ShapeCond: Fast Shapelet-Guided Dataset Condensation

Official PyTorch Implementation

 

---

✈️ Paper • 😊 Get Started • 🎈 Run • ✨ Results • 📜 Citation

Dataset condensation reduces large training sets into compact, informative subsets, but existing methods, mostly designed for images, fail to capture the temporal structures critical in time series data. We propose ShapeCond, an efficient framework that leverages shapelet-guided optimization to synthesize compact datasets while jointly preserving both local discriminative patterns and global temporal structures. ShapeCond is highly efficient—up to 29× faster than the prior state-of-the-art CondTSC, and up to 10,000× faster than naïve shapelet-based methods on long sequences (e.g., the Sleep dataset with 3,000 timesteps). By balancing local and global information, ShapeCond consistently outperforms all existing state-of-the-art time series dataset condensation methods.

Shapelet-guided optimization for Synthesizing Condensed Dataset.

git clone https://github.com/lunaaa95/ShapeCond.git
cd ShapeCond

Python version: 3.9.21

Install required packages:

pip install -r requirements.txt

All datasets should be placed in the ./data directory. Our experiments are conducted on seven time series classification datasets: HAR, Electric, Sleep, TwoPatterns, FacesUCR, Tiselac, and Pedestrian.

• HAR, Electric, and Sleep are downloaded by following the instructions provided in the TimeSeriesCond repository and are preprocessed using preprocessing.py.

• TwoPatterns, FacesUCR, and Tiselac are obtained via sktime.datasets. Run uea.py to automatically download and preprocess these datasets.

• Pedestrian is downloaded from the Monash Scalable Time Series Evaluation Repository and is preprocessed using preprocessing.py.

Run experiments using the following command:

./scripts/<framework>/<dataset>.sh

Examples:

./scripts/CNNBN/har.sh

./scripts/CNNBN/Pedestrian.sh

./scripts/Transformers/electric.sh

Fast shapelet discovery is automatically triggered the first time a dataset is processed, and the extracted shapelets are saved in the sc/ directory for reuse. This significantly reduces computational cost, enabling shapelet extraction on large-scale datasets. Pre-extracted shapelets are also provided in this repository for convenience.

@article{peng2026shapecond,
  title={ShapeCond: Fast Shapelet-Guided Dataset Condensation for Time Series Classification},
  author={Peng, Sijia and Xiong, Yun and Chen, Xi and Xie, Yi and Li, Guanzhi and Yu, Yanwei and Zhu, Yangyong and Shen, Zhiqiang},
  journal={arXiv preprint arXiv:2602.09008},
  year={2026}
}