Federated Dynamic Model for Spatiotemporal Data Forecasting in Transportation

This repository contains the implementation of the Federated-LSTM-DSTGCRN model, proposed in the manuscript:

"Federated Dynamic Modeling and Learning for Spatiotemporal Data Forecasting", Thien Pham, Angelo Furno, Faïcel Chamroukhi, Latifa Oukhellou, arXiv:2503.04528. March 2025

Overview

This project brings Federated Learning to spatiotemporal forecasting, making it more accurate, and privacy-friendly. To keep things robust and efficient, we introduce a Client-Side Validation mechanism, in which, clients check updates before applying them, so only the best improvements make it into the model. This means better accuracy, faster convergence, and no junk updates.

This model is built for real-world spatiotemporal forecasting tasks like multimodal transport demand and OD matrix prediction, all while maintaining data privacy and decentralization.

Due to privacy constraints, we are unable to share OD data in this repository. However, we do provide datasets for multimodal transport demand, including Chicago taxi, New York taxi, and New York bike data, as well as weather data for the corresponding locations.

Installation

1. Clone the Repository

git clone https://github.com/nhat-thien/Federated-LSTM-DSTGCRN
cd Federated-LSTM-DSTGCRN  

2. Install Dependencies

Ensure you have Python 3.8+ installed. Then, install the required dependencies:

pip install -r requirements.txt  

Usage

1. Prepare Your Data

• Place your dataset in the DATA/ folder.
• Ensure it follows the required format for training.

2. Configure Your Experiment

Modify the following files to set up your experiment:

• TestCase.py → Define the test case and dataset configurations.
• Hyperparameters.py → Set the base model and federated learning scheme.

3. Run the Model

Execute the training script:

python Experiments.py  

Repository Structure

📂 Federated-LSTM-DSTGCRN  
│── 📂 DATA/                    # Folder for datasets  
│── 📂 FL_HELPERS/              # Handle the federated learning
│── 📂 MODELS/                  # The models
│── Hyperparameters.py          # Configuration for base model & FL scheme  
│── TestCase.py                 # Define test cases  
│── Experiments.py              # Implement the experiment(s)  
│── requirements.txt            # Dependencies  
│── README.md                   # Project documentation  

Citing This Work

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

@misc{pham2025federateddynamicmodelinglearning,
      title={Federated Dynamic Modeling and Learning for Spatiotemporal Data Forecasting}, 
      author={Thien Pham and Angelo Furno and Faïcel Chamroukhi and Latifa Oukhellou},
      year={2025},
      eprint={2503.04528},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2503.04528}, 
}