Fraud risk assessment in car insurance using claims graph features in machine learning

Abstract

The article proposes a process for claims assessment in car insurance, which makes it possible to calculate the fraud rate on the annual set of claims using a reduced set of attributes and graph vertex properties. This approach improves the security of insurance companies’ assets against fraudulent attacks. A method for constructing a claims graph and extracting additional features from it for evaluation is described. It is shown that in order to build a graph, it is not necessary to have data on the connection of the claim participants. Two tests were carried out on a real opensource datasets with labelling of fraudulent cases. The results of the first one show the increase in classification metrics when using attributes obtained from the graph. The application of the proposed approach resulted in doubling the area under the Precision-Recall curve. The experimental results demonstrated high quality metrics for fraud detection, with a Recall rate of 83.33% and a Specificity rate of 91.05%. The second test confirmed the possibility of determining the insurance fraud level based on decision rule, which includes the condition of claims being connected to each other. The rule is able to detect claim groups with a high concentration of fraud, in which every second participant is a fraudster.