PriSec-CN: a leakage-controlled link-prediction benchmark for covert criminal networks

PriSec-CN is a leakage-controlled link-prediction benchmark for covert criminal networks, together with the complete analysis pipeline that produced it. The benchmark comprises 104,335 candidate actor pairs derived from 37 published covert criminal networks (3,828 actors, 18,962 ties) spanning terrorism, drug trafficking, mafia-type organisations, street gangs, corruption, organised crime and fraud. Every candidate pair is scored on a graph from which that pair has been deleted, so that no feature can encode its own label. Each pair carries eighteen relational scores, four structural attributes and a ground-truth label. The pipeline is seeded and deterministic: running the scripts in order reproduces every table and figure of the accompanying article. Provenance. The source graphs are network reconstructions already published by other researchers, principally the Covert Networks collection curated by the Mitchell Centre for Social Network Analysis, University of Manchester. Several of those collections identify actors by name; those names are part of the public record of the cases concerned and are not reproduced here, node labels in this release being within-network codes. Users of this benchmark should cite the original sources listed in the network inventory alongside this deposit. Intended use. This resource supports methodological research on the recovery of unobserved ties. It does not support, and must not be used for, the attribution of criminal association to identifiable individuals.