With the rapid development of society, criminal activities have become increasingly complex and diversified, and mass
cases become more frequent, which makes the traditional criminal organization analysis method based on criminology theory and
case analysis no longer meet the needs of intelligence work. Therefore, utilizing deep learning technology to analyze and explore the
characteristics of criminal organizations has become an essential approach for data policing efforts. This article utilizes the Varia‐
tional Graph Autoencoder (VGAE) model to predict relationships among members of criminal organizations. The encoder part of the
model extracts structural features of criminal organizations and generates feature vectors, while the decoder part reconstructs the
criminal organization structure using vector inner products. This process helps predict whether there is a connection between two
members of the criminal organization. To evaluate the experimental performance of VGAE in relationship prediction tasks, tests are
conducted on the open-source crime network dataset Montagna. The experimental results show that VGAE demonstrates high predic‐
tive performance and can effectively identify potential relationships among members of criminal organizations.
