A Survey of Generative Graph Self-Supervised Learning

Generative graph self-supervised learning is designed to harness the inherent information within graphs by crafting predictive tasks or structure/feature reconstruction tasks to produce supervisory signals. This process results in the generation of new graphs that closely mimic the original graph data in both structure and features. Thus, it has shown superior performance in addressing challenges such as data scarcity or insufficient labeling, attracting widespread attention from researchers. Despite significant progress in this field, there is still a lack of systematic organization and summarization. To this end, this paper aims to make a comparative study on the research achievements of generative graph self-supervised learning in recent years. It first introduces relevant background knowledge and provides formal definitions. Subsequently, it sorts out graph autoencoders in generative graph self-supervised learning and categorizes and analyzes graph mask autoencoders. In addition, this paper compiles commonly used datasets and evaluation metrics. Finally, it discusses the current challenges faced by generative graph self-supervised learning and offers prospects for future research directions.