Source-linked AI summary
Improving Multivariate Time Series Classification with Class-Wise Training and Model Aggregation
Mouhamadou Mansour Lo, Gildas Morvan, Mathieu Rossi, Fabrice Morganti, David Mercier
TL;DR
Multivariate time-series classifiers may combine dimensions that are informative for some classes but irrelevant for others. This paper selects dimensions independently for each class, trains dedicated class-wise models, and fuses their predictions; the class-wise variant achieves the best average rank and significantly outperforms original MiniRocket, with gains varying by problem category.
Problem
Random or global dimension handling may combine dimensions that are useful for some classes but irrelevant for others, potentially losing information in high-dimensional multivariate time-series classification.
Source: §1 · p. 2Method
The framework selects dimensions independently for each class, trains a dedicated classifier on each selected subset, and fuses the classifiers’ predictions.
Source: §1 · pp. 2–3, §5 · p. 13Results
The class-wise variant achieves the best average rank and significantly outperforms original MiniRocket, with gains depending on the problem category.
Source: §5 · pp. 13–14Takeaways & Limitations
Class-specific subsets explicitly indicate which dimensions distinguish each class, supporting interpretability, while the reported performance gains span heterogeneous datasets.
Source: §5 · pp. 13–14
Abstract
from arXiv · showhide
In this paper, we propose a class-wise dimension (channel) selection framework for Multivariate Time Series Classification (MTSC). Rather than applying a single global dimension selection process, the proposed approach independently identifies informative dimensions for each class. A dedicated learning process is subsequently performed for each class, followed by a fusion stage for final prediction. The objective is to improve the generation of discriminative feature representations while reducing the influence of noisy or non-informative dimensions. The proposed framework is evaluated using MiniRocket, a random kernel-based baseline method. Experimental results indicate that class-wise dimension selection improves the quality of extracted representations and can enhance classification performance, particularly in high-dimensional settings. These findings suggest that incorporating class-specific information into the training process represents a promising direction for MTSC, improving robustness through consistent gains across heterogeneous datasets, and interpretability through the explicit identification of class-relevant dimensions.
1 Introduction
The paper addresses the challenge of classifying multivariate time series when dimensions may be informative for some classes but irrelevant for others. It proposes selecting dimensions independently by class, training class-wise, and fusing predictions.
1 Introduction
- Many widely adopted time-series classification approaches began in the univariate setting, with multivariate extensions often relying on straightforward adaptations.The paper describes time series as ordered observations and the multivariate case as having at least two variables.Source: §1 · pp. 1–2
- Random convolution-kernel methods including ROCKET, MiniRocket, MultiRocket, and HYDRA offer a reported trade-off between classification performance and computational efficiency.The introduction contrasts these methods with natively multivariate deep learning and heterogeneous ensembles that typically have high computational cost.Source: §1 · p. 2
- Dimensions useful for some classes may be irrelevant for others, and combining them can potentially cause information loss.The paper proposes focusing on dimensions relevant to a specific class to increase the probability of generating informative, discriminative combinations.Source: §1 · p. 2
- The framework selects dimensions independently for each class, trains using a class-wise learning strategy, and fuses predictions.The paper compares the approach with a baseline method and analyzes its results.Source: §1 · pp. 2–3
2 Related Work
Related work covers random convolution-kernel methods and external dimension-reduction techniques for multivariate time-series classification. The paper highlights the trade-offs of random channel selection and reviews approaches that retain dimensions using class separability or predictive value.
2.1 Multivariate ROCKET-Based Methods
- ROCKET uses random convolution kernels and pooling; MiniRocket improves efficiency with a deterministic design, while MultiRocket and HYDRA enrich or combine feature representations.ROCKET summarizes activation maps with PPV and GMP; MultiRocket adds first-order differences and pooling operators, while HYDRA uses grouped kernels and histograms of maximal responses.Source: §2.1 · p. 3
- MiniRocket’s multivariate extension randomly selects 1–9 dimensions per kernel, dilation, and bias triplet, then sums their convolution outputs into an activation map.The random selection strategy preserves computational efficiency but may become suboptimal in high-dimensional settings, motivating dimension reduction to retain relevant dimensions.Source: §2.1 · p. 4
2.2 Recent Dimension Reduction Methods for MTSC
- ECS scores dimensions globally using distances between class prototypes, while ECP unions dimensions selected pairwise to preserve class-specific discriminative dimensions.Both methods assess prototype separability using measures such as Euclidean distance or Dynamic Time Warping; ECS uses an elbow-based selection criterion.Source: §2.2 · p. 4
- TSelect estimates each dimension’s predictive performance using statistical features and validation-set AUROC, then removes redundant or non-informative dimensions.Its two stages use logistic regression on a learning/validation split, followed by an irrelevant channel selector that retains complementary dimensions.Source: §2.2 · pp. 4–5
3 Proposed Method
The framework selects dimensions separately for each target class, trains a dedicated one-versus-rest classifier on each selected subset, and fuses their scores for prediction. It adapts ECS and ECP for class-wise selection and uses class-specific MiniRocket transformations whose cost depends on the selected dimensions.
3.1 Framework Overview
- Each target class receives its own selected dimension subset and one-versus-rest classifier; the resulting class scores are fused for the final multiclass prediction.The framework permits subsets to vary across classes rather than sharing a single globally selected subset.Source: §3.1 · p. 5, Figure 2 · p. 6
3.2 Class-Wise Dimension Selection
- Class-wise ECS ranks dimensions by aggregating prototype distances involving the target class, while class-wise ECP unions elbow-selected dimensions from pairs involving that class.Both adaptations condition selection on the target class; ECP preserves its pairwise selection principle.Source: §3.2 · p. 6
3.3 Class-Wise Learning and Score Fusion
- For each selected subset, the framework forms class-specific training data and trains a classifier for that target class.At inference, the target-class models process their associated dimensions and their decision scores are combined.Source: §3.1 · p. 5, §3.3 · pp. 7–8
3.4 Complexity Analysis
- Class-wise MiniRocket transformation cost depends on each class’s selected dimensions, with at most nine dimensions involved in each convolution.The selection-cost analysis includes prototype computation, pairwise class-distance computation, and dimension ranking with elbow detection.Source: §3.4 · p. 8
4 Results of the Experiments
Across the selected UEA datasets, class-wise variants with local dimension selection generally outperform their corresponding baselines. Results vary by dataset category and dimensionality, while MiniRocket was chosen as the baseline after comparisons found no significant differences among the evaluated ROCKET-family methods.
4.1 Experimental Settings
- Experiments used 25 equal-length UEA datasets with sequence lengths of at least 9, resampled 30 times; dataset dimensions ranged from 2 to 1345.The selected datasets span six application-domain categories.Source: §4.1 · pp. 8–9
4.2 Baseline Classifier Selection
- No statistically significant differences were found among ROCKET, MiniRocket, MultiRocket, HYDRA, and HYDRA + MultiRocket; MiniRocket was selected for its performance–efficiency trade-off.MiniRocket and both MultiRocket and HYDRA + MultiRocket had a 12/1/12 win/draw/loss ratio.Source: §4.2 · p. 9, Figure 3 · p. 9
4.3 Comparison between the Different Variants
- CW + L variants outperform their respective ECS and ECP baselines on most datasets and across nearly all dataset categories.Table 2 compares the variants; the reported category-level pattern is consistent with the broader result.Source: §4.3 · pp. 10, 12, Table 2 · p. 11
- ECS CW + L improves performance on EEG and AUDIO, while ECP CW + L surpasses original MiniRocket on OTHER and AUDIO.The EEG and OTHER categories particularly benefit from the class-wise strategy.Source: Table 3 · p. 12, §4.3 · p. 12
- CW without dimensionality reduction outperforms original MiniRocket for datasets with ≤4 classes and ≥3 dimensions, and is comparable elsewhere.Table 4 compares performance by number of dimensions and classes.Source: Table 4 · p. 12, §4.3 · p. 12
- ECS and ECP Base reduce dimensions substantially, with ECS selecting markedly fewer dimensions than ECP.Table 5 reports average selected-dimension counts by problem category.Source: §4.3 · p. 12, Table 5 · p. 13
5 Conclusion and Prospects
The conclusion presents class-wise dimension selection and learning with classifier fusion as a framework that improves MiniRocket performance across heterogeneous datasets. It also highlights class-specific dimensions as interpretable and identifies extensions to other methods as future work.
5 Conclusion and Prospects
- The class-wise variant achieves the best average rank and outperforms original MiniRocket, with gains depending on problem category.The conclusion reports consistent improvements across heterogeneous datasets and particular benefits for methods relying on random dimension selection.Source: §5 · pp. 13–14
- The framework selects relevant dimensions for each class, trains a dedicated classifier on each subset, and fuses classifiers for final prediction.Source: §5 · p. 13
- Class-specific subsets indicate which dimensions distinguish each class, such as sensors characterizing an activity in HAR problems.Source: §5 · p. 14
- Future work includes extending the framework to other ROCKET-based and time-series classification methods and adapting ChannelScorer for local dimension selection.Source: §5 · p. 14