When Heterophily Meets Heterogeneous Graphs: Latent Graphs Guided Unsupervised Representation Learning

This is the source code of "When Heterophily Meets Heterogeneous Graphs: Latent Graphs Guided Unsupervised Representation Learning" (LatGRL), accepted by IEEE Transactions on Neural Networks and Learning Systems 2025 (TNNLS).

Paper Link: https://ieeexplore.ieee.org/document/10905047

Arxiv Link: https://arxiv.org/abs/2409.00687

To our knowledge, we are the first to explore the phenomenon of node-level heterophily in heterogeneous graphs and attempt to address this issue in unsupervised heterogeneous graph learning. The marginal and multi-relational joint distributions of node-level homophily ratios across different meta-paths in real-world heterogeneous graphs are shown below:

Available Data

Datasets: https://drive.google.com/file/d/12nWwcrufexpU1n6W7YTyb-6sL1wQ-WrY/view?usp=sharing

Place the 'data' folder from the downloaded files into the 'LatGRL' directory.

Requirements

This code requires the following:

• Python==3.9.16
• PyTorch==1.13.1
• Numpy==1.24.2
• Scipy==1.10.1
• Scikit-learn==1.2.1

Running

For different datasets, please run the following code:

ACM:

LatGRL:

python main.py -dataset acm

LatGRL-S:

python main_sampler.py -dataset acm

DBLP:

LatGRL:

python main.py -dataset dblp

LatGRL-S:

python main_sampler.py -dataset dblp

Yelp:

LatGRL:

python main.py -dataset yelp

LatGRL-S:

python main_sampler.py -dataset yelp

IMDB:

LatGRL:

python main.py -dataset imdb

LatGRL-S:

python main_sampler.py -dataset imdb

Ogbn-mag:

LatGRL-S:

python main_sampler.py -dataset mag

BibTeX

@article{10905047,
  author={Shen, Zhixiang and Kang, Zhao},
  journal={IEEE Transactions on Neural Networks and Learning Systems}, 
  title={When Heterophily Meets Heterogeneous Graphs: Latent Graphs Guided Unsupervised Representation Learning}, 
  year={2025},
  pages={1-14},
  publisher={IEEE},
  doi={10.1109/TNNLS.2025.3540063}}