MegaCRN

MegaCRN: Meta-Graph Convolutional Recurrent Network

[AAAI23] R. Jiang*, Z. Wang*, J. Yong, P. Jeph, Q. Chen, Y. Kobayashi, X. Song, S. Fukushima, T. Suzumura, "Spatio-Temporal Meta-Graph Learning for Traffic Forecasting", Proc. of Thirty-Seventh AAAI Conference on Artificial Intelligence (AAAI), 2023. (Accepted to Appear)

[AI23] R. Jiang*, Z. Wang*, J. Yong, P. Jeph, Q. Chen, Y. Kobayashi, X. Song, T. Suzumura, S. Fukushima, "MegaCRN: Meta-Graph Convolutional Recurrent Network for Spatio-Temporal Modeling", Artificial intelligence, 2023. (Extended Journal Version under Review)

Code and data are now available (more data will come). Please kindly cite the following bibtex.

@article{jiang2022spatio,
  title={Spatio-Temporal Meta-Graph Learning for Traffic Forecasting},
  author={Jiang, Renhe and Wang, Zhaonan and Yong, Jiawei and Jeph, Puneet and Chen, Quanjun and Kobayashi, Yasumasa and Song, Xuan and Fukushima, Shintaro and Suzumura, Toyotaro},
  journal={arXiv preprint arXiv:2211.14701},
  year={2022}
}
@article{jiang2022megacrn,
  title={MegaCRN: Meta-Graph Convolutional Recurrent Network for Spatio-Temporal Modeling},
  author={Jiang, Renhe and Wang, Zhaonan and Yong, Jiawei and Jeph, Puneet and Chen, Quanjun and Kobayashi, Yasumasa and Song, Xuan and Suzumura, Toyotaro and Fukushima, Shintaro},
  journal={arXiv preprint arXiv:2212.05989},
  year={2022}
}

Preprints

Performance on Traffic Speed Benchmarks

Performance on Traffic Flow Benmarks

Requirements

• Python 3.8.8 -> Anaconda Distribution
• pytorch 1.9.1 -> py3.8_cuda11.1_cudnn8.0.5_0
• pandas 1.2.4
• numpy 1.20.1
• torch-summary 1.4.5 -> pip install torch-summary https://pypi.org/project/torch-summary/ (must necessary)
• jpholiday -> pip install jpholiday (not must, but if you want onehottime)

General Description

• The directory is structured in a flat style and only with two levels.
• The datasets are stored in DATA directories, and the model codes are put in model_DATA directories.
• The training and testing function is merged into one file, we can just run "python traintest_MegaCRN.py" under each model directory.
• Also we can run "python MegaCRN.py" to simply check the model architecture without feeding the data.
• Also under model directory, metrics.py file contains the metric functions and utils.py file contains a set of supporting functions.

How to run our model (general command)?

• cd model
• python traintest_MegaCRN.py --dataset=DATA --gpu=GPU_DEVICE_ID
• DATA = {METRLA, PEMSBAY}
• For PEMSBAY dataset, please first upzip ./PEMSBAY/pems-bay.zip to get ./PEMSBAY/pems-bay.h5 file.

How to run our model on PEMS03,04,07,08?

• cd model_PEMS3478
• python traintest_MegaCRN.py --dataset=DATA --gpu=GPU_DEVICE_ID
• DATA = {PEMS03, PEMS04, PEMS07, PEMS08}

How to run our model on EXPY-TKY?

• cd model_EXPYTKY
• python traintest_MegaCRN.py --dataset=DATA --gpu=GPU_DEVICE_ID
• DATA = {EXPYTKY, EXPYTKY*}
• EXPYTKY with 1843 links is the data used in our paper; EXPYTKY* is a superset of EXPYTKY that contains all 2841 expy-tky links.

Arguments

The default hyperparameters used in our paper are written in model/traintest_MegaCRN.py as follows.

• argument('--dataset', type=str, choices=['METRLA', 'PEMSBAY'], default='METRLA', help='which dataset to run')
• argument('--trainval_ratio', type=float, default=0.8, help='the ratio of training and validation data among the total')
• argument('--val_ratio', type=float, default=0.125, help='the ratio of validation data among the trainval ratio')
• argument('--seq_len', type=int, default=12, help='sequence length of prediction')
• argument('--his_len', type=int, default=12, help='sequence length of historical observation')
• argument('--channelin', type=int, default=1, help='number of input channel')
• argument('--channelout', type=int, default=1, help='number of output channel')
• argument("--loss", type=str, default='MaskMAE', help="MAE, MSE, MaskMAE")
• argument("--epoch", type=int, default=200, help="number of epochs of training")
• argument("--batch_size", type=int, default=64, help="size of the batches")
• argument("--lr", type=float, default=0.001, help="adam: learning rate")
• argument("--patience", type=float, default=10, help="patience used for early stop")
• argument('--num_layers', type=int, default=1, help='number of layers')
• argument('--hiddenunits', type=int, default=32, help='number of hidden units')
• argument('--mem_num', type=int, default=10, help='number of meta-nodes/prototypes')
• argument('--mem_dim', type=int, default=32, help='dimension of meta-nodes/prototypes')
• argument("--memory", type=eval, choices=[True, False], default='True', help="whether to use memory: True or False")
• argument("--meta", type=eval, choices=[True, False], default='True', help="whether to use meta-graph: True or False")
• argument("--decoder", type=str, choices=['sequence', 'stepwise'], default='stepwise', help="decoder type: sequence or stepwise")
• argument('--lamb', type=float, default=0.01, help='lamb value for separate loss')
• argument('--lamb1', type=float, default=0.01, help='lamb1 value for compact loss')
• argument('--gpu', type=int, default=0, help='which gpu to use')
• argument('--seed', type=int, default=100, help='random seed.')

Arguments (PEMS03,04,07,08)

The hyperparameters for PEMS03,04,07,08 in model_PEMS3478/traintest_MegaCRN.py are the same as the above except the following. </br> Differing from train:valid:test=7:1:2 for METR-LA and PEMS-BAY, the setting for PEMS03,04,07,08 is train:valid:test=6:2:2.

• argument('--dataset', type=str, choices=['PEMS03', 'PEMS04', 'PEMS07', 'PEMS08'], default='PEMS03', help='which dataset to run')
• argument('--val_ratio', type=float, default=0.25, help='the ratio of validation data among the trainval ratio') # this makes 6:2:2.
• argument("--loss", type=str, default='MAE', help="MAE, MSE, MaskMAE")
• argument("--patience", type=float, default=200, help="patience used for early stop")

Arguments (EXPY-TKY)

The hyperparameters for EXPY-TKY in model_EXPYTKY/raintest_MegaCRN.py are the same as the default except the following. </br> EXPY-TKY data is structured by month, where '202110' and '202111' used as training and validation and '202112' used as testing. </br> By further setting val_ratio as 0.25 (that meas 25% data of '202110' and '202111' as valid data), the ratio for train:valid:test is roughly 3:1:2. </br> The time interval for EXPY-TKY is 10 minutes, thus observation/prediction horizon are both set to 6, to perform 1-hour-to-1-hour forecasting. </br>

• argument('--dataset', type=str, choices=['EXPYTKY', 'EXPYTKY*'], default='EXPYTKY', help='which dataset to run')
• argument('--month', type=str, default='202112', help='which experiment setting (month) to run as testing data')
• argument('--val_ratio', type=float, default=0.25, help='the ratio of validation data among the trainval ratio')
• argument('--seq_len', type=int, default=6, help='sequence length of prediction')
• argument('--his_len', type=int, default=6, help='sequence length of historical observation')
• argument("--loss", type=str, default='MAE', help="MAE, MSE, MaskMAE")
• argument('--seed', type=int, default=1234, help='random seed')