hardgat

Graph Hard Attention Networks (HardGAT)

• Paper link: https://dl.acm.org/doi/pdf/10.1145/3292500.3330897
• Popular pytorch implementation: https://github.com/dmlc/dgl/tree/master/examples/pytorch/hardgat

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

TL_BACKEND="paddle" python gat_trainer.py --dataset cora --lr 0.01 --l2_coef 0.004 --k 8 --drop_rate 0.7
TL_BACKEND="paddle" python gat_trainer.py --dataset citeseer --lr 0.01 --l2_coef 0.004 --k 8 --drop_rate 0.6
TL_BACKEND="paddle" python gat_trainer.py --dataset pubmed --lr 0.01 --l2_coef 0.0015 --k 8 --drop_rate 0.6

TL_BACKEND="torch" python gat_trainer.py --dataset cora --lr 0.01 --l2_coef 0.005 --k 8 --drop_rate 0.7 
TL_BACKEND="torch" python gat_trainer.py --dataset citeseer --lr 0.005 --l2_coef 0.003 --k 8 --drop_rate 0.55
TL_BACKEND="torch" python gat_trainer.py --dataset pubmed --lr 0.01 --l2_coef 0.0015 --k 8 --drop_rate 0.6

TL_BACKEND="tensorflow" python gat_trainer.py --dataset cora --lr 0.03 --l2_coef 0.004 --k 8 --drop_rate 0.7 
TL_BACKEND="tensorflow" python gat_trainer.py --dataset citeseer --lr 0.01 --l2_coef 0.006 --k 8 --drop_rate 0.7
TL_BACKEND="tensorflow" python gat_trainer.py --dataset pubmed --lr 0.01 --l2_coef 0.0015 --k 8 --drop_rate 0.6

Dataset	Paper	Our(pd)	Our(torch)	Our(tf)
cora	83.5(±0.7)	83.50(±0.75)	82.55(±0.40)	83.80(±0.96)
citeseer	72.7(±0.6)	72.50(±0.76)	71.00(±0.45)	72.30(±0.65)
pubmed	79.2(±0.4)	78.90(±0.76)	78.53(±0.85)	78.32(±0.33)