Beatrice Bevilacqua, “Subgraphs to the rescue: how to view graphs as bags of subgraphs to enhance the capabilities of GNNs”
In-person in Clark Hall, Room 110
OR
virtually over Zoom
Join Zoom Meeting:
https://wse.zoom.us/j/97055652302?pwd=dWFUUHRHS1lna2h5K0U1cEt4RDRrQT09
Beatrice Bevilacqua
PhD Student
Purdue University
“Subgraphs to the rescue: how to view graphs as bags of subgraphs to enhance the capabilities of GNNs”
Abstract: Despite their recent success, standard Graph Neural Networks (GNNs) have limited extrapolation capabilities, as well as expressive power. In this talk we aim to show how representing graphs as bags of subgraphs can improve upon these two.
In the first part of the talk we leverage the stability of subgraph densities in graphon random graph models to learn representations that are invariant to distribution shifts from training to test data. We show that these representations can provably extrapolate to out-of-distribution test data, even when learned from a single training environment. In the second part of the talk we focus on deconstructing a graph into subgraphs to increase the expressive power of existing GNNs, and we analyze the symmetry group of the resulting data. We propose a design space to build architectures following such symmetry group, and we provide an upper bound to their expressiveness. We conclude the talk with possible future directions.
Biography: Beatrice Bevilacqua is a PhD student at Purdue University, advised by Prof. Bruno Ribeiro and also working closely with Dr. Haggai Maron. Her research focuses on expressivity and extrapolation capabilities of Graph Neural Networks. Previously, she was a Research Scientist Intern at Meta FAIR and at DeepMind, working on out-of-distribution generalization.