Performance Evaluation of Classical Machine Learning Models for Insurance Fraud Detection Under Severe Class Imbalance

Abstract

Insurance claim fraud continues to pose substantial financial and operational challenges for insurance companies worldwide, particularly in auto insurance, where fraudulent cases form a small but highly impactful portion of overall claims. One of the significant technical difficulties in detecting such fraud is the severe class imbalance in real-world insurance datasets, where legitimate claims vastly outnumber fraudulent ones. This study presents a systematic performance evaluation of classical machine learning (ML) classification models for insurance fraud detection under conditions of extreme class imbalance. The proposed framework focuses on widely used supervised learning algorithms, including Support Vector Machine (SVM), K-Nearest Neighbor (KNN), and Random Forest (RF), with an emphasis on understanding their behavior when trained on imbalanced data. To mitigate class imbalance bias, the Synthetic Minority Oversampling Technique (SMOTE) is applied to the dataset before model training. Model performance is evaluated using multiple metrics such as accuracy, precision, recall, and F1-score, which provide a more reliable assessment than accuracy alone in fraud detection scenarios. Experimental results demonstrate that ensemble-based methods, particularly Random Forest, achieve superior performance in identifying minority class fraud cases while maintaining stable overall classification accuracy. This research provides practical insights into selecting suitable classical ML models for insurance fraud detection, supporting the development of reliable decision support systems for insurance providers operating with imbalanced data.