Comparing methods for predicting missing links in email and covert networks

Networks are rarely observed in full. When some connections between people or organisations go unrecorded, researchers must infer them from the incomplete graph that remains. A new study in EPJ Data Science systematically compares three families of methods for this task of missing-link inference: dyadic link prediction, hyperlink prediction, and estimation of exponential random graph models (ERGMs), evaluated across public email and covert-network datasets.

Link prediction serves as the primary baseline in the evaluation. It relies on classical heuristics computed directly on the observed graph to score which unobserved dyadic ties are most likely to exist. Hyperlink prediction extends this framework to higher-order structures: instead of considering only pairs of nodes, it scores candidate higher-order group structures such as cliques. This is done in two ways, by lifting dyadic scores to the higher-order level and through a dedicated method called the CHEbyshev Spectral HyperlInk pREdictor, or CHESHIRE.

To make the comparison fair, all methods are assessed under a common masking protocol. The protocol removes dyadic evidence induced by held-out hyperlinks, so that each method faces the same partially observed network and the results remain comparable across approaches and datasets.

The findings draw a clear division of labour. For recovering missing dyadic links, classical link prediction remains strong across the public email and covert-network data. When the inferential target is higher-order group structure rather than individual ties, hyperlink prediction provides gains, with CHESHIRE performing particularly well in that setting. ERGMs offer a complementary route: because they are estimated statistical models of network dependence, they can express the probability of a tie conditional on the rest of the observed network, giving an interpretable benchmark grounded in dependence assumptions.

The study’s contribution is a comparative, reproducible evaluation that clarifies when each of the three approaches is most appropriate under network missingness. Rather than crowning a single winner, it shows that the right choice depends on the inferential target: dyadic recovery, higher-order group structure, or interpretable conditional tie probabilities. For analysts working with incomplete social, communication or covert networks, the results provide a practical guide to matching method to task.

Subject of Research: Comparative evaluation of link prediction, hyperlink prediction and ERGM estimation for missing-link inference in partially observed networks

Article Title: Missing links in public email and covert networks: a comparative evaluation of link prediction, hyperlink prediction, and ERGM estimation

Article References: Boudourides, M. (2026). Missing links in public email and covert networks: a comparative evaluation of link prediction, hyperlink prediction, and ERGM estimation. EPJ Data Science. https://doi.org/10.1140/epjds/s13688-026-00708-4

Image Credits: AI Generated

DOI: 10.1140/epjds/s13688-026-00708-4

Keywords: link prediction, hyperlink prediction, ERGM, network missingness, higher-order networks, CHESHIRE, covert networks, email networks, cliques, network inference, Missing, links