Electronic Theses and Dissertation

Data tidak seimbang merupakan salah satu tantangan utama dalam klasifikasi transaksi fraud karena jumlah data fraud jauh lebih sedikit dibandingkan data non-fraud. Penelitian ini bertujuan menerapkan dan mengevaluasi kualitas data sintetis pada tiga metode penyeimbangan data, yaitu Synthetic Minority Over-sampling Technique (SMOTE), Conditional Tabular Generative Adversarial Network (CTGAN), dan Tabular Denoising Diffusion Probabilistic Model (TabDDPM), pada tiga dataset transaksi fraud dengan ukuran dan tingkat ketidakseimbangan yang berbeda. Evaluasi dilakukan terhadap kualitas data sintetis menggunakan TabSynDex Score serta performa klasifikasi menggunakan algoritma CatBoost dengan metrik balanced accuracy, F1-score, dan Area Under Curve (AUC). Hasil penelitian menunjukkan bahwa TabDDPM menghasilkan kualitas data sintetis tertinggi pada dataset D1 dengan TabSynDex Score sebesar 0,8518 dan D3 sebesar 0,8018, sedangkan pada D2 nilai tertinggi diperoleh CTGAN sebesar 0,7633. TabDDPM memberikan nilai balanced accuracy tertinggi pada D1 sebesar 0,8514 dan D3 sebesar 0,8096. Secara keseluruhan, TabDDPM paling konsisten dalam menghasilkan data sintetis berkualitas tinggi dan meningkatkan kemampuan deteksi fraud.

Imbalanced data is one of the main challenges in fraud transaction classification because the amount of fraud data is far smaller than non-fraud data. This study aims to apply and evaluate the quality of synthetic data across three data balancing methods: Synthetic Minority Over-sampling Technique (SMOTE), Conditional Tabular Generative Adversarial Network (CTGAN), and Tabular Denoising Diffusion Probabilistic Model (TabDDPM), on three fraud transaction datasets with varying sizes and imbalance levels. The evaluation assessed synthetic data quality using the TabSynDex Score and classification performance using the CatBoost algorithm, measured by balanced accuracy, F1-score, and Area Under the Curve (AUC). Results show that TabDDPM produces the highest synthetic data quality on dataset D1 with a TabSynDex Score of 0.8518 and D3 with 0.8018, while on D2 the highest value was achieved by CTGAN at 0.7633. TabDDPM achieved the highest balanced accuracy on D1 at 0.8514 and D3 at 0.8096. Overall, TabDDPM is the most consistent method in generating high-quality synthetic data and improving fraud detection capability.