| dc.description.abstract | Deep learning on graphs has garnered considerable attention across various machine learning applications, encompassing social science, transportation services, and biomedical informatics. Nonetheless, prevailing methods have predominantly focused on supervised learning, resulting in several limitations, such as heavy reliance on labels and subpar generalization.
To address the scarcity of labels, self-supervised learning (SSL) has emerged as a promising approach for graph data. Traditional SSL methods for graphs primarily concentrate on enhancing model performance through advanced data augmentation strategies and contrastive loss functions. Despite the significant progress made by existing studies, they encounter severe efficiency challenges when dealing with large-scale graphs and resource-limited applications, such as online services. To bridge this gap, I have developed a series of graph SSL models that systematically enhance the efficiency of self-supervised learning on graphs across the stages of model training, inference, and deployment. Firstly, to improve training efficiency, we propose automating the data augmentation process through Graph Personalized Augmentation (GPA) and conducting augmentation-free training via model perturbation (PerturbGCL). Secondly, to expedite inference efficiency, we suggest distilling the fine-tuned classification model into a lightweight model using reliable knowledge distillation (Meta-MLP). Finally, to enhance deployment efficiency, we propose the development of a universal graph model (S2GAE) that enables the learned representation to generalize across different types of downstream tasks in the graph system.
My research presents a significant contribution to the research community by advancing the efficiency and applicability of self-supervised learning on graphs, addressing challenges related to label scarcity and resource limitations. These innovations have the potential to revolutionize various machine learning applications across disciplines, ranging from social science to transportation services and biomedical informatics, ultimately paving the way for more effective and widespread adoption of deep learning techniques in real-world graph scenarios. | |
