A Stacking Ensemble for Anti-Money-Laundering Transaction Monitoring in Mobile Payment Networks

Mobile-payment ecosystems generate billions of micro-transactions with extreme class imbalance (suspicious-to-normal ratio ~1:100,000). We propose a two-tier stacking ensemble combining gradient-boosted trees, an autoencoder reconstruction-error detector, and a graph-based community-isolation score. On a 90-day Japanese mobile-wallet dataset (~430 million transactions), the ensemble reduces false-positive volume by 41% while maintaining recall above 95% on regulator-confirmed AML cases. We discuss operational lessons including reviewer-in-the-loop feedback and concept-drift handling.