Effective Detection of Out-of-Distribution Data in Graph Representation Learning

The School of EECS is hosting the following HDR Progress Review 1 Confirmation Seminar:

Effective Detection of Out-of-Distribution Data in Graph Representation Learning

Speaker: Danny Wang

Abstract: The proliferation of Graph Neural Networks (GNNs) across diverse domains and real-world applications has underscored the importance of robust and reliable predictive systems. However, GNN performance is heavily contingent on the premise that test data adheres to the same distribution as training data, which is often breached in real-world scenarios where graph data is characterised by out-of-distribution (OOD) instances. This leads to inaccurate predictions that can be detrimental in critical areas like medical diagnosis and drug discovery.

This work addresses fundamental limitations in graph OOD detection through two complementary research directions. First, we propose a pseudo-OOD generation framework that effectively exposes OOD scenarios exclusively from in-distribution training data, eliminating the need for external OOD datasets or pre-trained generative models. Second, we introduce an information-theoretic perspective with theoretical grounding for improving OOD detection, coupled with an information decomposition framework for learning robust graph representations. Together, these contributions provide both practical solutions for real-world deployment constraints and theoretical foundations for understanding how graph-specific characteristics influence OOD detection performance, advancing the field toward more reliable graph learning systems.

Bio: Danny Wang is a PhD student in the School of Electrical Engineering and Computer Science at the University of Queensland. He earned both his Bachelor of Computer Science and Master of Data Science degrees from UQ. His research centres on robust graph representation learning and out-of-distribution data detection, under the supervision of Professor Helen Huang, Associate Professor Guangdong Bai, and Dr. Ruihong Qiu.