图自监督学习综述

A Review on Graph Self-Supervised Learning

关键词:  自监督学习  图神经网络  图表示学习  深度学习  图分析

Key words:  self-supervised learning  graph neural networks  graph representation learning  deep learning  graph analysis

摘要:

针对人工标注的高昂成本和监督学习泛化能力不足的问题，图自监督学习(GSSL)因其通过精心设计的代理任务提取信息知识，而不依赖于人工标签，已经成为一种有前途的图数据学习范式。本文首先介绍了图自监督学习的基本概念，以及其与传统图学习方法的区别和联系；然后详细回顾了当前图自监督学习的主要方法，并将其分为图自预测、对比学习、不变挖掘和混合方法等4个大类，同时探讨和总结了图自监督学习在医学图像上的应用；最后，指出了未来研究的方向，包括提高算法的可解释性、使得算法适用于更大的图规模和使用算法处理结构更为复杂的异质图，以期进一步推动图自监督学习的发展。

Abstract:

Given the high cost of manual annotation and the limited generalization capability of supervised learning,Graph Self-Supervised Learning(GSSL),which extracts information knowledge through carefully designed pretext tasks without relying on manual labels,has emerged as a promising paradigm in graph deep learning.The basic concepts of GSSL,including its differences and connections with conventional graph learning methods,were introduced.The current main methods of GSSL,categorizing them into four types:graph self-prediction,contrastive learning,invariance mining,and hybrid methods,were then reviewed.The applications of GSSL in medical imaging were discussed and summarized.Finally,the future research directions were proposed,including improving algorithm interpretability and extending the application of algorithms to larger graphs and heterogeneous graphs with more complex structures.The review aims to further promote the development of GSSL.