Close
Skip to content
Sanjay Melkote

Sanjay Melkote

Sanjay Melkote

Machine Learning Manager in U.S. Bank’s Financial Intelligence Unit

Sanjay Melkote is a Machine Learning Manager in U.S. Bank’s Financial Intelligence Unit and based in Columbus, Ohio. He oversees anti-money laundering-related machine learning development, iterative training, model comparison, and documentation for the Bank. His team focuses on the creation and evaluation of new models and tools to mitigate the risks of high-risk products, customers, and transaction channels.
Sanjay has tackled many problems in the private and public sectors using machine learning, predictive analytics, and optimization. He has served in several positions in industry, government, and academia, and holds a Ph.D. in Industrial Engineering and Management Sciences from Northwestern University.

Track: Artificial Intelligence & Healthcare

Monday, April 15, 10:30–11:20am

Identifying Suspicious Financial Activity Using Machine Learning

Detecting unusual and suspicious money laundering and terrorist financing activity is an important problem faced by all financial institutions. The obligation of banks to report suspicious activity related to potentially suspicious wire transfers and cash structuring is a regulatory requirement. Traditionally, banks have used rules-based programs to monitor transactions for suspicious activity. Although simple, these programs may not result in good accuracy, can generate redundant alerts, need to be manually updated, and cannot easily adapt to changes in the data. In this talk, we explore a machine learning-based approach to detecting suspicious financial activity. Inspired by the literature on class imbalance learning, we develop a hybrid method called EasyEnsembleRF that deeply explores the data while retaining fast training speeds. Using 1-2 years of transaction data for training and testing, we compare EasyEnsembleRF to several benchmark predictive models, including random forests, logistic regression, and neural networks. The results show EasyEnsembleRF has by far the lowest false negative rates of all the models tested, while maintaining low false positive rates. Production versions of EasyEnsembleRF are currently being piloted and are expected to result in enhanced detection of suspicious wire transfer and cash structuring activity.