Citation

Caruana MA, Grech L. Commun. Stat. Case Stud. Data Anal. Appl. 2021; 7(4): 520-535.

Copyright

(Copyright © 2021, Informa - Taylor and Francis Group)

DOI

10.1080/23737484.2021.1986169

PMID

unavailable

Abstract

The risk of incurring financial losses from fraudulent claims is an issue concerning all insurance companies. The detection of such claims is not an easy task. Moreover, a number of old-school methods have proven to be inefficient. Statistical techniques for predictive modelling have been applied to detect fraudulent claims. In this article, we compare two techniques: Artificial neural networks and the Naïve Bayes classifier. The theory underpinning both techniques is discussed and an application of these techniques to a dataset of labelled automobile insurance claims is then presented. Fraudulent claims only constitute a small percentage of the total number of claims. As a result, datasets tend to be unbalanced. This in turn causes a number of problems. To overcome such issues, techniques which deal with unbalanced datasets are also discussed. The suitability of Neural Networks and the Naïve Bayes classifier to the dataset is discussed and the results are compared and contrasted by using a number of performance measures including ROC curves, Accuracy, AUC, Precision, and Sensitivity. Both classification techniques gave comparable results with the Neural network giving slightly better results than the Naïve Bayes classifier on the training dataset. However, when applied to the test data, the Naïve Bayes classifier slightly outperformed the artificial neural network.

Language: en

Keywords

Artificial neural networks; Fraud detection; Insurance claims; Naïve Bayes classifier; Unbalanced datasets