acceptodds
Under review as a conference paper at ICLR 2027

Bridging Resolutions for Domain-Invariant Representation Learning in Time Series

Abstract

Unsupervised time-series domain adaptation (TSDA) aims to mitigate distribution shifts between labeled source domains and unlabeled target domains. Existing methods typically rely on two extreme views of time series, namely fine-grained temporal representations and coarse global frequency representations, while overlooking crucial intermediate-resolution structures for progressive domain-invariant learning. To address this limitation, we investigate the previously unexplored intermediate-resolution hierarchy induced by the Stationary Wavelet Transform and reveal two key properties: different resolutions exhibit distinct transferability and discriminability characteristics under domain shift, while different resolutions exhibit distinct category-level decision strengths. Based on these properties, we propose **PRISM**, a TSDA framework that bridges progressive representations across resolutions for domain-invariant learning. PRISM introduces Cross-Resolution Knowledge Transfer to bridge multi-resolution representations by progressively exchanging complementary knowledge across resolutions, and Committee-based Consensus Training to bridge the cross-resolution decision hierarchy, enabling more consistent domain-invariant decisions. Extensive experiments on six datasets demonstrate that PRISM consistently outperforms existing methods, achieving an average accuracy improvement of 4.21% in cross-domain settings. Code is available at the anonymous link: https://anonymous.4open.science/r/PRISM-E7BC/.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series, Domain Adaptation, Muti-Resolution, Wavelet Transform

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_dP9eOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the issue that existing unsupervised time-series domain adaptation methods overlook intermediate-resolution temporal structures. Based on the Stationary Wavelet Transform, it reveals differences across resolutions in transferability, discriminability, and category-level decision-making capability, and proposes PRISM, a time-series domain adaptation framework. PRISM progressively integrates multi-resolution features and decision information through cross-resolution knowledge transfer and committee-based consensus training, thereby improving cross-domain consistency and generalization. Experiments on six datasets demonstrate that the proposed method outperforms existing approaches.

Critical strengths
  • This paper introduces a stationary wavelet transform to construct an intermediate-resolution hierarchy. Three types of experiments systematically validate two key insights: the trade-off between resolution, transferability, and discriminability, and the class-specific decision advantages of different resolutions. These findings provide clear empirical justification for the proposed method.
  • Two dedicated mechanisms are designed to address these insights. Cross-Resolution Knowledge Transfer (CRT) employs asymmetric KL divergence to transfer discriminative information from fine to coarse resolutions in the source domain and regularize transferability from coarse to fine resolutions in the target domain, aligning with the characteristics of the multi-resolution hierarchy. Committee-based Consensus Training (CCT) uses the original resolution as the anchor and the remaining resolutions as an auxiliary committee, incorporating confidence masks to filter noisy predictions.
Critical weaknesses
  • CRT and CCT are arranged in a sequential architecture and share multi-resolution representations and classifier outputs, but lack explicitly designed interaction interfaces, feedback pathways, or joint loss terms. Although the combined use of the two modules achieves better performance than either module alone, the source of their synergistic gains has not been disentangled. Therefore, it remains unclear whether the improvement stems from positive interactions between the two mechanisms or merely from the additive gains contributed by each module independently.

  • The Stationary Wavelet Transform, knowledge distillation, committee learning, and adversarial domain adaptation are all well-established techniques. However, the paper does not clearly articulate whether its core contribution lies in proposing a novel domain adaptation principle or paradigm, or merely in engineering a combination of existing techniques for multi-resolution time-series data.

Questions for authors
  • The two key multi-resolution properties identified in the paper—the trade-off between transferability and discriminability, and the class-specific decision advantages of different resolutions—are validated only within the DANN framework and on activity recognition datasets. Have the authors investigated whether these patterns also hold under other domain adaptation paradigms, across different task types, and with alternative multiscale decomposition methods? If not, is there any theoretical or empirical evidence supporting the generalizability of these findings to a broader range of time-series domain adaptation scenarios?
4 comments
AuthorsOpenReview ↗

Looking forward to your feedback

Dear Reviewer dP9e,

Thank you very much for your insightful and valuable comments!

In our previous response, we have addressed your queries by conducting additional experiments to explicitly demonstrate the positive interactions between CRT and CCT (W1), further clarifying the core contributions and novelty of our work (W2), and providing additional empirical evidence supporting the generalizability and broader applicability of the two key multi-resolution properties identified in our paper (Q1).

We would be happy to provide further responses. We look forward to your feedback.

Thanks for your comments again!

Best Regards,

All Authors

AuthorsOpenReview ↗

Looking Forward to Your Feedback Before the Discussion Ends

Dear Reviewer dP9e,

Thank you again for your insightful and valuable comments!

We would like to kindly follow up on our previous rebuttal, in which we conducted additional experiments to explicitly demonstrate the positive interactions between CRT and CCT (W1), further clarified the core contributions and novelty of our work (W2), and provided additional empirical evidence supporting the generalizability and broader applicability of the two key multi-resolution properties identified in our paper (Q1).

As one week has passed since we submitted our rebuttal and fewer than 48 hours remain in the author-reviewer discussion period, we would greatly appreciate it if you could share your thoughts on our responses. We would be happy to provide any further clarification if some concerns remain.

Thank you again for your time, consideration, and constructive feedback. We look forward to hearing from you.

Best Regards,

All Authors

I appreciate the authors’ response. I have no further questions and have raised my score.

AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer dP9e,

We would like to express our sincere gratitude for your careful consideration of our rebuttal and your positive assessment of our responses. We are particularly grateful that our clarifications have addressed your concerns and that you have decided to raise your rating. Your thoughtful feedback, time, and effort throughout the review process are deeply appreciated and have been invaluable in helping us improve our work.

Thank you again for your thoughtful consideration and support!

Best Regards,

All Authors

Reviewer_SWvpOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies unsupervised domain adaptation for time-series classification (TSDA), aiming to mitigate the distribution shift between labeled source domains and unlabeled target domains. The authors propose the PRISM framework, which consists of two core modules: Cross-Resolution Knowledge Transfer, which bridges multi-resolution representations by progressively exchanging complementary knowledge across adjacent resolutions, and Committee-based Consensus Training, which bridges the cross-resolution decision hierarchy through committee consensus, enabling more consistent domain-invariant decisions. Extensive experiments on six datasets demonstrate that PRISM consistently outperforms existing methods, achieving an average accuracy improvement of 4.21% in cross-domain settings.

Critical strengths
  1. This paper systematically investigates the intermediate-resolution hierarchy induced by the Stationary Wavelet Transform and identifies two key properties through empirical analysis, different resolutions exhibit distinct transferability/discriminability characteristics and distinct category-level decision strengths under domain shift, providing a clear rationale for designing effective domain-invariant learning mechanisms.

  2. PRISM introduces two dedicated modules that directly correspond to the identified properties: Cross-Resolution Knowledge Transfer (CRT) progressively exchanges complementary knowledge across adjacent resolutions to bridge the representation level, while Committee-based Consensus Training (CCT) bridges the decision hierarchy through anchor-guided committee consensus. The modules map one-to-one onto the motivations rather than being arbitrarily combined.

  3. Across six datasets spanning diverse application scenarios and multiple SOTA baselines, PRISM achieves consistent improvements (4.21% on average), supported by ablation studies and sensitivity analyses that verify the effectiveness of each module.

Critical weaknesses
  1. The CLUDA row and the AdaMatch row in Table 1 are numerically identical across all 10 tasks. It is virtually impossible for two different methods to produce exactly matching results on every single task. Please verify the correctness of the experimental results.

  2. The two key insights in Figure 1 are based on only 3 datasets, a single DANN backbone, and A-distance measurements taken prior to adaptation. It remains unclear whether the conclusions, the ordering of transferability/discriminability across resolutions and the category-level complementarity generalize to PRISM's own architecture and to long-sequence datasets such as EEG and FD.

  3. Multi-scale/multi-resolution time-series modeling is already well established (e.g., TimeMixer, TimesNet, which the authors themselves cite). CRT is essentially cross-view consistency distillation, closely related to ACON's temporal–frequency mutual learning, while CCT is essentially confidence-thresholded self-training with consensus weighting. Therefore, the novelty of this paper is limited, and the claim of being the first TSDA framework to move beyond the temporal–frequency extremes appears to be an overstatement.

Questions for authors

Please answer the questions in the weakness section.

Reviewer_umqEOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PRISM, a framework for unsupervised time-series domain adaptation (UTSDA). PRISM leverages the Stationary Wavelet Transform (SWT) to construct multi-scale approximation views and learns temporal–frequency representations across scales. The approach is evaluated across six benchmarks involving ten source–target transfer tasks per dataset and five random seeds, demonstrating performance gains over recent domain adaptation baselines.

Critical strengths

Comprehensive Empirical Evaluation: The evaluation spans multiple domain types (human activity recognition, fault diagnosis, sleep staging), incorporates multiple transfer tasks per benchmark, and reports multi-run averages across five seeds.

Critical weaknesses
  1. Unclear Role and Necessity of Multi-Scale Hierarchy: The necessity of intermediate resolution scales is not adequately demonstrated. As shown in Table 5, four of the six datasets employ only a single decomposition level (L=1L=1). To justify the multi-resolution design, the paper needs a controlled comparison contrasting the full hierarchy against its endpoints (the original and coarsest views), as well as capacity-matched multi-branch baselines and simple low-pass filtered controls.

  2. Misalignment Between A\mathcal{A}-Distance Metrics and Transferability: A lower A\mathcal{A}-distance demonstrates reduced domain distinguishability in feature space, but it does not inherently guarantee improved transferability or semantic preservation. Wavelet smoothing can remove domain-specific noise while inadvertently discarding task-discriminative features. Demonstrating true category-level alignment requires direct class-wise metrics or explicit pseudo-label quality analysis.

  3. Inconsistencies Across Ablation Tables and Missing Implementation Details:

    • Table / Text Discrepancies: Appendix C.2 states that a multi-resolution baseline scores 83.40%83.40\% on WISDM, increasing to 86.01%86.01\% with CRT. However, Table 3 reports 86.01%86.01\% as "CRT-only", while "MRL+CRT" achieves 86.65%86.65\%. This mismatch makes it difficult to verify what constitutes the exact baseline configuration.
    • Underspecified Ablation Mechanics: The paper does not clarify how CRT or CCT functions when Multi-Resolution Learning (MRL) is disabled, nor whether the teacher/reference distribution in KL-divergence calculations is detached during optimization (sg[⋅]\text{sg}[\cdot]).
    • Hyperparameter Selection Criteria: While the text asserts that hyperparameters are chosen without target label supervision, the exact target-free criteria used to determine dataset-specific decomposition depths and loss weights are omitted, impacting reproducibility.
Questions for authors
  1. Necessity of Intermediate Scales: What empirical evidence demonstrates that intermediate resolution levels contribute meaningfully to transfer performance? Can you provide an ablation comparing the full multi-scale hierarchy against a two-view control (original + coarsest view) under a fixed maximum decomposition depth?

  2. Isolation of Wavelet Decomposition Benefits: How much of the framework's improvement is specifically attributable to Stationary Wavelet Decomposition? How does PRISM perform when compared against (i) single-resolution temporal-frequency baselines, (ii) capacity-matched multi-branch architectures without wavelet decomposition, and (iii) simple Gaussian/low-pass smoothed views?

  3. Reconciliation of Ablation Figures: How should the baseline metrics in Appendix C.2 / Table 6 be reconciled with Table 3? What architectural components remain active when MRL is disabled?

  4. CRT Implementation & Stability: Is the target/teacher-side probability distribution detached during the KL-divergence step in CRT? How stable is the proposed directional KL setup across different transfer tasks and random seeds?

  5. Impact of CCT on Pseudo-Label Quality: How does CCT quantitatively affect target pseudo-label accuracy? Could you report precision and recall metrics for pseudo-labels, particularly on sample subsets where the original resolution prediction is incorrect but auxiliary resolutions yield correct classifications?

  6. Target-Free Hyperparameter Tuning: What specific unsupervised validation metrics (e.g., source risk, entropy-based criteria) were used to select decomposition depths and loss parameters? Please distinguish this operational selection process from post-hoc parameter sensitivity analyses.

3 comments
AuthorsOpenReview ↗

Looking forward to your feedback

Dear Reviewer umqE,

Thank you very much for your insightful and valuable comments!

In our previous response, we have addressed your queries regarding the Multi-resolution Hierarchy and the Isolation of Wavelet Decomposition Benefits (W1, Q1, Q2), the relationship between transferability and A-distance (W2), the role of MRL and whether it is enabled in the relevant experimental settings, the gradient flow of CRT, and additional experiments validating the effectiveness and stability of CRT and CCT (W3, Q3, Q4, Q5), as well as our Target-Free Hyperparameter Tuning protocol (Q6).

We would be happy to provide further responses. We look forward to your feedback. Thanks for your comments again!

Best Regards,

All Authors

Thank you for the thorough rebuttal. I also read the authors' rebuttal to the other reviewers. I think my concerns have been substantially addressed. Thus, I raise my rating from 2 to 3.

AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer umqE,

We would like to express our sincere gratitude for your careful consideration of our rebuttal, including taking the time to review our responses to the other reviewers. We are particularly grateful for your recognition that our responses have substantially addressed your concerns and for your decision to raise the rating of our paper. Your thoughtful feedback, time, and effort throughout the review process are deeply appreciated and have been invaluable in helping us improve our work. If you have any additional questions or suggestions, please do not hesitate to share them with us. We look forward to further discussions to enhance this work collaboratively.

Thank you again for your thoughtful consideration and support!

Best Regards,

All Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Manuscript Changes

Dear Reviewers,

We have uploaded a revised PDF of the manuscript incorporating the following updates in response to the reviewers' comments:

  1. Main Text – Table 1: Updated the results reported in the AdaMatch and CLUDA rows.

  2. Appendix C.2 – Table 6: Updated the results in the Ours row, with the accuracy corrected from 86.01% to 86.65%.

  3. Appendix C.3: Added Figure 4 and a corresponding paragraph analyzing the effect of CCT on anchor prediction quality.

  4. Main Text – Introduction: Revised the contribution statement from “the first TSDA framework that moves beyond the conventional temporal–frequency extremes” to “the first TSDA framework to systematically investigate the intrinsic transferability–discriminability relationships across wavelet-induced resolutions”.

For ease of identification, all newly added or modified content in the revised PDF is highlighted in red. We hope these revisions and additional experimental analyses help clarify and strengthen the empirical evidence supporting our method. Thank you for your consideration.

Regards Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.