On Addressing the Limitations of Graph Neural Networks

Abstract

This report gives a summary of two problems about graph convolutional networks (GCNs): over-smoothing and heterophily challenges, and outlines future directions to explore.

arXiv commentProposal report and a concise version of Sitao Luan's thesis. The weight initialization part is quite interesting but will not be included in Sitao's formal thesis, thus Sitao put this preprint report online. arXiv admin note: substantial text overlap with arXiv:2109.05641, arXiv:2210.07606