Uncertainty Guarantees for Graph Neural Networks

• Applicant: Professor Dr. Aleksandar Bojchevski
• Subject Area: Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing
• Term: since 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 543963829

In a connected world graphs are everywhere. Almost all graph-learning tasks are routinely and effectively tackled with Graph Neural Networks (GNNs) which have emerged as a fundamental building block in many systems. Despite GNN's rise to prominence in both scientific and industry settings, there is a striking lack of research on quantifying their uncertainty, and even fewer studies aim to provide guarantees. Addressing this gap, our overarching goal is to provide uncertainty quantification for GNNs with rigorous distribution-free guarantees. We will adopt the conformal prediction framework which transforms the output of any model (any GNN) into a set that is guaranteed to contain the true label with any user-specified probability. We will cover exchangeable data under transductive and inductive setting, robustness to distribution shift and adversarial perturbations, and guarantees for arbitrary (potentially intersecting) subpopulations. Given the many diverse applications of GNNs we expect our results to be broadly relevant for science, industry, and society

• DFG Programme: Priority Programmes
• Subproject of: SPP 2298: Theoretical Foundations of Deep Learning