Graph Representation Learning (GRL), which aims to transfer the graph data into vector formats so that subsequent graph analysis tasks can be performed, has attracted wide attention in the machine learning and data mining community. GRL have achieved state-of-the-art performance in many graph data analysis tasks, including citation recommendation, drug discovery, and spam detection. However, GRL still suffers from two limitations for graph representation learning. First, they exploit non-smoothing node features which may result in suboptimal embedding and degenerated performance for graph classification. Second, they can not scale to large dataset due to the complex operation for modelling topological relation.
This project aims to overcome existing limitations of graph representation learning. Novel, flexible, accurate, and end-to-end frameworks will be proposed for link prediction, node classification, and graph classification tasks. This project will also propose novel models that can scale to graphs with billion edges, such as social networks and transaction networks.