AN ENSEMBLE OF UNSUPERVISED DEEP LEARNING MODELS FOR CREDIT CARD FRAUD DETECTION

Abstract

As the use of credit cards for transactions increases, so do frauds too. Global networking offers criminals just as many new opportunities as it do for ordinary users. The good news is that technology for detecting and preventing credit card fraud is likewise getting better with time. Machine Learning models have been incorporated in this field to reduce the cost and time it takes in analyzing credit card transactions and detect the fraudulent ones. However, current machine learning models primarily rely on supervised learning, which requires labeled data and struggles with new fraud patterns and class imbalances. Despite this, unsupervised machine learning models tend to perform worse than the supervised learning models in detecting credit card fraud, which makes them less likely to be a go-to option in this field. This research proposes an ensemble of unsupervised deep learning models, specifically Deep Auto encoders (AEs) and Recurrent Neural Networks (RNNs), to improve the performance of unsupervised models for credit card fraud detection. This research employs a quantitative methodology, utilizing secondary data from credit card transactions. The methodology involves preprocessing the data, followed by training and evaluating the models. The individual models achieved AUC scores of 0.96 and 0.82 respectively, while the ensemble model from their combination achieved an AUC score of 0.95.