This is the Pytorch code for reproducing the results of the Paper GraphMix (https://arxiv.org/pdf/1909.11715.pdf). GraphMix is a simple and efficient regularization method for Graph Neural Networks (GNN) based Semi-Supervised node classification. We show that with this regularizer, even the simpler GNN architecture such as GCN (Kipf et. al.) can achiever state-of-the-art results on benchmark graph datasets such as Cora/Citeseer/Pubmed.
This code is tested with Python 3.6.3 and requires following packages (see requirements.txt list of all the packages):
torch==1.1.0
numpy==1.16.3
pandas==0.24.1
Pillow==5.3.0
scikit-learn==0.21.2
scipy==1.2.1
seaborn==0.9.0
six==1.12.0
tqdm==4.32.2
For reproducing results of GraphMix(GCN) of Table1 in the paper, go to directory GraphMix/semisupervised/codes and run the following commands:
python run_cora.py
python run_citeseer.py
python run_pubmed.py
This codebase is based on the github repo : https://github.com/DeepGraphLearning/GMNN
To all the people using this codebase: let us know in the "issues" if you are having some difficulity in reproducing the results.