Abstract

We relate properties of attributed random graph models to the performance of GNN architectures. We identify regimes where GNNs outperform feedforward neural networks and non-attributed graph clustering methods. We compare GNN performance on our synthetic benchmark to performance on popular real-world datasets. We analyze the theoretical foundations for weak recovery in GNNs for popular one- and two-layer architectures. We obtain an explicit formula for the performance of a 1-layer GNN, and we obtain useful insights on how to proceed in the 2-layer case. Finally, we improve the bound for a notable result on the GNN size generalization problem by 1.

Degree

College and Department

Physical and Mathematical Sciences; Mathematics

Rights

https://lib.byu.edu/about/copyright/