SURF: Sequential Undersampling-Refinement Framework for Two-Stage Anomaly Detection

Machine learning serves as an important instrument in financial fraud detection but extreme class imbalance in credit card datasets proves to be a significant obstacle for its effectiveness. Our framework - SURF incorporates two sequential steps where ROA-GAN performs anomaly-based undersampling at its first stage followed by a 1D CNN which carries out fraud classification. Using unsupervised learning SURF efficiently selects critical fraud patterns eliminating the need for synthetic oversampling. This research investigates under-sampling effects on class separation by utilizing PCA and UMAP and benchmarks against Random Under-sampling and NearMiss. The presented methodology shows its ability to detect anomalies accurately through testing performed on several datasets including the Kaggle European cardholders' dataset alongside a banking dataset and the annthyroid dataset.