Source-linked AI summary
FETERS: Few-Shot Early Time-Series Classification via Effective Ratio Selection
Chen-An Tai, Yujia Wu, Vincent S. Tseng
TL;DR
Few-shot ETSC remains insufficiently studied despite limited labeled data, so FETERS selects a dataset-level stopping ratio with support-set LOO evaluation and combines Rocket and frozen Chronos features. Across 69 datasets, it achieves the best average HM and HM rank in 5-shot settings while remaining competitive in full-shot settings.
Problem
Few-shot ETSC lacks systematic study, while limited supervision complicates stopping decisions and effective representation extraction from partial prefixes.
Method
FETERS selects a dataset-level stopping ratio via support-set class-wise LOO evaluation and combines Rocket features with frozen Chronos representations without an additional learned stopping module.
Results
FETERS achieves the best average HM in both settings, reaching 0.679 in 5-shot, and the best HM on 38 of 69 datasets in 5-shot.
Takeaways & Limitations
The results support dataset-specific ratio selection and show FETERS manages the accuracy–earliness trade-off under limited labeled data.
Takeaways & Limitations
Learning high-capacity representations from small support sets may remain challenging when classification uses partially observed prefixes.
Abstract
from arXiv · showhide
Early time-series classification (ETSC) aims to make accurate predictions from partially observed time series as early as possible. Although various stopping mechanisms and feature learning strategies have been developed for ETSC, most existing methods assume access to sufficient labeled training data, which may be unrealistic in applications with limited annotation. Under limited supervision, learning an additional sample-level stopping module and extracting effective classification features can both become challenging. In this paper, we propose FETERS, a few-shot ETSC framework that selects a dataset-level stopping ratio through class-wise leave-one-out (LOO) evaluation on the support set and uses a penalty-based reward function to manage the accuracy-earliness trade-off, thereby avoiding the need to train an additional stopping module. FETERS further combines Rocket-based features with frozen Chronos representations for classification. Extensive experiments on 69 public datasets spanning 14 domains show that FETERS achieves state-of-the-art (SOTA) performance in the 5-shot setting, with the highest average harmonic mean (HM) and the best HM on 38 datasets, while outperforming the current SOTA method on 44 datasets. FETERS also remains competitive in the full-shot setting, demonstrating its effectiveness in managing the accuracy-earliness trade-off.
Introduction
FETERS addresses few-shot early time-series classification, where limited labels complicate stopping decisions and prefix representation learning. It selects a dataset-level observation ratio through support-set evaluation and combines Rocket features with frozen Chronos representations.
- Motivation: ETSC predicts from partial time-series observations as early as possible while preserving classification accuracy, creating an accuracy–earliness trade-off.Existing approaches commonly use sample-level halting policies, confidence criteria, or future-classification estimates.
- Motivation: Few-shot ETSC is motivated by time-sensitive domains where labels may be limited by costly annotation, specialized data collection, or rare target events.Examples include healthcare, industrial monitoring, and human activity recognition.
- Challenges: Limited supervision makes sample-level stopping mechanisms complex and sensitive to sampled examples, while cross-domain temporal variation complicates effective prefix representation learning.These challenges arise when adapting ETSC to per-class few-shot supervision.
- FETERS: FETERS proposes dataset-level ratio selection through support-set class-wise leave-one-out evaluation, avoiding an additional learned stopping module while adapting to dataset-specific temporal characteristics.A penalty parameter controls the preferred accuracy–earliness trade-off.
- FETERS: FETERS combines dataset-fitted Rocket features with frozen Chronos representations for prefix classification.The framework uses both feature types to support classification from partially observed prefixes.
Related Work
Prior ETSC work uses several stopping paradigms and feature-extraction strategies, but these methods were primarily developed for full-shot settings. Time-series foundation models offer an additional source of representations learned from large external corpora.
- Stopping mechanisms: ETSC stopping mechanisms include confidence-based, agent-based, and anticipation-based methods.Confidence-based methods use current prediction confidence, while agent-based methods learn whether to halt or continue as observations arrive.
- Feature extraction: ETSC feature extractors include non-deep approaches such as WEASEL and MiniRocket-based pipelines.WEASEL encodes subsequences as discrete words using Fourier-based representations, whereas MiniRocket extracts convolution features with largely fixed kernels, dilations, and biases.
- Few-shot limitation: Existing ETSC approaches were primarily developed in full-shot settings, making high-capacity representation learning from small support sets particularly challenging for partial prefixes.The challenge is especially pronounced when classification must be performed from partially observed prefixes.
- Foundation models: Time-series foundation models provide representations learned from large external corpora, with Chronos pretrained for probabilistic forecasting and demonstrating zero-shot generalization to unseen forecasting datasets.Recent studies have also explored time-series foundation models for time-series classification.
Proposed Method
FETERS is a few-shot ETSC framework with two components—dataset-level ratio selection and hybrid feature extraction—that jointly manage accuracy and earliness. It selects an observation ratio through class-wise leave-one-out evaluation and combines Rocket-based with Chronos-based prefix representations.
- Framework overview: FETERS comprises ratio selection and hybrid feature extraction for managing the accuracy–earliness trade-off in few-shot ETSC.The ratio-selection and feature-extraction components jointly form the framework.
- Dataset-level ratio selection: FETERS evaluates candidate observation ratios with class-wise leave-one-out folds on the support set instead of learning an additional sample-level stopping module.This avoids the added learning complexity and potential sensitivity to sampled support examples associated with sample-level stopping.
- Dataset-level ratio selection: The selected ratio maximizes Reward(r) = Acc(r)(1 − 0.95r^p), combining support-fold accuracy with a penalty for later decisions.Here, Acc(r) is averaged across support-set folds, p controls the penalty strength, and 0.95 keeps the penalty factor positive at r = 1.0.
- Hybrid feature extraction: For each observed prefix, FETERS extracts Rocket-based and Chronos-based features because prefixes at different ratios may have distinct temporal characteristics.Rocket provides efficient convolution-based representations, while Chronos supplies additional temporal representations from a time-series foundation model.
- Hybrid feature extraction: MiniRocket is used for datasets with fewer than five classes, whereas MultiRocket is used otherwise under a fixed class-count rule.The extractor is fitted on truncated support samples and then applied to support and test prefixes.
Experiments
FETERS is evaluated on 69 datasets across 14 domains under consistent few-shot and full-shot protocols. It achieves the strongest reported average accuracy and harmonic mean, while analyses support its stopping-ratio selection, hybrid representation, and penalty-based reward.
- Datasets: 69 datasets spanning 14 UCR-defined domains are used, including 45 classical ETSC benchmarks and 24 additional datasets.The benchmarks cover 14 domains in total, with the classical datasets spanning six domains and the additional datasets spanning eight.
- Evaluation protocol: FETERS is compared with four representative ETSC methods using original implementations and recommended configurations.In the 5-shot setting, all methods use fixed configurations, identical random seeds, and 100 runs averaged over seeds 40–139.
- Main results: 62.8% and 72.5% are FETERS’s highest average accuracies in the 5-shot and full-shot settings, with 26 and 22 dataset-level accuracy wins.FETERS also achieves the highest average HM: 0.679 in 5-shot and 0.749 in full-shot, with best HM on 38 and 30 of 69 datasets, respectively.
- Stopping behavior: FETERS analyzes whether dataset-level stopping ratios remain consistent across supervision settings while preserving an effective accuracy–earliness trade-off.The stopping analysis compares 5-shot and full-shot behavior and examines whether selected ratios remain reasonably concentrated.
- Feature ablation: Removing either frozen Chronos or Rocket features yields more losses than wins in accuracy and HM, while removing Rocket underperforms the complete model on 58 of 69 datasets in 5-shot.The results support combining complementary temporal information from frozen Chronos representations and dataset-fitted Rocket features.
- Penalty and ratio analysis: Increasing p generally permits later stopping and higher accuracy, whereas removing the penalty causes substantially later predictions and a large HM decrease.FETERS also achieves the highest average HM against globally fixed ratios, including r = 0.20 with similar average earliness, supporting dataset-specific ratio selection.
Conclusion
FETERS is a few-shot ETSC framework that selects a dataset-level stopping ratio through support-set class-wise LOO evaluation and combines Rocket features with frozen Chronos representations. Across 69 datasets in 14 domains, it achieves the best average HM and HM rank in the 5-shot setting while remaining competitive under full-shot supervision.
- Framework: FETERS selects a dataset-level stopping ratio using support-set class-wise LOO evaluation and performs prefix classification with dataset-fitted Rocket features and frozen Chronos representations.This design integrates stopping-ratio selection with complementary learned and pretrained representations.
- Evaluation: 69 datasets across 14 domains were used to evaluate FETERS in few-shot and full-shot settings.The experiments assess the framework across a broad, multi-domain benchmark.
- Results: In the 5-shot setting, FETERS achieves the best average HM and HM rank among the evaluated methods.Both aggregate performance measures favor FETERS under limited supervision.
- Results: FETERS remains competitive under full-shot supervision, indicating effectiveness across both limited- and sufficient-label regimes.The conclusion reports competitiveness rather than a best-performing result for the full-shot setting.