Skip to main content

Luigi Fracassetti: Beyond Message Passing: Neural Sheaf Diffusion and the Geometry of Information Flow on Graphs

Time: Mon 2026-11-23 16.00 - 18.00

Location: Mötesrum 12 - Cramérrummet, Albano Hus 1, Vån 3

Participating: Luigi Fracassetti (KTH/SU)

Export to calendar

Abstract: Graph Neural Networks (GNNs) often struggle to propagate information over long distances, a limitation commonly associated with oversquashing. This talk explores cellular sheaves as a geometric framework for addressing this challenge, focusing on Neural Sheaf Diffusion (NSD) and its ability to learn how information is transported across graphs.
Combining mathematical foundations with experiments on synthetic graphs and real-world urban networks, we investigate long-range information propagation through influence analysis, routing behaviour, and discrete curvature.
Our findings suggest that learned sheaf structures can enhance long-range sensitivity and induce direction-dependent information flow, although these effects cannot be fully explained by conventional influence metrics or scalar graph curvature. The talk illustrates both the potential and the limitations of geometric approaches to graph learning.

For further details about the seminar, please check out our website . We are welcoming everyone that is interested.