<![CDATA[Abstrak - Deteksi penipuan transaksi digital menghadapi tantangan besar berupa ketidakseimbangan kelas data (imbalanced data), di mana jumlah transaksi normal jauh melebihi transaksi penipuan. Ketimpangan ini menyebabkan algoritma machine learning cenderung berpihak pada kelas mayoritas. Penelitian ini bertujuan mengevaluasi kinerja algoritma Random Forest sebelum dan sesudah penerapan Synthetic Minority Over-sampling Technique (SMOTE) menggunakan dataset Credit Card Fraud Detection. Untuk mencegah kebocoran data (data leakage) dan memastikan stabilitas model, evaluasi dilakukan menggunakan Stratified 10-Fold Cross-Validation yang diintegrasikan dengan metode Pipeline. Hasil penelitian menunjukkan bahwa SMOTE secara signifikan meningkatkan kepekaan model dalam mengenali transaksi fraud, dibuktikan dengan kenaikan rata-rata Recall dari 0,7886 menjadi 0,8271 dan ROC-AUC dari 0,9497 menjadi 0,9719. Namun, peningkatan ini memicu fenomena trade-off, di mana rata-rata Presisi menurun drastis dari 0,9596 menjadi 0,8916 akibat lonjakan jumlah alarm palsu (False Positives). Dalam konteks industri finansial, trade-off ini dinilai sepadan demi meminimalkan risiko kerugian fatal akibat lolosnya transaksi penipuan (False Negatives). Kata kunci: Deteksi Penipuan; Random Forest; SMOTE; Imbalanced Data; Cross-Validation; Abstract - Digital transaction fraud detection faces a major challenge in the form of class data imbalance, where normal transactions far outnumber fraudulent ones. This imbalance causes machine learning algorithms to be biased toward the majority class. This study evaluates the performance of the Random Forest algorithm before and after the application of the Synthetic Minority Over-sampling Technique (SMOTE) using the Credit Card Fraud Detection dataset. To prevent data leakage and ensure model stability, the evaluation was conducted using Stratified 10-Fold Cross-Validation integrated with a Pipeline method. The results show that SMOTE significantly increases the model's sensitivity in recognizing fraud, evidenced by an increase in the average Recall from 0.7886 to 0.8271 and ROC-AUC from 0.9497 to 0.9719. However, this improvement triggers a trade-off phenomenon, where the average Precision drops drastically from 0.9596 to 0.8916 due to a surge in false alarms (False Positives). In the financial industry context, this trade-off is considered acceptable to minimize the fatal risk of undetected fraudulent transactions (False Negatives). Keywords: Fraud Detection; Random Forest; SMOTE; Imbalanced Data; Cross-Validation;]]>
