Leveraging Unsupervised Machine Learning and Deep Learning for Enhanced Card Fraud Detection

Abstract

The modern financial ecosystem is highly reliant on digital transactions, and despite the convenience, they have a significant flaw with high rates of fraudulent activities. Card fraud poses a severe risk to consumers and financial institutions, leading to economic losses, data breaches, and trust erosion. Traditional fraud detection methods mainly use supervised learning models or heuristics that fail to keep up with evolving fraud patterns and often have too many false positives and negatives. This research study aims to develop a standard detection framework to identify fraudulent transactions without relying on labelled fraud data. This is done by leveraging unsupervised learning techniques, namely isolation forest, One-Class SVM and deep learning-based Autoencoder using the OPENML ID: 45955 dataset referred to as ‘dataset.csv’. We show that unsupervised methods can be effective when fraudulent labels are sparse, while deep learning approaches offer notable improvements in balancing detection (recall) and correctness (precision). Empirical results reveal that Autoencoders yield a superior F1-score compared to Isolation Forest and OneClass SVM, with an Accuracy of up to 92.51%, Precision of 56.01%, Recall of 66.48%, and ROCAUC of 0.8074.