Exploring Consistency in Graph Representations: from Graph Kernels to Graph Neural Networks

Paper: Xuyuan Liu, Yinghao Cai, Qihui Yang, Yujun Yan. Exploring Consistency in Graph Representations: from Graph Kernels to Graph Neural Networks. Conference on Neural Information Processing Systems (NeurIPS), 2024.

Link:

Citation (bibtex):

@inproceedings{
liu2024exploring,
title={Exploring Consistency in Graph Representations: from Graph Kernels to Graph Neural Networks},
author={Xuyuan Liu and Yinghao Cai and Qihui Yang and Yujun Yan},
booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems},
year={2024},
url={https://openreview.net/forum?id=dg0hO4M11K}
}

Requirements

• torch==2.3.0
• torch-geometric==2.5.3
• ogb==1.3.1

File Overview

• run_dist.py: main script to train the model with consistency loss

• exp_run.py: main script to train the model without consistency loss

• model.py: models used in the experiments

• dataset.py&dataset.ogb.py: datasets used in the experiments

Usage

To run all experiments, run.sh is provided:

bash ./run.sh

To train a specified model with consistency loss on a given dataset: Take GCN and PROTEINS as an example

python run_dist.py --model=GCN --dataset=PROTEINS --device=0

To train a specified model without consistency loss on a given dataset:

python exp_run.py --model=GCN --dataset=PROTEINS --device=0

Since consistency loss can be computed using only layer outputs, and related functions are incorporated via Python decorators, it is easy to test your own model with consistency loss. To do so, simply load your model in the get_model function within train_dist.py, set the appropriate parameters, and the consistency loss will be calculated automatically.

Before testing, please ensure that:

• Your model has a defined layer structure
• Your model can return layer outputs, which represent the hidden features of each graph

Need help?

Email [email protected] or [email protected]