acceptodds
Under review as a conference paper at ICLR 2027

DTW4TS: Learning Time Series Representations from Reliable DTW Relations

Abstract

Real-world time series inherently exhibit complex temporal misalignments, posing challenges for current representation learning methods that primarily rely on rigid, point-to-point temporal matching. While Dynamic Time Warping (DTW) naturally accommodates such misalignments through nonlinear alignment, existing DTW-based methods often use DTW relations directly as supervision and mainly rely on scalar similarities, leaving relation reliability and fine-grained warping path correspondences comparatively underexplored. To address these limitations, we propose DTW4TS, a reliability-aware framework for time series representation learning. Rather than utilizing all scalar scores uniformly, DTW4TS first mitigates the impact of unreliable DTW relations by assessing local neighborhood consistency and alignment path deviation. We further introduce a lightweight DTW Adapter that encodes reliable warping paths and injects alignment information into intermediate encoder layers. In addition, an elastic contrastive objective constructs cross-sequence positive pairs from reliable temporal correspondences, encouraging the learned representations to capture elastic relationships across time series. Experiments across classification, anomaly detection, and forecasting benchmarks show that DTW4TS achieves the highest average classification accuracy and anomaly-detection F1, and the lowest average forecasting MSE, among the evaluated baselines.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series, Representation Learning, Dynamic Time Warping, Reliability Modeling, Contrastive Learning.

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_2PALOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

DTW4TS augments a TS2Vec-style representation learner with reliability-filtered Dynamic Time Warping (DTW) supervision. Candidate neighbors are first selected by DTW distance. Their reliability is estimated from neighborhood overlap and normalized alignment-path deviation, after which selected paths support two training mechanisms: a gated intermediate-layer adapter encoding path geometry and an elastic contrastive objective using cross-sequence temporal correspondences. At inference, the adapter is disabled and each instance is encoded independently.

The paper evaluates classification on 125 UCR and 27 UEA datasets, anomaly detection on five datasets, and forecasting on six datasets. Table 1 reports average classification accuracies of 0.835 and 0.715 for UCR and UEA. Table 2 reports mean anomaly-detection F1 of 0.433 versus 0.385 for SoftCLT. Table 3 reports the lowest overall forecasting MSE, 0.195 versus 0.221 for iTransformer, but not the lowest overall MAE: DTW4TS has 0.292 versus 0.287 for iTransformer (pp. 7-8).

Critical strengths
  1. A coherent use of alignment information. Reliability selection, path encoding, and elastic supervision address complementary aspects of exploiting DTW relations beyond a scalar distance.
  2. Broad evaluation with informative controls. The classification archive is substantial. Table 4 includes raw-DTW, individual-reliability-cue, adapter, elastic-loss, and shuffled-path variants. Distance- and rank-conditioned analyses further test whether reliability carries information beyond DTW proximity (p. 9 and Appendix E, p. 17).
  3. Existing robustness and sensitivity evidence. Appendix F varies the sampled window pool from 64 to 1,024. Appendix G evaluates controlled temporal warping across representative tasks. These are relevant positive results and should not be omitted when assessing the paper (pp. 18-19).
  4. No alignment machinery at inference. The intended downstream representation is produced by the standalone backbone. DTW distances and paths are constructed using the training split only, and the appendix reports preprocessing, training, inference, and cache costs (Appendices D and H, pp. 16 and 18-20).
Critical weaknesses
  1. Reliability is a useful heuristic rather than a universal semantic criterion. Large deviations from the diagonal can be valid, and shared-neighborhood consistency may disadvantage rare or isolated patterns. The controlled-warping results mitigate a generic robustness concern, but do not directly measure false rejection of valid relations by class frequency or deformation magnitude. This is a specific remaining question, not a demonstrated failure.
  2. The adapter mechanism merits additional diagnosis. Table 4 already shows that training with the adapter benefits downstream evaluation after it is removed. It would therefore be inaccurate to imply that useful information has not transferred to the backbone. What remains unclear is how much the backbone relies on the training-time residual, and whether branch dropout or distillation improves the resulting representation.
  3. Scaling beyond the tested alignment pools remains uncertain. The 64-1,024-window sensitivity analysis and timing table already address practical pool selection. The remaining issue is coverage and preprocessing at substantially larger instance counts or longer sequences. For example, Table 7 reports approximately 890 seconds for pairwise DTW preprocessing on UGLA with 896 instances of length 945, in addition to candidate-path/reliability construction (p. 20).
  4. The aggregate gains need stronger statistical and task-level interpretation. UCR average accuracy rises from 0.829 for TS2Vec to 0.835, a gain of 0.6 percentage points. Per-dataset paired analyses would clarify robustness. MSL recall of 0.075 is low for several methods, not a unique defect of DTW4TS; it limits absolute application claims rather than refuting the relative improvement. Forecasting gains are stronger in MSE than in MAE.
Questions for authors
  1. Beyond the existing controlled-warping experiment, how often does reliability filtering discard same-class relations in minority classes or strongly deformed but valid pairs?
  2. Given the positive backbone-only results in Table 4, can an adapter-branch dropout or distillation experiment diagnose the training-inference interaction? This is a request for mechanism analysis, not a missing baseline already supplied by the paper.
  3. Do mutual-neighbor selection or an explicitly constrained warping band provide comparable performance at similar pair budgets?
  4. Extending the reported 64-1,024-window study, how does rare-pattern coverage change for much larger pools or approximate DTW-neighbor construction? Please separate preprocessing time from neural training time.
  5. Are the classification improvements significant in paired comparisons across datasets, and are forecasting gains robust across both MSE and MAE?
Reviewer_coFXOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DTW4TS, a self-supervised time series representation learning framework built upon TS2Vec. The method selects DTW-based neighbor relations using a reliability score that combines shared-neighborhood consistency with average warping-path deviation from the diagonal. It then exploits the selected paths through two components: a DTW Adapter that injects encoded path geometry into intermediate encoder features, and an elastic contrastive objective that constructs cross-instance positives from aligned timestamps. DTW preprocessing and the adapter are used only during training. Experiments cover classification, anomaly detection, and forecasting, with additional reliability analyses, ablations, and controlled temporal-warping experiments.

Critical strengths

The paper addresses a meaningful limitation of scalar DTW supervision. A DTW distance summarizes an alignment into a single value, discarding its temporal correspondences. Using these correspondences to construct cross-instance temporal positives is a natural extension of representation learning objectives based on corresponding timestamps within augmented views.

The reliability mechanism receives useful empirical validation. The paper examines same-class probability and AUROC while controlling for DTW neighbor rank and approximately controlling for distance. These analyses support the claim that the proposed indicators contain information beyond raw DTW proximity.

The evaluation covers diverse downstream tasks. The classification experiments span 152 UCR/UEA datasets, and the paper also evaluates anomaly detection and forecasting. The component ablations, shuffled-path control, and hyperparameter analyses provide evidence that the gains involve both relation selection and path utilization.

Critical weaknesses

1. The path-deviation measure does not clearly distinguish incorrect alignment from legitimate temporal deformation. The proposed Rij=1∣Aij∣∑(a,b)∈Aij∣a−b∣T−1R_{ij}=\frac{1}{|A_{ij}|}\sum_{(a,b)\in A_{ij}}\frac{|a-b|}{T-1} measures average displacement from the diagonal. A correct alignment involving a substantial temporal shift can therefore receive a large penalty, whereas an unstable path that stays near the diagonal may receive a small one. This creates tension with the goal of learning elastic relationships. The empirical correlations support its usefulness as a heuristic, but do not establish that it measures alignment correctness.

2. The reliability validation concerns sequence-level semantics rather than timestamp-level correspondence accuracy. Same-class membership is a reasonable proxy for selecting related sequences, but it does not guarantee that their DTW-aligned timestamps are suitable positives. Since the elastic objective directly relies on these correspondences, validation against known temporal mappings would provide stronger support for the central mechanism.

3. The training-only adapter introduces an insufficiently explained training–inference discrepancy. Training representations include path-conditioned residuals, while evaluation representations are produced with these residuals removed. The paper reports favorable performance under this procedure, but does not explain or directly measure how the backbone retains the information supplied by the adapter. A consistency or distillation analysis would help establish why this design works.

Questions for authors

Main Concern: 1. How does reliability-based selection differ from selection using DTW distance alone? Please quantify how much the selected neighbor sets differ, for example through their overlap and same-class pair precision. A direct comparison between selecting the top �0��0� neighbors by reliability and selecting the same number by DTW distance would help isolate the benefit of the proposed score.

2. How is path deviation defined for sequences of unequal lengths? The current formulation uses �1��1�, assuming a shared length �2��2�. When the lengths differ, even a proportional alignment can deviate substantially from the main diagonal. Would the method instead measure deviation using normalized coordinates, such as �3��3�? Please clarify how unequal lengths are handled.

3. What are the average improvements over individual baselines across datasets? Beyond reporting average accuracy and other aggregate metrics, please report the mean and median of dataset-wise improvements over TS-TCC, TS2Vec, and other competitive baselines. Win/tie/loss counts and paired statistical tests would further clarify whether the improvements are broadly distributed and statistically meaningful. Please distinguish relative improvements from absolute percentage-point gains.

4. How are unreliable candidate neighborhoods handled? The method always selects �4��4� neighbors, even when all candidates have low reliability. Furthermore, selected relations become hard positives in the elastic loss without direct weighting by $ρij\rho_{ij} $. Please discuss this design and compare it with reliability thresholds, abstention, or weighted positive supervision.

4 comments
AuthorsOpenReview ↗

Response to Weaknesses: Reliability and Adapter Consistency

Thank you for your thoughtful review. We address the three weaknesses below and have revised the manuscript accordingly.

W1

We agree that path deviation is a heuristic indicator rather than an exact measure of alignment correctness. DTW4TS jointly considers path deviation and local neighborhood consistency, leveraging their complementary information to guide reliable DTW relation selection.

(1) Empirical Evidence. Figure 5 demonstrates the association between reliability measures and same-class probability, while Table 4 validates their effectiveness in representation learning. We further analyze same-class relation retention across path-deviation quintiles (Appendix E.2). Compared with DTW-only selection, reliability-guided selection increases retention in low-deviation Q1 (+14.39 pp UCR; +27.11 pp UEA) while reducing it in high-deviation Q5 (−19.95 pp UCR; −38.10 pp UEA).

DeviationUCR DTWUCR ReliableΔ(pp)UEA DTWUEA ReliableΔ(pp)
Q1 (low)71.9286.31+14.3965.7492.85+27.11
Q263.6078.76+15.1561.7681.35+19.59
Q361.2967.98+6.6956.8462.95+6.10
Q456.7353.64−3.0955.1443.53−11.61
Q5 (high)52.4632.50−19.9555.4517.35−38.10

These results support the effectiveness of path deviation as a structural reliability prior for selecting informative DTW relations. We also acknowledge that path deviation may assign lower reliability to legitimate temporal shifts. More precise reliability modeling remains a direction for future work, as noted in the revised Conclusion (line 524).

W2

We agree that sequence-level semantic similarity does not necessarily ensure accurate timestamp-level correspondences. DTW establishes elastic alignments by minimizing alignment costs [1,2]. In DTW4TS, reliability modeling selects reliable DTW relations and their associated paths without modifying the alignments themselves. Thus, the selected DTW correspondences provide fine-grained elastic supervision for learning cross-sequence temporal relationships.

Since ground-truth temporal correspondences are generally unavailable in real-world benchmarks, we provide complementary evidence: reliability analyses (Figure 5; Appendix E) demonstrate the effectiveness of reliable relation selection, while component ablations (Table 4) validate the downstream benefits of DTW-guided temporal learning.

W3

We added two complementary analyses in Section 4.4 (line 485) and Appendix F.

(1) Representation Consistency. We compare adapter-on/off representations using identical inputs and trained checkpoints, evaluating residual ratio, linear CKA, cosine similarity, and RFD. The results show high consistency (CKA: 0.978–0.999; cosine: 0.969–0.998), indicating that the backbone preserves highly consistent representations without the Adapter. Combined with Table 4, these findings support effective adapter-free inference.

TaskDatasetResidualCKACosineRFD
Cls.UGLA0.0230.9990.9970.079
Cls.SRSCP10.0310.9990.9970.075
Ano.GECCO0.1090.9960.9890.146
Ano.ECG0.2440.9780.9690.252
Fore.ETTm10.1370.9930.9780.218
Fore.Exchange Rate0.0480.9990.9980.064

(2) Self-Distillation. We further evaluate adapter-to-backbone self-distillation using an adapter-on teacher and adapter-off student with shared backbone parameters. Although distillation reduces RFD on four datasets, the standard strategy achieves better downstream performance on five of six datasets.

TaskDatasetMetricStandardDistilled
Cls.UGLARFD0.0790.078
Acc ↑0.9580.949
SRSCP1RFD0.0750.075
Acc ↑0.8020.798
Ano.GECCORFD0.1460.156
F1 ↑0.7580.779
ECGRFD0.2520.245
F1 ↑0.2810.278
Fore.ETTm1RFD0.2180.213
MSE ↓0.0750.079
Exchange RateRFD0.0640.061
MSE ↓0.2830.300

These findings indicate that minimizing representation discrepancy does not necessarily translate into better task performance, supporting the effectiveness of our original training-only Adapter design.

[1] Sakoe & Chiba. Dynamic programming algorithm optimization for spoken word recognition. IEEE TASSP, 1978.

[2] Cuturi & Blondel. Soft-DTW: A differentiable loss function for time-series. ICML, 2017.

AuthorsOpenReview ↗

Response to Questions: Additional Analyses and Experimental Evidence

Thank you for your insightful questions. We provide the following clarifications and additional experimental evidence.

Q1

We conduct a controlled comparison between DTW-based and reliability-guided selection. Both select $ K^+=4 $ neighbors from the same $ K=7 $ nearest DTW candidates, differing only in ranking criteria.

ArchiveOverlap@4DTW PrecisionReliability PrecisionΔ
UCR (125)0.6650.6290.649+0.020
UEA (27)0.6370.5740.578+0.004

Reliability-guided selection replaces 33.5% and 36.3% of DTW-selected neighbors on UCR and UEA, respectively, while improving same-class precision by 2.0 and 0.4 percentage points. These results demonstrate that reliability modeling meaningfully refines neighbor selection and improves the semantic consistency of selected relations (Appendix E.2).

Q2

DTW4TS is designed for equal-length inputs, consistent with common TSRL benchmarks. Unequal-length sequences can be converted into equal-length inputs through padding or interpolation-based resampling [1]. We also recognize normalized temporal coordinates as a promising extension for directly modeling variable-length sequences. We have clarified the equal-length assumption in the revised manuscript (line 223).

Q3

We added Appendix J, reporting mean/median absolute (pp) and relative (%) improvements, win/tie/loss counts, and paired Wilcoxon signed-rank tests with Holm correction. Classification results across 152 datasets are summarized below:

BaselineMean ΔAcc(pp)Median ΔAcc(pp)Mean Gain(%)Median Gain(%)W/T/LHolm p
TSTCC5.4352.45016.7943.054107/10/351.33×10⁻⁸
TS2Vec0.9650.5002.6670.52188/22/422.01×10⁻⁴
TimesURL4.5532.3508.5102.892107/15/305.84×10⁻¹⁴
SoftCLT3.1801.0505.7471.25099/23/303.81×10⁻¹⁰
PlanTS7.1333.20016.4934.719119/13/209.36×10⁻¹⁷
TimesNet11.7868.45035.34810.809125/9/184.13×10⁻¹⁹
MOMENT14.1829.00048.51213.403117/4/312.31×10⁻¹⁴
UniTS40.27237.550242.930100.101149/1/23.28×10⁻²⁵
FusAD10.3265.85025.2148.242126/7/192.31×10⁻¹⁸
AimTS13.7678.20053.91410.941120/9/232.79×10⁻¹⁶
FreRA14.80711.25032.37116.861130/7/154.78×10⁻²¹

DTW4TS achieves statistically significant improvements over all 11 baselines after Holm correction ($ p<0.05 $), providing strong evidence of consistent classification gains. For anomaly detection and forecasting, results also show empirical improvements, although the smaller dataset counts (5 and 6) limit statistical power [2].

Q4

We added experiments comparing reliability thresholding, abstention, and weighted supervision (Appendix E.3):

VariationUGLA Acc↑SRSCP1 Acc↑GECCO F1↑ECG F1↑ETTm1 MSE↓Exchange MSE↓
Threshold0.9570.7850.7440.2780.0760.300
Abstention0.9540.8110.7260.2600.0740.312
Weighted0.9560.7750.7930.2800.0780.294
DTW4TS0.9580.8020.7580.2810.0750.283

Thresholding shows no consistent improvement. Abstention benefits SRSCP1 and ETTm1, while weighted supervision improves GECCO, but neither consistently outperforms DTW4TS.

Design Rationale. Our fixed-$ K^+ $ strategy ranks neighbors by relative reliability, avoiding dataset-dependent threshold calibration while maintaining stable positive supervision. The results support this design as an effective balance between relation quality and supervision availability.

[1] Dau et al. The UCR Time Series Archive. IEEE/CAA JAS, 2019.

[2] Cohen. Statistical Power Analysis for the Behavioral Sciences, 2nd ed., 1988.

Thank you again for your time!

Dear Reviewer coFX,

We have carefully addressed your concerns in our rebuttal and revised the manuscript accordingly. In particular, we added empirical analyses to validate our reliability modeling, investigated the role of the DTW Adapter through representation consistency and self-distillation experiments, and extended our evaluations with statistical significance tests and comparisons of alternative reliability strategies.

We hope these additional results help address your concerns and clarify our design choices. We would be happy to provide further explanations if needed.

We sincerely appreciate your thoughtful questions and constructive feedback, which have helped us strengthen both the methodology and experimental validation of our work. Thank you again for your time and consideration!

Best regards,

Authors

The rebuttal addressed my concerns, and this edited version indeed shows more extensive experiments against other baselines across different datasets. I am therefore raising my score to weak acceptance. By the way, it would be better if a different color were used in the edited version to highlight the differences from the previous version.

Reviewer_EyiAOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

DTW4TS is a representation learning method for time series with reliable DTW-neighbor selection, an alignment-conditioned training adapter, and a cross-instance temporal contrastive loss. The main contribution of the method is to exploit reliable DTW relations in the model architecture and objective function. The proposed method is validated on a wide range of downstream tasks spanning classification, anomaly detection, and forecasting, and covers popular benchmarks such as the UCR, UEA, and ETT.

Critical strengths

One of the major strengths of this work is its comprehensive evaluation. The method is evaluated on the 125 UCR and 25 UEA datasets and achieves competitive or superior performance, demonstrating that it excels in both univariate and multivariate classification tasks. In addition, the method is evaluated for Anomaly detection and Forecasting, achieving equally good performance. Also, the method used for comparison in all experiments combines established and recent SOTA methods for time series representation learning.

Critical weaknesses

The presentation of the paper and method needs substantial improvement. It was difficult to follow the paper or to isolate its central contribution. Specifically, the abstract (lines 011-013) highlights that temporal misalignment is problematic for current representation learning methods that primarily rely on "rigid, point-to-point temporal matching". It would be helpful for the authors to be explicit about what they mean by rigid, point-to-point matching. In contrastive learning, two time series can be compared using their global embeddings without requiring that timestamp tt in one sequence match timestamp tt in the other. Methods can also use cropping, temporal augmentations, or other strategies that do not impose strict timestamp-level correspondence.

So the statement is too broad unless the authors identify which existing methods suffer from this limitation and how.

Similarly, the method section introduces several notations and components that make it hard to keep up or properly understand the proposed method.

Regarding the experiments, the evaluation on 125 UCR and 27 UEA and the competitive performance are convincing; however, given the computational requirements of precomputing this DTW relation during training, it is hard to see whether this extra gain is worth the cost. Also, ideally, the ablation should include all or most of the datasets used in your experiment. Specifically, the classification ablation should be performed on the 125 UCR and 27 UEA datasets. That way, it would be clear whether all the extra components in the method are necessary on average across all datasets rather than on these carefully selected subsets.

Lastly, on line 053, the authors stated that "SoftCLT uses DTW-derived distances to construct supervisory signals" and emphasized that a scalar similarity score alone provides limited information. It is important to note that SoftCLT uses these as soft weights for contrastive views generated by other methods on which SoftCLT is built, such as TS2Vec, rather than as a standalone supervisory signal.

Questions for authors
  1. Can you elaborate on what you mean by rigid, point-to-point matching and existing methods in TS that apply this in representation learning?
  2. Is this work focused on positive and negative pair mining, or enriching the scalar DTW similarity weights in methods like SoftCLT with reliability-aware DTW relations and warping paths?
  3. It is interesting to see that DTW4TS, which has a contrastive objective that inherently has an inductive bias that is well-suited for classification, clustering, and retrieval tasks, also performs competitively in forecasting and anomaly detection. Do you have any intuition as to why this is the case?
3 comments
AuthorsOpenReview ↗

Clarifications on Method Presentation, Computational Cost, and Extended Ablation Studies

Thank you for the constructive suggestions. We address each point below and have revised the manuscript accordingly.

W1&W2&Q1

We agree that global contrastive learning and temporal augmentations do not necessarily require rigid timestamp-level alignment. We have removed this overly broad claim from the abstract (lines 011–014) and clarified our focus on reliable fine-grained elastic correspondences across sequences. We also revised the abstract and introduction to emphasize three contributions: (i) DTW-relation reliability modeling via neighborhood consistency and path deviation, (ii) a DTW Adapter that injects reliable alignment paths into encoder representations, and (iii) elastic contrastive learning that constructs cross-sequence temporal positives.

W3

We revised the Method section to remove redundant notation and clarify key variables. Figure 2 now includes explicit data-flow annotations and clearer component connections, while the Method Overview better describes the workflow.

W4

(1) Computational Cost. As in SoftCLT [1], DTW relations are precomputed offline and cached for reuse, benefiting repeated training such as hyperparameter optimization [2]. On UGLA (Table 13), preprocessing takes 889.7s versus 302s for training. The amortized preprocessing cost across $ N $ runs is $ T_{\mathrm{amortized}}(N)=889.7/N $:

Training Runs ((N))Amortized Cost (s/run)Relative Overhead
1088.9729.46%
5017.795.89%
1008.902.95%
5001.780.59%
1,0000.890.29%

Thus, overhead decreases with repeated runs. DTW also provides useful elastic supervision: DTW4TS outperforms TS2Vec on all six datasets in Table 4.

(2) Extended Ablations. We extend classification ablations to 10 stratified datasets from each archive (UCR and UEA), covering diverse sequence lengths, sample sizes, and temporal characteristics. Mean accuracies are below; per-dataset results appear in Appendix K. We will extend these evaluations to the full benchmarks in the camera-ready version.

VariationUCR Avg.UEA Avg.
w/o Reliability0.8790.862
w/o R0.8780.859
w/o S0.8790.855
w/o Elastic Loss0.8760.855
w/o DTW Adapter0.8790.857
Shuffled Path0.8770.858
TS2Vec0.8810.836
DTW4TS0.8900.878

W5

We revised line 73 to clarify that SoftCLT uses DTW-derived distances as soft weights in existing contrastive objectives. DTW4TS instead models relation reliability and exploits fine-grained warping paths via its Adapter and elastic contrastive objective.

Q2

DTW4TS goes beyond pair mining and scalar DTW weighting in two ways:

(1) Beyond Pair Mining. Reliable DTW relations, selected through neighborhood consistency and path deviation, are not merely used to select positive/negative pairs: their alignment paths also guide representation encoding and temporal contrastive learning.

(2) Beyond Scalar Weighting. Unlike SoftCLT's distance-based weights, DTW4TS uses fine-grained temporal correspondences to inject alignment information into intermediate features and construct cross-sequence temporal positives.

Together, the Adapter and elastic objective exploit reliable alignments at both architectural and objective levels.

Q3

We attribute downstream gains to three factors:

(1) Elastic Temporal Learning. Reliability modeling, the Adapter, and elastic contrastive learning capture recurring local dynamics across misaligned sequences while preserving informative temporal structures.

(2) Downstream Benefits. For anomaly detection, these representations accommodate normal temporal variation while remaining sensitive to anomalous local deviations. For forecasting, they capture historical patterns despite temporal misalignments.

(3) Evidence. DTW4TS achieves the best average anomaly-detection F1 (0.433) and forecasting MSE (0.195) among evaluated methods. Controlled temporal-warping experiments (Appendix G) further show improved robustness over TS2Vec.

[1] Lee, S., Park, T., & Lee, K. Soft Contrastive Learning for Time Series. ICLR 2024.

[2] Bergstra, J., & Bengio, Y. Random Search for Hyper-Parameter Optimization. JMLR, 2012.

Thank you again for your time and thoughtful review!

Dear Reviewer EyiA,

We have carefully addressed your constructive comments and suggestions in our rebuttal and revised the manuscript accordingly. In particular, we have clarified the motivation and contributions of DTW4TS, improved the presentation of our methodology, and provided additional analyses and ablation experiments to address your concerns regarding computational overhead and the effectiveness of our proposed components.

As the discussion period progresses, we would greatly appreciate it if you could take a moment to review our responses and the corresponding revisions. We would be happy to provide any further clarification or address additional questions you may have.

Your feedback is invaluable in helping us improve the quality of our work. Thank you again for your time, thoughtful comments, and consideration.

Best regards,

Authors

Reviewer_EyiAOpenReview ↗

Response to Authors' Rebuttal

Thank you for the detailed rebuttal and for running additional experiments. While I appreciate the additional experiments on 10 datasets from each archive, this does not fully address my original suggestion to evaluate the ablation variants across the complete UCR and UEA benchmarks. The full evaluation would provide stronger evidence that the individual components consistently contribute to the overall performance rather than primarily benefiting the selected subsets. I understand the authors' intention to include these experiments in the camera-ready version, but I would have preferred to see this evidence in the current submission.

In addition, I still find that Q3 is insufficiently addressed. The authors attribute this performance to the ability to accommodate normal temporal variation while remaining sensitive to anomalous local deviations, and that the learned representations captures historical patterns despite temporal misalignments. While this is a plausible explanation, I find it somewhat generic and not sufficiently grounded in the learning objective.

In all, the rebuttal still hasn't resolved two of my main concerns:

  1. Ablations on the entire UCR and UEA archive
  2. Grounded clarification on why the learned representations transfer effectively to tasks such as Forecasting and Anomaly detection.

Comments on the paper

1 comment

Thank you for your rebuttal. All of my concerns are addressed. I think this work presents a good approach to thinking about temporal relation reliability and alignment, which provides a simple and effective methodological insight for practical application.

I will raise my score to support the acceptance of this work.


Update: I have raised my rating to:

  • Rating: 4: Clear acceptance
  • Confidence: 4: I have substantial direct research experience on this topic.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.