G3N: Graph Neighbourhood Neural Networks

Requirements

As listed in requirements.txt:

networkx==2.7.1
numpy==1.21.5
ogb==1.3.3
scikit_learn==1.1.0
scipy==1.7.3
torch==1.11.0
torch_geometric==2.0.4
torch_scatter==2.0.9
tqdm==4.64.0

Due to an issue associated with importing the ogb package, you may need to run in your environment pip uninstall setuptools

Synthetic datasets

For isomorphism tasks on EXP, SR25, graph8c and CSL, run

python3 iso.py --dataset exp --t <t> --d <d>
python3 iso.py --dataset sr25 --t <t> --d <d>
python3 iso.py --dataset graph8c --t <t> --d <d>
python3 iso.py --dataset csl --t <t> --d <d>

where d and t are parameters of G3N denoting neighbourhood size and neighbourhood subgraph dimension.

For substructure counting tasks run

python3 counting.py --ntask <n_task> --t <t> --d <d>

where n_task selects what substructure you want to count: 0: triangle, 1: tailed_triangle; 2: star; 3: 4-cycle.

Real world datasets

For TU datasets, you may look into tu.py or run the grid search by

python3 grid_tu.py --t 2 --d 1

For graph classification on MolHIV, run

python3 mol.py --t 3 --d 3

For graph regression on ZINC, run

python3 zinc.py --t 3 --d 3

Note that the default parameters in the .py files may not be the optimal hyperparameter configurations. Please refer to the paper or supplementary material for more information on hyperparameter selection.