K-plex cover pooling for graph neural networks

Abstract Graph pooling methods provide mechanisms for structure reduction that are intended to ease the diffusion of context between nodes further in graph, and typically leverage community discovery or node edge pruning heuristics. In this paper, we introduce a novel technique which borrows from classical results graph theory is non-parametric generalizes well graphs different nature connectivity patterns. Our method, named KPlexPool , builds on concepts covers k -plexes, i.e. pseudo-cliques where each can miss up links. The experimental evaluation benchmarks molecular social classification shows achieves state art performances against both parametric literature, despite generating pooled based solely topological information.