Robust subgraph counting with distribution-free random graph analysis

TL;DR

This paper introduces a method for calculating robust subgraph counts in networks that depend only on mean and MAD, providing tight bounds applicable to various real-world networks.

Contribution

It develops a distribution-free approach for subgraph counting that only requires mean and MAD, offering sharp bounds and identifying extremal graph structures.

Findings

Robust bounds hold for real-world network data.

Dense power-law networks have higher subgraph counts than sparse ones.

The method provides scalable laws for subgraph growth with network size.

Abstract

Subgraphs such as cliques, loops and stars form crucial connections in the topologies of real-world networks. Random graph models provide estimates for how often certain subgraphs appear, which in turn can be tested against real-world networks. These subgraph counts, however, crucially depend on the assumed degree distribution. Fitting a degree distribution to network data is challenging, in particular for scale-free networks with power-law degrees. In this paper we develop robust subgraph counts that do not depend on the entire degree distribution, but only on the mean and mean absolute deviation (MAD), summary statistics that are easy to obtain for most real-world networks. By solving an optimization problem, we provide tight (the sharpest possible) bounds for the subgraph counts, for all possible subgraphs, and for all networks with degree distributions that share the same mean and MAD. We identify the extremal random graph that attains the tight bounds as the graph with a specific three-point degree distribution. We leverage the bounds to obtain robust scaling laws for how the numbers of subgraphs grow as function of the network size. The scaling laws indicate that sparse power-law networks are not the most extreme networks in terms of subgraph counts, but dense power-law networks are. The robust bounds are also shown to hold for several real-world data sets.