Online payment fraud has been steadily increasing in recent years. Our focus is on installment payments for e-commerce, which pose a significant risk of customers failing to repay the full amount owed. To manage this risk, BNP Paribas Personal Finance has developed a system that combines graph databases and artificial intelligence, achieving a 20% reduction in fraud. In this article, we propose an extension of this system using a graph neural network (GraphSAGE) combined with an ensemble method (such as Random Forest or XGBoost). We demonstrate the performance improvements of this combined approach over the initial system using a real anonymized dataset made available to the community.