AHEG: Attention-based Heterogeneous Graph Convolutional Networks

MPhil Thesis Defence


Title: "AHEG: Attention-based Heterogeneous Graph Convolutional Networks"

By

Mr. Yifeng ZHANG


Abstract

Graph Convolutional Networks (GCNs) are network architectures that operate on 
graph data. They efficiently employ spectral or non-spectral approaches to 
aggregate information from different number of neighbors of a node in a graph. 
Existing GCNs often assume homogeneous graphs which cannot capture the rich 
semantics of the data, leading to unsatisfactory performance. Many datasets can 
be naturally modeled as heterogeneous graphs which reflects explicitly the rich 
semantical information between nodes. Designing a GCN on such graph is 
challenging, mainly due to the composite structure of its semantics.

We propose AHEG, an Attention-Based Heterogeneous Graph Convolutional Network. 
AHEG retrieves multiple kinds of relationships between different objects by a 
meta-path generation mechanism. Furthermore, it assigns values according their 
importance by means of a two-stage attention-based convolution to form node 
embeddings. As some graph-level machine learning tasks require pooling 
operations, AHEG has an optional pooling layer to downsample the features while 
preserving structural information. We conduct extensive experimental study of 
AHEG on two transductive graph datasets (DBLP and ACM) and one inductive 
dataset (the PPI). AHEG is shown to substantially outperform the state-of-the 
art schemes in terms of node-level or graph-level classification tasks, 
increasing the accuracy for 3.6% in best case. Besides, AHEG demonstrates much 
higher NMI/ARI values and better interpretability in clustering analysis.


Date:			Tuesday, 23 July 2019

Time:			9:30am - 11:30am

Venue:			Room 3494
 			Lifts 25/26

Committee Members:	Prof. Gary Chan (Supervisor)
 			Dr. Raymond Wong (Chairperson)
 			Dr. Kai Chen


**** ALL are Welcome ****