Geodesic Graph Neural Network for Efficient Graph Representation Learning

This is the official codebase of the paper

Geodesic Graph Neural Network for Efficient Graph Representation Learning

Lecheng Kong, Yixin Chen, Muhan Zhang

The package is developed based on the GNN toolbox gnnfree.

Installation

We recommend installation from Conda:

git clone https://github.com/woodcutter1998/gdgnn.git
cd gdgnn
sh setup.sh

Usage

--gd_type controls the type of geodesics, it can either be VerGD for vertical geodesics or HorGD for horizontal geodesics.

--num_layers controls the number of layers in the GNN.

To run different datasets, do python run_**.py with the parameters specified.

To search for hyperparameters, modify the hparams variable in the run_**.py files to specify the list of potentail hyperparameters. e.g.

hparams = {'num_layers':{2,3,4,5},
            'gd_type':{'VerGD, HorGD'},
            'dropout':{0.5, 0.7, 0.9}}

and do python run_**.py --psearch True. The program performs grid-search on the hyperparameters specified.

An example command to reproduce our results on OGBG-MOLHIV dataset is:

python run_MOL.py --train_data_set ogbg-molhiv --psearch True