Att upptäcka kreditkortsbedrägerier: En utvärdering av Gaussian Mixture Model, Isolation Forest och Local Outlier Factor

Abstract

This thesis evaluates the performance of three unsupervised machine learning models — Gaussian Mixture Model (GMM), Isolation Forest (IF), and Local Outlier Factor (LOF) — for detecting fraudulent credit card transactions. Fraud detection presents significant challenges due to the continuously evolving nature of fraudulent behavior, extreme class imbalance in datasets, and the need to minimize both false positives and false negatives. The models were tested on a large, synthetic dataset designed to simulate real-world credit card transactions. Principal Component Analysis (PCA) and Multiple Correspondence Analysis (MCA) were applied for dimensionality reduction. Model performance was assessed using AUC-ROC, AUC-PR, precision, and recall in two separate scenarios: threshold optimization for 80% recall and threshold optimization for F1-score. Results indicate that while all three models are capable of achieving high recall, their precision is extremely low, leading to a high false positive rate. LOF demonstrated the best overall performance, suggesting that local deviations in transaction patterns may be more informative than global anomalies. However, it still lacked sufficient precision for practical implementation. These findings suggest that unsupervised methods alone are insufficient for effective fraud detection and highlight the necessity of hybrid approaches to improve fraud detection. Additionally, integrating fraud prevention strategies can reduce reliance on post-factum fraud detection and enhance overall security.