This repository is the implementation of ACE-HGNN in PyTorch.
[ICDM 2021] ACE-HGNN: Adaptive Curvature Exploration Hyperbolic Graph Neural Network
Below are some essential pypi packages we use on our own experiment environment. You may install their dependencies at the same time.
python==3.6.8
pytorch==1.6.0
nashpy==0.0.21
networkx==2.2
scikit-learn==0.20.3
numpy==1.16.2
pandas==0.24.2
scipy==1.2.1
- Clone this repo
- Create a virtual environment using conda or virtualenv.
conda env create -f environment.yml virtualenv -p [PATH to python3.6 binary] ace-hgnn - Enter the virtual environment and run
pip install -r requirements.txt.
- Run
set_env.shin the command line. (Linux) - Please refer to
config.pyfor our Model's full parameters and their default values. - Run
python train.py [--param param_value]to train our model, with setting custom parameters. - Some training examples:
| Dataset | Task | Command | Test AUC/F1 |
|---|---|---|---|
| Cora | LP | python -u train.py --task lp --dataset cora --model HGCN --lr 0.005 --dim 16 --num-layers 2 --act relu --bias 1 --dropout 0.5 --weight-decay 0.001 --manifold PoincareBall --log-freq 5 --cuda 0 --c 1.0 --epochs 2000 --lr-reduce-freq 200 --lr-q 0.5 | 93.94 |
| Cora | NC | python -u train.py --task nc --dataset cora --model HGCN --lr 0.01 --dim 16 --num-layers 2 --act leaky_relu --bias 1 --dropout 0.1 --weight-decay 0.0005 --manifold PoincareBall --log-freq 5 --cuda 0 --c 1.0 --epochs 2000 --lr-reduce-freq 200 --lr-q 0.5 | 81.80 |
| PubMed | LP | python -u train.py --task lp --dataset pubmed --model HGCN --lr 0.01 --dim 16 --num-layers 2 --act relu --bias 1 --dropout 0.5 --weight-decay 0.001 --manifold PoincareBall --log-freq 5 --cuda 0 --c 1.0 --lr-q 0.5 --save 1 --theta 0.5 --lr-reduce-freq None --gamma 0.2 --seed 4789 | 95.11 |
| WebKB | LP | python -u train.py --task lp --dataset webkb --model HGCN --lr 0.005 --dim 16 --num-layers 2 --act relu --bias 1 --dropout 0.5 --weight-decay 0.001 --manifold PoincareBall --log-freq 5 --cuda 0 --c 1.0 --epochs 2000 --lr-reduce-freq 200 --lr-q 0.5 | 94.47 |
| PPI | LP | python -u train.py --task lp --dataset ppi --model HGCN --lr 0.0005 --dim 16 --num-layers 2 --act leaky_relu --bias 1 --dropout 0.05 --weight-decay 0.0005 --manifold PoincareBall --log-freq 5 --c 1.0 --epochs 5000 --lr-reduce-freq 500 --lr-q 0.5 | 92.03 |
| PPI | NC | python -u train.py --task nc --dataset ppi --model HGCN --dim 16 --num-layers 2 --bias 1 --dropout 0.05 --weight-decay 0.0005 --manifold PoincareBall --log-freq 5 --c 1.0 --epochs 1000 --lr-reduce-freq 100 --lr-q 0.5 --lr 0.0005 --act tanh | 67.54 |
Some of the code was forked from the following repositories:
We deeply thanks for their contributions to the open-source community.
We also thanks for the computing infrastructure provided by Beijing Advanced Innovation Center for Big Data and Brain Computing (BDBC).
@inproceedings{fu2021ace,
title={ACE-HGNN: Adaptive Curvature Exploration Hyperbolic Graph Neural Network},
author={Fu, Xingcheng and Li, Jianxin and Wu, Jia and Sun, Qingyun and Ji, Cheng and Wang, Senzhang and Tan, Jiajun and Peng, Hao and Philip, S Yu},
booktitle={2021 IEEE International Conference on Data Mining (ICDM)},
pages={111--120},
year={2021},
organization={IEEE Computer Society}
}