Introduction

This repository contains the source code for the publication in ICML 2023 titled Leveraging Label Non-Uniformity for Node Classification in Graph Neural Networks in PyTorch.

Scheme of the proposed model.

There are two separate modules (in the dashed blue boxes corresponding to the two algorithms. They can be applied either separately or jointly to the base model.

All models are trained for node classification task. The following datasets are included:

• Cora
• Citeseer
• Pubmed
• Texas
• Chameleon
• Wisconsin
• CS
• Photo

The following base models are includes:

• MLP
• GCN
• GAT

Setup

First, unzip the data.zip file for the datasets required during training.

We used Python 3.8.13.

The environment requirements are in the requirements.txt file and can be installed as follows:

conda create --name <envname> --file requirements.txt

Alternatively,

pip install -r pip-requirements.txt

Usage

We provide examples of training commands used to train WGNN for node classification.

• Cora dataset (Test acc: 83.19 +/- 0.53)

python wgnn.py --dataset=cora --model=gcn --n-hidden=16 --self-loop --early-stop --eta1=30 --eta2=40

To search for the best eta1 and eta2 hyperparameters for ${\color{blue} \text{Algo. 2}}$, we run all combinations using the --all-combination flag.

python wgnn.py --dataset=cora --model=gcn --n-hidden=16 --self-loop --early-stop --save --all-combination

Citation

If you find this code useful, please cite the following paper:

F. Ji, S. H. Lee, H. Meng, K. Zhao, J. Yang, and W. P. Tay, “Leveraging label non-uniformity for node classification in graph neural networks,” in Proc. International Conference on Machine Learning, Hawaii, USA, Jul. 2023.