nhuang37/spectral-inspired-gnn

Code for Paper "From Local to Global: Spectral-Inspired Graph Neural Networks"

Spectral-Inspired Graph Neural Networks

This repo contains implementations of spectral-inspired GNNs, a simple layer-wise normalization technique to inject global information in message-passing GNNs, motivated from spectral graph embedding.

Dependencies

• Python 3.7+
• Pytorch 1.10+
• pytorch-geometric
• dgl-cu111

You can follow the code below to install pytorch-geometric

import os
import torch
os.environ['TORCH'] = torch.__version__
pip install -q torch-scatter -f https://data.pyg.org/whl/torch-${TORCH}.html
pip install -q torch-sparse -f https://data.pyg.org/whl/torch-${TORCH}.html
pip install -q torch-cluster -f https://data.pyg.org/whl/torch-${TORCH}.html
pip install -q git+https://github.com/pyg-team/pytorch_geometric.git

Follow the code below to install dgl

pip install dgl-cu111 -f https://data.dgl.ai/wheels/repo.html

PowerEmbed

Run PowerEmbed on the 10 benchmark graphs with the default implementation:

python power.py --datasets wiki --loop --data_path YOUR_DATA_PATH --result_path YOUR_RESULT_PATH

datasets include options with: wiki / planetoid / webkb / amazon / coauthor

Baseline

Run GCNII, GPR-GNN and JKNet on the 10 benchmark graphs with the default implementation:

python /baseline/gcn_variants_real_graph.py --datasets wiki --loop --model GCNII --data_path YOUR_DATA_PATH --result_path YOUR_RESULT_PATH

datasets include options with: wiki / planetoid / webkb / amazon / coauthor

models include options with: GCNII / GPRGNN / JKNet

Simulations and visualizations

• Convergence of PowerEmbed
  • Fig 2: power_convergence.ipynb
• SBM simulations
  • Fig 3, 5: /ablation study/sbm_boxplot.py
  • Fig 6: /ablation study/sbm_varying_density.py