Graphlet MPNNs: Extending Message-Passing Neural Networks with Graphlet Information

Anonymous

Keywords: graph neural networks, graphlets, GNN, MPNN

TL;DR: We describe Graphlet MPNNs, a general unifying class of models that augments MPNNs with graph structural information, and fully characterize their distinguishing power.

Abstract:

We propose Graphlet Message-Passing Neural Networks (MPNNs) as an extension
of classical MPNNs in which vertex and edge graphlet information is taken
into account. In this way, the distinguishing power of MPNNs is increased in a
natural way. We introduce Graphlet MPNNs in quite some generality, hereby encompassing
recent proposals. Our main result is a complete characterization of the
distinguishing power of Graphlet MPNNs. We conclude this paper by outlining
some interesting directions for future research.