Learning Multiaspect Traffic Couplings by Multirelational Graph Attention Networks for Traffic Prediction

Data

1. Download from:
2. Unzip and move to ./data

Usage

1. Train

   • Docker container (recommended)

     # PEMSD3
     docker run -it --rm --gpus=all --shm-size=512m -v /path/to/checkpoints:/ms-gat/checkpoints luokn/ms-gat -d pemsd3 -i 1,2,3,24 -w 8
     # PEMSD4
     docker run -it --rm --gpus=all --shm-size=512m -v /path/to/checkpoints:/ms-gat/checkpoints luokn/ms-gat -d pemsd4 -w 8
     # PEMSD7
     docker run -it --rm --gpus=all --shm-size=512m -v /path/to/checkpoints:/ms-gat/checkpoints luokn/ms-gat -d pemsd7 -b 32 -w 8
     # PEMSD8
     docker run -it --rm --gpus=all --shm-size=512m -v /path/to/checkpoints:/ms-gat/checkpoints luokn/ms-gat -d pemsd8 -w 8

   • Physical machine:

     # PEMSD3
     python3 src/main.py -d pemsd3 -o checkpoints/pemsd3 -i 1,2,3,24 -w 8
     # PEMSD4
     python3 src/main.py -d pemsd4 -o checkpoints/pemsd4 -w 8
     # PEMSD7
     python3 src/main.py -d pemsd7 -o checkpoints/pemsd7 -b 32 -w 8
     # PEMSD8
     python3 src/main.py -d pemsd8 -o checkpoints/pemsd8 -w 8

2. Evaluate

   python3 src/main.py --eval -d pemsd4 -o checkpoints/pemsd4 -c checkpoints/pemsd4/xx_xxx.xx.pkl

Checkpoints

• PEMSD3: MAE = 15.60 MAPE = 16.36% RMSE = 26.36
• PEMSD4: MAE = 19.59 MAPE = 13.34% RMSE = 31.58
• PEMSD7: MAE = 20.44 MAPE = 8.85% RMSE = 34.11
• PEMSD8: MAE = 14.58 MAPE = 10.10% RMSE = 23.94

Citation

    @ARTICLE{9780244,
        author    ={Huang, Jing and Luo, Kun and Cao, Longbing and Wen, Yuanqiao and Zhong, Shuyuan},
        journal   ={IEEE Transactions on Intelligent Transportation Systems},
        title     ={Learning Multiaspect Traffic Couplings by Multirelational Graph Attention Networks for Traffic Prediction},
        year      ={2022},
        volume    ={},
        number    ={},
        pages     ={1-15},
        doi       ={10.1109/TITS.2022.3173689}
    }