This is the implementation of GFN used in paper "Are Powerful Graph Neural Nets Necessary? A Dissection on Graph Classification", previously titled "Dissecting Graph Neural Networks on Graph Classification". GFN is a simple lightweight neural set function with graph augmented features. It augments nodes with graph structural and propagated features, and then defines the function directly on a set of nodes. See our paper for more details.
This code was developed under version 1.1.0. So if it doesn't work for your pytorch_geometric, you may want to install the right version.
Replace the benchmark experiment below with whichever you want to run, see all exps in main.py.
python main --exp benchmark
Please cite our paper if you find it helpful in your own work:
@article{gfn2019,
title={Are Powerful Graph Neural Nets Necessary? A Dissection on Graph Classification},
author={Ting Chen, Song Bian, Yizhou Sun}
journal={CoRR},
volume={abs/1905.04579},
year={2019},
}