An Enhanced Focal Loss Function to Mitigate Class Imbalance in Auto Insurance Fraud Detection with Explainable AI

Imbalanced data is a real problem in fraud detection: The paper demonstrates that standard machine learning algorithms struggle when fraudulent claims are rare compared to legitimate ones, causing them to miss fraud cases entirely—a critical issue since even small percentages of fraud represent significant financial losses in insurance.

A staged training approach outperforms simpler methods: The proposed three-stage framework (using different versions of focal loss sequentially) achieved better detection of fraudulent claims compared to conventional approaches, suggesting that how models are trained matters as much as what data they use.

Explainability is built into the solution, not added later: The paper emphasizes that SHAP analysis provides "interpretable feature-attribution patterns," meaning fraud investigators and actuaries can understand why the system flagged a claim as suspicious—essential for responsible deployment in high-stakes financial decisions.

Mathematical rigor supports practical reliability: By deriving specific conditions for when their loss function remains mathematically stable, the authors ensure the model trains reliably rather than unpredictably, reducing the risk of unexplained failures in production systems.

Transparency enables accountability in insurance decisions: By combining improved fraud detection with explainable AI, this approach allows insurance companies to justify their fraud determinations to customers and regulators, addressing fairness concerns in automated decision-making.