TSVP Talk: Bringing Shape to Graph Representations: From Topological Contrastive Learning to Joint-Embedding Predictive Architectures by Yulia Gel

Title: Bringing Shape to Graph  Representations: From Topological Contrastive Learning to Joint-Embedding Predictive Architectures

Speaker: Yulia Gel, Department of Statistics, Virginia Tech

Abstract: Graph contrastive learning (GCL) enables us to obtain rich representations in a wide variety of applications which involve abundant unlabeled information.  However, many GCL approaches tend to overlook the important latent information on higher-order graph substructures. To address this limitation, we introduce concepts of shape into GCL, specifically, topological invariance and extended persistence on graphs.  We then propose a contrastive learning framework that aligns the topological representations of two augmented views of the same graph. These representations are obtained by extracting latent shape characteristics at multiple resolutions and summarizing them through extended persistence landscapes (EPLs). Our extensive numerical results on molecular and chemical compound datasets show that the Topological Graph Contrastive Learning delivers significant performance gains in self-supervised graph classification and also exhibits robustness under noisy scenarios. Finally, we discuss how shape concepts rooted in persistent homology can be integrated into the Joint-Embedding Predictive Architectures (JEPAs) for graphs.

Profile: Yulia R. Gel is a professor of statistics at Virginia Tech, USA. She earned her doctorate in mathematics at Saint Petersburg State University, Russia. Prior to joining Virginia Tech, she was on stint as Program Director at National Science Foundation of USA during 2021-2025. She received her tenure at University of Waterloo, Canada and also held visiting appointments in NASA Jet Propulsion Lab (USA), The Isaac Newton Institute for Mathematical Sciences at the University of Cambridge (UK), Johns Hopkins University, and University of California at Berkeley. Yulia is a Fellow of the American Statistical Association. Her research interests are at the intersection of statistical topological learning and geometric deep learning.

Language: English

Target audience: General audience/everyone at OIST and beyond.
Freely accessible to all OIST members and guests without registration.

This talk will also be broadcast online via Zoom:
Meeting ID: 993 1216 5065
Passcode: 603487

※ Please note that this event may be recorded and the videos uploaded. In addition, photos may be taken during the event. These are intended for publication online (the OIST website, social media, etc.)※