hongjoon0805/HALO

Descent Steps of a Relation-Aware Energy Produce Heterogeneous Graph Neural Networks

This repo is official code for NeurIPS 2022 paper "Descent Steps of a Relation-Aware Energy Produce Heterogeneous Graph Neural Networks"

Execution Details

Requirements

Python 3.7.10
Pytorch: 1.10.0
DGL: 0.8.0
CUDA: 11.4

Downloading dataset

Knowledge graphs: Downloaded from dgl.data.rdf
HGB datasets: Pre-defined class "HGBDataset" tries to download the datasets
Academic datset: Pre-defined class "AcademicDataset" tries to download the datasets

All the datasets are automatically downloaded if you run one of the following command.

Note

Since it is not available to get labels of test nodes in HGB dataset, we instead evaluate our methods using validation set. All the evaluation procedures are implemented in our repo.

For more details on the HGB dataset, please refer to this Link

Reproducing Table 1

Execution command

# AIFB
$ python main.py --date NeurIPS2022 --seed 1 --mlp_bef=1 --mlp_aft=1 --prop_step=16 --num_epoch=1000 --lam=1.0 --alp=0.1 --dropout 0.5 --inp_dropout 0.5 --learn_emb=16 --hidden_size=16 --lr 0.001 --weight_decay 1e-05 --data=AIFB

# MUTAG
$ python main.py --date NeurIPS2022 --seed 1 --mlp_bef=1 --mlp_aft=1 --prop_step=16 --num_epoch=1000 --lam=0.01 --alp=1.0 --dropout 0.5 --inp_dropout 0.5 --learn_emb=16 --hidden_size=16 --lr 0.001 --weight_decay 0.0001 --data=MUTAG

# BGS
$ python main.py --date NeurIPS2022 --seed 1 --mlp_bef=1 --mlp_aft=1 --prop_step=8 --num_epoch=1000 --lam=0.1 --alp=1.0 --dropout 0.5 --inp_dropout 0.5 --learn_emb=16 --hidden_size=16 --lr 0.01 --weight_decay 1e-05 --data=BGS

# AM
$ python main.py --date NeurIPS2022BinFeat --seed 1 --mlp_bef=1 --mlp_aft=1 --prop_step=4 --num_epoch=1000 --lam=1.0 --alp=1.0 --dropout 0.5 --inp_dropout 0.5 --learn_emb=16 --hidden_size=16 --lr 0.01 --weight_decay 0.0001 --data=AM

# DBLP
$ python main.py --date NeurIPS2022 --seed 1 --mlp_bef=1 --mlp_aft=1 --prop_step=8 --num_epoch=1000 --lam=1.0 --alp=1.0 --dropout 0.5 --inp_dropout 0.5 --learn_emb=256 --hidden_size=256 --lr 0.0001 --weight_decay 1e-05 --data=DBLP

# IMDB
$ python main.py --date NeurIPS2022 --seed 1 --mlp_bef=1 --mlp_aft=1 --prop_step=32 --num_epoch=1000 --lam=1.0 --alp=1.0 --dropout 0.5 --inp_dropout 0.5 --learn_emb=64 --hidden_size=64 --lr 0.001 --weight_decay 1e-05 --multilabel --data=IMDB

# ACM
$ python main.py --date NeurIPS2022 --seed 1 --mlp_bef=1 --mlp_aft=1 --prop_step=32 --num_epoch=1000 --lam=0.1 --alp=0.1 --dropout 0.5 --inp_dropout 0.5 --hidden_size=32 --lr 0.01 --weight_decay 0.0001 --data=ACM

# Freebase
$ python main.py --date NeurIPS2022NodeFeat --seed 1 --mlp_bef=1 --mlp_aft=1 --prop_step=4 --num_epoch=1000 --lam=1.0 --alp=1.0 --dropout 0.5 --inp_dropout 0.5 --hidden_size=32 --lr 0.01 --weight_decay 0.001 --data=Freebase

Reproducing Table 2

# DBLP-reduced
$ python main.py --date NeurIPS2022 --seed 1 --mlp_bef=1 --mlp_aft=1 --prop_step=8 --num_epoch=1000 --lam=1.0 --alp=1.0 --dropout 0.5 --inp_dropout 0.5 --learn_emb=256 --hidden_size=256 --lr 0.0001 --weight_decay 1e-05 --data=DBLP_reduced --multicategory

$ python main.py --date NeurIPS2022 --seed 1 --prop_step=1000 --num_epoch=1 --ZooBP --data=DBLP_reduced --multicategory

# Academic-reduced
$ python main.py --date NeurIPS2022 --seed 1 --mlp_bef=1 --mlp_aft=1 --prop_step=8 --num_epoch=1000 --lam=1.0 --alp=1.0 --dropout 0.5 --inp_dropout 0.5 --learn_emb=256 --hidden_size=256 --lr 0.0001 --weight_decay 1e-05 --data=Academic_reduced --multicategory

$ python main.py --date NeurIPS2022 --seed 1 --prop_step=1000 --num_epoch=1 --ZooBP --data=Academic_reduced --multicategory

hongjoon0805 / HALO