UCLA

Deep learning models for graph structured data are popular recent developments in representation learning. These models are essential for numerous knowledge-driven applications including knowledge completion, and semantic applications like semantic search, question answering, and recommender systems. In this dissertation, we will focus on methods for modeling the rich information in heterogeneous networks, by delving into our proposed state-of-the-art graph neural network architecture, knowledge graph embedding model, question answering system, and social network embedding model. While prior works have merely scratched the surface to exploit the comprehensive data available from heterogeneous information in graphs, in this work, we will thoroughly examine systematic methods for mining and modeling different levels of heterogeneity to build the next generation of powerful deep learning models.