Automobile insurance fraud detection

Abstract

One of the most concerning issues for insurance companies is the risk of making financial losses from fraudulent claims. Detection of fraud is not an easy task and most old-school methods prove to be inefficient, with costly investigations resulting in further losses for the company. Statistical techniques for data mining and predictive modelling have been effectively applied by many researchers to uncover fraudulent claims. Fraud detection in this case will be formulated as a binary classification problem. Particularly, the Artificial Neural Network known as the Multi-layer Perceptron and the Naive Bayes classifier are popular in this area of study. By studying the relevant theory behind these two techniques, both classifiers will be applied to a data set of labelled automobile insurance claims to study their suitability to this application. The classifiers are compared by a number of performance measures including ROC curves and AUC, while attempting to minimise the number of false alarms and maximise the number of hits.