This is the code of the paper Breaking the Expression Bottleneck of Graph Neural Networks.

• Requirements
• Run Experiments
  • ogbg-ppa
  • ogbg-code
  • ogbg-molhiv & ogbg-molpcba
  • QM9
  • TU
• License

The code is built upon the PyTorch Geometric, which has been tested running under Python 3.6.9.

The following packages need to be installed manually:

• torch==1.5.1
• torch-geometric==1.5.0

The following packages can be installed by running

python3 -m pip install -r requirements.txt

• ogb==1.2.1
• numpy
• easydict
• tensorboardX

You can run experiments on multiple datasets as follows.

To run experiments on ogbg-ppa, change directory to ogbg/ppa:

cd ogbg/ppa

You can set hyper-parameters in ogbg-ppa.json.

You can change CUDA_VISIBLE_DEVICES and output directory in run_script.sh.

Then, run the following script:

./run_script.sh

To run experiments on ogbg-code, change directory to ogbg/code:

cd ogbg/code

You can set hyper-parameters in ogbg-code.json.

You can change CUDA_VISIBLE_DEVICES and output directory in run_script.sh.

Then, run the following script:

./run_script.sh

To run experiments on ogbg-mol*, change directory to ogbg/mol:

cd ogbg/mol

You can set dataset name and hyper-parameters in ogbg-mol.json. Dataset name should be either ogbg-molhiv or ogbg-molpcba.

You can change CUDA_VISIBLE_DEVICES and output directory in run_script.sh.

Then, run the following script:

./run_script.sh

To run experiments on QM9, change directory to qm9:

cd qm9

You can set hyper-parameters in QM9.json.

You can change CUDA_VISIBLE_DEVICES and output directory in run_script.sh.

Then, run the following script:

./run_script.sh

To run experiments on TU, change directory to tu:

cd tu

There are several datesets in TU. You can set dataset name in run_script.sh and set hyper-parameters in configs/<dataset>.json.

You can change CUDA_VISIBLE_DEVICES and output directory in run_script.sh.

Then, run the following script:

./run_script.sh

@ARTICLE {yang2022breaking,
	author = {Yang, Mingqi and Wang, Renjian and Shen, Yanming and Qi, Heng and Yin, Baocai},
	journal = {IEEE Transactions on Knowledge & Data Engineering},
	title = {Breaking the Expression Bottleneck of Graph Neural Networks},
	year = {2022},
	doi = {10.1109/TKDE.2022.3168070},
	address = {Los Alamitos, CA, USA}
}

MIT License