dropedge

DropEdge

• Paper link: https://arxiv.org/abs/1907.10903
• Author's code repo: https://github.com/DropEdge/DropEdge.

Dataset Statics

Dataset	# Nodes	# Edges	# Classes
Cora	2,708	10,556	7
Citeseer	3,327	9,228	6
Pubmed	19,717	88,651	3
Refer to Planetoid.

Results(cora & 4 layer GCN)

# available dataset: "cora", "citeseer", "pubmed"
TL_BACKEND="paddle" python dropedge_trainer.py --dataset cora --lr 0.01 --l2_coef 0.0005 --drop_rate 0.8 --n_layers 4 --sampling_percent 0.7
TL_BACKEND="paddle" python dropedge_trainer.py --dataset cora --lr 0.01 --l2_coef 0.0005 --drop_rate 0.8 --n_layers 4 --sampling_percent 1
TL_BACKEND="tensorflow" python dropedge_trainer.py --dataset cora --lr 0.01 --l2_coef 0.0005 --drop_rate 0.8 --n_layers 4 --sampling_percent 0.7
TL_BACKEND="tensorflow" python dropedge_trainer.py --dataset cora --lr 0.01 --l2_coef 0.0005 --drop_rate 0.8 --n_layers 4 --sampling_percent 1
TL_BACKEND="torch" python dropedge_trainer.py --dataset cora --lr 0.01 --l2_coef 0.0005 --drop_rate 0.8 --n_layers 4 --sampling_percent 0.7
TL_BACKEND="torch" python dropedge_trainer.py --dataset cora --lr 0.01 --l2_coef 0.0005 --drop_rate 0.8 --n_layers 4 --sampling_percent 1

Method	Paper	Our(tf)	Our(th)	Our(pd)
Original	80.40	80.42±1.2	81.16±1.14	80.44±0.66
DropEdge	82.00	81.18±2.22	81.94±1.56	81.20±1.4

• The model performance is the average of 5 tests