New Study Compares Link Prediction Methods for Covert Networks – HP Wins for Groups

Researchers have long struggled with missing data in network analysis—whether in public email archives or covert communication channels. A new study by Moses Boudourides, posted on arXiv (2605.22606), tackles this by comparing three approaches for inferring missing links: classical dyadic link prediction (LP), hyperlink prediction (HP) that targets higher-order cliques, and exponential random graph models (ERGM). The key innovation is a common masking protocol that removes dyadic evidence induced by held-out hyperlinks, ensuring fair comparison. Across multiple datasets, LP using classical heuristics (e.g., common neighbors, Jaccard) excels at recovering dyadic edges. But when the goal is to predict group structures (cliques of 3+ nodes), HP—especially the CHESHIRE (CHEbyshev Spectral HyperlInk Predictor) algorithm—delivers superior performance. ERGMs provide a statistically interpretable complement, modeling tie probabilities conditioned on network dependencies.

The practical implications are significant. For cybersecurity analysts monitoring covert networks, HP can reveal hidden subgroups that LP would miss. The study also releases reproducible code, allowing others to apply and extend the benchmarks. The findings suggest that no single method dominates; instead, the choice depends on the inferential target—dyadic vs. hyperlink recovery. This comparative evaluation is a valuable reference for researchers and practitioners in social network analysis, intelligence, and fraud detection.