Diversity-Driven Ensemble Learning for Time Series Classification

This repository provides PyTorch implementations of a diversity-driven ensemble learning framework for time series classification (TSC). The approach builds on the LITE architecture to create ensembles of neural networks whose feature representations are explicitly decorrelated to improve generalization and classification performance.

The framework has been evaluated on the UCR Time Series Archive datasets, demonstrating state-of-the-art accuracy with fewer ensemble members compared to traditional methods.

Features

• LITE Model: Lightweight Inception-based time series classifier.
• Decorrelated Training: Penalizes redundancy in learned features across ensemble members.
• Flexible Ensemble Sizes: Includes scripts to train ensembles of different sizes (2 to 5 models).
• Reproducible Experiments: Configurable via command-line arguments.

Repository Structure

.
├── base.py              # Base model training script
├── cotrain.py           # Decorrelated training with 1 reference model
├── cotrain_2.py         # Decorrelated training with 2 reference models
├── cotrain_3.py         # Decorrelated training with 3 reference models
├── cotrain_4.py         # Decorrelated training with 4 reference models
├── cotrain_5.py         # Decorrelated training with 4 reference models (variant)
├── lite.py              # LITE model definition
├── utils.py             # Data loading, preprocessing, and utilities

Installation

1. Clone this repository

   git clone https://github.com/YOUR_USERNAME/YOUR_REPO_NAME.git
   cd YOUR_REPO_NAME

2. Install dependencies

   These scripts require:

   • Python 3.8+
   • PyTorch
   • NumPy
   • Pandas
   • scikit-learn
   • Matplotlib

   You can install them via pip:

   pip install torch numpy pandas scikit-learn matplotlib

Usage

All training scripts support command-line arguments:

• --classifier: Currently supports only `LITE`
• --datasets: Name of the dataset (or multiple datasets) from the UCR archive
• --runs: Number of runs to perform
• --output-directory: Where to store results

Example to train a single LITE model (baseline):

python base.py --datasets ECGFiveDays --runs 1 --output-directory results/base/

Example to train a decorrelated ensemble with one reference model:

python cotrain.py --datasets ECGFiveDays --runs 1 --output-directory results/ensemble_1/

Example to train a decorrelated ensemble with two reference models:

python cotrain_2.py --datasets ECGFiveDays --runs 1 --output-directory results/ensemble_2/

...and so on up to `cotrain_5.py`.

Note: You must set up your UCR Archive dataset folder paths in `utils.py` (see `load_data()`).

Results

This framework was evaluated on 128 UCR datasets and achieved:

• Improved accuracy compared to standard LITE ensembles
• Higher feature diversity as measured by feature orthogonality metrics

See the paper for detailed quantitative and qualitative analysis:

Enhancing Time Series Classification with Diversity-Driven Neural Network Ensembles

Citation

If you use this code, please cite:

@inproceedings{abdullayev2025diversity,
  title={Enhancing Time Series Classification with Diversity-Driven Neural Network Ensembles},
  author={Abdullayev, Javidan and Devanne, Maxime and Meyer, Cyril and Ismail-Fawaz, Ali and Weber, Jonathan and Forestier, Germain},
  booktitle={Proceedings of the International Joint Conference on Neural Networks (IJCNN)},
  year={2025}
}

License

This repository is released under the MIT License.

Acknowledgments

This work was conducted at Université de Haute Alsace and supported by the IRIMAS laboratory.