Graph representation learning with attribute information

Abstract

The ubiquitous graph-structured data is a very important form of Big Data and widely used for modelling complex linked data with a broad spectrum of applications such as bioinformatics, web search, social network, road network, etc. Over the last decade, tremendous research efforts have been devoted to many fundamental problems in managing and analyzing (large) graph data. Recently, graph embedding or graph representation learning techniques are attracting a large amount of research attention, which aim to learn low-dimensional vector representations for vertices such that the structure of the graph can be well preserved.This thesis investigates state-of-the-art graph representation learning methods and promotes the research front-line with proposed cutting-edge graph modelling methods and related applications in real-world graph-structured data (i.e., social link prediction, recommendation). This thesis focuses on different attribute information of a graph (i.e., edge attributes, node attributes, and graph attributes). The key contributions of the thesis are three-fold. First, for edge attributes, we propose Projected Metric Embedding (PME) that fills the research gap that existing homogenous graph embedding methods cannot distinguish the heterogeneity between different types of nodes and links. The second contribution of this thesis is the exploration of exploiting centrality information (node attributes) with Graph Convolutional Networks for graph representation learning, which improves the training efficiency and effectiveness in several downstream tasks on graphs. Last but not least, we focus on a special kind of graph bipartite graphs and propose a novel embedding method for general bipartite graphs featured a sequential modelling paradigm along with recent advances in Graph Convolutional Networks (GCNs). Our proposed method is shown beneficial of capturing higher-order implicit dependencies between same typed vertices in a bipartite graph. We also apply the proposed method to enhance social recommendation performance by enriching higher-order implicit user to user relationships in a user-item interaction bipartite graph.