Comparison of machine learning models in fake damage detection in the insurance industry

The issue of fraud is a very important and important issue in insurance. Abuse can be explained as false claim claims in insurance. False damage declarations cause a lot of costs for insurance companies. In order to eliminate these negative effects, damage teams in insurance companies need to spend a lot of workforce to separate abusive and non-abusive files among hundreds of damage reports. In addition, as the person committing the abuse develops different techniques, it becomes more difficult to catch the abuse. All these reasons can lead to material and moral losses for insurance companies. In this study aims to develop machine learning models to assist damage teams in the insurance industry in predicting fraudulent claims and to determine which machine learning methods are more suitable for this problem. In this study, automobile and traffic insurance data of an insurance company serving in our country were used. K-NN, Decision tree, Logistic Regression, CART and Artificial Neural Networks algorithms were used in model creation. In addition, the results of these algorithms were compared again using the feature selection method with better features. As a result of all these methods, it is seen that artificial neural networks are the most successful models with 86.4 percent using feature selection, CART model with 86.1 percent and logistic regression model with 85.7 percent. It is thought that the results obtained from the study may help the damage teams in insurance companies to detect abusive damage files.