Prototype Learning with Concept Bottleneck Models for Interpretable Insurance Fraud Detection

Technical Summary: Prototype Learning with Concept Bottleneck Models for Interpretable Insurance Fraud Detection

Problem Statement
The paper addresses the critical need for interpretability in high-stakes insurance fraud detection. Current state-of-the-art approaches often rely on black-box models (e.g., XGBoost, Random Forests) paired with post-hoc explanation methods like SHAP or LIME. The author argues that these post-hoc methods suffer from a lack of faithfulness; they approximate the model's decision boundary via local surrogates that may diverge from the true logic, are sensitive to input perturbations, and offer no guarantee that the explanation reflects the model's internal state. In the context of insurance, where errors can lead to significant financial loss or wrongful blacklisting of customers, the inability to verify the "truth" of an explanation is a fundamental limitation.

Methodology
The proposed solution is an intrinsically interpretable framework that combines Prototype Learning with Concept Bottleneck Models (CBM). The architecture operates on a "concept–prototype" two-layer design:

1. Concept Bottleneck Layer (Bottom):

   • Instead of mapping raw features directly to a fraud probability, the model first predicts a vector of seven human-understandable high-level concepts: time anomaly, amount anomaly, text anomaly, relation anomaly, history anomaly, cooperation anomaly, and a composite anomaly score.
   • These concepts are defined within a specific ontology derived from domain interviews.
   • The task predictor (the final classification layer) is a linear model ($g$) that maps these predicted concepts ($\hat{c}$) to the fraud probability. Because the layer is linear, the weights directly indicate the importance of each concept.

2. Prototype Layer (Top):

   • Built upon the concept embeddings, this layer assigns each claim to the nearest "fraud-type prototype" (e.g., short-term claim, forged document, overtreatment).
   • This provides example-level explanations, allowing auditors to see that a specific claim resembles a known fraud prototype based on its concept activation.

3. Training Strategy:

   • The model is trained in three stages: (i) pretraining concept predictors using weakly supervised, rule-generated concept labels; (ii) training the task predictor using fraud labels; and (iii) joint fine-tuning using a composite loss function that includes classification loss, concept loss, prototype loss, and a concept–prototype consistency loss ($L_{cp}$) to ensure the concept vector aligns with the assigned prototype.

Theoretical Contributions
The paper establishes a theoretical foundation for this architecture, addressing three specific gaps identified in the literature:

• Interpretability Axioms: Formal definitions for prototype assignment, order preservation, and exemplar interpretability are provided.
• Convergence Rate: A theorem is derived proving that the prototype estimation converges at a rate of $O(\sqrt{\log(1/\delta)/N})$, assuming the embedding is Lipschitz continuous.
• Performance Decomposition: A bound is proven for CBMs showing that the total error ($\epsilon_{CBM}$) is bounded by the task predictor error on true concepts plus a term proportional to the concept prediction error ($\epsilon_h$). This theoretically justifies that if concepts are predicted accurately, the CBM performance approaches that of a black-box model.
• Information-Theoretic Basis: The framework utilizes the data-processing inequality to formalize the trade-off between concept informativeness and compactness.

Experimental Results
Due to the proprietary nature of real insurance data, the study utilizes a simulated dataset of 5,000 claims featuring a weak, noisy fraud signal and realistic label noise (3%).

• Predictive Performance: The Concept Bottleneck Model achieved a test AUC of 0.806. This is statistically indistinguishable from black-box baselines (Random Forest: 0.814; Gradient Boosting: 0.819) and actually outperformed the cross-validated AUCs of the black-box models (0.801 and 0.788).
• Concept Accuracy: The model predicted the seven interpretable concepts with a mean accuracy of 0.801 and an AUPRC of 0.438 (significantly above the 0.123 chance level).
• Prototype Attribution: The prototype layer attributed fraud types with 36.4% accuracy, a 46% relative improvement over a 25% random baseline, demonstrating that the concept embeddings retain genuine structural information about fraud types.
• Trade-off: The framework sacrifices at most 1–2 percentage points in AUC compared to black boxes while upgrading auditability from "low" (post-hoc) to "high" (intrinsic).

Significance and Claims
The paper positions itself as the first combination of prototype learning and concept bottleneck models specifically for insurance fraud detection. Its primary significance lies in offering a "white-box" decision process that satisfies regulatory transparency requirements without a substantial sacrifice in predictive power.

The author explicitly frames this as a feasibility study. They modestly acknowledge that:

• The concept ontology and numerical results must be re-validated on real, de-identified claims before deployment.
• The theoretical convergence bounds are analytical and require finite-sample verification.
• The concept–prototype consistency loss has not yet been ablated.
• Future work could explore automatic concept discovery to reduce reliance on manually curated ontologies.

Ultimately, the paper argues that in high-stakes domains like insurance, the trade-off of a marginal performance drop for guaranteed, intrinsic interpretability is strongly favorable.