A Comparative Study of Generative Oversampling Methods for Imbalanced Fraud Classification

Park, Jisoo (2025) A Comparative Study of Generative Oversampling Methods for Imbalanced Fraud Classification. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (1MB) | Preview
Preview	PDF (Configuration Manual) Download (735kB) | Preview

Abstract

Credit card fraud detection poses a significant challenge due to extreme class imbalance, where fraudulent transactions make up less than 0.2% of the data. While generative oversampling methods—particularly Generative Adversarial Networks (GANs)—have shown promise in addressing this issue, many existing approaches lack classifier-awareness, rely on limited evaluation metrics, and fail to align synthetic data with real-world decision boundaries.

This study proposes an Enhanced Conditional GAN that integrates feature matching loss, gradient penalty, and conditional generation to create classifier-aligned synthetic fraud samples. We conduct a comprehensive comparison of this model against traditional oversample method (SMOTE, ADASYN) and generative models (CTGAN, TVAE, Gaussian Copula) using a real-world imbalanced fraud dataset.

Results show that the proposed model achieves an F1-score of 0.8619, PR-AUC of 0.8591, and ROC-AUC of 0.9620, achieving competitive PR-AUC and competitive precision among all generative approaches, and closely matching the no-oversampling baseline in overall balance. Statistical and visual analyses reveal that the Enhanced GAN prioritizes decision-relevant sample generation over exact distributional matching, resulting in better classification outcomes.

This work highlights the importance of classifier-aware oversampling and provides a robust framework for evaluating synthetic data quality in high-risk, imbalanced domains such as fraud detection. Future directions include real-time deployment, model compression, and benchmarking against diffusion and LLM-based tabular generators.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Subhnil, Shubham UNSPECIFIED
Subjects:	H Social Sciences > HA Statistics H Social Sciences > HG Finance H Social Sciences > HG Finance > Credit. Debt. Loans.
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	02 Jul 2026 14:39
Last Modified:	02 Jul 2026 14:39
URI:	https://norma.ncirl.ie/id/eprint/9445

Actions (login required)

View Item