SSTVM: Score-based Spatio-Temporal Variational Model

This repository provides the implementation of SSTVM, a modular framework for spatiotemporal forecasting. It integrates Variational Graph Autoencoders (VGAE) and Score-based Diffusion in latent space, enabling robust and efficient modeling of complex spatiotemporal dynamics under noisy or sparse data conditions.

---

📁 Supported Datasets

This project supports nine publicly available benchmark datasets:

• Los-loop
• PeMS08
• T-Drive
• Didi_chengdu
• AIR-BJ
• ECG5000
• Electricity
• Solar
• Exchange-Rate

Please create a data/ folder under each dataset directory and place the corresponding dataset files inside it.

---

🔧 Environment Requirements

For graph_generation:

• tensorflow
• numpy, scipy, matplotlib, pandas, seaborn, sklearn
• argparse, configparser, time, math

For prediction:

• torch
• tensorflow
• numpy, matplotlib, pandas, sklearn
• argparse, csv, time, shutil

---

🚀 Run the Demo

Go to the corresponding dataset folder and run:

python main.py

---

🙏 Acknowledgment

We sincerely thank the original authors of DVGNN for their excellent open-source contribution, which served as the foundation of this repository.

We build upon their framework and extend it into the proposed SSTVM model, including new latent-space diffusion mechanisms, enhanced robustness, and broader dataset support.