Hypergraphs are powerful tools for modeling complex interactions across various domains, including biomedicine. However, learning meaningful node representations from hypergraphs remains a challenge. Existing supervised methods often lack generalizability, thereby limiting their real-world applications. We propose a new method, Pre-trained Hypergraph Convolutional Neural Networks with Self-supervised Learning (PhyGCN), which leverages hypergraph structure for self-supervision to enhance node representations. PhyGCN introduces a unique training strategy that integrates variable hyperedge sizes with self-supervised learning, enabling improved generalization to unseen data. Applications on multi-way chromatin interactions and polypharmacy side-effects demonstrate the effectiveness of PhyGCN. As a generic framework for high-order interaction datasets with abundant unlabeled data, PhyGCN holds strong potential for enhancing hypergraph node representations across various domains.
This package is developed and tested on Linux.
- Linux: Ubuntu 16.04
conda create -n myenv python=3.6.8
conda activate myenv
pip install -r requirements.txt
The package should take less than 5 min to install.
Data is provided in HNHN_data and pre-trained model checkpoints are provided in checkpoints.
Example:
python pretrain.py --data pubmed --f conv --num-epoch 300 --dropedge 0.7 --layers 2
Note that our pre-trained checkpoints are provided and can be directly used for note classification.
Example:
python main.py --f conv --data cora-cite --num-epoch 300 --layers 1 --split 1
To get the results on 10 random train-test splits, use
chmod +x script.sh
./script.sh
Use CUDA_VISIBLE_DEVICES=0 if multiple GPUs are available. The test accuracy on the 10 splits will be collected in result.txt
This repo is built upon Hyper-SAGNN.