Explore contrastive learning on graph representation learning

• Main Author: Liu, Qiuyu
• Other Authors: Lihui Chen
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/158301
• Institution: Nanyang Technological University

Graph is a type of structured data to describe the multiple objects as well as their relationships, and is attracting increasing attention in recent years as it can represent various types of data structures in the real life, such as the social networks, etc. Hence, it is important to research graph analysis and explore effective methods of graph processing. Graph representation learning methods has been shown to be promising on graph analysis by converting the graphs or part of the graphs into a lower-dimensional representations while preserving all critical information. However, graph representation learning still faces various challenges due to the diversity of graph types, objects, and their corresponding applications. For the node-level graph representation learning, traditional methods focus more on homogeneous graphs and have achieved impressive performance. However, for the heterogeneous information networks (HINs) which have various types of nodes and edges, the methods designed for homogeneous graphs are no longer optimal to handle the semantic incompatibility issues. A novel co-contrastive learning mechanism called HeCo [1] was proposed to address such issue by constructing a cross-view mechanism to effectively capture both local and high-level information. Though HeCo achieved remarkable performance on HIN data, inspiring by various contrastive learning studies [2-6], more contrastive learning methods can be further explored based on [1]. This thesis explores the potential of contrastive methods on HIN data and introduces two research works to improve the current graph representation learning methods of HINs. 1. Methodology 1 as a simple contrastive augmentation method. 2. Decoupled contrastive loss function with Methodology 1. The Methodology 1 reported in this thesis is under preparation for publication, this part of work will not be presented.