Publication Type
PhD Dissertation
Version
publishedVersion
Publication Date
7-2025
Abstract
Graph representation learning has become fundamental in various domains, from social networks to molecular structures, enabling extraction of meaningful patterns from graph-structured data. While deep learning approaches, particularly graph neural networks, have shown promising results, their effectiveness is often limited by the scarcity of labeled data. This challenge is particularly acute in graph domains where annotation requires specialized expertise and is prohibitively expensive. Self-supervised learning has emerged as a promising direction to address this limitation by creating auxiliary tasks from unlabeled data, with augmentation strategies playing a crucial role in their success.
Current graph self-supervised learning methods face several critical challenges limiting their practical effectiveness. First, existing methods heavily rely on manually crafted augmentation strategies, requiring significant domain expertise and computational resources for hyperparameter tuning, with different graph domains often requiring distinct strategies for optimal performance. Second, there exists a fundamental gap between contrastive and generative learning paradigms. While contrastive methods have demonstrated effectiveness in graph learning, they face challenges with robust augmentation design and computational overhead. Generative methods excel at capturing fine-grained structural details through reconstruction but may struggle with learning discriminative features essential for downstream tasks. Current attempts to combine these paradigms maintain a strict separation rather than fundamentally unifying their objectives. Third, integrating graph learning with large language models, particularly for text-attributed graphs, presents unique challenges. Traditional graph augmentations may disrupt the semantic coherence, while text-focused techniques often fail to preserve critical structural information. Furthermore, existing methods typically require labeled data from source domains, which significantly limiting their adaptability across applications.
To address these challenges, this dissertation presents three key contributions through the lens of augmentation: (1) adaptive augmentation selection strategies that eliminate manual selection and tuning, (2) a unified framework bridging contrastive and generative paradigms through feature masking, and (3) a quantization-based approach for integrating structural and semantic information in text-attributed graphs. Our empirical evaluation demonstrates consistent performance improvements across various graph learning tasks, particularly in challenging scenarios involving limited labeled data, multi-modal integration, and cross-domain transfer. These contributions advance both the theoretical understanding and practical capabilities of graph self-supervised learning.
Keywords
Graph representation learning, Self-supervised learning, Large language model
Degree Awarded
PhD in Computer Science
Discipline
Graphics and Human Computer Interfaces
Supervisor(s)
FANG, Yuan
First Page
1
Last Page
168
Publisher
Singapore Management University
City or Country
Singapore
Citation
BO, Jianyuan.
Enhancing graph representation learning through self-supervision: An augmentation perspective. (2025). 1-168.
Available at: https://ink.library.smu.edu.sg/etd_coll/790
Copyright Owner and License
Author
Creative Commons License
This work is licensed under a Creative Commons Attribution-NonCommercial-No Derivative Works 4.0 International License.