Ensemble of Synthetic and Balancing Techniques for Credit Card Fraud Detection

Abstract

Globally, the problem of credit card fraud continues to plague the finance sector, with losses estimated to reach over $43 billion in the near future. The proliferation of online payments has brought about new difficulties in the identification of fraud due to overwhelming class imbalance in fraud data. Most traditional machine learning approaches do not consider this imbalance, resulting in the development of biased classifiers with low generalizability. This dissertation proposes an innovative ensemble-based framework EA-CT (Ensemble ADASYN-CTGAN) that applies ADASYN and Conditional Tabular Generative Adversarial Networks (CTGAN) to improve accuracy and robustness of fraud detection. The study begins with the class imbalance challenge using ADASYN, an oversampling technique that creates synthetic samples for the minority class by focusing on more difficult instances to learn. After this step, the dataset is enriched with realistic synthetic fraud samples using CTGAN to ensure diversity and representativeness. The enriched dataset was analyzed using different classifiers. Results show that the fraud detection capabilities using EA-CT have been enhanced with the proposed framework. XGBoost reached the highest AUC-ROC score of 95.20% and recall of 87.99%. ANN achieved the highest accuracy of 88.86%. KNN exhibited an impressive F1-score at 87.05%.  In comparison with other advanced approaches, EA-CT was proven to outperform, in lowering false negative rates as well as improving detection rates.