Improving Multivariate Time Series Classification with Class-Wise Training and Model Aggregation
Mouhamadou Mansour Lo, Gildas Morvan, Mathieu Rossi, Fabrice Morganti, David Mercier
cs.LG
Quick verdict
Proposes per-class channel selection + one-vs-rest MiniRocket for MTSC; solid UEA benchmarks show small but significant gains, yet only vs MiniRocket variants.
Summary
The paper proposes a class-wise framework for multivariate time series classification (MTSC) in which dimension (channel) selection is performed independently for each class rather than globally. For each class K, a class-wise adaptation of the ECS/ECP prototype-distance selection methods identifies a subset of informative dimensions; a binary one-vs-rest MiniRocket classifier is trained on that subset; and the C per-class decision scores are fused by argmax of sigmoid-transformed scores. The method is evaluated on 25 equal-length UEA archive datasets with 30 resamples, using critical difference diagrams and multiple comparison matrices. Without dimension reduction, class-wise training (CW) achieves the best average rank (2.92 vs 3.72 for standard MiniRocket) and significantly outperforms it, with notable gains on EthanolConcentration (47.12→49.89) and FaceDetection (62.55→65.51). With reduction, the class-wise local variants (CW+L) outperform their global-selection baselines on most datasets (e.g., ERing under ECS: 87.72→96.09), and local selection also cuts retained dimensions (e.g., AUDIO under ECS: 111.53→96.35). Main limitations: comparison is restricted to MiniRocket variants — MultiRocket and HYDRA are not re-fitted with the framework; absolute gains are modest and dataset-dependent (e.g., AtrialFibrillation stays near 26%); no runtime measurements despite the C-fold training overhead; and no code release is mentioned.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
4 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Class-wise (per-class) dimension selection produces more discriminative representations than a single global selection for MTSC.
- The class-wise variant without dimension reduction (CW) achieves the best average rank (2.92) and significantly outperforms standard MiniRocket (3.72) across 25 UEA datasets with 30 resamples.
- Class-wise local selection (CW+L) outperforms global-selection baselines on most datasets while reducing retained dimensions (e.g., ECS AUDIO 111.53→96.35).
- Gains concentrate in EEG and OTHER categories and on datasets with at most four classes, and reduction effectiveness is category-dependent.
- Explicit class-specific subsets provide interpretability by identifying which dimensions distinguish each class.
Strengths
Above-average evaluation protocol
25 equal-length UEA datasets (2-1345 dimensions, lengths 30-17,984), 30 resamples each with standard errors, plus both critical difference diagrams (Fig. 4) and multiple comparison matrices (Fig. 3) — more rigorous than typical MTSC papers.
Clean ablation chain isolates the source of gains
Table 1 separates one-vs-rest class-wise training (CW vs Base); Table 2 crosses {ECS, ECP} x {Base, CW+G, CW+L}, so the contributions of class-wise training vs class-wise (local) selection are individually measurable.
Principled, fully specified selection adaptation
Heterogeneity-aware reporting
Weaknesses
Single-backbone generality claim
Selection machinery does not improve on no-selection CW
Uneven statistical testing
Unstated experimental configuration and no code
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (arXiv:2609.07493, cs.LG) proposes a class-wise framework for Multivariate Time Series Classification: instead of one global channel-selection step, dimensions are selected independently per class using a class-conditioned adaptation of the prototype-distance selectors ECS/ECP (Algorithm 1); a binary one-vs-rest MiniRocket classifier is trained per class on its selected subset (Eqs. 3-4); predictions are made by argmax over sigmoid-transformed per-class scores (Eq. 5). Three central claims: (i) class-wise selection yields more discriminative representations than global selection; (ii) the class-wise variant without reduction (CW) significantly outperforms standard MiniRocket; (iii) the framework adds interpretability by exposing class-relevant dimensions. Scope is deliberately narrow: one backbone (MiniRocket) and two selectors; TSelect is explicitly excluded (Section 3.2) because its per-dimension classifier scoring does not adapt naturally to a class-conditioned form.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does this paper propose for multivariate time series classification?
It proposes selecting informative input dimensions (channels) separately for each class instead of once globally: a class-conditioned version of the ECS/ECP prototype-distance selectors picks a dimension subset per class, a binary one-vs-rest MiniRocket classifier is trained on that subset, and the per-class scores are fused by argmax into the final multiclass prediction.
Does the method beat standard MiniRocket?
On 25 UEA archive datasets with 30 resamples each, the class-wise variant without dimension reduction achieved the best average rank (2.92 vs 3.72) and significantly outperformed standard MiniRocket in rank terms. Per-dataset gains are modest — e.g., EthanolConcentration 47.12→49.89 and FaceDetection 62.55→65.51 — and results with dimension reduction are mixed but mostly positive versus their global-selection baselines.
Is code or data available for this paper?
The paper does not mention a code release. The UEA archive and MiniRocket implementations are public, and Algorithm 1 is fully specified, but the exact prototype/distance configuration of ECS/ECP used in experiments is not stated, so exact replication requires reimplementation and some guesswork.
When does class-wise dimension selection help the most?
According to the paper's Tables 3-4, gains concentrate on EEG and miscellaneous (OTHER) dataset categories and on datasets with at most four classes; on high-dimensional datasets (50+ dimensions) the ECS-based local variant performs best, and per-class selection also reduces the number of retained dimensions in AUDIO and OTHER categories.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- In the extracted Table 2, the final ECP CW+L cell for UWave renders as '93.93±0.2412345678' because PDF extraction merged the cell with adjacent CDD axis digits; the value was read as 93.93±0.24.
- Figures 1, 2, 5 and 6 are graphical; only their inline numeric labels (e.g., CDD ranks in Fig. 4, axis values of Figs. 5-6) were legible in the extraction, so pictorial aspects were not assessed.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.