Contrastive learning for heterogeneous graph neural networks

• Main Author: Dong, Renzhi
• Other Authors: Lihui Chen
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/164475
• Institution: Nanyang Technological University

The challenge of node classification in a heterogeneous graph has generated a lot of research interests in recent years. HeCo, as a novel and popular contrastive learning-based model, performs as a leading method in this field. Therefore, the technical details of HeCo model is reviewed and re-implemented in this dissertation, and some of the components were modified in order to better understand HeCo model and to explore possible new ideas based on it. This report summarizes the existing methods of data augmentation for contrastive learning, and a data augmentation method is applied. Substituting HeCo’s cross-view mechanism with the summarized method we got a revised new model called HeCo*. Then a control experiment is designed and carried out to show the performance of cross-view mechanism in HeCo, it demonstrated that HeCo does outperform the traditional contrastive learning methods in node classification and node clustering significantly. Also, a re-implementation of node clustering is done, a line chart of the variation of NMI and ARI is obtained to better illustrate the training process and the performance of HeCo and HeCo* in downstream applications.