Heterogeneous graph attention network

• Main Author: Wang, Zhixing
• Other Authors: Lihui CHEN
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2020
• Subjects: Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/141153
• Institution: Nanyang Technological University

A graph is a data structure which contains connection information in edges and feature information in nodes. Graph neural network applies deep learning to process graph and it becomes more and more popular by its powerful representation ability. The researches on graph neural network are divided into two orientations: homogeneous graph and heterogeneous graph. Traditional graph representation methods mainly deal with problems in homogeneous information network where types of targeted nodes and connections are all same. However, these methods cause the loss of different semantic information because relationships between nodes and their types are always different in the real world. For the researches on heterogeneous graph, graph representation learning is made in heterogeneous information network (HIN) and there are two questions should be considered: (1) What kind of model can learn the new graph embedding by fusing the information with features in different types. (2) As nodes’ features in different meta paths (defined in HIN) have different impact on graph embedding, how the model learns the importance weight of each meta path. Recently, a model named heterogeneous graph attention network (HAN) is proposed to solve above two questions. It includes two main parts: node-level attention and semantic-level attention. Node-level attention is using the attention mechanism for nodes and their neighbors in one meta-path to get the attention weighs, which helps to distinguish the importance among nodes. For semantic-level attention, it measures the importance among different meta-paths and assign weights to each of them. In this project, the published work HAN is comprehended and analyzed thoroughly, and extensive experiments are designed to verify whether the HAN has the superior performance in representation learning and can learn the weights of different meta paths through the attention mechanism at the semantic level.