A neuro-symbolic fraud detection system uses a symbolic rule layer to detect concept drift before F1 scores drop, without requiring labels. The key insight: the MLP path compensates for gradual drift in feature-fraud relationships, but the symbolic rule layer cannot adapt, making it a sensitive early-warning signal. A FIDI Z-Score metric — which normalizes feature contribution changes against their own window history rather than a fixed threshold — fires in 5 of 5 seeds for concept drift, averaging 0.40 windows ahead of F1. The system has clear blind spots: covariate drift is completely invisible to the symbolic layer (0/5 seeds), and prior drift fires after F1 by design. The alert system is ~50 lines of code using PyTorch and scikit-learn, requires no labels at inference time, and needs only 3 clean windows of history to operate.
Table of contents
TL;DR: What You Will Get From This ArticleThe Story So FarThree Ways Fraud Can ChangeThe Problem With the First Three MetricsThe Metrics: Building a Label-Free Drift Detection SystemResults: What Each Metric DidThe Alert Demo: Window 3Why FIDI Z-Score Sees It Before F1 DoesWhat This System Cannot DoResults SummaryBuilding ItV14: Three Articles, One FeatureWhat to Do With ThisThree Things That Will Catch You Using This Concept Drift Early Warning SystemClosingSeriesDisclosureReferencesSort: