iehgcn

Interpretable and Efficient Heterogeneous Graph Convolutional Network (ieHGCN)

• Paper link: https://arxiv.org/abs/2005.13183v3
• Author's code repo: https://github.com/kepsail/ie-HGCN. Note that the original code is implemented with Pytorch for the paper.

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	4

DBLP dataset refer to HGBDataset.

IMDBdataset refer to IMDB.

Performance

For the DBLP dataset: train test val = 974, 1420, 243 about 37% for training.

For the IMDB dataset: train test val = 400, 3478, 400, about 9% for training.

Dataset	Paper(80% training)	Paper(60% training)	Paper(40% training)	Paper(20% training)	Our(tf)	Our(th)	Our(pd)
DBLP	96.29	95.25	93.83	93.85	92.30±0.49%	90.90±0.74%	91.18±0.66%
IMDB	58.35	60.84	59.81	56.60	58.10±0.42%	55.22±1.21%	56.08±2.13%

TL_BACKEND="tensorflow" python3 iehgcn_trainer.py --dataset DBLP --n_epoch 30 --lr 0.01 --num_layers 3 --hidden_channels [64, 32] --l2_coef 0.0005 --drop_rate 0.2
TL_BACKEND="torch" python3 iehgcn_trainer.py --dataset DBLP --n_epoch 30 --lr 0.005 --num_layers 4 --hidden_channels [64, 32, 16] --l2_coef 0.0005 --drop_rate 0.0
TL_BACKEND="paddle" python3 iehgcn_trainer.py --dataset DBLP --n_epoch 30 --lr 0.01 --num_layers 4 --hidden_channels [64, 32, 16] --l2_coef 0.0005 --drop_rate 0.1

TL_BACKEND="torch" python3 iehgcn_trainer.py --dataset IMDB --n_epoch 25 --lr 0.01 --num_layers 3 --hidden_channels [64, 32] --l2_coef 0.0005 --drop_rate 0.2
TL_BACKEND="tensorflow" python3 iehgcn_trainer.py --dataset IMDB --n_epoch 25 --lr 0.005 --num_layers 3 --hidden_channels [64, 32] --l2_coef 0.0005 --drop_rate 0.2
TL_BACKEND="paddle" python3 iehgcn_trainer.py --dataset IMDB --n_epoch 25 --lr 0.005 --num_layers 3 --hidden_channels [64, 32] --l2_coef 0.0005 --drop_rate 0.2