acceptodds
Under review as a conference paper at ICLR 2027

TimeCED: Contribution-Based Evidence Discovery for Multivariate Time Series Classification

Abstract

Multivariate time series classification (MTSC) models typically integrate temporal and cross-channel information into sequence-level representations, which may obscure important local information. Existing methods explicitly model or emphasize local regions, but do not directly quantify how each region contributes to the classification objective. Motivated by this, we propose TimeCED, a Contribution-Based Evidence Discovery framework that separates sequence-level representation learning from local evidence discovery. TimeCED directly quantifies the contribution of each channel-temporal region through the change in classification loss induced by candidate removal. Rather than computing this effect exactly, we use a first-order Taylor expansion to efficiently approximate it. High-contribution candidates are consolidated and reused in subsequent training to improve classification performance. We evaluate TimeCED on all 30 datasets in the UEA multivariate time series classification archive, where it achieves strong performance against nine baseline methods. Further analyses confirm the effectiveness of the defined contribution score and the reliability of the Taylor approximation.

Primary area:
learning on time series and dynamical systems
Keywords:
Multivariate Time Series Classification, Local Evidence Discovery, Contribution-Based Learning, Taylor Approximation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_TVrLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TimeCED proposes a multivariate time-series classifier that separates sequence-level representation learning from local candidate selection. It defines a candidate's contribution as the change in classification loss when that candidate's learned embedding is removed, approximates this effect with a first-order Taylor expansion, aggregates scores over samples and discovery epochs, and reuses high-scoring channel-temporal candidates for subsequent training. On 30 UEA datasets, the reported mean accuracy is 0.744 versus 0.718 for MiniROCKET in the paper's comparison table. Removing the whole candidate pathway lowers the mean by 3.60 percentage points. A Top-100 versus random/bottom selection comparison on eleven eligible datasets yields means of 0.659, 0.633, and 0.625. The Taylor scores agree closely with exact removal of the same learned embeddings, with mean Spearman correlation 0.9954 across those eleven datasets.

Critical strengths

S1. The paper addresses the potential loss of local discriminative information during global temporal aggregation by explicitly identifying channel-temporal regions that contribute to the classification objective. The two-pathway design is conceptually coherent, and the contribution-based selection mechanism provides a more direct connection to the classification loss than purely attention-based importance weighting. The distinction between global sequence modeling and local evidence discovery is a meaningful design choice.

S2. The authors evaluate all 30 UEA datasets and provide relevant structural ablations. The evidence-selection experiments show that top-ranked candidates outperform random and bottom-ranked candidates by 2.6 and 3.4 percentage points, respectively, on the evaluated subset. The Taylor approximation is also validated against exact candidate removal, achieving an average Spearman correlation of 0.9954 and a Top-100 overlap of 97.7%. These results provide credible evidence that the approximation faithfully reproduces the proposed removal-based ranking and that the resulting ranking is useful for classification.

Critical weaknesses

W1. The paper shows that the Taylor approximation is accurate, but it does not adequately compare the proposed score with other importance measures, such as gradient-based scores or absolute contribution values. Therefore, it remains unclear whether the proposed scoring method is responsible for the performance improvements.

W2. TimeCED selects the same important channel-temporal locations for all samples. However, different classes or samples may contain discriminative patterns at different time positions. The paper does not investigate whether this shared selection strategy remains effective in such cases. Additional analysis of computational overhead would also strengthen the evaluation.

Questions for authors

Q1. Have the authors evaluated gradient-based scores, absolute contribution values, or removal-based importance scores under the same Top-M selection budget? Such results would help establish the advantage of the proposed scoring method.

Q2. How does it compare with simpler importance scores under the same architecture and selection budget? Additionally, does the Top-100 selection strategy consistently outperform random selection on the remaining 19 UEA datasets? These results would help determine whether the proposed scoring mechanism is responsible for the reported improvements.

Reviewer_RFy4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes TimeCED, which combines sequence-level representation learning with contribution-based evidence discovery for multivariate time-series classification. TimeCED quantifies the contribution of each channel-temporal region through the change in classification loss caused by candidate removal, and utilizes high-contribution regions for classification. To improve computational efficiency, the authors estimate the contribution of each channel-temporal region using a first-order Taylor approximation. Experimental results on the UEA benchmark demonstrate that TimeCED achieves competitive performance compared to existing multivariate time-series classification methods.

Critical strengths

The proposed method combines sequence-level information with important local information for multivariate time-series classification. However, the novelty is limited compared to existing methods, and the validity of the contribution estimation approach raises several concerns.

Critical weaknesses
  1. The difference between the proposed method and existing approaches for measuring the contributions of local regions is unclear
  2. Replacing a candidate embedding with a zero vector may not have the same effect as removing the candidate embedding
  3. The authors should justify why they average the contribution score from all samples and all epochs
  4. Sensitivity analysis is not conducted for the hyperparameters introduced by the proposed method.
  5. The performance gains in Tables 1 and 3 appear marginal
  6. Figure 3 only demonstrates the consistency of contribution scores across random seeds, without examining their consistency across individual samples
  7. The effectiveness of the candidate pathway should be further validated by examining whether it improves the performance of existing time-series classification models
  8. The notation consistency should be improved
Questions for authors
  1. How does the proposed contribution estimation method differ from Grad-CAM and existing local evidence modeling approaches?
  2. Does replacing a candidate embedding with a zero vector have the same effect as actually removing the embedding?
  3. Why are contribution scores averaged across all samples and epochs? The scores may be unreliable during early epochs, and the important local regions may vary across different samples
  4. Please provide sensitivity analysis results for the number of temporal patches SS, number of selected candidates MM, candidate sampling budget BB, target exposure count rr, and minimum discovery duration EminE_{min}
  5. Please verify whether the performance improvements in Tables 1 and 3 are statistically significant
  6. Are the important regions demonstrated in Figure 3 consistently important across individual samples?
  7. Can the proposed candidate pathway also improve the performance of existing time-series classification models?
  8. Please explain why the proposed method focuses on multivariate time series, not on univariate time series
  9. The "candidate encoder" in Figure 2 is referred to as the "event encoder" in the text. Please ensure notation consistency
Reviewer_nEPPOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents an MTSC model combining a sequence pathway with a local candidate pathway to quantify per-class contributions of each local region. It mainly estimates candidate contributions using a first-order Taylor approximation of the loss change induced by zeroing candidate embeddings. The scores are accumulated during stochastic discovery, and the channel-temporal candidates with the highest scores are kept for subsequent training. The evaluation is conducted on 30 UEA MTSC datasets, with ablation studies and comparisons against exact candidate removal.

Critical strengths
  1. The presented framework provides an explicit, task-dependent criterion for selecting local candidates.
  2. Compared with the repeated removal evaluations, reusing classification gradients makes region contribution estimation computationally attractive.
  3. The mean results reported across 3 runs on 30 UEA datasets make the claim convincing.
  4. The statistical results are well visualized, and the paper is easy to follow.
Critical weaknesses

Major:

  1. Zeroing a candidate embedding affects the candidate pathway only. The corresponding raw region remains accessible through the sequence pathway. The score measures a conditional contribution rather than the intrinsic importance of the original region.
  2. Section 4.3 reports only average accuracy changes across the benchmark. Please provide a more detailed analysis identifying datasets on which removing a component has little effect, improves performance, or causes substantial degradation. Relating these outcomes to dataset properties, such as sequence length, dimensionality, and temporal dependencies, would guide practical application and model optimization.
  3. The authors are encouraged to provide pairwise Wilcoxon signed-rank tests across datasets, with appropriate multiple-comparison correction, particularly for comparisons against the strongest baselines, as in the TSC-dedicated studies [1-6].
  4. The stability and class dependence of the selected candidates remain insufficiently characterized. Please quantify the consistency of selected candidate sets across seeds and examine whether different classes rely on different channel–temporal regions.
  5. The target number of temporal patches is fixed at 20, and both the candidate sampling budget and the number of retained candidates are capped at 100. However, the authors do not provide sensitivity analysis and practical selection guidelines based on dataset characteristics or temporal properties.

Minor:

  1. Please number the main equations to facilitate cross-referencing.
  2. Figure 2 appears to have sufficient space for larger text. Please increase the font size to improve readability.

[1] Küken, Jaris, et al. "TimEE: End-to-end Time Series Classification via In-Context Learning." arXiv preprint arXiv:2607.07500 (2026).

[2] Tang, Yingxia, et al. "Convolutional network integrated with frequency adaptive learning for multivariate time series classification." ACM Transactions on Knowledge Discovery from Data 19.8 (2025): 1-23.

[3] Le, Xuan-May, et al. "Shapeformer: Shapelet transformer for multivariate time series classification." Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 2024.

[4] Liu, Huaiyuan, et al. "Todynet: temporal dynamic graph neural network for multivariate time series classification." Information Sciences 677 (2024): 120914.

[5] Wang, Jingyuan, et al. "WHEN: A Wavelet-DTW hybrid attention network for heterogeneous time series analysis." Proceedings of the 29th ACM SIGKDD conference on knowledge discovery and data mining. 2023.

[6] Tang, Wensi, et al. "Omni-scale cnns: a simple and effective kernel size configuration for time series classification." arXiv preprint arXiv:2002.10061 (2020).

Questions for authors
  1. How should candidate contribution be interpreted when the corresponding raw regions remain accessible through the sequence pathway?
  2. Do authors provide an analysis on dataset characteristics when removing a component improves or substantially degrades performance?
  3. Are improvements over the strongest baselines statistically significant under pairwise Wilcoxon signed-rank tests with multiple-comparison correction?
  4. How stable are the selected candidate sets across seeds, and how do they differ across classes?
  5. How sensitive are performance and computational cost to patch granularity, sampling budget, and retained-set size? How should these parameters be chosen for customized datasets?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.