Machine Learning-Based Vehicle Insurance Claim Fraud Detection and Risk Assessment

Abstract

Insurance fraud remains a major concern for the global insurance industry, resulting in substantial financial losses each year. Detecting fraudulent vehicle insurance claims is particularly challenging because of high dimensional data, imbalanced class distributions, and subtle behavioral patterns associated with fraud. This paper presents a unified machine learning framework designed to improve the reliability and consistency of fraud detection across vehicle insurance claims. The proposed framework is tested using the VehicleInsuranceClaim Fraud Detection Dataset from Kaggle which has 15420 samples and 33 attributes. Data preprocessing involves handling missing values, removing duplicates, performing One-Hot Encoding, Z-score normalization and class balancing using ADASYN, and splitting the data into 70:30 train-test sets with stratification. Two ensemble models, namely, Random Forest (RF) and Extreme Gradient Boosting (XGBoost), are created to classify fraudulent claims and to estimate risk. Experimental results demonstrate that the proposed RF achieves 97.95% accuracy, 97.98% precision, 97.91% recall, and 97.93% F1-score, while XGBoost achieves 97.96% accuracy, 97.94% precision, 97.89% recall, and 97.95% F1-score. The findings show that both of the suggested ensemble models provide very reliable classification performance and may facilitate effective, data-driven risk assessment and auto insurance fraud detection.