Pith. sign in

REVIEW 3 major objections 5 minor 56 references

Adding evidential uncertainty to domain adaptation improves time-series accuracy and calibration

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A UDA framework with multi-scale input mixing and Dirichlet-prior uncertainty estimation improves F1 and calibration on five time-series benchmarks.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Useful engineering contribution pairing evidential uncertainty with multi-scale mixing for time-series UDA, but the headline SOTA claim rests on target-label-based hyperparameter selection and a blanket claim its own Table I contradicts. the 3 major comments →

arxiv 2508.18630 v1 pith:5G42ID6F submitted 2025-08-26 cs.LG cs.CV

Uncertainty Awareness on Unsupervised Domain Adaptation for Time Series Data

Weide Liu , Xiaoyang Zhong , Lu Wang , Jingwen Hou , Yuemei Luo , Jiebin Yan , Yuming Fang This is my paper
classification cs.LG cs.CV
keywords unsupervised domain adaptationtime series classificationevidential learningDirichlet prioruncertainty estimationhuman activity recognitionmulti-scale architecturecalibration
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 argues that unsupervised domain adaptation for time series classification should explicitly quantify prediction uncertainty, not just align source and target feature distributions. It proposes adding an evidential-learning head—a Dirichlet prior over class probabilities—to standard UDA losses, together with a multi-scale mixing architecture that downsamples each input into several temporal resolutions. The central empirical claim is that the uncertainty term alone improves F1 scores across all twelve baseline UDA methods on five benchmark datasets (UCIHAR, WISDM, HHAR, SSC, MFD), and that adding the multi-scale mixer pushes performance to state-of-the-art. The paper also reports lower Expected Calibration Error, meaning prediction confidence is better calibrated. A reader should care because the uncertainty estimates are closed-form and parameter-free, and they make adapted models more trustworthy in sensor-based healthcare and fault-diagnosis applications.

Core claim

The paper's claim is that in UDA for time series, aligning features is not enough: the model must also know when it is uncertain. The authors model classifier outputs as evidence parameters α of a Dirichlet distribution over class probabilities, giving an uncertainty u = K/S where S is total evidence, and train with an evidential Bayesian-risk loss plus KL regularization. They add this evidential term to the source classification and domain-alignment losses. They find that every one of the twelve baselines, when augmented with the Dirichlet-based uncertainty term, beats its counterpart without uncertainty, and the improvement is largest where uncertainty matters most—the noAdapt setting gain

What carries the argument

Two mechanisms carry the argument. (1) Evidential uncertainty: instead of a softmax point estimate, the classifier outputs α = softplus(network output)+1, parameters of a Dirichlet prior over class probabilities; the evidential loss (Bayesian risk with cross-entropy, plus a KL term that anneals with the epoch) estimates these α, and uncertainty is the closed-form u = K/S. (2) Multi-scale mixing architecture: the input series is downsampled M times into coarse views x0...xM, each encoded by a CNN with auxiliary classification heads, and the concatenated features feed the final classifier; the paper tests five downsampling strategies, with learnable-1D-CNN and max/avg pooling variants performi

Load-bearing premise

The load-bearing premise is that a labeled subset of target-domain samples is available to select the uncertainty-loss weight λ3; if target labels are truly unavailable, the reported improvements and the chosen λ3 may not hold.

What would settle it

Run the same evidential-UDA pipeline with λ3 fixed to a constant (or selected by cross-validation on the source domain only) and measure F1 and ECE on the five datasets; if the gains over baselines disappear or reverse, the method's dependence on target labels for validation is confirmed. A complementary check: on a dataset where the uncertainty–F1 correlation is weak, the method should lose its advantage.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Uncertainty estimation can be bolted onto existing UDA pipelines without adding parameters; the paper shows consistent F1 gains across 12 different alignment losses.
  • The uncertainty score is predictive of error: across five datasets it is roughly inversely linear with F1, so it can flag unreliable target-domain predictions.
  • Better calibration (lower ECE) follows from the same mechanism, which matters for safety-related deployment of activity-recognition and fault-diagnosis models.
  • Multi-scale input mixing reduces measurable domain discrepancy (MMD and Wasserstein distance) as well as improving accuracy.
  • The best pooling strategy is dataset-dependent, but the averaged improvements are consistent across all folds and datasets.

Where Pith is reading between the lines

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

  • The paper tunes λ3 using a labeled subset of target samples for validation (Target Risk); a strictly label-free protocol might need a fixed or source-only heuristic, so the reported gains describe a weakly supervised setting rather than pure UDA.
  • The same evidential term could serve as an active-learning or selective-prediction signal: samples with high uncertainty (u ≈ K/S) are exactly the ones whose adaptation is poor, suggesting where to collect labels.
  • Because the uncertainty estimate is closed-form and parameter-free, the mechanism likely transfers beyond HAR to any categorical time-series classifier; the SSC and MFD results already demonstrate this, but the paper does not explore combining uncertainty with test-time adaptation.
  • The uncertainty–F1 correlation suggests a direct extension: using uncertainty as a loss weight for target samples during adaptation, which the current method does not do.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes an unsupervised domain adaptation (UDA) framework for time-series classification that combines a multi-scale mixing architecture with an evidential-learning-based uncertainty penalty. The method is evaluated on five time-series datasets (UCIHAR, WISDM, HHAR, SSC, MFD) with twelve UDA baselines, reporting F1 improvements, lower Expected Calibration Error, and qualitative analyses of uncertainty, features, and domain discrepancy. The central claim is that the combined multi-scale and uncertainty-aware approach achieves state-of-the-art performance.

Significance. If the empirical claims were established under a fully unsupervised protocol, the paper would make a useful contribution by showing that evidential uncertainty can be plugged into a range of UDA losses and improve accuracy and calibration. The evaluation is broad (5 datasets, 12 baselines, multiple down-sampling variants) and the code is released, which are concrete strengths. However, the central empirical claim currently rests on a hyperparameter-selection procedure that uses labeled target data, and one of the paper's generalizing statements about uncertainty is falsified by the paper's own table. The uncertainty mechanism itself is a direct application of Sensoy et al. (2017), so the novelty lies in the combination and in the multi-scale architecture; this is reasonable but requires the experiments to be clean.

major comments (3)
  1. [III-D, Eq. (10)] The final loss uses λ3 as a weight on the evidential loss, and the paper states that 'optimal λ3 is found empirically by the Target Risk (TGT) [44] by leaving out a subset of target domain samples and their labels as a validation set.' This means labeled target instances are used for model selection, which is inconsistent with the 'unsupervised' setting and can inflate reported gains if λ3 is tuned to the evaluation distribution. The paper reports no λ3 values and no sensitivity analysis. Please re-run the experiments with source-only validation (or with a fixed λ3 chosen a priori) and report the resulting F1 scores; also compare to the current target-validation results.
  2. [IV-B.1, Table I] The claim that 'All baseline models when incorporated with uncertainty outperform their counterparts' is contradicted by Table I. For example, on SSC, DDC drops from 60.88 to 60.02 with +DIRce, HoMM drops from 60.81 to 60.05, DANN drops from 60.80 to 59.39; on WISDM, DDC drops from 53.78 to 53.57 with +NIG. This undermines the generality claim for the uncertainty component and must be revised, with an analysis of failure cases.
  3. [III-C, Eq. (8)] The evidential loss Levi is written as a sum over Ns (source) samples, but the section title is 'Target Modeling with Uncertainty' and Fig. 1 appears to apply the evidential loss to both source and target branches. It is unclear whether the loss is computed on target samples, and if so, what labels or pseudo-labels are used. This ambiguity is load-bearing because the paper's mechanism for improving target-domain predictions through uncertainty is otherwise underspecified. Please state explicitly which samples contribute to Levi and how target uncertainty (Eq. (9)) is used to affect training.
minor comments (5)
  1. [III-A] The target domain is denoted X^s instead of X^t in the task definition; this is likely a typo.
  2. [V] In the Conclusion, 'better generation ability' should be 'better generalization ability'.
  3. [Table V] The notation table lists 'Llevi' while the text uses 'Levi'; please make the notation consistent.
  4. [II-C] Equations (4)-(6) are taken directly from [17]; please make the notation self-contained (e.g., define y_ik) and explicitly note whether any modification is made for the UDA setting.
  5. [Tables I and II] The tables report only F1 averages; please add standard deviations or significance tests over the five folds, and state how many random seeds are used.

Circularity Check

0 steps flagged

No equation-level circularity; target-label hyperparameter selection weakens the unsupervised protocol but does not make the derivation self-referential.

full rationale

The paper's derivation chain is not circular in the sense defined by the rubric. The evidential learning component (Eqs. 2-9) is taken from Sensoy et al. [17] and Amini et al. [18] as an external building block; the uncertainty measure u=K/S and the evidential losses are not re-derived from the paper's own results. The multi-scale mixing architecture is defined operationally (downsampling, auxiliary headers) and evaluated on external benchmarks, and the final loss Eq. (10) is a weighted combination, not an identity. The only substantive concern is in Section III-D: 'optimal λ3 is found empirically by the Target Risk (TGT) [44] by leaving out a subset of target domain samples and their labels as a validation set.' This means the strength of the uncertainty term is selected using target labels, so the reported '+UN' gains are not established under a strictly unsupervised protocol; however, this is a validation/protocol weakness and a possible source of overfitting, not a case where a prediction equals its input by construction. The claim in Section IV-B.1 that 'All baseline models when incorporated with uncertainty outperform their counterparts' is also not supported by Table I (e.g., SSC DDC 60.88→60.02), but this is an internal empirical inconsistency, not circularity. The self-citation [20] is used only for provenance of the workshop version, not as a load-bearing justification. Therefore no circular step can be identified with the required specificity, and the honest finding is no significant circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The method is a composition of existing components (evidential learning, multi-scale pooling) with no new entities. The main 'free parameters' are loss weights and architectural choices, several of which are selected using target labels or after seeing results.

free parameters (5)
  • λ3 (evidential loss weight) = Not reported; selected via Target Risk on a labeled target validation set
    Controls the balance between UDA loss and evidential loss. Tuning on target labels is a leak in the unsupervised setting.
  • λ_i (multi-scale auxiliary loss weights) = {0.5, 0.25, 0.25}
    Hand-set weights for auxiliary classification losses at the three scales. Chosen without tuning or stated justification.
  • λ1, λ2 (UDA loss weights) = Inherited from AdaTime
    Set to the optimal values given in [44]; not re-tuned for this method.
  • Evidential loss type (ml/ce/mse) = ce selected as best
    Three Bayesian risk losses were tried; cross-entropy was chosen based on average F1. This is a post-hoc model selection.
  • Downsampling variant (ML/MM/MR/MA) = Varies by dataset; no single variant fixed
    Five pooling strategies are implemented and all results are shown. The best variant for each dataset could be selected after seeing results, which is a form of selection bias.
axioms (4)
  • domain assumption Ps(y|x) = Pt(y|x) (conditional label distribution identical across domains)
    Stated in Section III-A as a premise for UDA. If the conditional relationship differs (e.g., sensor placement changes the mapping), feature alignment may not transfer.
  • standard math Dirichlet prior is conjugate to the categorical distribution, yielding a closed-form posterior
    Used in Section III-C to derive Equations (2)-(6) from evidential learning [17].
  • domain assumption The KL regularization term (Equation 7) improves uncertainty calibration
    Added without new empirical justification in this paper; relies on the behavior reported in [17].
  • ad hoc to paper Target labels may be used for hyperparameter selection via Target Risk
    Section III-D uses a labeled target validation set to pick λ3. This is not a standard unsupervised assumption and can inflate performance.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Uncertainty Awareness on Unsupervised Domain Adaptation for Time Series Data." pith.science (2026). https://pith.science/paper/5G42ID6F

@misc{pith2026250818630,
  author       = {Pith},
  title        = {Pith review of: Uncertainty Awareness on Unsupervised Domain Adaptation for Time Series Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5G42ID6F}},
  note         = {Machine review of arXiv:2508.18630}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Unsupervised domain adaptation methods seek to generalize effectively on unlabeled test data, especially when encountering the common challenge in time series data that distribution shifts occur between training and testing datasets. In this paper, we propose incorporating multi-scale feature extraction and uncertainty estimation to improve the model's generalization and robustness across domains. Our approach begins with a multi-scale mixed input architecture that captures features at different scales, increasing training diversity and reducing feature discrepancies between the training and testing domains. Based on the mixed input architecture, we further introduce an uncertainty awareness mechanism based on evidential learning by imposing a Dirichlet prior on the labels to facilitate both target prediction and uncertainty estimation. The uncertainty awareness mechanism enhances domain adaptation by aligning features with the same labels across different domains, which leads to significant performance improvements in the target domain. Additionally, our uncertainty-aware model demonstrates a much lower Expected Calibration Error (ECE), indicating better-calibrated prediction confidence. Our experimental results show that this combined approach of mixed input architecture with the uncertainty awareness mechanism achieves state-of-the-art performance across multiple benchmark datasets, underscoring its effectiveness in unsupervised domain adaptation for time series data.

Figures

Figures reproduced from arXiv: 2508.18630 by Jiebin Yan, Jingwen Hou, Lu Wang, Weide Liu, Xiaoyang Zhong, Yuemei Luo, Yuming Fang.

Figure 1
Figure 1. Figure 1: The architecture of our method. Where Ld is domain loss, Lcls is classifier loss, and Levi is our evidential loss. f s and f s i are features of the images for the source domain and source domain at scale i. “Aux Headers” refers to the auxiliary classification heads for different scales and “Final Header” is the final classification head using all concatenated features of different scales. tions and long-t… view at source ↗
Figure 2
Figure 2. Figure 2: Different down-sampling method in our multi-scale mixing architecture, where (a) denotes down-sampling with the 1D [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The confusion matrix and uncertainty. Using the noAdapt method as the baseline, test on 5 different datasets. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: This figure illustrates the label distribution of source data across various dataset splits, along with the gap (indicated [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The PCA visualization of samples of different domains across 10 splits (source and target samples are shown in yellow [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The correlation between F1 score and uncertainty. The baseline is the basic model without domain adaption, randomly [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: The Expected Calibration Error (ECE) metric. The DDC method served as a baseline for experiments on five datasets. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Density of uncertainty. The baseline is the basic model without domain adaption, and add evidential to it to estimate [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Performance evaluation across different folds over 5 different datasets. The baseline is the basic model without domain [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: The figure illustrates the t-SNE visualization results of features extracted by different methods and across various [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: The principal component analysis (PCA) visualization results of each sample and the uncertainty of some categories. [PITH_FULL_IMAGE:figures/full_fig_p011_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: The correlation between F1 score and Monte Carlo uncertainty. The baseline is the basic model without domain [PITH_FULL_IMAGE:figures/full_fig_p014_12.png] view at source ↗

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

56 extracted references · 53 canonical work pages · 3 internal anchors

  1. [1]

    Deep learning for time series classification: a review,

    H. Ismail Fawaz, G. Forestier, J. Weber, L. Idoumghar, and P.-A. Muller, “Deep learning for time series classification: a review,” Data mining and knowledge discovery, vol. 33, no. 4, pp. 917–963, 2019

    work page 2019

  2. [2]

    A systematic study of unsupervised domain adaptation for robust human- activity recognition,

    Y . Chang, A. Mathur, A. Isopoussu, J. Song, and F. Kawsar, “A systematic study of unsupervised domain adaptation for robust human- activity recognition,” Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, vol. 4, no. 1, pp. 1–30, 2020

    work page 2020

  3. [3]

    Collecting com- plex activity datasets in highly rich networked sensor environments,

    D. Roggen, A. Calatroni, M. Rossi, T. Holleczek, K. F ¨orster, G. Tr¨oster, P. Lukowicz, D. Bannach, G. Pirkl, A. Ferscha et al., “Collecting com- plex activity datasets in highly rich networked sensor environments,” in Seventh international conference on networked sensing systems (INSS) . IEEE, 2010, pp. 233–240

    work page 2010

  4. [4]

    Adast: Attentive cross-domain eeg-based sleep staging framework with iterative self-training,

    E. Eldele, M. Ragab, Z. Chen, M. Wu, C.-K. Kwoh, X. Li, and C. Guan, “Adast: Attentive cross-domain eeg-based sleep staging framework with iterative self-training,” IEEE Transactions on Emerging Topics in Com- putational Intelligence, vol. 7, no. 1, pp. 210–221, 2022

    work page 2022

  5. [5]

    Physiobank, physiotoolkit, and physionet components of a new research resource for complex physiologic signals,

    A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff, P. C. Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C.-K. Peng, and H. E. Stanley, “Physiobank, physiotoolkit, and physionet components of a new research resource for complex physiologic signals,” Circulation, vol. 101, no. 23, pp. 215–220, 2000

    work page 2000

  6. [6]

    A decade survey of transfer learning (2010–2020),

    S. Niu, Y . Liu, J. Wang, and H. Song, “A decade survey of transfer learning (2010–2020),” IEEE Transactions on Artificial Intelligence , vol. 1, no. 2, pp. 151–166, 2020

    work page 2010

  7. [7]

    Self-supervised pose adaptation for cross- domain image animation,

    C. Wang, C. Xu, and D. Tao, “Self-supervised pose adaptation for cross- domain image animation,” IEEE Transactions on Artificial Intelligence , vol. 1, no. 1, pp. 34–46, 2020

    work page 2020

  8. [8]

    Multiple graphs and low-rank embedding for multi-source heterogeneous domain adaptation,

    H. Wu and M. K. Ng, “Multiple graphs and low-rank embedding for multi-source heterogeneous domain adaptation,” ACM Transactions on Knowledge Discovery from Data (TKDD), vol. 16, no. 4, pp. 1–25, 2022

    work page 2022

  9. [9]

    Deep coral: Correlation alignment for deep domain adaptation,

    B. Sun and K. Saenko, “Deep coral: Correlation alignment for deep domain adaptation,” in ECCV. Springer, 2016, pp. 443–450. 14 0.70 0.75 0.80 0.85 0.90 0.95 0.50 0.55 0.60 0.65 0.70 F1 Uncertainty UCIHAR 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.38 0.48 0.58 0.68 0.78 F1 Uncertainty WISDM 0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.50 0.60 0.70 0.80 0.90 F1 Unce...

    work page 2016

  10. [10]

    Homm: Higher-order moment matching for unsupervised domain adap- tation,

    C. Chen, Z. Fu, Z. Chen, S. Jin, Z. Cheng, X. Jin, and X.-S. Hua, “Homm: Higher-order moment matching for unsupervised domain adap- tation,” AAAI, 2020

    work page 2020

  11. [11]

    Deep domain confusion: Maximizing for domain invariance,

    E. Tzeng, J. Hoffman, N. Zhang, K. Saenko, and T. Darrell, “Deep domain confusion: Maximizing for domain invariance,” arXiv preprint arXiv:1412.3474, 2014

    Pith/arXiv arXiv 2014

  12. [12]

    Deep subdomain adaptation network for image classification,

    Y . Zhu, F. Zhuang, J. Wang, G. Ke, J. Chen, J. Bian, H. Xiong, and Q. He, “Deep subdomain adaptation network for image classification,” IEEE Transactions on Neural Networks and Learning Systems , vol. 32, no. 4, pp. 1713–1722, 2021

    work page 2021

  13. [13]

    Domain-adversarial training of neural networks,

    Y . Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Lavi- olette, M. March, and V . Lempitsky, “Domain-adversarial training of neural networks,” 2016

    work page 2016

  14. [14]

    Conditional adversarial domain adaptation,

    M. Long, Z. Cao, J. Wang, and M. I. Jordan, “Conditional adversarial domain adaptation,” in NeurIPS, 2018

    work page 2018

  15. [15]

    Self- supervised autoregressive domain adaptation for time series data,

    M. Ragab, E. Eldele, Z. Chen, M. Wu, C.-K. Kwoh, and X. Li, “Self- supervised autoregressive domain adaptation for time series data,” IEEE Transactions on Neural Networks and Learning Systems, pp. 1–11, 2022

    work page 2022

  16. [16]

    Adversarial spectral kernel matching for unsuper- vised time series domain adaptation,

    Q. Liu and H. Xue, “Adversarial spectral kernel matching for unsuper- vised time series domain adaptation,” in Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21 , Z.- H. Zhou, Ed. International Joint Conferences on Artificial Intelligence Organization, 8 2021, pp. 2744–2750

    work page 2021

  17. [17]

    Evidential deep learning to quantify classification uncertainty,

    M. Sensoy, L. M. Kaplan, and M. Kandemir, “Evidential deep learning to quantify classification uncertainty,” in NeurIPS, 2018, pp. 3179–3189

    work page 2018

  18. [18]

    Deep evidential regression,

    A. Amini, W. Schwarting, A. Soleimany, and D. Rus, “Deep evidential regression,” NeurIPS, vol. 33, pp. 14 927–14 937, 2020

    work page 2020

  19. [19]

    ADATIME: A Benchmarking Suite for Domain Adaptation on Time Series Data

    M. Ragab, E. Eldele, W. L. Tan, C.-S. Foo, Z. Chen, M. Wu, C. K. Kwoh, and X. Li, “Adatime: A benchmarking suite for domain adaptation on time series data,” arXiv preprint arXiv:2203.08321 , 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  20. [20]

    Uncertainty awareness for unsupervised domain adaptation on human activity recognition,

    W. Liu, X. Zhong, L. Wang, J. Hou, Y . Luo, J. Yan, and Y . Fang, “Uncertainty awareness for unsupervised domain adaptation on human activity recognition,” in International Joint Conference on Artificial Intelligence. Springer, 2024, pp. 85–98

    work page 2024

  21. [21]

    Condition monitoring of bearing damage in electromechanical drive systems by using motor current signals of electric motors: A benchmark data set for data-driven classification,

    C. Lessmeier, J. K. Kimotho, D. Zimmer, and W. Sextro, “Condition monitoring of bearing damage in electromechanical drive systems by using motor current signals of electric motors: A benchmark data set for data-driven classification,” in PHM Society European Conference , vol. 3, 2016, pp. 05–08

    work page 2016

  22. [22]

    Return of frustratingly easy domain adaptation,

    B. Sun, J. Feng, and K. Saenko, “Return of frustratingly easy domain adaptation,” in AAAI, 2016

    work page 2016

  23. [23]

    Domain adaptation via transfer component analysis,

    S. J. Pan, I. W. Tsang, J. T. Kwok, and Q. Yang, “Domain adaptation via transfer component analysis,” IEEE transactions on neural networks, vol. 22, no. 2, pp. 199–210, 2010

    work page 2010

  24. [24]

    Multi-source deep do- main adaptation with weak supervision for time-series sensor data,

    G. Wilson, J. R. Doppa, and D. J. Cook, “Multi-source deep do- main adaptation with weak supervision for time-series sensor data,” in SIGKDD, 2020

    work page 2020

  25. [25]

    Fedcov: Enhanced trustworthy federated learning for machine rul prediction with continuous-to-discrete conversion,

    C. Cai, Y . Fang, W. Liu, R. Jin, J. Cheng, and Z. Chen, “Fedcov: Enhanced trustworthy federated learning for machine rul prediction with continuous-to-discrete conversion,” IEEE Transactions on Industrial Informatics, 2024

    work page 2024

  26. [26]

    Bayesian uncertainty calibration for federated time series analysis,

    C. Cai, W. Liu, X. Xia, Z. Chen, and Y . Fang, “Bayesian uncertainty calibration for federated time series analysis,” IEEE Transactions on Multimedia, 2024

    work page 2024

  27. [27]

    Harmonizing base and novel classes: A class-contrastive approach for generalized few-shot segmentation,

    W. Liu, Z. Wu, Y . Zhao, Y . Fang, C.-S. Foo, J. Cheng, and G. Lin, “Harmonizing base and novel classes: A class-contrastive approach for generalized few-shot segmentation,” International Journal of Computer Vision, vol. 132, no. 4, pp. 1277–1291, 2024

    2024

  28. [28]

    Cau: A causality attention unit for spatial-temporal sequence forecast,

    B. Qin, F. Meng, X. Fang, G. Dai, S. Yuan, and B. Mu, “Cau: A causality attention unit for spatial-temporal sequence forecast,” IEEE Transactions on Multimedia, 2023

    work page 2023

  29. [29]

    Crnet: Cross-reference networks for few-shot segmentation,

    W. Liu, C. Zhang, G. Lin, and F. Liu, “Crnet: Cross-reference networks for few-shot segmentation,” in CVPR, 2020, pp. 4165–4173

    work page 2020

  30. [30]

    Variational recurrent adversarial deep domain adaptation,

    S. Purushotham, W. Carvalho, T. Nilanon, and Y . Liu, “Variational recurrent adversarial deep domain adaptation,” in ICLR, 2017

    work page 2017

  31. [31]

    Multi-source deep domain adaptation with weak supervision for time-series sensor data,

    G. Wilson, J. R. Doppa, and D. J. Cook, “Multi-source deep domain adaptation with weak supervision for time-series sensor data,” Proceed- ings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , 2020

    work page 2020

  32. [32]

    Time-Series Domain Adaptation via Sparse Associative Structure Alignment: Learning Invariance and Variance

    R. Cai, J. Chen, Z. Li, W. Chen, K. Zhang, J. Ye, Z. Li, X. Yang, and Z. Zhang, “Time series domain adaptation via sparse associative structure alignment,” ArXiv, vol. abs/2205.03554, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  33. [33]

    Domain adaptation for time-series classification to mitigate covariate shift,

    F. Ott, D. R ¨ugamer, L. Heublein, B. Bischl, and C. Mutschler, “Domain adaptation for time-series classification to mitigate covariate shift,” Proceedings of the 30th ACM International Conference on Multimedia , 2022

    work page 2022

  34. [34]

    Domain adaptation for time series forecasting via attention sharing,

    X. Jin, Y . Park, D. Maddix, H. Wang, and Y . Wang, “Domain adaptation for time series forecasting via attention sharing,” in Proceedings of the 39th International Conference on Machine Learning , ser. Proceedings of Machine Learning Research, K. Chaudhuri, S. Jegelka, L. Song, C. Szepesvari, G. Niu, and S. Sabato, Eds., vol. 162. PMLR, 17–23 Jul 2022, pp...

    work page 2022

  35. [36]

    Calda: Improving multi-source time series domain adaptation with contrastive adversarial learning,

    G. Wilson, J. R. Doppa, and D. J. Cook, “Calda: Improving multi-source time series domain adaptation with contrastive adversarial learning,” 2021

    work page 2021

  36. [37]

    Toward practical factory activity recognition: unsupervised understanding of repetitive as- sembly work in a factory,

    T. Maekawa, D. Nakai, K. Ohara, and Y . Namioka, “Toward practical factory activity recognition: unsupervised understanding of repetitive as- sembly work in a factory,” inProceedings of the 2016 ACM International Joint Conference on Pervasive and Ubiquitous Computing , 2016, pp. 1088–1099

    work page 2016

  37. [38]

    G. Yue, S. Zhang, T. Zhou, B. Jiang, W. Liu, and T. Wang, “Pyramid 15 TABLE VI: Class distribution of different training and test sets in the UCIHAR dataset. The correspondence between class indices and class names is as follows: 0-‘walk’, 1-‘upstairs’, 2-‘downstairs’, 3-‘sit’, 4-‘stand’, 5-‘lie’. UCIHAR TrainClass IndicesSumTest Class IndicesSum0 1 2 3 4...

    work page 2024

  38. [39]

    Modality- aware feature matching: A comprehensive review of single-and cross- modality techniques,

    W. Liu, W. Zhou, J. Liu, P. Hu, J. Cheng, J. Han, and W. Lin, “Modality- aware feature matching: A comprehensive review of single-and cross- modality techniques,” arXiv preprint arXiv:2507.22791 , 2025

    Pith/arXiv arXiv 2025

  39. [40]

    Simple and scalable predictive uncertainty estimation using deep ensembles,

    B. Lakshminarayanan, A. Pritzel, and C. Blundell, “Simple and scalable predictive uncertainty estimation using deep ensembles,” in NeurIPS, 2017, pp. 6402–6413

    work page 2017

  40. [41]

    Depth uncertainty in neural networks,

    J. Antor ´an, J. U. Allingham, and J. M. Hern ´andez-Lobato, “Depth uncertainty in neural networks,” in NeurIPS, 2020, pp. 10 620–10 634

    work page 2020

  41. [42]

    Uncertainty estimation using a single deep deterministic neural network,

    J. van Amersfoort, L. Smith, Y . W. Teh, and Y . Gal, “Uncertainty estimation using a single deep deterministic neural network,” in ICML, 2020, pp. 9690–9700

    work page 2020

  42. [43]

    Induction of multiscale temporal structure,

    M. C. Mozer, “Induction of multiscale temporal structure,” Advances in neural information processing systems , vol. 4, 1991

    work page 1991

  43. [44]

    Adatime: A benchmarking suite for domain adaptation on time series data,

    M. Ragab, E. Eldele, W. L. Tan, C.-S. Foo, Z. Chen, M. Wu, C.- K. Kwoh, and X. Li, “Adatime: A benchmarking suite for domain adaptation on time series data,” ACM Transactions on Knowledge Discovery from Data , vol. 17, no. 8, pp. 1–18, 2023

    work page 2023

  44. [45]

    A public domain dataset for human activity recognition using smartphones,

    D. Anguita, A. Ghio, L. Oneto, X. Parra, and J. L. Reyes-Ortiz, “A public domain dataset for human activity recognition using smartphones,” in European Symposium on Artificial Neural Networks, 2013, pp. 437–442

    work page 2013

  45. [46]

    Activity recognition using cell phone accelerometers,

    J. R. Kwapisz, G. M. Weiss, and S. A. Moore, “Activity recognition using cell phone accelerometers,” Sigkdd Explorations , vol. 12, no. 2, pp. 74–82, 2011

    work page 2011

  46. [47]

    Smart devices are different: Assessing and mitigatingmobile sensing heterogeneities for activity recognition,

    A. Stisen, H. Blunck, S. Bhattacharya, T. S. Prentow, M. B. Kjærgaard, A. Dey, T. Sonne, and M. M. Jensen, “Smart devices are different: Assessing and mitigatingmobile sensing heterogeneities for activity recognition,” in Proceedings of the 13th ACM Conference on Embedded Networked Sensor Systems , 2015, pp. 127–140

    work page 2015

  47. [48]

    Physiobank, physiotoolkit, and physionet: components of a new re- search resource for complex physiologic signals,

    A. L. Goldberger, L. Amaral, L. Glass, J. M. Hausdorff, R. G. Ivanov, P.and Mark, J. E. Mietus, G. B. Moody, C. Peng, and H. E. Stanley, “Physiobank, physiotoolkit, and physionet: components of a new re- search resource for complex physiologic signals,” Circulation, vol. 101, no. 23, pp. e215–e220, 2000

    work page 2000

  48. [49]

    An attention-based deep learning approach for sleep stage classification with single-channel eeg,

    E. Eldele, Z. Chen, C. Liu, M. Wu, C.-K. Kwoh, X. Li, and C. Guan, “An attention-based deep learning approach for sleep stage classification with single-channel eeg,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 29, pp. 809–818, 2021

    work page 2021

  49. [50]

    Condition monitoring of bearing damage in electromechanical drive systems by using motor current signals of electric motors: A benchmark data set 16 for data-driven classification,

    C. Lessmeier, J. K. Kimotho, D. Zimmer, and W. Sextro, “Condition monitoring of bearing damage in electromechanical drive systems by using motor current signals of electric motors: A benchmark data set 16 for data-driven classification,” in PHM Society European Conference , vol. 3, no. 1, 2016

    work page 2016

  50. [51]

    Correlation alignment for unsuper- vised domain adaptation,

    B. Sun, J. Feng, and K. Saenko, “Correlation alignment for unsuper- vised domain adaptation,” in Domain Adaptation in Computer Vision Applications. Springer, 2017, pp. 153–171

    work page 2017

  51. [52]

    Domain-adversarial training of neural networks,

    Y . Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Lavio- lette, M. Marchand, and V . Lempitsky, “Domain-adversarial training of neural networks,” JMLR, vol. 17, no. 1, pp. 1–35, 2016

    work page 2016

  52. [53]

    M. M. Rahman, C. Fookes, M. Baktashmotlagh, and S. Sridharan, On Minimum Discrepancy Estimation for Deep Domain Adaptation. Cham: Springer International Publishing, 2020, pp. 81–94

    work page 2020

  53. [54]

    A dirt-t approach to unsu- pervised domain adaptation,

    R. Shu, H. Bui, H. Narui, and S. Ermon, “A dirt-t approach to unsu- pervised domain adaptation,” in International Conference on Learning Representations, 2018

    work page 2018

  54. [55]

    Adversarial spectral kernel matching for unsuper- vised time series domain adaptation,

    Q. Liu and H. Xue, “Adversarial spectral kernel matching for unsuper- vised time series domain adaptation,” in Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21 , 08 2021, pp. 2744–2750

    work page 2021

  55. [56]

    Time series domain adaptation via sparse associative struc- ture alignment,

    R. Cai, J. Chen, Z. Li, W. Chen, K. Zhang, J. Ye, Z. Li, X. Yang, and Z. Zhang, “Time series domain adaptation via sparse associative struc- ture alignment,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 35, 2021, pp. 6859–6867

    work page 2021

  56. [57]

    Contrastive Learning for Unsupervised Domain Adaptation of Time Series

    Y . Ozyurt, S. Feuerriegel, and C. Zhang, “Contrastive learning for unsupervised domain adaptation of time series,” arXiv preprint arXiv:2206.06243, 2022. 17 TABLE VII: Class distribution of different training and test sets in the WISDM dataset. The correspondence between class indices and class names is as follows: 0-‘walk’, 1-‘jog’, 2-‘sit’, 3-‘stand’, ...

    work page internal anchor Pith review Pith/arXiv arXiv 2022

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.