Credit Card Fraud Detection using novelty detection techniques

Abdullah Barayan


Supervised by Yuhua Li; Moderated by Alun D Preece

Credit card fraud is a growing issue affecting the financial industry and cardholders worldwide, resulting in billions of dollars in losses annually. A considerable amount of research has been done to detect credit card fraud. Most proposed machine and deep learning fraud detection approaches use a supervised learning methodology that requires appropriately labelled and balanced training datasets. Organising these datasets takes a great deal of time and effort. It is difficult for these approaches to deal with imbalanced or unlabeled datasets. However, few studies have employed unsupervised novelty detection methods, focusing on addressing the imbalance and lack of unlabeled data issues that supervised methods suffer from them. On the other hand, high data dimensionality has not received much attention, mainly when using unsupervised techniques, which are challenging, unlike supervised approaches. This study proposes unsupervised novelty detection techniques with the help of dimensionality reduction methods to detect credit card fraud. The European cardholder dataset has been used to evaluate this approach. In this study, One-Class SVM, Isolation Forest, and Autoencoder novelty detection techniques have been utilised with the help of Autoencoder to reduce dimensionality for the detection of credit card frauds, and the performance of these techniques are compared mainly based on the AUC and FNR. The experimental results show that using unsupervised novelty detection techniques and Autoencoder as a feature reduction method has a promising potential to detect credit card fraud. The results also demonstrate that the proposed approach deals effectively with imbalanced and unlabeled datasets and can reduce training and prediction time. AE-AE achieved the highest AUC score and the lowest FNR, with 93% and 10.169%, respectively.

Final Report (20/10/2022) [Zip Archive]

Publication Form