UCLA

Graph structures are a powerful abstraction of many real-world data, such as human interactions and information networks. Despite the powerful abstraction, graphs are challenging to model due to the high-dimensional, irregular and heterogeneous characteristics of many real-world graph data. An essential problem arose is how to effectively and efficiently learn the representation for objects in graphs. In this thesis, both the effectiveness as well as efficiency aspects of the graph representation learning problem are addressed. Specifically, we start by proposing an effective approach for learning heterogeneous graph embedding in an unsupervised setting. Then this is generalized to semi-supervised scenario where label guidance is leveraged. The effective graph representation learning models are followed by efficient techniques, where we propose efficient sampling strategies to improve the training efficiency for content-rich graph embedding models. Finally, to reduce storage and memory cost of the embedding table used in various models, we introduce a framework based on KD code, which can compress the embedding table in an end-to-end fashion. We conduct extensive experiments on various real-world tasks on graph data (e.g. anomaly detection, recommendation and text classifications), and the empirical results validate both effectiveness as well as efficiency of our proposed algorithms.