Klasifikasi Fraud Pada Transaksi Finansial Melalui Integrasi TabTransformer dan Oversampling Generatif CTGAN

Abstract

Extreme class imbalance in the BankSim dataset (1.2% fraud) is a major hurdle to building reliable detection systems. This study proposes the integration of the TabTransformer architecture with the Conditional Tabular GAN (CTGAN) oversampling technique to address majority class bias. Data quality evaluations indicate that CTGAN produces synthetic data with an overall quality score of 90.05% and a column pair correlation trend of 91.63%. Experimental findings prove the proposed model delivers superior performance, achieving an F1-Score of 85.34%, a Recall of 81.39%, and a Balanced Accuracy of 90.64%. These results significantly outperform the SMOTE technique, which recorded an F1-Score of 83.99% but suffered from probability calibration failure with an extreme optimal threshold of 0.98. In contrast, the CTGAN scenario demonstrates efficient decision threshold stability at 0.46. Validation through SHAP analysis confirms that engineered variables such as merchantRisk, custStepDiff, and amtZScoreByCat provide dominant contributions to model predictions. This research concludes that the synergy of the Data-Centric AI paradigm facilitates the creation of robust, precise, and highly accountable classification models for digital banking protection within financial transaction systems.