Network Analysis for Organized Fraud Detection in Automobile Insurance With Graph Theory and Poisson Process

Abstract

Fraudulent claims in the automobile industry pose a significant threat to the financial stability of insurance companies and erode the trust between policyholders and insurers. Organized fraud, which involves intricate schemes and multiple parties, presents a substantial challenge in detection due to imbalanced datasets. While existing techniques such as over-sampling and under-sampling have been proposed to address this issue, they often lead to overfitting, loss of information, and reduced accuracy. However, assigning a suspicious label to each policyholder is more changeable, as it can identify potential risks and prevent fraudulent activities before they occur. In response to these challenges, we propose a novel heuristic approach called OrFGP that identifies suspiciously organized fraud groups within an accident network and provides credibility levels for accidents and associated individuals. We first demonstrate that car accidents follow a Poisson random process. We then combine this process with graph theory to introduce an accident network. In the network, our objective is to identify regular behavior between accidents, which, based on the stochastic nature of accidents, can indicate organized fraud. OrFGP uses probabilistic concepts in conjunction with local network connectivity metrics to evaluate the credibility of accidents and individuals. The results indicate that OrFGP outperforms state-of-the-art approaches, particularly in imbalanced datasets. In fact, OrFGP achieves an accuracy of 98% and improves the F1-score by at least 3%.