Graph AI for Criminal Network Analysis

Abstract

Criminal Networks represent a class of complex, dynamic, and covert systems whose analysis poses significant challenges to traditional graph-based and Machine Learning

(ML) methods. Existing approaches predominantly rely on single-layer static represen- tations, which fail to capture the heterogeneous and multi-relational nature of criminal

interactions, and suffer from accumulated uncertainty inherent in incomplete law en- forcement data. This thesis proposes a framework that combines Heterogeneous Graph

Neural Networks with Fuzzy Logic techniques to address two fundamental tasks in criminal network analysis: node classification and link prediction. The heterogeneous graph representation preserves the semantic distinctions between multiple relational

layers, including social, operational, and communication ties, while fuzzy logic is em- ployed to enhance the reliability of gathered evidence by modeling and mitigating the

uncertainty inherent in criminal network data. The proposed framework is evaluated on two real-world criminal network datasets, namely the Noordin Top Terrorist Network and the Sicilian Mafia Network. Experimental results demonstrate the effectiveness of the approach, achieving a Macro-F1 score of 71% on the node classification task, outperforming existing single-layer and homogeneous graph-based baselines.