Hyper-SAGNN: a self-attention based graph neural network for hypergraphs

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)

Requirements

python >= 3.6.8

Tensorflow >= 1.0.0 (< 2.0.0)

Pytorch >= 1.0

Usage

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.')

Cite

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}
}