Spatial-Temporal Interactive Dynamic Graph Convolution Network for Traffic Forecasting

This is the original pytorch implementation of STIDGCN in the following paper ⚡️ : Spatial-Temporal Interactive Dynamic Graph Convolution Network for Traffic Forecasting, 2022.05. I hope these codes are helpful to you, 🌟!

Abstract

Accurate traffic forecasting is essential for urban traffic control, route planning, and flow detection. Although many spatial-temporal methods are currently proposed, they are still deficient in synchronously capturing the spatial-temporal dependence of traffic data. In addition, most methods ignore the hidden dynamic associations that arise between the road network nodes as it evolves over time. We propose a neural network-based Spatial-Temporal Interactive Dynamic Graph Convolutional Network (STIDGCN) to address the above challenges for traffic forecasting. Specifically, we propose an interactive dynamic graph convolution structure which divides the traffic data by intervals and synchronously captures the divided traffic data‘s spatial-temporal dependence through an interactive learning strategy. The interactive learning strategy motivates STIDGCN effective for long-range forecasting. We also propose a dynamic graph convolution module through a novel dynamic graph generation method to capture the dynamically changing spatial correlations in the traffic network. Based on a priori knowledge and input data, the dynamic graph generation method can generate a dynamic graph structure, which allows exploring the unseen node connections in the road network and simulating the dynamic associations between nodes over time. Extensive experiments on four real-world traffic flow datasets demonstrate that STIDGCN outperforms the state-of-the-art baselines.

Requirements

python
numpy
pandas
torch
matplotlib
scipy
argparse

Data Preparation

Download the dataset(PEMS03, PEMS04, PEMS07, PEMS08) from here, Baidu Drive, and the password is 1s5t. You can put them in the "data" folder. The data here is generated using generate_datasets.py and gen_adj_mx.py, you don't need to do any further processing. If you want to see the details of how the data is processed, check out generate_datasets.py and gen_adj_mx.py.

Train Commands

# PEMS03
nohup python -u train.py --data PEMS03 > PEMS03.log 2>&1 &

# PEMS04
nohup python -u train.py --data PEMS04 > PEMS04.log 2>&1 &

# PEMS07
nohup python -u train.py --data PEMS07 > PEMS07.log 2>&1 &

# PEMS08
nohup python -u train.py --data PEMS08 > PEMS08.log 2>&1 &

Results

Cite

If you make use of this code in your own work, please cite our paper:

@misc{liu2022spatialtemporal,
     title={Spatial-Temporal Interactive Dynamic Graph Convolution Network for Traffic Forecasting}, 
     author={Aoyu Liu and Yaying Zhang},
     year={2022},
     eprint={2205.08689},
     archivePrefix={arXiv},
     primaryClass={cs.LG}
}

Acknowledgments

Our model is built based on model of Graph WaveNet and SCINet.

P13LIAM / STIDGCN