Document Type : Original Research Paper

Authors

• Asma Hamzeh ¹
• Mohammad Javad Nadjafi-Arani ²

¹ Assistant professor, Department of New Insurance Technologies, Insurance Research Institute, Tehran, Iran

² Associate Professor, Department of Computer Science, Mahallat institute of higher education, Mahallat, Iran

Abstract

BACKGROUND AND OBJECTIVES: Insurance fraud presents a persistent challenge within the insurance industry, leading to substantial financial losses and eroding public trust. Financial institutions are actively seeking accurate methods to identify the activities of fraudsters and scammers. Due to its direct effect on serving the clients of institutions, this will lead to the reduction of operating costs, gaining the trust of other insurers, and maintaining and improving the market share of insurers as reliable financial service providers. One of the most prevalent forms of fraud occurs in auto insurance, where organized and opportunistic fraudulent activities are rampant. Fabricated accidents, especially those involving groups, staged injuries, and orchestrated scenes, are among the common fraudulent practices in this realm. Opportunistic fraud is typically committed by an individual who simply seizes an opportunity to inflate a claim or receive an exaggerated estimate for damages or repairs from their insurance companies. In contrast, professional fraud is often carried out by organized groups. These rings typically target multiple fake identities, organizations, or even brands. These criminal networks frequently rely on insiders to help them defraud companies, simultaneously using various schemes. Although the amounts involved in professional fraud cases are much larger, they occur less frequently than opportunistic insurance fraud. Combating insurance fraud is a challenging issue. Most traditional systems can detect opportunistic fraud; however, due to the significant financial losses involved, insurance companies are particularly focused on identifying organized fraud rings. Consequently, insurers need to adopt advanced technologies and sophisticated systems to effectively address this problem.
METHODS: Network analysis is a valuable technique for fraud detection, enabling the evaluation of communications among individuals and entities (both real and legal) to uncover new dimensions of these interactions. This paper introduces a mathematical model, based on graph theory, to identify suspicious clusters associated with organized fraud. A network, termed the “accident network,” was first introduced using graph theory in this research. This network demonstrates characteristics of a random graph. Subsequently, suspicious clusters within this network are identified using a graph theory-based algorithm. The occurrence probability of such clusters in a random accident network is then examined by defining a binomial distribution over its edges.
FINDINGS: This process assigns a label (indicating fraudulent or non-fraudulent) to each accident and individual involved. Given the algorithm's structure and complexity, the proposed method is capable of efficiently analyzing large datasets.
CONCLUSION: Insurance fraud is an act committed to defraud insurers for financial gain. Insurance fraud has existed since the formation of commercial enterprises and has so far imposed billions of dollars in costs on insurance companies annually. Insurance fraud comes in various forms and occurs in all insurance domains, covering a wide range of claims from exaggerated ones to fabricated accidents and damages.Auto insurance fraud, particularly the organized fraud studied in this research, is often carried out through group structures. This structure leads to significant cost increases for insurers and consequently higher insurance premiums. Today, given the necessity of fraud detection in various fields, data mining and machine learning techniques such as artificial neural networks, fuzzy logic, and genetic algorithms have become common tools for fraud detection due to their high capabilities in modeling and navigating complex problems.Another tool used for detecting organized fraud is graph theory. In this approach, the problem is first mathematically modeled. This means the accident network is first modeled as a graph, and then organized fraud is detected using available tools. Then, computer science concepts are utilized to more precisely identify networks suspected of fraud. More accurately, in structures like a country's accident data, the amount of available data is very large finding relationships among them is quite difficult.While using tools like data mining, machine learning, neural networks, fuzzy logic, genetic algorithms, etc., whose main purpose is to find relationships among data, is very useful, they have some shortcomings. These tools, if meta-heuristic algorithms are used, will have inaccuracies or overfitting in imbalanced data. In heuristic algorithms, finding relationships among large amounts of data has very high computational complexity, which in some cases may take weeks or more to execute.In this research, the researchers have tried to address this shortcoming using mathematical models while accurately examining the probability of suspicious events. Therefore, in this research, first, the accident network was modeled using graph theory, and then it was shown that this model is a random process, and the presence of regular elements in the model indicates sets of vehicles suspected of fraud. Subsequently, based on an algorithm for finding suspicious subgraphs written as an m-file script in MATLAB, suspicious vehicles were extracted from all vehicles.Finally, it was proven that the accident network is a Poisson process, and its occurrence probability can be determined. This reasoning, based on graph modeling structure, helps assign a credibility degree to each accident and each vehicle regarding suspicion of fraud. For future research, it is suggested that a more extensive network, including all stakeholders in organized fraud, should be created and examined. More precisely, the network should examine the label assignment of main beneficiaries who profit from an accident based on their profit shares. Specifically, future work should investigate labeling main beneficiaries based on their profit shares from an accident, enabling insurers to adopt tailored policies for various stakeholders (e.g., policyholders, vehicle occupants, repair shops) to reduce financial losses and restore public trust.

Keywords

• Car insurance
• Graph theory
• Poisson distribution
• labeling

Main Subjects

• Financial / Applied Mathematics