han

Heterogeneous Graph Attention Network (HAN)

This is an implementation of HAN for heterogeneous graphs.

• Paper link: https://arxiv.org/abs/1903.07293
• Author's code repo: https://github.com/Jhy1993/HAN. Note that the original code is implemented with Tensorflow for the paper.

Usage

python han_trainer.py for reproducing HAN's work on IMDB.

Note: this scripts only support IMDB, which means command python han_trainer.py --dataset ACM will not run on ACM. If you want to test the performance of other datasets, you are suggested to make some modification of the trainer script.

Performance

Reference performance numbers for the IMDB dataset: (0.01, 200, 0.0001, 8, 0.8, 0.58178, 0.002811689883326394)

train test val = 400, 3478, 400, about 9% for trianing

Dataset	Paper(80% training)	Paper(60% training)	Paper(40% training)	Paper(20% training)	Our(tf)	Our(th)	Our(pd)
IMDB	58.51	58.32	57.97	55.73	57.78(±0.51)	55.66(±1.05)	56.58(±0.51)

TL_BACKEND="tensorflow" python3 han_trainer.py --n_epoch 200 --lr 0.01 --l2_coef 0.0001 --heads 8 --drop_rate 0.8

TL_BACKEND="torch" python3 han_trainer.py --n_epoch 200 --lr 0.01 --l2_coef 0.0001 --heads 16 --drop_rate 0.4

TL_BACKEND="paddle" python3 han_trainer.py --n_epoch 200 --lr 0.01 --l2_coef 0.0001 --heads 16 --drop_rate 0.4