Join us for a Ph.D proposal defence of Aida Sadeghi on her thesis: Static and Temporal Graph Neural Networks for Anomaly Detection.

The Ph.D proposal will be online (Microsoft Teams).

Microsoft Teams Need help?
Join the meeting now
Meeting ID: 335 175 594 544
Passcode: mD6Uaf

Committee:

• Christoph Lohrmann, Supervisor, Assistant Professor, Reykjavik University, Iceland
• María Óskarsdottir, Supervisor, Associate Professor, Reykjavik University, Iceland
• Anna Sigríður Islind, Supervisor, Associate Professor, Reykjavik University, Iceland
• Yngvi Björnsson, Committee Member, Professor, Reykjavik University, Iceland Bart 
• Baesens, Committee Member, Professor, KU Leuven, Belgium

Abstract:

Anomaly detection in complex graphs concerns finding hidden, rare patterns in the data. It may offer benefits for different applications, such as fraud detection in financial sectors through flagging suspicious behaviors. Introducing scalable and explainable methods for anomaly detection has been the subject of some research in data science over the past decades. Graph Neural Networks (GNNs) are one of the most widely used techniques in Machine Learning due to their capacity for capturing complex dependencies in graphs. Yet, the questions of how to improve anomaly detection via GNNs by considering graph topologies, temporal aspects, explainability techniques, and feature selection methods remain unanswered. This research proposes a GNN-based framework to address these challenges through multiple interconnected studies. First, we focus on a static approach to GNNs, emphasizing the integration of graph topology attributes. We then extend this approach by introducing temporal encoding techniques to capture temporal dependencies, enhancing anomaly detection performance compared to static models. Furthermore, we study how GNNs determine anomalies using explainability techniques, improving detection stability and decision transparency. Finally, we integrate the feature selection and graph sampling methods within GNN frameworks, leading to enhanced scalability in anomaly detection. This research project benefits domain experts seeking anomalies in finance, providing higher transparency and explainability in detection.