生产中的可解释人工智能：用于实时欺诈检测的神经符号模型

This study presents a neuro-symbolic model for real-time fraud detection that generates deterministic and human-readable explanations in 0.9 milliseconds as a by-product of the forward pass. This represents a 33× speedup compared to the traditional SHAP method, which requires 30 milliseconds to explain a fraud prediction and produces stochastic explanations that run after the decision, relying on a maintained background dataset at inference time. The model achieves identical fraud recall rates on the Kaggle Credit Card Fraud dataset, demonstrating its effectiveness and efficiency in production environments.