<![CDATA[Fraud detection in Quick Response Code Indonesian Standard (QRIS) transaction environments is challenging because fraud labels are often limited, delayed, or incomplete, making it difficult to apply supervised learning effectively. In addition, high transaction volumes create operational constraints for investigation teams, requiring an efficient mechanism to prioritize high-risk merchants for review. To address this problem, this study proposes a segmentation-aware Isolation Forest framework for fraud investigation prioritization in QRIS transactions. Transaction-level data from a national bank in Indonesia were transformed into merchant-level behavioral representations using 38 engineered features capturing transaction intensity, temporal patterns, nominal concentration, burst behavior, and transactional irregularities. Anomaly scores were then used as ranking signals for prioritizing selective investigation. Model performance was evaluated using proxy-based ranking metrics, capacity simulation investigations, selective alert optimization, transaction-level interpretability, and expert validation. The results show that only 1.03% of merchants were classified as very high anomalies, while the model achieved strong ranking performance with an Area Under the Curve (AUC) of 0.8794, Precision@Top 5% 0.3306, and Lift of 6.5942. Selective alerting reduced investigation workload by up to 80.24%, with higher precision and lower FAR, although at the cost of reduced recall. Expert validation further indicated 94.12% prioritized merchants were operationally relevant.]]>
Copyrights © 2026
