[AAAI 2026] TSGDiff: Rethinking Synthetic Time Series Generation from a Pure Graph Perspective

📋 Table of Contents

• Installation
• Dataset
• Configuration
• Results
• Citation

🔧 Installation

See requirements.txt for details.

📊 Dataset

The model is designed for time series generation, and the data set used is shown in the datasets directory.

⚙️ Configuration

Parameter configuration

Argument	Type	Default	Description
--data_file	str	datasets/{filename}.csv	Path to input data file (ETTh.csv)
--output_dir	str	output	Directory to save output results (including visualizations and metrics)
--slide_win	int	48	Sliding window size for time series slicing
--slide_stride	int	1	Sliding stride for time series slicing
--hidden_dim	int	1600	Graph encoder hidden layer dimension
--embed_dim	int	1600	Graph encoder embedding dimension
--use_period_index	int	2	Index of period used for graph edge generation
--batch_size	int	128	Batch size for training
--epochs	int	500	Number of training epochs
--learning_rate	float	0.01	Learning rate for optimizer
--normalize_data	bool	True	Whether to normalize input data
--feature_range	tuple	(-1, 1)	Feature range for Min-Max normalization
--kl_weight	float	0.2	Weight for KL divergence loss in total loss calculation
--diffusion_weight	float	1.0	Weight for Diffusion model loss in total loss calculation
--num_timesteps	int	1000	Number of timesteps for Diffusion model
--beta_start	float	0.0001	Starting value of beta for Diffusion model noise schedule
--beta_end	float	0.02	Ending value of beta for Diffusion model noise schedule
--nblocks	int	3	Number of intermediate blocks in Diffusion Block
--nunits	int	64	Number of units per Diffusion Block
--k	int	2	Multiplier for extending edges with remaining periods (graph construction)
--ff_weight	float	sqrt(slide_win)/5	Weight for Fourier loss (scaled by sliding window size)

📊 Results

Results are automatically saved to:

• Original data: output/original_data.npy
• Generated data: output/generated_data.npy
• Visualizations: output/visualizations/
• Evaluation metrics: output/metrics.txt

📁 Project Structure

TSGDiff/
├── main.py                          # Main entry point for training/inference/evaluation
├── model.py                         # Core implementation of TSGDiff Model
├── config.py                        # Global hyperparameter & path configurations
├── train_utils.py                   # Training utilities
├── data_utils.py                    # Data loading & preprocessing for temporal-spatial graph
├── graph_metric.py                  # Topo-FID calculations
├── evaluation_utils.py              # Evaluation metrics
├── visualization_utils.py           # Visualization tools
├── requirements.txt                 # Python dependencies
├── config/
│   └── *.yaml                       # Scene/dataset-specific configuration files
└── datasets/
    └── *.csv                        # Temporal-spatial graph datasets

📝 Citation

If you use this code in your research, please cite:

@inproceedings{TSGDiff,
  title={TSGDiff: Rethinking Synthetic Time Series Generation from a Pure Graph Perspective},
  author={Shen, Lifeng and Li, Xuyang and Long, Lele},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2026}
}

📄 License

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

📧 Contact

For questions or issues, please contact:

• Xuyang Li : lixuyang.lee@foxmail.com