agnn

Attention-based Graph Neural Network(AGNN)

• Paper link: http://arxiv.org/abs/1803.03735:

• Popular pytorch implementation: https://github.com/dawnranger/pytorch-AGNN

Dataset Statics

Refer to Planetoid.

Results

TL_BACKEND=tensorflow python agnn_trainer.py --dataset cora --lr 0.01 --l2_coef 0.002 --drop_rate 0.7
TL_BACKEND=tensorflow python agnn_trainer.py --dataset citeseer --lr 0.005 --l2_coef 0.006 --drop_rate 0.5
TL_BACKEND=tensorflow python agnn_trainer.py --dataset pubmed --n_att_layers 4 --lr 0.008 --l2_coef 0.005 --drop_rate 0.5
TL_BACKEND=torch python agnn_trainer.py --dataset cora --lr 0.01 --l2_coef 0.005 --drop_rate 0.7
TL_BACKEND=torch python agnn_trainer.py --dataset citeseer --lr 0.001 --l2_coef 0.02 --drop_rate 0.5
TL_BACKEND=torch python agnn_trainer.py --dataset pubmed --n_att_layers 4 --lr 0.01 --l2_coef 0.002 --drop_rate 0.7
TL_BACKEND=paddle python agnn_trainer.py --dataset cora --lr 0.02 --l2_coef 0.001 --drop_rate 0.5
TL_BACKEND=paddle python agnn_trainer.py --dataset citeseer --lr 0.005 --l2_coef 0.01 --drop_rate 0.7
TL_BACKEND=paddle python agnn_trainer.py --dataset pubmed --n_att_layers 4 --lr 0.02 --l2_coef 0.002 --drop_rate 0.6

Dataset	Paper	Our(tf)	Our(pd)	Our(torch)
cora	83.1	83.28(±0.64)	83.48(±0.35)	83.0(±0.65)
citeseer	71.7	71.7(±0.53)	73.24(±0.71)	72.52(±1.13)
pubmed	79.9	79.02(±0.82)	78.94(±0.43)	79.10(±0.20)