Boosting Fine-Grained Urban Flow Inference via Lightweight Architecture and Focalized Optimization

Overview

PLGF (Progressive Local-Global Fusion Network) is a deep learning model designed for fine-grained urban flow inference. The model combines local and global feature fusion mechanisms with progressive upsampling and feature modulation to achieve high-precision flow density map prediction. Also, it adopts a focalized optimization strategy to address the imbalanced distribution of flow density values, improving the model's generalization ability.

Project Structure

.
├── data/                           # Data storage directory
│   └── TaxiBJ/                     # TaxiBJ dataset directory
├── util/                           # Utility modules and model components
│   ├── __init__.py                 # Package initialization file
│   ├── PLGF.py                     # PLGF model core implementation
│   ├── dataset.py                  # Dataset loading and building 
│   ├── metric.py                   # Evaluation metrics (MSE, MAE, MAPE)
│   └── focalloss.py                # Focalized Loss implementation
├── main.py                         # Main training script
└── README.md                       # Project documentation

Requirements

This code is implemented in Python and based on the PyTorch framework. To ensure compatibility, please install the following dependencies:

Basic Environment

• Python: 3.8+
• PyTorch: 1.8.0+

Running

you can run the code by running the following command:

python main.py

📁 Dataset

TaxiBJ datasets can be obtained in baseline UrbanFM's repo.

📝 Citation

If you find our work useful in your research, please consider citing our paper:

@inproceedings{zhu2026boosting,
  title={Boosting Fine-Grained Urban Flow Inference via Lightweight Architecture and Focalized Optimization},
  author={Zhu, Yuanshao and Zhao, Xiangyu and Zhang, Zijian and Wei, Xuetao and Yu, James Jianqiao},
  booktitle={Proceedings of the AAAI conference on artificial intelligence},
  volume={40},
  year={2026}
}