Event Date
Abstract
Machine learning over graphs exceeds in its scalability and stability in processing structured data. In this talk I will address the question of how graph machine learning can generalize across different graph scales and how they can remain stable on large-scale graphs. I do so by considering manifolds as graph limit models. In this talk, first we will explain how to build analogous learning architectures over manifolds as the limit objects of GNNs and graph transformers when the graphs are sampled from manifolds. These limit models provide a unifying framework to analyze size generalization of GNNs and graph transformers. I will further discuss the extension to include transferability discussion for GNNs over sparse random geometric graphs. These findings offer practical guidelines for designing graph learning architectures, particularly by imposing constraints on the spectral properties of filter functions. Theoretical results are verified in real-world scenarios, including point cloud analysis, wireless resource allocation, and terrain data analysis.
Bio
Zhiyang Wang is a postdoc scholar at UCSD and an incoming assistant professor at Washington University in St Louis. She earned her Ph.D. degree at the University of Pennsylvania in the Electrical and Systems Engineering Department. Previously, she received her B.E. and M.E. degrees in 2016 and 2019, respectively, from the University of Science and Technology of China. Her research interests include graph signal processing, graph neural networks, geometric deep learning, and multi-agent systems. She received the best student paper award at the 29th European Signal Processing Conference and the Bruce Ford Memorial Fellowship at the University of Pennsylvania. She was chosen as a Rising Star in Signal Processing and an EECS Rising Star in 2023, as well as a Rising Star in Data Science in 2024.