rgcn

Relational Graph Convolutional Network

• Paper: Modeling Relational Data with Graph Convolutional Networks

• Author's code for entity classification: https://github.com/tkipf/relational-gcn

• Author's code for link prediction: https://github.com/MichSchli/RelationPrediction

Dataset Statics

Dataset	#Nodes	#Edges	#Relations	#Labeled
AIFB	8,285	58,086	90	176
MUTAG	23,644	148,454	46	340
BGS	333,845	1,832,398	206	146
AM	1,666,764	11,976,642	266	1000

Dateset	#Nodes	#Node Types	#Edges	#Relations	Target	#Classes
IMDB	11,616	3	102,804	4	movie	3

Results

TL_BACKEND="torch" python rgcn_trainer.py --dataset aifb  --l2 5e-5
TL_BACKEND="torch" python rgcn_trainer.py --dataset mutag --l2_coef 5e-2
TL_BACKEND="torch" python rgcn_trainer.py --dataset bgs --lr 0.0001 --l2_coef 5e-2
TL_BACKEND="torch" python rgcn_trainer.py --dataset imdb

TL_BACKEND="tensorflow" python rgcn_trainer.py --dataset aifb 
TL_BACKEND="tensorflow" python rgcn_trainer.py --dataset mutag --l2_coef 5e-2
TL_BACKEND="tensorflow" python rgcn_trainer.py --dataset bgs --l2_coef 5e-2
TL_BACKEND="tensorflow" python rgcn_trainer.py --dataset imdb

TL_BACKEND="paddle" python rgcn_trainer.py --dataset aifb 
TL_BACKEND="paddle" python rgcn_trainer.py --dataset mutag --l2_coef 5e-2
TL_BACKEND="paddle" python rgcn_trainer.py --dataset bgs --l2_coef 5e-2
TL_BACKEND="paddle" python rgcn_trainer.py --dataset imdb

Dataset	Paper	Our(th)	Our(tf)	Our(pd)
AIFB	95.83	96.11(±1.52)	94.17(±2.05)	95.56(±2.3)
MUTAG	73.23	85.0(±0.66)	85.29(±1.20)	85.00(±1.9)
BGS	83.10	74.1(±1.7)	73.79(±1.9)	73.56(±3.8)
AM	89.29
IMDB		48.54(±0.62)	48.30(±1.20)	48.44(±1.10)