Evaluating Unsupervised Clustering for Credit Card Fraud Detection Under Extreme Class Imbalance

Abstract

The exponential rise in digital payments has elevated the difficulty of identifying fraudulent credit card activities, especially considering the extreme class imbalance inherent in financial records, where illicit actions typically represent a minuscule fraction of overall traffic. This research aims to assess the efficacy of unsupervised machine learning techniques for anomaly recognition within a public, anonymized dataset. The proposed methodology establishes K-Means clustering as a foundational baseline to understand broader structural patterns. Subsequently, Density-Based Spatial Clustering of Applications with Noise (DBSCAN) is deployed as the principal mechanism to isolate dense anomalous regions. To enhance processing speed and determine optimal hyperparameters, specifically epsilon and minimum points, initial tuning occurs on a representative data sample, followed by a comprehensive evaluation across the entire dataset. System performance is systematically evaluated through confusion matrix metrics, prioritizing the accurate classification of minority fraud cases. Experimental outcomes reveal that the DBSCAN algorithm attains an 88.5% detection rate for illegitimate transactions, substantially exceeding the 42.3% threshold achieved by the baseline model. Nevertheless, this heightened sensitivity introduces a trade-off, generating a 10.2% false-positive rate regarding legitimate operations. Ultimately, the density-based approach proves robust for isolating rare fraudulent behaviors in massive data environments, demonstrating substantial viability for practical deployment despite the slight increase in false alarms