The lunch at NR starts at 12:00, while the talk starts at 12:15.
Speaker: Fredrik Johannessen (DNB)
Location: Norsk Regnesentral + Zoom
Title: Finding Money Launderers Using Heterogeneous Graph Neural Networks
Abstract: Money laundering is the activity of securing proceeds from criminal acts by concealing their origins. It is a vast global problem, as it is essential for all types of crime where the goal is to make profit. Anti money laundering (AML) laws consist of measures that must be taken by banks to hinder them from being exploited for money laundering purposes. One of these measures is electronic surveillance of transactions, put in place to detect illicit activities so that they can be stopped. Essentially, these surveillance systems are made up of multiple manually created rules, each of which consists of a few if/else statements. Such simple rules have fallen short of providing efficient surveillance systems, because they fail to detect actual money laundering with high precision, and are easy to circumvent for experienced criminals.
To overcome the shortcomings of rule-based systems, machine learning methods that automatically learn when to generate alerts are becoming more popular. In addition, since groups of criminals typically collaborate to launder money, network science has high potential to further advance the surveillance systems. The work done so far on applying network science to AML uses networks to create features that capture various aspects of an entity’s role in the network. These features are then combined with entity-level features in a downstream machine learning task. A drawback from this approach is that information on the full network is lost before it arrives at the machine learning task that produces the output of interest. Graph Neural Network (GNN) is a group of methods that overcome this drawback by applying machine learning directly on the network data. Because of their applicability on large networks and ability to combine entity- and network data, they have rapidly gained popularity during the past few years.
In the present work, we explore the use of graph neural networks to detect money laundering activities in a large heterogeneous network created from real world bank transactions and business role data belonging to Norway’s largest bank, DNB. Using the Python library PyTorch Geometric, we experiment with different model architectures that are able to utilize the rich semantics present in a heterogeneous graph. In particular, we extend the homogeneous GNN method of Message Passing Neural Network (MPNN) to a heterogeneous graph, and present a new method for how messages across the various relations are aggregated. Further, our results emphasize the importance of utilizing the relation features when present, which many of the heterogeneous GNNs proposed in the literature do not support.
The results of the model’s performances show that the approach has great potential to increase the quality of electronic surveillance systems. To our knowledge, there exists no published work on applying graph neural networks on a large real-world heterogeneous network for the purpose of AML.
Join Zoom Meeting:
https://uio.zoom.us/j/63199595088?pwd=Y05GYkJwR1dCUHFhUnlGMFpsdGc3UT09
Meeting ID: 631 9959 5088
Passcode: 302551
Welcome!
Best regards,
Thea Roksvåg and Lars Henry Berge Olsen.
