Evaluating The Effectiveness of Augmentation and Classifier Algorithms for Fraud Detection: Comparing CGAN and SMOTE with Random Forest and XGBoost

Abstract

Fraud detection in imbalanced datasets, where fraudulent transactions represent a small fraction of total data, presents a major challenge for machine learning models. Traditional classifiers often perform poorly in such scenarios due to their bias toward the majority class. This study investigates the effectiveness of two data augmentation techniques, Synthetic Minority Over-sampling Technique (SMOTE) and Conditional Generative Adversarial Networks (CGAN) in improving fraud detection performance. Both methods are applied to balance the dataset, and their impact is evaluated using two classifiers: Random Forest (RF) and XGBoost. The models are tested across three versions of the dataset: the original imbalanced data, the SMOTE-augmented data, and the CGAN-augmented data. Evaluation metrics include accuracy, precision, recall, F1 score, and ROC-AUC. Results indicate that both augmentation techniques enhance the models' ability to detect fraudulent transactions compared to the original dataset. Notably, CGAN outperforms SMOTE in terms of recall and F1 score, suggesting its ability to generate more diverse and realistic synthetic samples. While SMOTE creates new samples through interpolation, CGAN uses an adversarial process involving a generator and a discriminator, resulting in more complex data representations. The study also finds that XGBoost combined with CGAN yields the highest performance, effectively capturing intricate fraud patterns. In contrast, SMOTE, though beneficial, shows limited capacity in improving recall. This research highlights the importance of advanced augmentation techniques like CGAN in addressing class imbalance and improving fraud detection systems. It also opens pathways for future exploration of deep learning-based augmentation and classification methods in fraud detection.