Moving beyond local development requires a structured lifecycle including ideation, proof of value, scaling, and continuous monitoring.
Traditional rule-based approaches for anti-money laundering suffer from high false-positive rates and fail to capture complex laundering patterns.
AI can lower false positives and identify new threats, but success requires involving business experts and using explainable AI to build trust.
Machine learning models analyze peer group behavior to identify outliers, such as a jobless person with high cash transactions, which rule-based systems miss.
The process involves selecting relevant features, creating a master data table, using dimensionality reduction to find outliers, and scoring customers by reconstruction error.
Principal Component Analysis (PCA) uses linear transformations while autoencoders use non-linear transformations for dimensionality reduction and reconstruction.
The unsupervised models successfully identified over 100 suspicious cases, with one-third being new discoveries not caught by the existing rule-based engine.
The end-to-end system uses Hadoop for the data mart, PySpark for transformation, and Python with scikit-learn and MLflow for model development.
AI enhances AML efforts by detecting novel patterns, focusing agents on high-risk alerts, and providing transparent results to build trust with regulators.
29:40
22:06Lukas Kölbl
34:21Aarno Aukia
57:46Adrian Schmitt
48:07Lina Weichbrodt
23:50Aditi Godbole
58:00Vijay Krishan Gupta & Gauravdeep Singh Lotey
2:11:02Boris Krumrey & Andreas Palfi & Radu Pruna
.gif?w=240&auto=compress,format)
Jobs that call for the skills explored in this talk.
Deutsches Medizinrechenzentrum GmbH