CAREER: Modern Machine Learning on Graphs: From Theory to Practice

Georgia Tech Research Corporation, Atlanta GA

Investigators

Abstract

Graphs or networks provide a fundamental tool to model data throughout the sciences. For instance, graphs can be used to represent people and their social relations, proteins and their interactions, and brain neurons and the connectome. Many real-world applications involve solving prediction tasks based on graph-structured data such as network anomaly detection, recommendation, drug design, material property analysis, and particle collision denoising. Therefore, it is crucial to develop robust and trustworthy algorithm tools that may learn from graph-structured data. This project aims to address the challenges posed by the emerging complex features of modern graph-structured data. Firstly, the project will focus on learning from higher order relations that are often missed by traditional graph models. Secondly, it will explore deep learning models for graph-structured data. Both directions pursue the approaches that are both principled in theory and useful in practice. The success of this project will lead to an improvement of the understanding of social and biological networks and the quality of data processing for high-energy physics. Moreover, this project will foster collaborations across college students in data science, the computing industry, and domain scientists. The technical contributions of this project are organized into two interrelated themes: (i) Developing expressive hypergraph models with their fast-learning algorithms. This research theme proposes to study hypergraph diffusion problems, where each hyperedge is associated with a nonlinear diffusion function that can be either handcrafted or learned to model complex higher-order relations. (ii) Developing provably expressive and generalizable graph neural network (GNN) models. This research theme proposes to design provably more expressive GNN architectures based on a framework of structural feature augmentation, analyze and improve their generalization capability. The developed approaches will be evaluated in the graph learning tasks such as node classification, link prediction, graph anomaly detection, and emerging problems in high-energy physics. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.