Using Credit Card Fraud Data to Evaluate Binary and One-Class Classification  
Author Joffrey L. Leevy


Co-Author(s) John Hancock; Taghi M. Khoshgoftaar; Azadeh Abdollah Zadeh


Abstract The rapid growth of e-commerce has led to a yearly increase in credit card fraud incidents, highlighting the need for effective fraud detection methods. Our study focuses on the Credit Card Fraud Detection Dataset, a popular dataset due to its real-world transaction data, high class imbalance, and potential to be a benchmark for credit card fraud detection. In this paper, we compare binary classification and one-class classification by evaluating six classifiers, with five being binary and four being Decision Tree-based. The metrics used are Area Under the Receiver Operating Characteristic Curve (AUC) and Area Under the Precision-Recall Curve (AUPRC). Our results show that binary classification is more effective for credit card fraud detection than one-class classification. We note that our One-Class Support Vector Machine (SVM) classifier produced the worst performance scores (0.9085 for AUC and 0.3788 for AUPRC). In addition, we determined that our Decision Treebased classifiers are the top performers. Out of this group of binary classifier ensembles, CatBoost yields the best performance scores (0.9834 for AUC and 0.8592 for AUPRC).


Keywords one-class classification, binary classification, credit card fraud, class imbalance
    Article #:  RQD28-402

Proceedings of 28th ISSAT International Conference on Reliability & Quality in Design
August 3-5, 2023