EvoTSC: Evolving Feature Learning Models for Time Series Classification via Genetic Programming
New model evolves lightweight classifiers with 11 benchmark wins...
Researchers at Victoria University of Wellington have introduced EvoTSC, a novel genetic programming (GP) framework designed to automatically evolve feature learning models for time series classification. Published on arXiv (2604.25499), the method addresses two key pain points: the scarcity of labeled time series data and the high computational costs of traditional deep learning approaches. EvoTSC uses a carefully crafted multi-layer program structure that embeds diverse forms of prior expert knowledge directly into the evolutionary search process, steering it toward operations proven effective for time series analysis.
To combat overfitting, a common issue with small datasets, the team developed a tailored Pareto tournament selection strategy that favors models with consistent performance across varying training data subsets. Extensive experiments on univariate time series classification datasets demonstrate that EvoTSC significantly outperforms 11 benchmark methods in most comparisons. Further analysis confirms the contribution of each component and the resource efficiency of the evolved models, making EvoTSC a promising alternative for domains where labeled data is limited and computational budgets are tight.
- EvoTSC uses genetic programming to automatically evolve lightweight feature learning models, reducing reliance on labeled data and computational resources.
- Its multi-layer program structure embeds prior expert knowledge to guide evolution toward effective time series operations.
- A Pareto tournament selection strategy mitigates overfitting by favoring models consistent across data subsets, outperforming 11 benchmark methods on univariate datasets.
Why It Matters
EvoTSC offers efficient, generalizable time series classification for data-scarce domains like healthcare and finance.