Janet Layne – Computer Science
Title: Unsupervised Methods for Learning Structural Graph Representations
Node representation learning methods generate vectorial representations of the nodes in a network for use in standard machine learning models. These methods project nodes into a low-dimensional representation space while preserving information about relationships between them in the graph.
Approaches largely fall into one of two categories: those that capture information about connectivity between nodes, and those that capture a node’s structural information. For tasks where node structural role is important, connectivity-based methods show poor performance. Compared to connectivity-based methods, relatively few approaches exist that generate structural node representations. A review of the common structural methods will be presented to highlight the continued need for development of new approaches.