Community Detection in Temporal Graphs for Capturing Pass-through Money Laundering Behaviors

Company Description

TMNL (Transaction Monitoring Netherlands) was established by the five largest Dutch banks; and is the 1st initiative in the world fighting financial crime in a multi-bank setting. TMNL combines transaction data from the various banks and makes meaningful connections between this data. TMNL is creating smart models to detect these potentially unusual transactions (see example in Fig. 1). These models are used effectively and responsibly, while excluding risks — such as the risk of discrimination — in the process.
The links that these models make provide new insights into potential money laundering and the financing of terrorism. Small multidisciplinary teams, consisting of AML experts, data scientists, data engineers, and machine learning engineers, are creating these models and working together using a fixed approach.

Project Description

In our previous work [1] we introduced a framework for unearthing money laundering flows
from a massive transaction data. The framework utilises temporal and higher order graphical
representations. In the experimental evaluation, one of our main focuses was on validating the technical feasibility of dealing with massively large temporal graphs. For the functional part we focused on a very specific modus operandi for money laundering. We now want to expand our work to deal with unknown-unknown money laundering schemes. We propose the following potential expansions and contributions:

  1. Synthetic data generation
    Synthetic data generated using a model trained on the actual data. The data would also include injected money laundering flows. This helps in having much more controllable experiments where several types of injected money laundry flows can be studied.
    Challenges:
    • Incorporating chained interactions
    • Retaining anomalous behavior
  2. Novel weighting criteria
    A weighting method to capture the rapid movement of funds and shell companies’ behaviors. The current method only quantifies the 2nd order interactions.
    Challenge: Calculating passthrough rate, on a massive temporal graph, is computationally expensive.
  3. Novel community detection algorithm
    A (scalable) community detection algorithm optimized for dependent flows rather than modularity, density, etc. We also believe that within the context of AML modelling an algorithm that detects overlapping communities is more suitable.
  4. Novel anomaly detection algorithm
    A parameter free algorithm for detecting anomalous communities. The algorithm should only consider the graphical structure and (money flow) interactions within a community. Can we include timeseries (entity level) patterns?

Contact

For more information, contact dr. ing. Marwan Hassani.

Reference

[1] Haseeb Tariq, Marwan Hassani, Topology-Agnostic Detection of Temporal Money Laundering Flows in Billion-Scale Transactions. in PKDD 2023 (https://doi.org/10.48550/arXiv.2309.13662)

Leave a Reply