LEARNING ON GRAPHS

Abstract

Learning on graphs mainly aims to analyze the property of entities from graphs where entities and entity relations are represented as nodes and edges, respectively. Most of the existing graph-based learning methods mainly model graph structure based on a local smoothness assumption, that is closely connected nodes have similar predictions.

In this thesis, we investigate techniques to thoroughly mine the graph data so as to enhance local smoothness modelling (LSM). In particular, we improve LSM by considering relation types. We encode domain knowledge to guide LSM. We adjust the strength of smoothness in a time-aware manner. We develop adversarial training to enhance the robustness of LSM.

The proposed methods leverage different types of additional information suitable for different kinds of graph applications. Therefore, we test the methods in different applications: university ranking, popularity prediction, stock ranking, and text classification. Extensive experiments demonstrate the effectiveness of the proposed method.