TOWARDS LABEL-EFFICIENT DEEP LEARNING ON GRAPH-STRUCTURED DATA

Abstract

Deep Learning (DL) has become a major tool in artificial intelligence. Modern DL models are mainly developed for regular-structured data like images. However, many real-world data do not have a regular structure, such as social networks. A good choice for representing these data is graph, a generic data structure that succinctly represents a set of objects and their relations. A key factor for DL to achieve high performance is abundant annotation. However, collecting sufficient annotation for graph-structured data can be quite difficult. In this thesis, we develop label-efficient DL methods that achieve superior performance with a few labels. First, we build very deep graph neural networks that can capture rich semantic information from unlabeled data. Second, we propose two novel active learning strategies for selecting the most valuable samples to label.