jeongwhanchoi / stg-ncde

"Graph Neural Controlled Differential Equations for Traffic Forecasting", AAAI 2022

Graph Neural Controlled Differential Equations for Traffic Forecasting

Introduction

This is the repository of our accepted AAAI 2022 paper "Graph Neural Controlled Differential Equations for Traffic Forecasting". Paper is available on arxiv.

Citation

If you find this code useful, you may cite us as:

@inproceedings{choi2022STGNCDE,
  title={Graph Neural Controlled Differential Equations for Traffic Forecasting},
  author={Jeongwhan Choi AND Hwangyong Choi AND Jeehyun Hwang AND Noseong Park},
  booktitle={AAAI},
  year={2022}
}

Setup Python environment for STG-NCDE

Install python environment

$ conda env create -f environment.yml 

Reproducibility

Usage

In terminal

• Run the shell file (at the root of the project)

$ bash run.sh

• Run the python file (at the model folder)

$ cd model

$ python Run_cde.py --dataset='PEMSD4' --model='GCDE' --model_type='type1' --embed_dim=10 --hid_dim=64 --hid_hid_dim=64 --num_layers=2 --lr_init=0.001 --weight_decay=1e-3 --epochs=200 --tensorboard --comment="" --device=0

Star History

stg-ncde's Issues

Config file for other datasets

Hi jeongwhanchoi,

Thanks for your great work!
Can I have the config file you used for all other datasets in your paper? And can you share the dataset PeMSD7(M) and PeMSD7(L) ?
Thanks!

Problems considering irregular data

Hi, very great work!
Could you please show the code and detailed experimental setting related to irregular data experiments?
Thank you for consideration!

STG-NRDE

Hello, where is your code of STG-NRDE？

About irregular data

Hi, could you please show the code related to irregular data generation? Or, could you describe how you generate the irregular data, please?
Thank you.

Excuse me, can you provide the data set of PEMSD7(L) dataset?

I can't search this dataset online. Can you provide it?
Thank you.

Longer time-series input may performs bad?

Congratudate for your AAAI 2022 work. It's a nice work, but i met some problem when i use your model as a sota in my experiment. It performed very bad when i try to input 16, 24 or 144 steps in electricity and bikeNYC dataset. Even though i try to adjust some parameters like node_embedd dim, Chenb_k (1,2,3), hidden channel (16 may be the best), loss function (L2 and L1), it stilll does not make. I would like to inform you for some good solutions to solve it. I will appreciate it very much if there are some replies.

PEMS03 MAPE

Hello, why does MAPE have this problem on PEMS03?