Every day, BNP Paribas Personal Finance (BNPP PF) receives thousands of requests for installment payments, which are processed in a matter of seconds. This speed is what defines the quality of the service, but it also makes it vulnerable: fraudsters exploit it to obtain financing with no intention of repaying it. To detect these cases, BNPP PF links the requests in a graph, where each request becomes a node connected to others that share identifying information. This relational structure is a treasure trove of information that has yet to be fully tapped. This thesis demonstrates how to leverage it to better detect fraud, while meeting the service’s real-time constraints. It makes three contributions. First, a multi-level identity resolution method that reveals links between applications, including a method adapted for split payments. Second, a hybrid architecture combining graph neural networks and ensemble methods to exploit both individual and relational signals. Finally, the publication of an anonymized and distorted dataset derived from real-world data. The results confirm the value of the graph approach and the hybrid coupling, the benefits of which increase when relational information is lacking in the data.