This is an implementation of "Hyper-SAGNN: a self-attention based graph neural network for hypergraphs" (ICLR2020)
The datasets included in this repo are originally from DHNE (https://github.com/tadpole/DHNE)
python >= 3.6.8
Tensorflow >= 1.0.0 (< 2.0.0)
Pytorch >= 1.0
To run the code:
cd Code
python main.py --data wordnet -f adj
Change the following arguments to reproduce corresponding results from the manuscript,
The --data argument can take "GPS", "drug", "MovieLens", "wordnet". This argument is case sensitive
The -f, --feature argument can take "adj" or "walk" represents encoder based approach and random walk based approach respectively.
Other arguments are as followed:
parser.add_argument('--dimensions', type=int, default=64,
help='Number of dimensions. Default is 64.')
parser.add_argument('-l', '--walk-length', type=int, default=40,
help='Length of walk per source. Default is 40.')
parser.add_argument('-r', '--num-walks', type=int, default=10,
help='Number of walks per source. Default is 10.')
parser.add_argument('-k', '--window-size', type=int, default=10,
help='Context size for optimization. Default is 10.')
If you want to cite our paper:
@inproceedings{
zhang2020hypersagnn,
title={Hyper-{SAGNN}: a self-attention based graph neural network for hypergraphs},
author={Zhang, Ruochi and Zou, Yuesong and Ma, Jian},
booktitle={International Conference on Learning Representations (ICLR)},
year={2020}
}