Graph-structured data are the backbone of numerous real-world machine learning tasks, such as social networks, recommender systems, traffic networks, and so on. The fundamental challenge in solving these tasks is to find a way to encode graph structures as well as to incorporate various node or edge information so that machine learning models can easily exploit them. In this dissertation, we explore deep learning architectures, especially the graph neural networks for multiple graph learning applications, i.e., node classification, link prediction, spatiotemporal graph forecasting on irregular grid, and supervised sequence learning problems.
