Fast nonparametric inference of network backbones for graph sparsification

Abstract

A network backbone provides a useful sparse representation of a weighted network by keeping only its most important links, permitting a range of computational speedups and simplifying complex network visualizations. There are many possible criteria for a link to be considered important, and hence many methods have been developed for the task of network backboning for graph sparsification. These methods can be classified as global or local in nature depending on whether they evaluate the importance of an edge in the context of the whole network or an individual node neighborhood. A key limitation of existing network backboning methods is that they either artificially restrict the topology of the backbone to take a specific form (e.g. a tree) or they require the specification of a free parameter (e.g. a significance level) that determines the number of edges to keep in the backbone. Here we develop a completely nonparametric framework for inferring the backbone of a weighted network that overcomes these limitations by automatically selecting the optimal number of edges to retain in the backbone using the Minimum Description Length (MDL) principle from information theory. We develop two encoding schemes that serve as objective functions for global and local network backbones, as well as efficient optimization algorithms to identify the optimal backbones according to these objectives with runtime complexity log-linear in the number of edges. We show that the proposed framework is generalizable to any discrete weight distribution on the edges using a maximum a posteriori (MAP) estimation procedure with an asymptotically equivalent Bayesian generative model of the backbone. We compare the proposed method with existing methods in a range of tasks on real and synthetic networks.