Simplifying Graph Convolutional Networks

Authors:

• Felix Wu*
• Tianyi Zhang*
• Amauri Holanda de Souza Jr.*
• Christopher Fifty
• Tao Yu
• Kilian Q. Weinberger

*: Equal Contribution

Overview

In good faith of facilitating future research, this branch contains additional codebase of our ablation studies. Notably, in normalization.py we provide more functions on normalizing the adjacency matrix. These design choices corresponds to different fixed spectral filters.

Specifically, we can design a low pass filter (A' = I - 1/2 * D^-1/2 * A * D^-1/2). This low pass filter performs competitively, in particular on the Reddit dataset. Such a result corresponds well with our spectral analysis in the paper. For reference:

Dataset	Metric	Filter
Cora	Acc: 80.4 %	Low-Pass, degree=2
Citeseer	Acc: 71.9 %	Low-Pass, degree=2
Pubmed	Acc: 78.5 %	Low-Pass, degree=2
Reddit	F1: 95.4 %	Low-Pass, degree=4

This branch is still in progress and may change in the future. New results based on this branch will be included in the supplementary materials of the next version of our paper.

if you find this repo (branch) useful, please cite:

@article{sgc,
  title={Simplifying Graph Convolutional Networks},
  author={Wu, Felix and Zhang, Tianyi and Souza Jr., Amauri Holanda and Fifty, Christopher and Yu, Tao and Weinberger, Kilian Q.},
  journal={arXiv preprint arXiv:1902.07153},
  year={2019}
}

Usage

See possibile normalization functions by python citation.py --help. Example commands include:

$ python citation.py --dataset cora --normalization AugRWalk --degree 2
$ python citation.py --dataset citeseer --normalization AugNormAdj --concat
$ python citation.py --dataset pubmed --normalization LowPass --degree 2

Reddit:

$ python reddit.py --inductive --test --normalization LowPass --degree 4