Business Analytics Seminar: Çağrı Bilgi (TU Delft)
"Advancements in graph learning for financial crime"
Event details from: Business Analytics Seminar: Çağrı Bilgi (TU Delft)
- Date
- 29 September 2025
- Time
- 12:00 – 13:00
Abstract:
Graph neural networks (GNNs) have become the standard tool for machine learning tasks on graph-structured data and have applications in various fields, including fraud detection and Anti-Money Laundering (AML). Transactions between bank accounts can be seen as a transaction graph, which reveals money laundering patterns. In this talk, we will discuss two topics.
- Transaction graphs allow for parallel edges between the same nodes, as the same two accounts may have multiple money transfers. Existing methods disregard these parallel edges and collapse multigraphs into simple graphs. We present MEGA-GNN, a novel architecture we developed for edge-attributed multigraphs. MEGA-GNN preserves permutation equivariance through a two-stage aggregation process: first combining parallel edges, then aggregating across neighbors. Our experiments on financial datasets demonstrate that MEGA-GNN significantly outperforms existing solutions.
- Centralized methods, such as MEGA-GNN, assume the global transaction graph is collected in one place. Practical and privacy concerns make this assumption a problematic hurdle to real-life application. To address this problem, we are exploring Federated Learning. Banks collectively train the same model on their local data, only sharing the model's parameters. We will share some challenges and preliminary results.
General information:
This seminar will be organised in a hybrid setup. If you are interested in joining this seminar, please send an email to the secretariat of Amsterdam Business School at secbs-abs@uva.nl.
