Hybrid Neuro-Symbolic Fraud Detection: Guiding Neural Networks with Domain Rules

A recent exploration into enhancing fraud detection through neuro-symbolic methods revealed that incorporating simple domain rules into the loss function initially appeared to significantly improve model performance on highly imbalanced datasets. However, subsequent testing across multiple random seeds and after fixing a threshold bug demonstrated that these gains were fragile and largely dependent on specific evaluation conditions, highlighting the pitfalls of relying solely on metrics like ROC-AUC in rare-event scenarios. This experience underscores the importance of robust evaluation strategies in fraud detection models, especially when integrating domain knowledge via rule-based constraints. While the domain rules provided a slight and consistent improvement in ranking metrics, the overall