This thesis studies scalable graph learning from the perspective that learning performance depends not only on model design, but also on how useful information is organized, preserved, and enriched during training.
Three graph learning settings are considered: node classification on large homogeneous graphs, link prediction on knowledge graphs, and knowledge graph entity typing. The proposed contributions address structural graph partitioning, semantic partitioning for knowledge graph, and selective enrichment for context-limited entities.
The results show that data organization and context construction significantly affect both scalability and prediction quality. More generally, the thesis argues that scalable graph learning is also a problem of information organization.