Dr. Johannes Lutzeyer
Ecole Polytechnique
15th May 2024, 03:00pm – 04:00pm (GST)
| Title: | Recent Advances in Graph Neural Network Robustness |
| Abstract: | Graph Neural Networks (GNNs) have celebrated many academic and industrial successes in the past years; providing a rich ground for theoretical analysis and achieving state-of-the-art results in several learning tasks. As the industrial use of GNNs becomes pervasive, the study of their robustness is of increasing importance. In this talk, I will review recent work we have been doing on the robustness of GNNs. We will begin by giving an accessible introduction to the area of Graph Representation Learning with a focus on GNNs. We will then explore our recent, simple and effective approach, in which we add noise to GNN hidden states to yield more robust GNNs (AAAI, 2024). Finally, we will discuss an effort in which we used an iterative orthonormalisation algorithm to work with approximately orthonormal weight matrices in GNNs, an approach we call GCORN (ICLR, 2024). We will observe how GCORN models are more robust to primarily feature, but also structural attacks, on GNNs both in theory and in practice. The presented work was done in collaboration with Yassine Abbahaddou*, Sofiane Ennadir*, Henrik Boström and Michalis Vazirgiannis (* denotes equal contribution). |
| Bio: | Johannes Lutzeyer is an Assistant Professor in the Data Science and Mining Group at the Computer Science Department of École Polytechnique in France. Previously, he completed a 2.5 year postdoc, under the supervision of Prof. Michalis Vazirgiannis, at École Polytechnique and a PhD thesis on the spectral properties of the adjacency and Laplacian matrices under the supervision of Prof. Andrew Walden in the Mathematics Department at Imperial College London. Johannes' current research is in the area of Graph Representation Learning with a focus on Graph Neural Networks. |