LightHGNN

This repository contains the source code for the paper "LightHGNN: Distilling Hypergraph Neural Networks into MLPs for 100x Faster Inference" published in IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 2024 by Yifan Feng, Yihe Luo, Shihui Ying, Yue Gao*. This paper is available at here.

Introduction

In this repository, we provide the implementation of our LightHGNNs, including LightHGNN and LightHGNN+, which is based on the following environments:

• python 3.9: basic programming language.
• dhg 0.9.4: for hypergraph representation and learning.
• torch 1.12.1: for computation.
• hydra-core 1.3.2: for configuration and multi-run management.

Installation

1. Clone this repository.
2. Install the required libraries.

pip install -r requirements.txt

Usage

For transtive setting, you can run the following command:

python trans_train.py

For multi-run and obtain the average results, you can run the following command:

python trans_multi_exp.py

For production setting, you can run the following command:

python prod_train.py

For multi-run and obtain the average results, you can run the following command:

python prod_multi_exp.py

Change Models You can change the teacher by modifying the teacher in trans_config.yaml and prod_config.yaml as following:

model:
    teacher: hgnn # hgnn, hgnnp, hnhn, unigcn

Also, you can change the student by modifying the student in trans_config.yaml and prod_config.yaml as following:

model:
    student: light_hgnn # light_hgnn, light_hgnnp

Change Datasets In our paper, 13 grpah/hypergraph datasets are adopted for evaluation.

• Graph datasets: cora, pubmed, citeseer
• Hypergraph datasets: news20, ca_cora, cc_cora, cc_citeseer, dblp4k_conf, dblp4k_paper, dblp4k_term, imdb_aw, recipe_100k, recipe_200k

You can change the dataset by modifying the dataset in trans_config.yaml and prod_config.yaml as following:

data:
    name: dblp4k_paper # cora, pubmed, news20, ca_cora, dblp4k_term, imdb_aw, ...

Important Note Since the recipe_100k and recipe_200k datasets are too large and contains more than 10k vertices, the two dataset can only be used under the production setting. Please do not use the two datasets for the transitive setting.

Citation

If you find this repository useful in your research, please cite the following papers:

@inproceedings{feng2024lighthgnn,
title={Light{HGNN}: Distilling Hypergraph Neural Networks into {MLP}s for 100x Faster Inference},
author={Feng, Yifan and Luo, Yihe and Ying, Shihui and Gao, Yue},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024},
}

@article{gao2022hgnn+,
  title={HGNN+: General hypergraph neural networks},
  author={Gao, Yue and Feng, Yifan and Ji, Shuyi and Ji, Rongrong},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  volume={45},
  number={3},
  pages={3181--3199},
  year={2022},
  publisher={IEEE}
}

@inproceedings{feng2019hypergraph,
  title={Hypergraph neural networks},
  author={Feng, Yifan and You, Haoxuan and Zhang, Zizhao and Ji, Rongrong and Gao, Yue},
  booktitle={Proceedings of the AAAI conference on artificial intelligence},
  volume={33},
  number={01},
  pages={3558--3565},
  year={2019}
}