Amaretto: An Active Learning Framework for Money Laundering Detection

Monitoring financial transactions is a critical Anti-Money Laundering (AML) obligation for institutions. In recent years, machine learning-based transaction monitoring systems have successfully complemented traditional rule-based to reduce the high number of false positives and effort needed review all alerts manually. Unfortunately, solutions also disadvantages: while unsupervised models can detect novel anomalous patterns, they are usually characterized by alarms; supervised models, instead, offers higher detection rate but require large amount labeled data achieve such performance. this paper, we present Amaretto, an active learning framework money laundering that combines techniques support processes improving performance reducing compliance management costs. Amaretto exploits selection strategies target subset investigation, making more efficient use feedback provided analyst. We perform experimental evaluation on synthetic dataset industrial partner, which simulates profiles clients trading in international capital markets. show outperforms state-of-the-art risk particular, compare commonly used AML domain with ones implemented work. Isolation Random Forests best task under analysis, AUROC 0.9 first (20% better average) 0.793 second (30 % average). addition, lower investigation costs computed terms daily be examined negatives. Finally, against fraud system, achieving performances analyzed scenarios. Worth mentioning, improves up 50 reduces overall cost 20% most realistic scenario analysis.