Enhancement of catboost algorithm applied on insurance fraud detection

ABSTRACT: Detecting insurance fraud remains one the main areas of concern, causing great financial loss to the insurers and an increase in premium rates for the genuine policyholders. In most situations, traditional fraud detection systems have a hard time achieving a proper balance between their accuracy, generalization, and interpretability. The focus of the research is to enhance the CatBoost algorithm for fraud detection while addressing hyperparameter sensitivity, overfitting, and diminished explainability. To achieve that, the proposed method is the integration of Sherpa to enhance the hyperparameter selection, apply Stratified Bootstrapping to avoid overfitting, and implement SHAP (Shapley Additive Explanation) for enhanced transparency of the issues found on the model and it undertakes the evaluation of these proposed refinements on a publicly accessible fraud detection dataset. Results have shown a significant improvement in fraud classification precision, increasing from 83% to 96%, whereas the overfitting gap only decreases from 13.13% to 3.5%. Besides, the model also enhances the recall of fraud detection, thus becoming a more reliable and explainable option. The new framework provides insights into key fraud drivers via SHAP, allowing insurance companies to make better data-informed decisions, eventually resulting in reduced fraudulent claims and financial exposures.

Filipiniana