Prototype Learning with Concept Bottleneck Models for Interpretable Insurance Fraud Detection

Abstract Post-hoc explanation methods (SHAP, LIME, GNNExplainer) are effective at describing already-trained models, but their faithfulness cannot be guaranteed, and they are sensitive to input perturbations. Intrinsically interpretable models embed interpretability into the model architecture from the outset, fundamentally avoiding the faithfulness problem. This paper proposes an intrinsically interpretable insurance-fraud attribution framework built on prototype learning and concept bottleneck models (CBM). At the theory level, we establish interpretability axioms for prototype learning, derive a convergence rate for prototype estimation, and prove a performance-decomposition bound for the concept bottleneck. At the method level, we design a "concept--prototype" two-layer interpretable architecture: the bottom concept bottleneck decomposes the prediction into human-understandable high-level concepts (e.g., time anomaly, amount anomaly, relation anomaly), while the top prototype layer maps each case to a fraud-type prototype, providing example-level explanations. On simulated health-insurance claim data with a realistically weak and noisy fraud signal, the concept bottleneck model attains a test AUC of 0.806, within 1-2 points of black-box baselines, while predicting seven interpretable concepts at 0.801 accuracy and attributing fraud types at 36.4% versus a 25% random baseline. The core contribution is the first combination of prototype learning and concept bottleneck models for insurance fraud detection, offering an interpretable alternative that satisfies regulatory transparency requirements.