TMRC: Temporal Multi-features Representation Learning-Based Clustering for Time-Series Data

A novel deep learning approach for time series clustering that combines Variational Mode Decomposition (VMD) with LSTM Autoencoders for temporal multi-features representation learning and enhanced clustering performance.

📄 Paper: IEEE Xplore - Temporal Multi-features Representation Learning-Based Clustering for Time-Series Data

🏗️ TMRC Model Architecture

The TMRC model consists of three main components:

1. VMD Decomposition: Breaks down the input time series into K temporal features (intrinsic mode functions)
2. LSTM Autoencoders: Multiple autoencoders learn representations from each temporal feature independently
3. Multi-features Fusion & Clustering: Concatenates learned representations and performs TimeSeriesKMeans clustering

The TMRC approach enables learning of complementary temporal features at different frequency scales, providing richer representations for more accurate time series clustering.

🌟 Features

• Temporal multi-features decomposition: TMRC decomposes time series into multiple temporal features using VMD
• Deep representation learning: Employs LSTM Autoencoders to learn meaningful representations from each temporal feature
• Multi-features fusion: Combines learned representations from all temporal features for comprehensive clustering
• Advanced time series clustering: Performs clustering using TimeSeriesKMeans on fused multi-features representations
• Comprehensive evaluation: Uses Rand Index and Normalized Mutual Information metrics
• Automated grid search: Tests multiple hyperparameter combinations automatically
• UCR/UEA dataset support: Compatible with standard time series benchmarks

🚀 Quick Start

For Windows Users

1. Run all 36 datasets (full grid search):

   run_experiments.bat

2. Quick test (single run per dataset):

   run_quick_test.bat

3. Single dataset experiment:

   run_single_dataset.bat ECG200

Manual Python Execution

Basic usage:

python main.py --dataset ECG200 --device cuda

Grid search:

python main.py --dataset ECG200 \
                --K_list 2 3 4 5 6 \
                --hidden_size_list 10 50 100 400 800 \
                --num_epochs 1001 \
                --output_dir results

📊 Supported Datasets

The scripts are configured to run on these 36 UCR datasets:

ArrowHead, Beef, BeetleFly, BirdChicken, Car, ChlorineConcentration, 
Coffee, DiatomsizeReduction, Dist.phal.outl.agegroup, 
Dist.phal.outl.correct, ECG200, ECGFiveDays, GunPoint, Ham, 
Herring, Lighting2, Meat, Mid.phal.outl.agegroup, 
Mid.phal.outl.correct, Mid.phal.TW, MoteStrain, OSULeaf, 
Plane, Prox.phal.outl.ageGroup, Prox.phal.TW, SonyAIBORobotSurface, 
SonyAIBORobotSurfaceII, SwedishLeaf, Symbols, ToeSegmentation1, 
ToeSegmentation2, TwoLeadECG, TwoPatterns, Wafer, Wine, WordsSynonyms

⚙️ Command Line Arguments

Data Arguments

• --dataset: Dataset name from UCR/UEA archive (default: 'GunPoint')
• --data_dir: Path to dataset directory (default: 'datasets/UCR')
• --dataset_list: CSV file containing list of datasets to run

Model Hyperparameters

• --K_list: List of VMD decomposition modes (default: [2,3,4,5,6])
• --hidden_size_list: List of LSTM hidden sizes (default: [10,50,100,400,800,1200,1600,2000])
• --num_epochs: Number of training epochs (default: 1001)
• --lr: Learning rate (default: 0.0002)
• --num_layers: Number of LSTM layers (default: 1)

VMD Parameters

• --vmd_alpha: VMD bandwidth constraint (default: 2000)
• --vmd_tau: VMD noise-tolerance (default: 0.0)
• --vmd_tol: VMD convergence tolerance (default: 1e-7)

Experiment Settings

• --output_dir: Directory to save results (default: 'results')
• --device: Device to use ('cuda' or 'cpu')
• --seed: Random seed for reproducibility (default: 42)
• --verbose: Print detailed progress
• --single_run: Run single experiment instead of grid search

📝 Citation

If you use TMRC in your research, please cite our paper:

@article{lee2024temporal,
  title={Temporal multi-features representation learning-based clustering for time-series data},
  author={Lee, Jaehoon and Kim, Dohee and Sim, Sunghyun},
  journal={IEEE Access},
  volume={12},
  pages={87675--87690},
  year={2024},
  publisher={IEEE},
  doi={10.1109/ACCESS.2024.3417888},
  url={https://ieeexplore.ieee.org/abstract/document/10565866}
}

📜 License

This project is licensed under the MIT License - see the LICENSE file for details.

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

📧 Contact

For questions or issues, please open an issue on GitHub or contact [dynamic97312@naver.com].

---

Keywords: Time Series Clustering, Deep Learning, LSTM Autoencoders, Variational Mode Decomposition, Temporal Features, Multi-scale Analysis