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

A hybrid neuro-symbolic approach is proposed for credit card fraud detection on highly imbalanced datasets. This technique enhances a standard neural network by adding a differentiable rule-based loss function to the training objective. The rule loss penalizes the model for not assigning high fraud probability to transactions with suspicious characteristics, such as unusually large amounts or atypical PCA signatures. While the hybrid model shows a marginal improvement in ROC-AUC over a pure neural baseline, the analysis emphasizes that on imbalanced data, the threshold selection strategy is as critical as the model architecture for meaningful evaluation.