欢迎访问中国科学院大学学报,今天是

中国科学院大学学报 ›› 2024, Vol. 41 ›› Issue (6): 830-841.DOI: 10.7523/j.ucas.2023.048

• 电子信息与计算机科学 • 上一篇    

基于自监督异质图神经网络的图分类框架

袁鸣1, 赵彤1,2   

  1. 1. 中国科学院大学数学科学学院, 北京 100049;
    2. 中国科学院大数据挖掘与知识管理重点实验室, 北京 100049
  • 收稿日期:2023-03-06 修回日期:2023-05-05 发布日期:2023-06-12
  • 通讯作者: 赵彤,E-mail:zhaotong@ucas.ac.cn
  • 基金资助:
    国家自然科学基金(12271504,11991022)和国家重点研发计划(2022YFA1003900,2021YFA1000403)资助

A framework of graph classification with self-supervised heterogeneous graph neural network

YUAN Ming1, ZHAO Tong1,2   

  1. 1. School of Mathematical Sciences, University of Chinese Academy of Sciences, Beijing 100049, China;
    2. Key Laboratory of Big Data Mining and Knowledge Management, Chinese Academy of Sciences, Beijing 100049, China
  • Received:2023-03-06 Revised:2023-05-05 Published:2023-06-12

摘要: 图数据以各种各样的形式广泛存在着,图分类任务对于许多问题有重要指导意义。然而图分类任务依然面临很多挑战,包括如何充分利用图结构蕴含的语义信息、进一步降低计算复杂度及获取标签的成本。提出一种超节点异质网络的构建方法,并由此提出可应用于图分类问题的新型框架GChgnn。该框架通过引入双视角的图表示机制以及自监督的对比学习,实现了:1)对大规模图分类任务目标间的相似性进行度量;2)借鉴图匹配方法,通过跨图思想提高相似性度量的准确度,并弥补其无法给出图嵌入显式表达式的不足;3)规避了在网络中设计复杂的卷积与池化算子。通过在一些公开数据集上的测试证明,该框架的综合效果优于现有的解决图分类问题的其他方法。

关键词: 图分类, 异质图神经网络, 自监督学习, 对比学习

Abstract: Graph data exists widely in various forms, and the tasks of graph classification have great significance for many problems. However, the tasks of graph classification still face many challenges, including how to make full use of the semantic information contained in the graph structures, and how to further reduce the computational complexity and the cost of obtaining labels. In this paper, a construction method of a hyper-node heterogeneous network is proposed for the first time, along with a new framework, GChgnn, which can be applied to graph classification. The GChgnn framework achieves the following goals through the introduction of a double-view graph representation mechanism and self-supervised contrastive learning: 1)measuring the similarity between the objectives of the large-scale graph classification tasks; 2)inspired by the graph matching methods, improving the accuracy of similarity measurement by the cross-graph idea, and making up for the lack of the explicit expression of graph embedding; 3)avoid the need to design complicated convolution and pooling operators in the network. After testing on some public datasets, the framework outperforms other existing methods.

Key words: graph classification, heterogeneous graph neural network, self-supervised learning, contrastive learning

中图分类号: