PatchNet

Implementation of "Handling Feature Heterogeneity with Learnable Graph Patches" which is accepted by KDD' 25.

Citation

If you find our work useful in your research or applications, please kindly cite:

@inproceedings{sun2025handling,
  title={Handling Feature Heterogeneity with Learnable Graph Patches},
  author={Sun, Yifei and Yang, Yang and Feng, Xiao and Wang, Zijun and Zhong, Haoyang and Wang, Chunping and Chen, Lei},
  year={2025},
  booktitle={Proceedings of the 31th ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
}

The code details will be finalized within two weeks, stay tuned!

Patching Process

Please see Patching for detailed visualization of extracting learnable patches.

Note

This repository is intended for review purposes only. The full version will be released upon acceptance.

Installation

• python version = 3.7.12
• Environment Construction: conda env create --file F4G.yml
• Mole-BERT is needed, so please refer to here for detailed information.

Pre-training and fine-tuning

1. Run Mole-BERT's tokenizer training step using their default settings.

python vqvae.py

2. Start self-supervised pre-training.

• Multi-GPU

CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node=4 main_multi.py --batch_size=256 --output_model_dir=saves/

• Single-GPU

python main_single.py --batch_size=256 --output_model_dir=saves/ --pretrain_dataset zinc

1. Fine-tuning

python molecule_finetune.py --dataset=$dataset --input_model_file=saves/Multi_model.pth --epochs=100

Reproducing results in the paper

Our results in the paper can be reproduced using a random seed ranging from 0 to 9 with scaffold splitting.