BigST: Linear Complexity Spatio-Temporal Graph Neural Network for Traffic Forecasting on Large-Scale Road Networks
This is a refactored implementation of BigST model as described in the following paper: [BigST: Linear Complexity Spatio-Temporal Graph Neural Network for Traffic Forecasting on Large-Scale Road Networks, VLDB 2024].
python3, numpy, pandas, scipy, torch
Intel(R) Xeon(R) Gold 5220R CPU @ 2.20GHz NVIDIA A40 48GB
The California dataset is downloaded from the Caltrans Performance Measurement System (PeMS) website, we use Station 5-Minute traffic speed data ranging from 2022-04-01 to 2022-06-31. We provide the processed California dataset link for public use. The Beijing dataset is obtained from DiDi Chuxing, a commercial company. It cannot be made available to the public before we get a formal permission.
To preprocess the long historical traffic time series:
python preprocess/preprocess.py
You can use the following command for model training:
python run.py --use_residual --use_bn
If you want to use preprocessed features at training stage, please specify the "use_long" argument:
python run.py --use_residual --use_bn --use_long
If you find our work is useful for your research, please consider citing:
@article{hanbigst,
title={BigST: Linear Complexity Spatio-Temporal Graph Neural Network for Traffic Forecasting on Large-Scale Road Networks},
author={Han, Jindong and Zhang, Weijia and Liu, Hao and Tao, Tao and Tan, Naiqiang and Xiong, Hui},
journal={Proceedings of the VLDB Endowment},
volume={17},
number={5},
pages={1081--1090},
year={2024},
publisher={VLDB Endowment}
}
We thank the authors of the following repository for code reference: Graph WaveNet, BasicTS, Nodeformer, and Performer.