Uncovering hidden alliances in organized crime networks with machine learning: from node similarity to graph neural networks

Covert or dark networks, such as those formed by organized crime or illicit alliances, pose unique challenges for link prediction research due to their inherent secrecy and data incompleteness. In this study, we compare a suite of advanced methods for inferring missing links in covert networks, including classical node similarity indices, graph embedding approaches, and GNN-based methods. We evaluate them on two real-world datasets: the Lantia dataset (a 2021 network of alliances between criminal organizations) and the 2020 Bacrim dataset (an open dataset on alliances among Mexican criminal groups in 2020). Our experiments focus on how these algorithms handle missing or partially observed links, a common reality in covert networks. Performance is measured using Area Under the ROC Curve (AUC) and F1 scores. Although classical similarity measures and embedding algorithms offer meaningful predictions, GNN-based models, particularly the Graph Convolutional Network, consistently achieve near-perfect AUC and F1 scores. These results highlight that learning from the broader graph topology via GNNs can effectively uncover hidden ties in incomplete and inherently noisy covert network data. Our findings provide practical guidance for researchers, analysts, and policymakers looking to identify future or missing alliances in large, fragmented, and illicit networks.