interactive segmentation. arXiv preprint arXiv:1805.04398 (2018)

    2018 arXiv

  15. [23]

    Lakmal Meegahapola, Hamza Hassoune, and Daniel Gatica-Perez. 2024. M3BAT: Unsupervised Domain Adaptation for Multimodal Mobile Sensing with Multi- Branch Adversarial Training. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 8, 2 (2024), 1–30

    2024

  16. [24]

    Masatoshi Nagano, Tomoaki Nakamura, Takayuki Nagai, Daichi Mochihashi, Ichiro Kobayashi, and Wataru Takano. 2019. HVGH: unsupervised segmentation for high-dimensional time series using deep neural compression and statistical generative model. Frontiers in Robotics and AI 6 (2019), 115

    2019

  17. [25]

    Christoforos Nalmpantis and Dimitris Vrakas. 2020. On time series represen- tations for multi-label NILM. Neural Computing and Applications 32 (2020), 17275–17290

    2020

  18. [26]

    Ashwin Narayan, Francisco Anaya Reyes, Meifeng Ren, and Yu Haoyong. 2021. Real-time hierarchical classification of time series data for locomotion mode detection. IEEE Journal of Biomedical and Health Informatics 26, 4 (2021), 1749– 1760

    2021

  19. [27]

    Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam

    Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. 2023. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. In ICLR. OpenReview.net

    2023

  20. [28]

    Francisco Javier Ordóñez and Daniel Roggen. 2016. Deep convolutional and lstm recurrent neural networks for multimodal wearable activity recognition. Sensors 16, 1 (2016), 115

    2016

  21. [29]

    Ting-Yun Ou, Ching Chang, and Wen-Chih Peng. 2024. COKE: Causal Discovery with Chronological Order and Expert Knowledge in High Proportion of Missing Manufacturing Data. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management . 4803–4810

    2024

  22. [30]

    Fengchao Peng, Qiong Luo, and Lionel M Ni. 2017. ACTS: an active learning method for time series classification. In 2017 IEEE 33rd International Conference on Data Engineering (ICDE) . IEEE, 175–178

    2017

  23. [31]

    Mathias Perslev, Michael Jensen, Sune Darkner, Poul Jørgen Jennum, and Chris- tian Igel. 2019. U-time: A fully convolutional network for time series segmentation applied to sleep staging. Advances in Neural Information Processing Systems 32 (2019)

    2019

  24. [32]

    Huy Phan, Oliver Y Chén, Philipp Koch, Zongqing Lu, Ian McLoughlin, Alfred Mertins, and Maarten De Vos. 2020. Towards more accurate automatic sleep staging via deep transfer learning. IEEE Transactions on Biomedical Engineering 68, 6 (2020), 1787–1798

    2020

  25. [33]

    Marco S Reis. 2019. Multiscale and multi-granularity process analytics: A review. Processes 7, 2 (2019), 61

    2019

  26. [34]

    Attila Reiss and Didier Stricker. 2012. Introducing a new benchmarked dataset for activity monitoring. In 2012 16th International Symposium on Wearable Computers. IEEE, 108–109

    2012

  27. [35]

    Pranab Sahoo, Ayush Kumar Singh, Sriparna Saha, Vinija Jain, Samrat Mondal, and Aman Chadha. 2024. A systematic survey of prompt engineering in large language models: Techniques and applications. arXiv preprint arXiv:2402.07927 (2024)

    2024 arXiv

  28. [36]

    Patrick Schäfer, Arik Ermshaus, and Ulf Leser. 2021. Clasp-time series segmenta- tion. In Proceedings of the 30th ACM international conference on information & knowledge management. 1578–1587

    2021

  29. [37]

    Burr Settles. 2009. Active learning literature survey. (2009)

    2009

  30. [38]

    Yanru Sun, Zongxia Xie, Dongyue Chen, Emadeldeen Eldele, and Qinghua Hu

  31. [39]

    Chengyu Wang, Kui Wu, Tongqing Zhou, and Zhiping Cai. 2023. Time2state: An unsupervised framework for inferring the latent states in time series data. Proceedings of the ACM on Management of Data 1, 1 (2023), 1–18

    2023

  32. [40]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems 35 (2022), 24824–24837

    2022

  33. [41]

    Ning Xu, Brian Price, Scott Cohen, Jimei Yang, and Thomas S Huang. 2016. Deep interactive object selection. In Proceedings of the IEEE conference on computer vision and pattern recognition . 373–381

    2016

  34. [42]

    Yu Yang, Sthitie Bom, and Xiaotong Shen. 2024. A hierarchical ensemble causal structure learning approach for wafer manufacturing. Journal of Intelligent Manufacturing 35, 6 (2024), 2961–2978. PromptTSS: A Prompting-Based Approach for Interactive Multi-Granularity Time Series ...

    2024

  35. [43]

    Mi Zhang and Alexander A Sawchuk. 2012. USC-HAD: A daily activity dataset for ubiquitous activity recognition using wearable sensors. In Proceedings of the 2012 ACM conference on ubiquitous computing . 1036–1043

    2012

  36. [44]

    Xu Zhang, Kailun Yang, Jiacheng Lin, Jin Yuan, Zhiyong Li, and Shutao Li. 2024. PVPUFormer: Probabilistic Visual Prompt Unified Transformer for Interactive Image Segmentation. IEEE Transactions on Image Processing (2024)

    2024

  37. [2019]

    Data mining and knowledge discovery 33 (2019), 96–130

    Domain agnostic online semantic segmentation for multi-dimensional time series. Data mining and knowledge discovery 33 (2019), 96–130

    2019

  38. [2023]

    In Proceedings of the IEEE/CVF International Conference on Computer Vision

    Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 4015–4026

  39. [2024]

    In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

    Calibration of time-series forecasting: Detecting and adapting context- driven distribution shift. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 341–352

  40. [2025]

    In Proceedings of the AAAI Conference on Artificial Intelligence, Vol

    Hierarchical classification auxiliary network for time series forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 20743–20751

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.