HLCL: Graph Contrastive Learning under Heterophily via Graph Filters

Official minimal implementation of HLCL (UAI 2024). HLCL introduces a homophily–heterophily separation and applies low-pass and high-pass graph filters to augmented views, enabling state-of-the-art graph contrastive learning on heterophilic graphs.

Installation

pip install -r requirements.txt

Usage

python HLCL.py \
  --device 2 \
  --dataset pubmed \
  --low_k 0.8 \
  --high_k 0.2 \
  --augmentation PPRDiffusion \
  --haug1 0.4 \
  --runs 3 \
  --num_parts 3

Results and checkpoints are saved in:

• ./result/
• ./model/

Citation

@inproceedings{yang2024hlcl,
  title={Graph Contrastive Learning under Heterophily via Graph Filters},
  author={Yang, Wenhan and Mirzasoleiman, Baharan},
  booktitle={Uncertainty in Artificial Intelligence (UAI)},
  year={2024}
}