Latent representations of transaction network graphs in continuous vector spaces as features for money laundering detection

Abstract

This paper explores the construction of network graphs from a large ban ktransaction dataset and draws from Ąndings in language modeling and unsupervised learning to transform these graphs into multidimensional vector representations. Such latent representations encode relationships and community structures within the transaction network. Three classiĄers with varying complexity are trained on these latent representations to detect suspicious behavior with respect to money laundering. The specific challenges accompanying highly imbalanced classes are discussed as well and two strategies to overcome these challenges are compared.