Deep Learning Method for Detecting Fraudulent Motor Insurance Claims Using Unbalanced Data

This study investigates the use of deep learning methods in detecting fraudulent claims in the motor insurance industry. To help combat these problems, many researchers have come out with various machine learning algorithms to detect fraudulent claims. However, these claims suffer from the problem of imbalanced datasets. This is because fraudulent claims cases are rare compared to genuine ones. Therefore, the training data will have less instances of fraud cases as compared to the genuine cases. On the other hand, classification algorithms tend to learn more on the majority classes compared to the minority ones, and consequently fail to perform well on relatively rare cases of the minority classes. This research aims at using sampling techniques in the preprocessing stage to balance the datasets and minimize the issue class imbalances in deep learning. The outcome of deep learning with sampling is compared to unsampled cases. Results show that the deep learning approach performs perfectly well on the balanced datasets and the performance on imbalanced datasets is average.