计算机科学与探索 ›› 2023, Vol. 17 ›› Issue (1): 179-186.DOI: 10.3778/j.issn.1673-9418.2104082

• 人工智能·模式识别 • 上一篇    下一篇

融合结构和特征的图层次化池化模型

马涪元,王英,李丽娜,汪洪吉   

  1. 1. 吉林大学 计算机科学与技术学院,长春 130012
    2. 符号计算与知识工程教育部重点实验室(吉林大学),长春 130012 
    3. 吉林大学 人工智能学院,长春 130012
  • 出版日期:2023-01-01 发布日期:2023-01-01

Structure and Feature Fusion Graph Hierarchical Pooling Model

MA Fuyuan, WANG Ying, LI Lina, WANG Hongji   

  1. 1. College of Computer Science and Technology, Jilin University, Changchun 130012, China 
    2. Key Laboratory of Symbol Computation and Knowledge Engineering of Ministry of Education (Jilin University), Changchun 130012, China 
    3. College of Artificial Intelligence, Jilin University, Changchun 130012, China
  • Online:2023-01-01 Published:2023-01-01

摘要: 作为深度神经网络向非欧式数据上的扩展,图神经网络(GNN)已经在图节点分类任务、链接预测任务和图分类任务中取得了显著成就。在图分类任务上,当前方法一般通过层次化的池化过程同时考虑图的局部和全局结构信息以学习高层次的图表示。在对当前的图分类模型进行对比分析后,考虑当前方法的不足,结合不同方法的优势,提出结构和特征融合池化模型(SAFPool)。SAFPool模型在池化时使用了两个聚类分配矩阵生成模块,分别是基于结构的聚类学习和基于特征的聚类学习模块,基于结构的聚类学习根据图结构信息对结构相似的节点聚类,基于特征的聚类学习则根据图节点特征对特征相似的节点聚类。二者的聚类结果加权聚合后便能获取实现聚类策略的聚类分配矩阵以同时利用图结构和节点特征信息。最后,在多个图分类数据集上通过对比实验和可视化说明了同时显式地利用图节点特征信息和图结构信息实现聚类策略的有效性。

关键词: 图神经网络, 图分类, 图池化, 聚类分配矩阵, 层次化模型

Abstract: Graph neural networks (GNN), which extend deep neural networks to non-Euclidean data, have been proven to be powerful for numerous graph related tasks such as node classification, link prediction, and graph classification. On the task of graph classification, recent studies aim to learn graph-level representation through a hierarchical poo-ling procedure using the local and global structure information of the graph. After comparing and analyzing the cur-rent graph classification model, this paper proposes the structure and feature fusion pooling model (SAFPool) consi-dering shortcomings of the current method and combining the advantages of different methods. SAFPool utilizes two assignment matrix generation modules during the pooling process, which are structure-based cluster learning and feature-based cluster learning modules. Structure-based cluster learning module clusters nodes with similar struc-tures based on graph structure information, and feature-based cluster learning clusters nodes with similar features based on graph node features. Then the two clustering assignment matrices are weighted and aggregated to obtain the clustering assignment matric which implements the clustering strategy to utilize graph structure and node feature information at the same time. Finally, comparative experiments and visualization on multiple graph classification datasets demonstrate the effectiveness of using graph node information and structure information to implement a clustering strategy in graph classification.

Key words: graph neural network, graph classification, graph pooling, clustering assignment matrix, hierarchical model