From Nodes to Networks: Statistical Perspectives on Graph Representation Learning

Networks provide a natural framework to model interactions and dependencies across complex systems—from biological and social networks to financial and communication infrastructures. This talk presents a statistical perspective on learning from networked data, focusing on how nodes, subgraphs, and entire graphs can be efficiently represented in low-dimensional vector spaces. Starting from classical graph statistics and shallow embeddings, we move toward modern encoder–decoder architectures and graph neural networks (GNNs), highlighting their strengths, limitations, and interpretability challenges. We also explore recent advances in whole-graph embedding and graph transformer models, discussing their applications in real-world scenarios such as biomedical network analysis. Finally, open research directions in expressivity, generalization, and optimization are outlined, connecting theoretical insights with emerging opportunities in graph machine learning.