Pith. sign in

REVIEW 5 major objections 5 minor 55 references

InstructTime++: Time Series Classification with Multimodal Language Modeling via Implicit Feature Enhancement

T0 review · 5 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read Time series classification can be cast as a text-generation task: a language model that reads discretized signals plus statistical and visual descriptions as text outperforms specialized classifiers, according to this paper.

desk verdict Clever framework, shaky evidence: duplicated tables, a mismatched ablation control, and ECG 'implicit' features that look hand-engineered — send to review, but with a major-revision request. read the letter →

arxiv 2601.14968 v2 pith:QJHA5VUM submitted 2026-01-21 cs.LG cs.AI

Mingyue Cheng , Xiaoyu Tao , Huajian Zhang , Qi Liu , Zhiding Liu , Yucong Luo , Yiheng Chen , Enhong Chen This is my paper · ORCID
classification cs.LGcs.AI
keywords timeseriesclassificationmultimodallanguagemodelgenerativevectorquantizationimplicitfeatureextractioninstructiontuningcross-domainpre-trainingvisualgrounding
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

The paper tries to establish that a generative language model, fed vector-quantized time series tokens along with textual instructions and context, can classify time series by generating the label as text—and that performance improves further when the model is also given 'implicit' features, namely statistical summaries and vision-language captions of the signal's plot, rendered as sentences. The authors argue this paradigm fixes two limits of classic classifiers: it lets contextual side information (age, gender, sensor type) enter the decision, and it lets the model exploit semantic relationships among class labels. If right, it suggests a general recipe: bridge continuous data into the language model's text world through discretized tokens, align embeddings with generative pre-training, and enrich the prompt with domain toolkits. The key evidence is the reported accuracy gains of InstructTime++ over discriminative baselines on five benchmarks, with ablations showing both statistical and visual text contribute.

What carries the argument

The argument rests on three mechanisms working together. First, a vector-quantized discretization module converts continuous signal patches into codebook tokens, turning numbers into discrete symbols a language model can consume. Second, an alignment MLP plus generative self-supervised pre-training anchors these temporal tokens near text embeddings, so the model can reason across modalities. Third, the implicit feature toolkit—predefined statistical functions computed from the signal and a vision-language model that captions a 2D plot of the signal—produces textual descriptions injected into the prompt. These textified features are the load-bearing addition in InstructTime++ and the reason t

What would settle it

Train a simple classifier (logistic regression or gradient boosting) on exactly the same statistical features used for each dataset, without any language model. If that classifier matches or exceeds InstructTime++'s accuracy on ECG, EEG, or RWC, the claimed contribution of the multimodal generative framework is not supported.

Watch

Extended reading notes

Core claim

The central claim is that time series classification can be reformulated as a multimodal generative task, and that explicitly textualizing implicit signal structure—statistical descriptors (mean, variance, entropy, trend, periodicity) and vision-language captions of the plotted signal—is what compensates for a language model's missing temporal inductive bias. The paper demonstrates that the full model, InstructTime++-Adapt, achieves the best or near-best accuracy and F1 on EEG, ECG, HAR, FD, and RWC datasets, and that removing either the statistical or the visual text descriptions degrades performance, showing the two feature types are complementary. The model also transfers to two unseen do

Load-bearing premise

The paper attributes the reported gains to the generative LLM framework, but the implicit features are selected per dataset and may encode label-defining information; if they do, the causal claim that the model's reasoning compensates for limited temporal inductive bias collapses.

Editorial extensions

If this is right

  • If the paradigm holds, time series classifiers can be replaced by generative models that output text, enabling free-form explanations and natural incorporation of side information such as patient demographics or sensor metadata.
  • Cross-domain generative pre-training on multiple time series domains transfers to entirely unseen domains (EP and SAD), suggesting a foundation-model route for time series classification rather than per-dataset training.
  • The ablation results imply that statistical descriptors and visual captions contribute complementary information; a model that fuses both is more robust than one relying on a single feature type.
  • Representing labels as language rather than one-hot vectors gives a way to handle multi-label classification and semantically related classes (e.g., walking vs. jogging), which the paper shows improves multi-label ECG accuracy.
  • The toolkit-based prompt design provides a template for grounding other non-linguistic modalities (e.g., audio, sensor streams) into language models by rendering their structure as text.

Reading between the lines

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

  • Editorial inference: The paper does not compare against a baseline that feeds the same statistical features to a conventional classifier (e.g., logistic regression or gradient boosting). If such a simple model matches InstructTime++'s accuracy, the generative language framework itself may add little beyond the hand-selected features.
  • Editorial inference: The per-dataset choice of statistical functions (e.g., ECG rhythm.rr_mean_s) may be informed by domain knowledge of which features are diagnostically decisive; this makes the causal attribution to 'compensating for limited inductive bias' fragile.
  • Editorial inference: A testable extension would replace the mined features with deliberately label-free features (e.g., random noise summaries); if performance collapses, it would show the features, not the model architecture, carry the discriminative signal.
  • Editorial inference: The framework suggests a general recipe for multimodal grounding, but its success likely depends on the quality and discriminativeness of the toolkits; a fair evaluation should fix the feature set across datasets to measure the genuine contribution of the language model.
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

5 major / 5 minor

Summary. This paper proposes InstructTime, a generative framework for time series classification that discretizes input series with VQ networks, projects the discrete tokens into an LM embedding space, and generates class labels as text conditioned on domain instructions and contextual features. InstructTime++ extends this by adding an 'implicit feature' channel: statistical features and VLM-derived visual descriptions are computed by 'specialized toolkits' and translated into text, then inserted into the prompt. Experiments are reported on EEG, ECG, HAR, FD, and RWC, plus transfer experiments on EP and SAD, with ablations over pre-training, instruction text, projector size, codebook size, patch size, and implicit feature components.

Significance. The paper targets a real and important limitation of LM-based time series classification: LMs have weak temporal inductive bias, and contextual/structural information is difficult to inject. The idea of translating implicit statistical and visual features into text is clean and potentially reusable, and the authors provide code and a broad ablation structure. However, the current evidence is not reliable enough to support the central claim: there is an internal contradiction between the described generic toolkit and the ECG features actually used, key experimental tables are duplicated under different names, and the comparisons lack statistical rigor and the closest baselines.

major comments (5)
  1. [§4.2.1 and Fig. 5] The generic statistical toolkit F is specified as mean, variance, skewness, kurtosis, sample/approximate entropy, linear trend, and periodicity. The ECG prompt in Fig. 5 contains rhythm.rr_mean_s, rhythm.rr_sd_s, rhythm.regular, frequency_hint.lf_hf, and dominant frequency. No rule in F produces these clinical indices; they must come from a per-dataset, domain-specific feature set. RR-interval statistics and LF/HF are standard cardiac diagnostics and are near-label markers for ECG abnormality. Consequently, the gains attributed to implicit feature modeling in Table 10 (ECG) may reflect hand-engineered diagnostic features rather than generic implicit pattern mining. The paper must provide a protocol for constructing F per dataset and an ablation using only the features actually derivable from the described toolkit.
  2. [§5.3.1 and §5.3.2, Tables 3 and 5] Tables 3 and 5 report identical numbers under different captions. For example, the EEG 'w/o Cross-domain' / 'w/ Cross-domain' rows in Table 3 are 0.7854/0.4854 and 0.8452/0.6240, exactly the same as the 'w/o Pre-training' / 'w/ Pre-training' rows in Table 5; the same pattern holds for ECG, HAR, FD, and RWC. Because one table is described as evaluating cross-domain pre-training and the other as evaluating autoregressive pre-training, the two effects are not separately identified. The corresponding claims in §5.3.1 and §5.3.2 are unsupported as written and need to be disentangled with clearly distinct experimental conditions.
  3. [Tables 2–12] All performance numbers are reported from a single run, with no standard deviations, confidence intervals, or significance tests. Many differences are small (e.g., HAR accuracy 0.9298 vs. 0.9294 in Table 2; EEG F1 0.6735 vs. 0.6240 in Table 2). Without repeated runs and error bars, the central claim of 'superior performance' is not statistically supported. At minimum, the main comparison table and the key ablation tables need multi-seed results and an appropriate significance test.
  4. [§5.1.2 and Table 2] The baseline set omits the closest LLM-based time series classifiers and standard feature-based pipelines. GPT-As-Classifier is a weak adaptation of GPT-2, and the self-supervised baselines are not multimodal/generative. The paper also does not compare against the authors' own hierarchical multimodal LLM work cited as [43], nor against feature-based pipelines such as tsfresh plus a classifier. Without these comparisons, the claim that InstructTime++ outperforms the relevant prior art is overstated.
  5. [§5.3.7, §5.4.3, Tables 9 and 12] Codebook size K and patch sizes are selected per dataset (e.g., Table 9 uses K ∈ {128,256,384,512,768} and Table 12 uses different patch tuples for different datasets). The paper does not describe a held-out validation procedure for selecting these hyperparameters, nor whether the same selection protocol was applied to all baselines. If per-dataset choices were made based on test-set performance, the reported gains could be an artifact of tuning. Please specify the model selection protocol and report sensitivity.
minor comments (5)
  1. [§5.1.1] The citations appear mismatched: ECG is described as derived from PhysioNet 2020, but reference [2] is the smartphone HAR dataset; HAR is cited as [12], which is the authors' cross-domain LM paper. Please correct the references.
  2. [§5.3.4] The sentence 'Most approaches achieve relatively strong performance in the multi-label classification, which is expected since multi-label classification is more challenging' is internally contradictory; Fig. 7 actually shows much lower accuracy for the multi-label setting. Please rephrase.
  3. [§3.3, Eq. (1)] The reconstruction term in Eq. (1) is written as a log-probability but the text implies an MSE-style reconstruction; please specify the exact objective, including how the straight-through estimator and codebook EMA are implemented.
  4. [§2.1 and §3.2.2] Output parsing is under-specified: how are 'predefined class-specific keywords' matched in the generated text, especially for the 27-class multi-label ECG dataset? Also, the prompt template in Fig. 3 does not actually list candidate labels despite the guideline in §3.2.2.
  5. [General] The text contains several typos and formatting errors: 'pre-fineded', 'V ectorQuantized', 'Table Table 9', and inconsistent table headers (e.g., Table 5 lacks a model-variants column while the text discusses both InstructTime-Universal and InstructTime-Adapt).

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the empirical claims are self-contained, self-citations are not load-bearing, and the ECG feature discrepancy is a leakage/validity concern rather than a definitional loop.

full rationale

InstructTime++ is an empirical pipeline, not a derivation; there is no analytic chain in which a predicted quantity is defined in terms of its inputs. The central claim of superior classification performance is supported by in-paper experiments (Table 2) against external baselines (PatchTST, MiniROCKET, TimeMAE, GPT-As-Classifier) that share no fitted parameters with the proposed method, so the headline benchmark result is not forced by construction. The authors cite their own prior work ([9], [12], [43], [44]) for scaffolding ideas such as VQ discretization and cross-domain pre-training, but these citations are not invoked as an external authority for the paper's empirical conclusion; the experiments and ablations are performed and reported here, so the self-citations are not load-bearing in the definitional sense. The one internal tension worth flagging explicitly: Section 4.2.1 describes a generic 'rule-driven statistical feature extraction' toolkit (mean, variance, skewness, kurtosis, sample/approximate entropy, linear trend, periodicity), while Figure 5's ECG prompt contains 'rhythm.rr_mean_s: 1.034', 'rhythm.rr_sd_s: 0.033', and 'frequency_hint.lf_hf' — standard clinical diagnostics that cannot be produced by the listed functions. This indicates per-dataset, domain-specific feature selection that may be informed by class definitions, meaning the ablation gains in Table 10 could stem from injecting near-label information rather than from 'enhancing multimodal reasoning capability' as claimed in Section 4.1. That is a genuine validity/leakage concern for the attribution story, but it is not circularity under the defined patterns: the class labels are not defined in terms of the features, the feature set is hand-selected rather than a fitted parameter whose prediction is statistically forced, and the reported accuracies are measured rather than derived. The score of 2 reflects the mild confirmation loop created by tuning the implicit-feature toolkits on the same datasets used to demonstrate gains and the density of same-author citations, without any exhibitable reduction of the central result to its own inputs.

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

The paper's empirical claims rest on several tuned components (codebook, patch, projector, feature toolkit, prompt template) and domain assumptions about LLM transfer and text-generation evaluation. No new physical or mathematical entity is introduced.

free parameters (5)
  • VQ codebook size (K) = 256 (selected per dataset from 128–768)
    Table 9 shows accuracy varies with token count; a 'sweet spot' is selected per dataset after evaluating classification performance.
  • Patch size per dataset = e.g., EEG (25,20,1,32,25); FD (40,25,2,40,32) etc.
    Table 12 tunes patch sizes per dataset; no held-out validation described, so these appear fit to the test sets.
  • Alignment projector architecture = MLP hidden sizes 64,128,256,512,768
    Table 8 selects the deepest MLP as best; this is a hand-chosen architecture.
  • Statistical feature set F = mean, variance, skewness, kurtosis, sample entropy, approximate entropy, linear trend, periodicity
    Hand-selected feature list in §4.2.1; no analysis of feature selection or leakage.
  • VLM image rendering rules = unspecified
    Rendering strategy is 'domain-informed' but not specified; a hand-designed component that absorbs dataset assumptions.
assumptions (5)
  • domain assumption VQ-VAE discretization with straight-through estimator preserves discriminative temporal information
    Section 3.2.1 assumes reconstruction objective yields codes useful for classification.
  • domain assumption A language model pre-trained on natural language can, after alignment, perform reasoning over numerical time-series tokens
    Section 3.4 assumes generative pre-training aligns temporal and textual modalities.
  • domain assumption Auto-regressive prediction of next temporal token is a beneficial auxiliary objective for classification
    Tables 3–5 present this as an empirical effect, but no theoretical grounding is given.
  • domain assumption The benchmark datasets and their standard train/test splits are appropriate and used correctly
    The results depend entirely on these datasets; preprocessing is not fully described.
  • domain assumption Generated text can be mapped to ground-truth classes via keyword identification
    Section 2.1 states labels are determined by identifying predefined class-specific keywords; parsing details are absent.

how reviews work

0 comments
Cite this review

Pith. "Pith review of InstructTime++: Time Series Classification with Multimodal Language Modeling via Implicit Feature Enhancement." pith.science (2026). https://pith.science/paper/QJHA5VUM

@misc{pith2026260114968,
  author       = {Pith},
  title        = {Pith review of: InstructTime++: Time Series Classification with Multimodal Language Modeling via Implicit Feature Enhancement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QJHA5VUM}},
  note         = {Machine review of arXiv:2601.14968}
}
read the original abstract

Most existing time series classification methods adopt a discriminative paradigm that maps input sequences directly to one-hot encoded class labels. While effective, this paradigm struggles to incorporate contextual features and fails to capture semantic relationships among classes. To address these limitations, we propose InstructTime, a novel framework that reformulates time series classification as a multimodal generative task. Specifically, continuous numerical sequences, contextual textual features, and task instructions are treated as multimodal inputs, while class labels are generated as textual outputs by tuned language models. To bridge the modality gap, InstructTime introduces a time series discretization module that converts continuous sequences into discrete temporal tokens, together with an alignment projection layer and a generative self-supervised pre-training strategy to enhance cross-modal representation alignment. Building upon this framework, we further propose InstructTime++, which extends InstructTime by incorporating implicit feature modeling to compensate for the limited inductive bias of language models. InstructTime++ leverages specialized toolkits to mine informative implicit patterns from raw time series and contextual inputs, including statistical feature extraction and vision-language-based image captioning, and translates them into textual descriptions for seamless integration. Extensive experiments on multiple benchmark datasets demonstrate the superior performance of InstructTime++.

Figures

Figures reproduced from arXiv: 2601.14968 by the authors.

Figure 1
Figure 1. Illustrating the overall pipeline of proposed InstructTime. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the overall framework of the proposed InstructTime. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Illustration of the newly proposed prompt template. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Illustration of the framework of extended InstructTime++. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Illustration of the newly proposed prompt template of extended InstructTime++. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Visualization of t-SNE embeddings for instances sampled from multiple datasets. [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Performance comparison of multi-label and single-label ECG classification. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Statistics of token usage frequency across multiple datasets. Each heatmap corresponds to one dataset [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Visualizing the reconstruction of ECG signals in the vector quantized networks. [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 4 linked inside Pith

  1. [43]

    Xiaoyu Tao, Tingyue Pan, Mingyue Cheng, and Yucong Luo. 2024. Hierarchical multimodal llms with semantic space alignment for enhanced time series classification.arXiv preprint arXiv:2410.18686(2024)

    arXiv 2024

  2. [47]

    Jiahao Wang, Mingyue Cheng, Qingyang Mao, Yitong Zhou, Feiyang Xu, and Xin Li. 2024. TableTime: Reformulating Time Series Classification as Training-Free Table Understanding with Large Language Models. InProceedings of the 34th ACM International Conference on Information and Knowledge Management. 3009–3019

    2024

  3. [1]

    Ralph G Andrzejak, Klaus Lehnertz, Florian Mormann, Christoph Rieke, Peter David, and Christian E Elger. 2001. Indications of nonlinear deterministic and finite-dimensional structures in time series of brain electrical activity: Dependence on recording region and brain state.Physical Review E64, 6 (2001), 061907

    2001

  4. [2]

    Davide Anguita, Alessandro Ghio, Luca Oneto, Xavier Parra, Jorge Luis Reyes-Ortiz, et al . 2013. A public domain dataset for human activity recognition using smartphones.. InEsann, Vol. 3. 3

    2013

  5. [3]

    Anthony Bagnall, Hoang Anh Dau, Jason Lines, Michael Flynn, James Large, Aaron Bostrom, Paul Southam, and Eamonn Keogh. 2018. The UEA multivariate time series classification archive, 2018.arXiv preprint arXiv:1811.00075 (2018)

    arXiv 2018

  6. [4]

    Yoshua Bengio, Nicholas Léonard, and Aaron Courville. 2013. Estimating or propagating gradients through stochastic neurons for conditional computation.arXiv preprint arXiv:1308.3432(2013)

    arXiv 2013

  7. [5]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners.Advances in neural information processing systems33 (2020), 1877–1901

    2020

  8. [6]

    Defu Cao, Furong Jia, Sercan O Arik, Tomas Pfister, Yixiang Zheng, Wen Ye, and Yan Liu. 2024. Tempo: Prompt-based generative pre-trained transformer for time series forecasting. InThe Twelfth International Conference on Learning Representations

    2024

Show all 55 references
  1. [7]

    Jiazhen Chen, Mingbin Feng, and Tony S Wirjanto. 2025. Prospective Multi-Graph Cohesion for Multivariate Time Series Anomaly Detection. InProceedings of the Eighteenth ACM International Conference on Web Search and Data Mining. 98–106

    2025

  2. [8]

    Xuanhao Chen, Liwei Deng, Yan Zhao, and Kai Zheng. 2023. Adversarial autoencoder for unsupervised time series anomaly detection and interpretation. InProceedings of the Sixteenth ACM International Conference on Web Search and Data Mining. 267–275

    2023

  3. [9]

    Mingyue Cheng, Qi Liu, Zhiding Liu, Zhi Li, Yucong Luo, and Enhong Chen. 2023. Formertime: Hierarchical multi-scale representations for multivariate time series classification. InProceedings of the ACM web conference 2023. 1437–1445

    2023

  4. [10]

    Mingyue Cheng, Qi Liu, Zhiding Liu, Hao Zhang, Rujiao Zhang, and Enhong Chen. 2023. TimeMAE: Self-Supervised Representations of Time Series with Decoupled Masked Autoencoders.arXiv preprint arXiv:2303.00320(2023)

    2023

  5. [11]

    Mingyue Cheng, Zhiding Liu, Xiaoyu Tao, Qi Liu, Jintao Zhang, Tingyue Pan, Shilong Zhang, Panjing He, Xiaohan Zhang, Daoyu Wang, et al. [n. d.]. A comprehensive survey of time series forecasting: Concepts, challenges, and future directions.Authorea Preprints([n. d.])

  6. [12]

    Mingyue Cheng, Xiaoyu Tao, Qi Liu, Hao Zhang, Yiheng Chen, and Defu Lian. 2025. Cross-domain pre-training with language models for transferable time series representations. InProceedings of the Eighteenth ACM International Conference on Web Search and Data Mining. 175–183

    2025

  7. [13]

    Mingyue Cheng, Jiqian Yang, Tingyue Pan, Qi Liu, Zhi Li, and Shijin Wang. 2025. Convtimenet: A deep hierarchical fully convolutional model for multivariate time series analysis. InCompanion Proceedings of the ACM on Web Conference J. ACM, Vol. 37, No. 4, Article 111. Publicati...

    2025

  8. [14]

    Maximilian Christ, Nils Braun, Julius Neuffer, and Andreas W Kempa-Liehr. 2018. Time series feature extraction on basis of scalable hypothesis tests (tsfresh–a python package).Neurocomputing307 (2018), 72–77

    2018

  9. [15]

    Angus Dempster, Daniel F Schmidt, and Geoffrey I Webb. 2021. Minirocket: A very fast (almost) deterministic transform for time series classification. InProceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining. 248–257

    2021

  10. [16]

    Houtao Deng, George Runger, Eugene Tuv, and Martyanov Vladimir. 2013. A time series forest for classification and feature extraction.Information Sciences239 (2013), 142–153

    2013

  11. [17]

    Hui Ding, Goce Trajcevski, Peter Scheuermann, Xiaoyue Wang, and Eamonn Keogh. 2008. Querying and mining of time series data: experimental comparison of representations and distance measures.Proceedings of the VLDB Endowment1, 2 (2008), 1542–1552

    2008

  12. [18]

    Emadeldeen Eldele, Mohamed Ragab, Zhenghua Chen, Min Wu, Chee Keong Kwoh, Xiaoli Li, and Cuntai Guan

  13. [19]

    Josif Grabocka, Nicolas Schilling, Martin Wistuba, and Lars Schmidt-Thieme. 2014. Learning time-series shapelets. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining. 392–401

    2014

  14. [20]

    Robert Gray. 1984. Vector quantization.IEEE Assp Magazine1, 2 (1984), 4–29

    1984

  15. [21]

    Ashish Gupta, Hari Prabhat Gupta, Bhaskar Biswas, and Tanima Dutta. 2020. Approaches and applications of early classification of time series: A review.IEEE Transactions on Artificial Intelligence1, 1 (2020), 47–61

    2020

  16. [22]

    Liangzhe Han, Bowen Du, Leilei Sun, Yanjie Fu, Yisheng Lv, and Hui Xiong. 2021. Dynamic and multi-faceted spatio- temporal deep learning for traffic speed forecasting. InProceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining. 547–555

    2021

  17. [23]

    G Hinton. 2014. Distilling the Knowledge in a Neural Network. InDeep Learning and Representation Learning Workshop in Conjunction with NIPS

    2014

  18. [24]

    Hassan Ismail Fawaz, Benjamin Lucas, Germain Forestier, Charlotte Pelletier, Daniel F Schmidt, Jonathan Weber, Geoffrey I Webb, Lhassane Idoumghar, Pierre-Alain Muller, and François Petitjean. 2020. Inceptiontime: Finding alexnet for time series classification.Data Mining and ...

    2020

  19. [25]

    Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. InProceedings of NAACL-HLT. 4171–4186

    2019

  20. [26]

    Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. InProceedings of the 58th...

    2020

  21. [27]

    Xiaocui Li, Guoliang Li, Xinyu Zhang, Yangtao Wang, Qingyu Shi, and Wei Liang. 2025. Incomplete Multi-view Clustering via Local Reasoning and Correlation Analysis. InProceedings of the Eighteenth ACM International Conference on Web Search and Data Mining. 484–492

    2025

  22. [28]

    Jessica Lin, Eamonn Keogh, Li Wei, and Stefano Lonardi. 2007. Experiencing SAX: a novel symbolic representation of time series.Data Mining and knowledge discovery15 (2007), 107–144

    2007

  23. [29]

    Jessica Lin, Rohan Khade, and Yuan Li. 2012. Rotation-invariant similarity in time series using bag-of-patterns representation.Journal of Intelligent Information Systems39 (2012), 287–315

    2012

  24. [30]

    Hanxiao Liu, Zihang Dai, David So, and Quoc V Le. 2021. Pay attention to mlps.Advances in neural information processing systems34 (2021), 9204–9215

    2021

  25. [31]

    Minghao Liu, Shengqi Ren, Siyuan Ma, Jiahui Jiao, Yizhou Chen, Zhiguang Wang, and Wei Song. 2021. Gated transformer networks for multivariate time series classification.arXiv preprint arXiv:2103.14438(2021)

    2021 arXiv

  26. [32]

    Zhiding Liu, Mingyue Cheng, Zhi Li, Zhenya Huang, Qi Liu, Yanhu Xie, and Enhong Chen. 2024. Adaptive normalization for non-stationary time series forecasting: A temporal slice perspective.Advances in Neural Information Processing Systems36 (2024)

    2024

  27. [33]

    Zhiding Liu, Jiqian Yang, Mingyue Cheng, Yucong Luo, and Zhi Li. 2024. Generative pretrained hierarchical transformer for time series forecasting. InProceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining. 2003–2013

    2024

  28. [34]

    Matthew Middlehurst, Patrick Schäfer, and Anthony Bagnall. 2024. Bake off redux: a review and experimental evaluation of recent time series classification algorithms.Data Mining and Knowledge Discovery38, 4 (2024), 1958–2031

    2024

  29. [35]

    Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient estimation of word representations in vector space. InInternational Conference on Learning Representations

    2013

  30. [36]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. J. ACM, Vol. 37, No. 4, Article 111. Publication dat...

    2022

  31. [37]

    François Petitjean, Germain Forestier, Geoffrey I Webb, Ann E Nicholson, Yanping Chen, and Eamonn Keogh. 2014. Dynamic time warping averaging of time series allows faster and more accurate classification. In2014 IEEE international conference on data mining. IEEE, 470–479

    2014

  32. [38]

    Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. 2018. Improving language understanding by generative pre-training. (2018)

    2018

  33. [39]

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners.OpenAI blog1, 8 (2019), 9

    2019

  34. [40]

    Steffen Rendle. 2010. Factorization machines. In2010 IEEE International conference on data mining. IEEE, 995–1000

    2010

  35. [41]

    Pavel Senin and Sergey Malinchik. 2013. Sax-vsm: Interpretable time series classification using sax and vector space model. In2013 IEEE 13th international conference on data mining. IEEE, 1175–1180

    2013

  36. [42]

    Ahmed Shifaz, Charlotte Pelletier, François Petitjean, and Geoffrey I Webb. 2020. TS-CHIEF: a scalable and accurate forest algorithm for time series classification.Data Mining and Knowledge Discovery34, 3 (2020), 742–775

    2020

  37. [44]

    Xiaoyu Tao, Shilong Zhang, Mingyue Cheng, Daoyu Wang, Tingyue Pan, Bokai Pan, Changqing Zhang, and Shijin Wang. 2025. From values to tokens: An llm-driven framework for context-aware time series forecasting via symbolic discretization.arXiv preprint arXiv:2508.09191(2025)

    2025 arXiv

  38. [45]

    Aäron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. 2016. WaveNet: A Generative Model for Raw Audio. InProc. SSW 2016. 125–125

    2016

  39. [46]

    Aaron Van Den Oord, Oriol Vinyals, et al. 2017. Neural discrete representation learning.Advances in neural information processing systems30 (2017)

    2017

  40. [48]

    Shiyu Wang. 2024. Neuralreconciler for hierarchical time series forecasting. InProceedings of the 17th ACM International Conference on Web Search and Data Mining. 731–739

    2024

  41. [49]

    Hyowon Wi, Yehjin Shin, and Noseong Park. 2024. Continuous-time autoencoders for regular and irregular time series imputation. InProceedings of the 17th ACM International Conference on Web Search and Data Mining. 826–835

    2024

  42. [50]

    Lexiang Ye and Eamonn Keogh. 2009. Time series shapelets: a new primitive for data mining. InProceedings of the 15th ACM SIGKDD international conference on Knowledge discovery and data mining. 947–956

    2009

  43. [51]

    George Zerveas, Srideepika Jayaraman, Dhaval Patel, Anuradha Bhamidipaty, and Carsten Eickhoff. 2021. A transformer- based framework for multivariate time series representation learning. InProceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining. 2114–2124

    2021

  44. [52]

    Xuchao Zhang, Yifeng Gao, Jessica Lin, and Chang-Tien Lu. 2020. Tapnet: Multivariate time series classification with attentional prototypical network. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 34. 6845–6852

    2020

  45. [53]

    Yi Zheng, Qi Liu, Enhong Chen, Yong Ge, and J Leon Zhao. 2014. Time series classification using multi-channels deep convolutional neural networks. InInternational conference on web-age information management. Springer, 298–310

    2014

  46. [54]

    Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. 2021. Informer: Beyond efficient transformer for long sequence time-series forecasting. InProceedings of the AAAI conference on artificial intelligence, Vol. 35. 11106–11115. Receive...

    2021

  47. [2021]

    InProceedings of the Thirtieth International Joint Conference on Artificial Intelligence

    Time-Series Representation Learning via Temporal and Contextual Contrasting. InProceedings of the Thirtieth International Joint Conference on Artificial Intelligence. International Joint Conferences on Artificial Intelligence Organization, 2352–2359

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.