Skip to content
airyuanshenPublic

About

Welcome!GEnSHIN is a graph neural network for traffic forecasting that captures spatio-temporal dependencies using attention-enhanced GCRU units, asymmetric graph learning, and dynamic memory banks. It achieves strong performance on METR-LA, especially during peak hours.

Resources

Stars

5 stars

Watchers

0 watching

Forks

Latest commit

 

History

7 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 

Repository files navigation

GEnSHIN: Graph Enhanced Spatio-temporal Hierarchical Inference Network

📌 Overview

GEnSHIN (Graph Enhanced Spatio-temporal Hierarchical Inference Network) is a novel deep learning architecture designed for traffic flow forecasting in intelligent transportation systems. The model effectively captures complex spatiotemporal dependencies in traffic networks through innovative graph-based enhancements and hierarchical reasoning mechanisms.You can view our paper in https://arxiv.org/abs/2601.04550.

🎯 Key Features

  • Attention-Enhanced GCRU Units: Integrates Transformer modules with Graph Convolutional Recurrent Units to capture long-term temporal dependencies.
  • Asymmetric Dual-Embedding Graph Generation: Learns data-driven asymmetric topological structures while preserving real road network information.
  • Dynamic Memory Bank: Stores learnable traffic pattern prototypes and provides personalized representations for each sensor node.
  • Adaptive Graph Updater: Dynamically adjusts graph structures during decoding to adapt to changing network conditions.

🏗️ Model Architecture

GEnSHIN consists of three main components:

  1. Encoder: Processes historical traffic data using attention-enhanced GCRU units and Transformer modules.
  2. Memory Bank Module: Stores traffic pattern prototypes and retrieves relevant patterns through attention mechanisms.
  3. Decoder: Generates predictions using enhanced representations and dynamically updated graph structures.

📊 Performance

The model has been evaluated on the METR-LA dataset and demonstrates state-of-the-art performance in traffic flow prediction:

Metric Performance
MAE 3.60
RMSE 7.69
MAPE 9.06%

GEnSHIN shows particularly strong performance during peak traffic hours (morning and evening rush periods), maintaining stable and accurate predictions under challenging conditions.

🔬 Technical Highlights

  • Non-Symmetric Graph Learning: Captures directional traffic flow patterns that often exhibit asymmetry (e.g., inbound vs. outbound traffic during rush hours).
  • Pattern-Aware Representations: Each sensor node receives personalized traffic pattern embeddings based on its historical behavior.
  • Multi-Scale Spatiotemporal Modeling: Combines local spatial dependencies with global temporal patterns through hierarchical processing.

📈 Applications

GEnSHIN is suitable for various intelligent transportation applications:

  • Urban traffic flow prediction
  • Congestion forecasting
  • Route optimization
  • Traffic management systems
  • Smart city infrastructure planning

📚 Citation

If you use GEnSHIN in your research, please cite our work:

@article{zhou2026genshin,
  title={GEnSHIN: Graph Enhanced Spatio-temporal Hierarchical Inference Network for Traffic Flow Forecasting},
  author={Zhou, Zhiyan and Liao, Junjie and Zhang, Wenhao and Liao, Yingyi and Wang, Ziai},
  journal={arXiv preprint},
  year={2026}
}

🔗 Repository

The complete implementation of GEnSHIN is available in this repository. The code provides:

  • Model architecture implementation
  • Data preprocessing utilities
  • Evaluation metrics
  • Pre-trained model weights (where applicable)

📄 License

This project is released under the MIT License. See the LICENSE file for details.

🤝 Contributing

We welcome contributions from the research community. Please feel free to submit issues, feature requests, or pull requests to help improve GEnSHIN.

🙏 Acknowledgments

We thank the Beijing Normal University (Zhuhai) Supercomputing Center for providing computational resources. We also acknowledge all researchers whose previous work has contributed to the development of this model.


Note: This implementation is for research purposes. For production deployment in critical transportation systems, additional validation and safety considerations are recommended.

About

Welcome!GEnSHIN is a graph neural network for traffic forecasting that captures spatio-temporal dependencies using attention-enhanced GCRU units, asymmetric graph learning, and dynamic memory banks. It achieves strong performance on METR-LA, especially during peak hours.

Resources

Stars

5 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages