Neural Graduated Assignment for Maximum Common Edge Subgraphs

Overview

This project implements a deep learning-based algorithm that can find the maximum common edge subgraph between two molecular graphs and compute their similarity. The project supports multiple datasets including AIDS, MCF-7, and MOLHIV.

Quick Start

Requirements

• Python 3.9+
• PyTorch
• PyTorch Geometric
• NetworkX
• NumPy
• Scipy

Installation

pip install -r requirements.txt

Data Generation

Use the following command to generate data:

bash data_gen.sh

Or you can use our pre-generated data by unzipping the data.py file.

Training

Use the following command to start training the model:

bash run.sh

Evaluation

Evaluate the results by the following command

python evaluate.py

Project Structure

NGA/
├── NGA_main.py          # Main training script
├── model.py             # Neural network model definition
├── data.py              # Data processing module
├── utils.py             # Utility functions
├── modular_product.py   # Association common graph construction
├── sinkhorn.py          # Sinkhorn algorithm implementation
├── evaluate.py          # Model evaluation script
├── combine_data.py      # Data combination script
├── data_gen.sh          # Data generation script
├── run.sh               # Training execution script
└── data/                # Data directory
    

Key Parameters

• --dataset: Dataset selection
• --hidden: Hidden layer dimension
• --lr: Learning rate
• --epochs: Number of training epochs
• --device: Computing device
• --retrieval: Compute MCES or graph similarity & retrieval

Citation

@inproceedings{
ying2026neural,
title={Neural Graduated Assignment for Maximum Common Edge Subgraphs},
author={Chaolong Ying and Yingqi Ruan and Xuemin Chen and Yaomin Wang and Tianshu Yu},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=ZVlTIyRe35}
}