readme.md

Heterogeneous Graph Transformer (HGT)

• Paper link: https://arxiv.org/pdf/2003.01332

Dataset Statics

Dataset	# Nodes	# Node Types	# Edges	# Edge Types	Target	# Classes
DBLP	26,128	4	239,566	6	author	4
IMDB	21,420	4	86,642	6	movie	5

DBLP dataset refer to HGBDataset.

IMDBdataset refer to IMDB.

Results

# available dataset: "DBLP", "IMDB"
TL_BACKEND="paddle" python hgt_trainer.py --dataset DBLP --lr 0.0001 --n_epoch 200 --hidden_dim 1024 --l2_coef 1e-6 --heads 4 --drop_rate 0.9
TL_BACKEND="torch" python hgt_trainer.py --dataset DBLP --lr 0.0001 --n_epoch 200 --hidden_dim 1024 --l2_coef 5e-6 --heads 4 --drop_rate 0.9
TL_BACKEND="paddle" python hgt_trainer.py --dataset IMDB --lr 0.0001 --n_epoch 200 --hidden_dim 1024 --l2_coef 1e-6 --heads 4 --drop_rate 0.9
TL_BACKEND="torch" python hgt_trainer.py --dataset IMDB --lr 0.0001 --n_epoch 150 --hidden_dim 1024 --l2_coef 1e-6 --heads 4 --drop_rate 0.5
TL_BACKEND="tensorflow" python hgt_trainer.py --dataset IMDB --lr 0.0001 --n_epoch 200 --hidden_dim 1024 --l2_coef 5e-6 --heads 4 --drop_rate 0.9

Dataset	Paper	Our(pd)	Our(tf)	Our(torch)
DBLP	93.01±0.23	92.4(±0.92)		90.89(±1.08)
IMDB	63.00±1.19	54.51(±1.99)	55.98(±2.09)	54.93(±1.34)

The experimental results under IMDB dataset are consistent with PyG.