Relational Learning for Account-Level Money-Mule Detection

This approach scores accounts directly using HeteroGraphSAGE message passing on heterogeneous bank relational schemas. By treating foreign keys as edges, the model identifies criminal networks through multi-hop neighborhood analysis rather than individual transaction classification.