Graph Neural Networks (GNNs) integrate machine learning models to understand graph-structured data that characterize relationships, and have achieved exceptional utility in various domains. The primary aim of this thesis is to study the prominent scalability issue: as the graph size increases, common GNNs exhibit prohibitive computational overhead and unsatisfactory performance. Therefore, scaling up GNNs is essential for processing massive real-world data and broadening their practical applications.
The scope of this thesis focuses on large-scale GNN designs and improvements. State-of-the-art methods are proposed for incorporating advanced graph data management techniques and handling diverse graph patterns efficiently. Novel theoretical frameworks and comprehensive benchmarks are also presented for understanding and tackling the scalability issue.
Overall, this thesis delivers multiple new GNN models capable of processing up to billion-scale graphs. It also contributes to fundamental analyses and practical insights advancing GNN developments and deployments at large scale.