Heterophily-informed Message Passing

Haishan Wang · Arno Solin . Vikas Garg

TMLR 2025

Paper

---

This is the public codes for TMLR2025 paper: "Heterophily-informed Message Passing"

Comparison of heterophiliy-informed MP with traditional GNN

Experiment 1: Node classification

Get into the experiment directory of node classification at first:

cd node_classification

1) Environment setup

The environment of node classification experiment is installed by the following command

conda env create -f environment.yml 
conda activate himp-nc

2) Dataset

The data is donwloaded automatically if you run the training script.

3) Training the model

This is how to train a GNN model with different versions of message passing

python main.py  --data_name ${dataname} --model_name ${gnn} --het_mode ${het_mode} --save_dir ${save_dir}

The arguments notes:

• --data_name, the name of dataset. Options: ( Cora CiteSeer PubMed Cornell Texas Wisconsin Computers Photo Chameleon Squirrel Roman-empire Amazon-ratings Minesweeper Tolokers Questions )
• --model_name, the GNN types. Options: ( GCN GAT GIN GraphSAGE )
• het_mode, the heterophily-preference of message passing. Options: ( original heterophily homophily mix )
• save_dir, the path of running result, including the metric of best model, and trained model. One example script for running GCN+het. on Cora dataset:

source scripts/example_nc.sh

Experiment 2: Molecular generation

Get into the experiment directory of molecular generation at first:

cd molecule_generation

1) Environment setup

conda env create -f environment.yml 
conda activate hetflow

2) Dataset preprocess

Preprocess SMILES strings to be molecular graphs

python data/data_preprocess.py --data_name qm9
python data/data_preprocess.py --data_name zinc250k

3) Training

4) Evaluation