This is the code repository for the ICML 2023 paper On the Connection Between MPNN and Graph Transformer. There are three folders corresponding to the three experiment subsections.
Experiments has three parts
- MPNN + VN is a surprisingly strong baseline, outperforming GT on the recently proposed Long Range Graph Benchmark (LRGB) dataset
- our MPNN + VN improves over early implementation on a wide range of OGB datasets
- MPNN + VN outperforms Linear Transformer and MPNN on the climate modeling task.
The lrgb contains the code for section 7.1 where we showed adding virtual node can significantly improve the result on lrgb datasets. The code is modified from the original lrgb paper. See Readme for instructions.
The GraphGPS contains the code for section 7.2, where we showed that by leveraging the the framework of GraphGPS, simple MPNN + VN can achieve competitive on ogb datasets. See Readme for instructions.
The climate-graph-main contains the code for section 7.3, where we experiment the MPNN + VN for sea-surface temperature prediction. See Readme for instructions.
@article{cai2023connection,
title={On the Connection Between MPNN and Graph Transformer},
author={Cai, Chen and Hy, Truong Son and Yu, Rose and Wang, Yusu},
journal={International Conference on Machine Learning},
year={2023}
}