Edge sampling using network local information

TL;DR

This paper investigates a simple local-information-based edge sampling method, demonstrating that strong local connectivity ensures the sampled network retains key spectral properties, with implications for efficient sampling of complex networks.

Contribution

It provides a theoretical analysis linking local connectivity to spectral properties of sampled networks, filling a gap in understanding heuristic edge sampling methods.

Findings

Sampled networks exhibit strong spectral properties under sufficient local connectivity.

Global parameters like clustering coefficient relate to spectral guarantees.

Conditions are identified for efficient sampling of random networks and hypergraphs.

Abstract

Edge sampling is an important topic in network analysis. It provides a natural way to reduce network size while retaining desired features of the original network. Sampling methods that only use local information are common in practice as they do not require access to the entire network and can be parallelized easily. Despite promising empirical performances, most of these methods are derived from heuristic considerations and therefore still lack theoretical justification. To address this issue, we study in this paper a simple edge sampling scheme that uses network local information. We show that when local connectivity is sufficiently strong, the sampled network satisfies a strong spectral property. We quantify the strength of local connectivity by a global parameter and relate it to more common network statistics such as the clustering coefficient and network curvature. Based on this result, we also provide sufficient conditions under which random networks and hypergraphs can be sampled efficiently.