Skip to yearly menu bar Skip to main content


Poster Session
in
Workshop: Scientific Methods for Understanding Neural Networks

Is Expressivity Essential for the Predictive Performance of Graph Neural Networks?

Fabian Jogl · Pascal Welke · Thomas Gärtner

[ ] [ Project Page ]
Sun 15 Dec 4:30 p.m. PST — 5:30 p.m. PST

Abstract:

Motivated by the large amount of research on the expressivity of GNNs, we study the impact of expressivity on the predictive performance of GNNs. By performing knowledge distillation from highly expressive teacher GNNs to less expressive student GNNs, we demonstrate that knowledge distillation reduces the predictive performance gap between teachers and students significantly. As knowledge distillation does not increase the expressivity of the student GNN, it follows that most of this gap in predictive performance cannot be due to expressivity.

Chat is not available.