Project: Synthetic data generator for temporal graph patterns related to transactional risk
Description
Research on transactional risk detection (fraud, chargebacks, money laundering) is held back by data. Real transaction data cannot be shared, and the synthetic generators that exist do not capture how risky behaviour appears over time in the graph of accounts, merchants and transactions. Patterns such as layering, circular transfers or bursts of coordinated activity are inherently temporal and relational, so a generator that cannot express them cannot produce data suitable for evaluating graph-based detection methods.
In this project you will catalogue temporal graph patterns representing different transactional risk modalities, and then build a generation framework that instantiates them into synthetic transaction data containing a controlled fraction of risky activity. The generator should be parameterizable, so that datasets with different pattern mixes, densities and degrees of occlusion can be produced and shared openly. You will evaluate the result by training a graph neural network on the generated data and measuring how well it detects the planted risky transactions.
Research questions: which risk patterns can be expressed as temporal graphs, and what are their structure and attributes? How can these patterns drive synthetic transaction generation? Should part of the data be occluded to mirror real settings, where one stakeholder does not observe everything? Given the resulting dataset and a GNN model, how accurately can fraudulent transactions be detected?
Prerequisites: solid programming skills; background in graph data management and/or machine learning on graphs; interest in data generation and benchmarking.
Details
- Supervisor
-
Nick Yakovets
- Secondary supervisor
-
JRJože Martin Rožanec (Jožef Stefan Institute, Slovenia)
- Interested?
- Get in contact