Blind Spot Decomposition Theory: A Mathematical Framework for Characterising and Correcting Missed Fraud in Machine Learning Systems

We introduce Blind Spot Decomposition Theory (BSDT), which decomposes machine learning fraud detection failures into four near-orthogonal, measurable components: Camouflage, Feature Gap, Activity Anomaly, and Temporal Novelty. The Missed Fraud Likelihood Score (MFLS) quantifies each transaction's susceptibility to being missed and enables post-hoc correction without model retraining. Three self-calibrating weight methods (mutual information, Fisher variance ratio, Bayesian online updating) eliminate manual hyperparameter tuning. Validated on seven datasets spanning six domains (blockchain, e-commerce, credit card, aerospace, medical imaging, digit recognition) with 36+ model–dataset combinations, the best variant achieves mean F1 of 0.787 with 88.1% false positive reduction and recall recovery of up to +84.4 percentage points. A strict four-way split protocol confirms cross-domain generalisation with no label leakage. This work builds on AMTTP (DOI: 10.36227/techrxiv.177220113.33816607/v1), a pre-settlement blockchain compliance platform.