Snowball and Truncated Krylov Graph Convolutional Networks

PyTorch and TensorFlow2 implementation of Snowball and Truncated Krylov Graph Convolutional Network (GCN) architectures for semi-supervised classification [1].

This repository contains the Cora, CiteSeer and PubMed dataset.

Performance Ranking

Results are collected through the PyTorch implementation, which are published in our NeurIPS paper.

There are slight differences between the 2 implementations, so you may have to redo the hyperparameter search for the TensorFlow2 implementation.

Please feel free to leave comments if you have trouble reproducing the results!

Cora

CiteSeer

PubMed

Requirements

• PyTorch 1.3.x or TensorFlow 2.x.x
• Python 3.6+
• Best with NVIDIA apex (we have used the NGC container with singularity)

Initialization

python initialize_dataset.py

Usage

python train.py

References

[1] Luan, et al., Break the Ceiling: Stronger Multi-scale Deep Graph Convolutional Networks, 2019

Cite

Please kindly cite our work if necessary:

@incollection{luan2019break,
title = {Break the Ceiling: Stronger Multi-scale Deep Graph Convolutional Networks},
author = {Luan, Sitao and Zhao, Mingde and Chang, Xiao-Wen and Precup, Doina},
booktitle = {Advances in Neural Information Processing Systems 32},
editor = {H. Wallach and H. Larochelle and A. Beygelzimer and F. d\textquotesingle Alch\'{e}-Buc and E. Fox and R. Garnett},
pages = {10943-10953},
year = {2019},
publisher = {Curran Associates, Inc.},
url = {https://arxiv.org/abs/1906.02174}
}