Credit Card Fraud Detection with Unsupervised Learning: A Performance Analysis

Abstract

Digital financial ecosystems face growing threats from fraud as online transaction volumes skyrocket. In this study, we compare four unsupervised learning techniques—One-Class SVM, Local Outlier Factor (LOF), Variational Autoencoder (VAE) and Generative Adversarial Network (GAN)—for spotting fraudulent patterns in structured credit-card data. We train each model on a real-world transaction dataset and evaluate them using precision, recall and AUC-ROC. Our findings show that the VAE delivers the highest recall, successfully uncovering the largest share of fraud cases, while the GAN achieves the best balance between precision and recall. In contrast, LOF and One-Class SVM struggle when fraud examples are extremely rare. Overall, these results highlight the promise of deep, generative approaches for detecting subtle anomalies and suggest that VAEs and GANs can form the core of scalable, label-efficient fraud-detection pipelines.