Electronic Theses and Dissertation

Transformasi digital dalam industri keuangan telah membawa kemudahan akses layanan keuangan sekaligus tantangan baru dalam deteksi penipuan. Penelitian ini mengembangkan model machine learning untuk mendeteksi anomali transaksi finansial dengan membandingkan kinerja LightGBM, XGBoost, dan teknik Stacking yang menggabungkan keduanya dengan Random Forest sebagai meta-learner. Dataset yang digunakan berasal dari catatan peminjaman fintech, mencakup 857.899 data training dan 367.702 data testing, dengan pra-pemrosesan meliputi feature selection, penyeimbangan data menggunakan SMOTE, dan normalisasi Standard Scaler. Hasil evaluasi menunjukkan model Stacking unggul dengan F1-Score 84,93%, Precision 74,06%, Recall 99,54%, dan Accuracy 82,28%, mengalahkan performa individual LightGBM dan XGBoost. Model ini berhasil mendeteksi 337.707 transaksi fraud dengan kesalahan minimal, membuktikan efektivitasnya dalam mengidentifikasi pola penipuan. Penelitian ini juga menemukan adanya pola aktivitas fraud yang terorganisir, ditunjukkan oleh kemunculan berulang user dengan reference contact dan loan type yang sama, mengindikasikan adanya potensi penipuan kolektif. Temuan ini tidak hanya memperkaya literatur deteksi fraud dengan kombinasi algoritma boosting dan ensemble learning, tetapi juga merekomendasikan model siap pakai untuk industri fintech guna meminimalkan risiko kerugian. Temuan ini membuktikan pendekatan Stacking mampu meningkatkan akurasi deteksi anomali finansial sekaligus menawarkan solusi efisien untuk tantangan ketidakseimbangan data dan kompleksitas komputasi, sehingga dapat memperkuat sistem keamanan transaksi digital di sektor fintech.
Kata kunci: deteksi penipuan, machine learning, LightGBM, XGBoost, Stacking, teknologi keuangan

The digital transformation in the financial industry has facilitated easier access to financial services while introducing new challenges in fraud detection. This study develops machine learning models to detect financial transaction anomalies by comparing the performance of LightGBM, XGBoost, and a Stacking technique that combines both with Random Forest as a meta-learner. The dataset, derived from fintech loan records, consists of 857.899 training data points and 367.702 testing data points, with preprocessing steps including feature selection, data balancing using SMOTE, and normalization with Standard Scaler. Evaluation results show that the Stacking model outperforms individual models, achieving an F1-Score of 84,93%, Precision 74,06%, Recall of 99,54% and Accuracy of 82,28%. This model successfully detected 337.707 fraudulent transactions with minimal error, demonstrating its effectiveness in identifying fraudulent patterns. The study also uncovered organized fraud activity patterns, indicated by recurring users sharing the same reference contact and loan type, suggesting potential coordinated fraud attempts. This research not only enriches the fraud detection literature through the integration of boosting algorithms and ensemble learning, but also recommends a deployable model for the fintech industry to mitigate financial loss risks. The findings affirm that the Stacking approach enhances the accuracy of financial anomaly detection while offering an efficient solution to data imbalance and computational complexity, thereby strengthening digital transaction security in the fintech sector. Keywords: fraud detection, machine learning, LightGBM, XGBoost, Stacking, financial technology