Transferability for Graph Convolutional Networks

Abstract

This work develops a general transferability theory for graph convolutional networks; applicable to architectures based on both undirected- as well as recently introduced directed convolutional filters. Transferability is considered between graphs that are similar from the perspective of information diffusion. A detailed theoretical investigation establishes which filters render networks stable with respect to this novel approach to transferability. Illustrative examples (including graph-coarsening) showcase how newly established results may inform the design of transferable architectures in practice. Numerical experiments on real-world data validate the theoretical findings and complement the mathematical analysis.