DGL Implementation of the GeniePath Paper

This DGL example implements the GNN model proposed in the paper GeniePath: Graph Neural Networks with Adaptive Receptive Paths.

Example implementor

This example was implemented by Kay Liu during his SDE intern work at the AWS Shanghai AI Lab.

The datasets used for node classification are Pubmed citation network dataset (tranductive) and Protein-Protein Interaction dataset (inductive).

If want to train on Pubmed (transductive), run

python pubmed.py

If want to use a GPU, run

python pubmed.py --gpu 0

If want to train GeniePath-Lazy, run

python pubmed.py --lazy True

If want to train on PPI (inductive), run

python ppi.py

Dataset: Pubmed (ACC)

Method	GeniePath
Paper	78.5%
DGL	73.0%

Dataset: PPI (micro-F1)

Method	GeniePath	GeniePath-lazy	GeniePath-lazy-residual
Paper	0.9520	0.9790	0.9850
DGL	0.9729	0.9802	0.9798