Kaixiong Zhou: Deep Graph Representation Learning: Scalability and Efficiency | UNC School of Data Science and Society

Large-scale graph data is ubiquitous in science and industry, including social networks, knowledge graphs, and biochemical molecules. Over the past five years, graph neural networks (GNNs) have emerged as de facto standard models to analyze the node features and graph topology. Despite the prominent effectiveness obtained in recent progress, the scalability and efficiency of GNNs are notoriously challenged throughout machine learning cycle. First, it is costly to scale GNNs on modern graphs with more than billions of nodes. Second, most of GNNs contain less than three layers and are vulnerable to be scaled towards deep models. The existing shallow GNNs prohibits the learning of global graph topology. Third, the training of deep GNNs on massive data takes expensive time cost in every new application.

In this talk, Kaixiong Zhou will discuss some recent results in improving the scalability and efficiency of GNN models. These results are (i) designing scalable graph machine learning algorithms to save memory usage and computation time on the large-scale graph data; (ii) answering how to develop deep graph neural networks from theoretical foundations and practical tools; (iii) proposing graph prompt learning to enhance modeling efficiency on massive data. Finally, Zhou will discuss his broader research vision in large-scale graph training system and hybrid quantum-classical graph computing.

About Kaixiong Zhou

Kaixiong Zhou is a fifth-year Ph.D. student from the Department of Computer Science at Rice University. His interests lie in the broad area of large-scale graph data mining and machine learning, particularly in deep graph analysis, efficient graph representation learning, graph quantum computing, and the real-world science applications in biochemical informatics.

Individual meetings with Kaixiong Zhou can be arranged by contacting Jocelyn Tsai at jct33@unc.edu.