STR-GODEs: Spatial–Temporal-Ridership Graph ODEs for Metro Ridership Prediction

This repository is the official implementation of STR-GODEs.

The setting is adapted from https://github.com/HCPLab-SYSU/PVCGN.

If our code is helpful to you, please cite our paper

@article{huang2021str,
  title={STR-GODEs: Spatial-Temporal-Ridership Graph ODEs for Metro Ridership Prediction},
  author={Huang, Chuyu},
  journal={arXiv preprint arXiv:2107.04980},
  year={2021}
}

Requirements

To install requirements:

• python3
• numpy
• yaml
• pytorch
• torch_geometric
• torchdiff

A anaconda environment file py36.yml is provided in folder "test". You can activate it by modifying the prefix in it

prefix: /home/XXXXXX/.conda/envs/py36

and run the command：

conda env create -f py36.yml

Note that CUDA version in $PATH may be different from 10.1 in py36.yml, you may need to install the corresponding version of the library file(torch-scatter, torch-cluster, torch-sparse) in the following link:

https://pytorch-geometric.com/whl/torch-1.6.0.html

pip install torch_cluster-latest+cu101-cp36-cp36m-linux_x86_64.whl

pip install torch_scatter-latest+cu101-cp36-cp36m-linux_x86_64.whl

pip install torch_sparse-latest+cu101-cp36-cp36m-linux_x86_64.whl

and reinstall torch-geometric：

pip install torch-geometric

Extract dataset

cd data && tar xvf data.tar.gz

Training

To train the model(s) in the paper, run this command:

• SHMetro

python str_godes_train.py --config data/config/sh.yaml

• HZMetro

python str_godes_train.py --config data/config/hz.yaml

By setting the "irregular" variable in the configuration file(.yaml) to "true", we can conduct irregular prediction experiments.

Evaluation

To evaluate my model, set the "save_path" variable in configuration file(.yaml) and run:

• SHMetro

python str_godes_evaluation.py --config data/config/sh.yaml

• HZMetro

python str_godes_evaluation.py --config data/config/hz.yaml

To evaluate my model in peak periods, run:

• SHMetro

python str_godes_evaluation_peak.py --config data/config/sh.yaml

• HZMetro

python str_godes_evaluation_peak.py --config data/config/hz.yaml

Test

The pretrained models are provided in the "test" folder:

• SHMetro

python str_godes_evaluation.py --config test/STR_GODEs_sh.yaml

• HZMetro

python str_godes_evaluation.py --config test/STR_GODEs_hz.yaml

Results

Our model achieves the following performance on :

Experiment1: conventional prediction experiment

Experiment2: conventional prediction experiment in peak periods(7:30-9:30 and 17:30-19:30)

Experiment3: irregular prediction experiment

Model STR-GODEs	metrics	15min	30min	45min	60min
	MAE	22.84	23.24	23.69	24.25
experiment1_HZMetro	RMSE	37.35	38.41	39.42	40.81
	MAPE	13.70%	13.87%	14.34%	15.37%
-----------------------	----------------	----------------	--------------	----------------	--------------
	MAE	23.21	23.63	24.65	25.56
experiment1_SHMetro	RMSE	44.58	46.28	49.93	53.39
	MAPE	16.99%	17.12%	17.58%	18.25%
-----------------------	----------------	----------------	--------------	----------------	--------------
	MAE	31.68	32.37	32.06	30.82
experiment2_HZMetro	RMSE	48.78	50.33	50.69	50.45
	MAPE	9.31%	9.18%	9.62%	10.20%
-----------------------	----------------	----------------	--------------	----------------	--------------
	MAE	35.17	35.81	36.21	35.44
experiment2_SHMetro	RMSE	61.93	64.34	67.16	66.99
	MAPE	13.03%	13.20%	13.78%	14.87%
-----------------------	----------------	----------------	--------------	----------------	--------------
	MAE	17.70	18.25	18.19	18.45
experiment3_HZMetro	RMSE	37.31	36.04	37.52	37.41
	MAPE	10.14%	10.75%	10.08%	10.74%
-----------------------	----------------	----------------	--------------	----------------	--------------
	MAE	16.11	15.69	14.96	15.66
experiment3_SHMetro	RMSE	38.66	37.36	36.33	39.72
	MAPE	11.94%	11.96%	11.19%	11.02%
-----------------------	----------------	----------------	--------------	----------------	--------------