Deep Support Vector Data Description dalam Penanganan Anomali pada Proses Pengajuan Klaim Asuransi Kredit

Abstract

This study proposes Deep Support Vector Data Description (Deep-SVDD) as an anomaly-detection approach for credit-insurance claim submissions processed via host-to-host systems. Operational tabular data (5,000 observations) were prepared through an anti-leakage pipeline (deduplication, standardization, outlier handling, categorical encoding, and numeric scaling) and a time-based split (Train/Validation/Test). The model was trained on a Train-Normal subset to learn normality patterns, while PCA and HDBSCAN were used as supporting analyses in the latent space to enhance interpretability. Anomaly scores were converted into decisions using a percentile-based threshold aligned with audit capacity and then frozen prior to testing. Results indicate strong performance under class imbalance reflected by PR-AUC = 0.9673 and operational effectiveness through Recall@20 ≈ 44.19%, positioning the model as a precision-oriented, efficient, and accountable first-line detector that reduces manual verification workload while maintaining decision transparency.