Graph Neural Tangent Kernel: Fusing Graph Neural Networks with Graph Kernels

This repository implements Graph Neural Tangent Kernel (infinitely wide multi-layer GNNs trained by gradient descent), described in the following paper:

Simon S. Du, Kangcheng Hou, Barnabás Póczos, Ruslan Salakhutdinov, Ruosong Wang, Keyulu Xu. Graph Neural Tangent Kernel: Fusing Graph Neural Networks with Graph Kernels. NeurIPS 2019. [arXiv] [Paper]

Test run

Unzip the dataset file

unzip dataset.zip

Here we demonstrate how to use GNTK to perform classification on IMDB-BINARY dataset. We set the number of BLOCK operations to be 2, the number of MLP layers to be 2 and c_u to be 1.

Compute the GNTK gram matrix

mkdir out
python gram.py --dataset IMDBBINARY --num_mlp_layers 2 --num_layers 2 --scale uniform --jk 1 --out_dir out

Classification with kernel regression

python search.py --data_dir ./out --dataset IMDBBINARY

Therefore we get the hyper-parameter search results at ./out/grid_search.csv.

Experiment for all datasets

To run the experiment described in our paper, please run bash run_gram.sh and bash run_search.sh in order.