ruiwu1990/correlation-gnn

Jia, J. and Benson, A.R., 2020, August. Residual correlation in graph neural network regression. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining (pp. 588-598).

PyTorch Implementation of Correlated Graph Neural Networks

Outcome Correlation in Graph Neural Network Regression

Junteng Jia and Austin Benson

arXiv:2002.08274, 2020.

Overview

Correlated Graph Neural Networks model the correlations among labels as well as features of each node:

• C-GNN models the correlation as a multivariate Gaussian and learns the correlation structure in O(m) per optimization step, where m is the number of edges.
• LP-GNN assumes positive correlation among neighboring vertices, and runs label propagation to interpolate GNN residuals on the testing vertices.

Requirements

• Python 3.7+
• PyTorch 1.2.0+
• DGL
• GPyTorch

Usage

Download this Repository

git clone this repo to your local machine.

Dataset

US Election dataset is used as a running example. The dataset is included in this repo.

Train and Test

We so far implemented 2 graph neural network structures: GCN and GraphSAGE, for LP-GNN as well as C-GNN.

To train and test GCN-based LP-GNN, use the following script:

scripts/run_lp_gcn.sh

To train and test GraphSAGE-based LP-GNN, use the following script:

scripts/run_lp_graphsage.sh

To train and test GCN-based C-GNN, use the following script:

scripts/run_cgnn_gcn.sh

To train and test GraphSAGE-based C-GNN, use the following script:

scripts/run_cgnn_graphsage.sh

The default hyper-parameters should give reasonably good results.

If you have any questions, feel free to open an issue.

References

C-GNN (original implementation in Julia)
GCN
GraphSAGE