Neuro-symbolic fraud detection systems can now identify concept drift at inference time using symbolic rules, eliminating reliance on labeled data. Researchers extended the FIDI (Feature Importance Drift Index) metric with Z-score normalization, creating FIDI Z, which detects anomalous shifts in rule behavior before traditional metrics like F1 scores degrade. In experiments on the Kaggle Credit Card Fraud dataset (284,807 transactions, 0.17% fraud rate), FIDI Z flagged drift in 5 of 5 simulated scenarios, sometimes weeks before performance metrics deteriorated. This contrasts with RWSS (Rule Weight Stability Score), which missed 3 of 5 seeds due to near-constant rule activations during early training phases.

The breakthrough lies in monitoring symbolic rules rather than input distributions or output labels. When concept drift occurs - such as when fraud patterns invert (e.g., V14's threshold flips from <−1.5σ to >−1.5σ) - FIDI Z detects the change through Z-score deviations in feature contributions to rules. This approach works purely on the symbolic layer, making it label-free and inference-time compatible. Covariate drift (input feature shifts) and prior drift (fraud rate changes) remain detectable via separate monitors, but concept drift requires this novel rule-based approach.

The system's alert mechanism triggers when FIDI Z exceeds 9.53 standard deviations from historical baselines, as seen in window 3 of the experiment. Unlike conventional methods requiring retraining schedules, this approach enables emergency retraining based on symbolic layer anomalies. The architecture builds on Hybrid Neuro-Symbolic AI, where neural networks learn rules like "IF V14 < −1.5σ → Fraud" autonomously, then monitors those rules for stability.

This development addresses a critical gap in production fraud systems, where concept drift often goes undetected until financial losses mount. By treating learned rules as canaries in coal mines, the method provides early warnings without sacrificing model explainability. The open-source implementation (50 lines of code) offers a practical path for financial institutions to enhance fraud detection resilience.