Missing Links in Public Email and Covert Networks: A Comparative Evaluation of Link Prediction, Hyperlink Prediction, and ERGM Estimation

TL;DR

This paper compares link prediction, hyperlink prediction, and ERGM estimation methods for inferring missing links in partially observed networks, highlighting their strengths in different network structures.

Contribution

It provides a systematic, reproducible evaluation of these methods, clarifying their applicability under various network missingness scenarios.

Findings

LP is effective for dyadic link recovery.

HP, especially CHESHIRE, improves higher-order structure inference.

ERGMs offer interpretable dependence-based insights.

Abstract

We study missing-link inference in partially observed networks by systematically comparing dyadic link prediction (LP) with hyperlink prediction (HP) and an estimation-based ERGM comparator. LP serves as the primary baseline, using classical heuristics computed on the observed graph. HP extends this framework by scoring candidate higher-order structures (cliques) via lifted dyadic scores and via the CHEbyshev Spectral HyperlInk pREdictor (CHESHIRE). All methods are evaluated under a common masking protocol that removes dyadic evidence induced by held-out hyperlinks to ensure comparability. Across public email and covert-network datasets, LP remains strong for dyadic recovery, while HP -- particularly CHESHIRE -- provides gains when the inferential target is higher-order group structure. ERGMs offer an interpretable dependence-based complement through conditional tie probabilities. The contribution is a comparative, reproducible evaluation clarifying when LP, HP, and ERGM estimation are most appropriate under network missingness.