融合基序信息的图同构注意力网络的图分类问题研究  

Graph Classification Via Graph Isomorphic Attention Networks with Motif Information

作　　者：衡红军[1] 曹莹莹 HENG Hongjun;CAO Yingying(College of Computer Science and Technology,Civil Aviation University of China,Tianjin 300300,China)

机构地区：[1]中国民航大学计算机科学与技术学院,天津300300

出　　处：《小型微型计算机系统》2025年第3期552-558,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(U1333109)资助.

摘　　要：基于频繁子图挖掘算法的图分类方法无法避免子图同构计算,算法的效率低且忽略了节点特征信息,而基于图神经网络的方法则关注节点特征信息.本文提出一种融合基序信息的图同构注意力网络的图分类方法.该方法首先利用图的拓扑结构和节点类别信息,提取数据集中的子图结构构成基序集合,再基于基序集合生成基序级图嵌入表示,避免了频繁子图挖掘;然后在图同构网络的池化操作中引入全局注意力机制,学习高质量的节点级图嵌入表示;最后将基序级和节点级图嵌入表示拼接起来用于图分类.该图嵌入表示不仅包含了图中节点的特征信息,也反映了图的结构特征信息.实验结果表明,所构建的网络模型在五个公开数据集上取得了优异的分类精度.The graph classification method based on frequent subgraph mining algorithm cannot avoid subgraph isomorphism calculation.These methods have low algorithm efficiency and ignore node feature information,while the methods based on graph neural networks focus on node feature information.We propose a graph isomorphic attention network that integrates motif information.This method first utilizes the graph′s topological structure and node category information to extract subgraph structures from the dataset,forming a set of motifs.Then,it generates motif-level graph embeddings based on this motif set,avoiding the need for frequent subgraph mining.Next,a global attention mechanism is introduced in the pooling operation of the graph isomorphism network to learn high-quality node-level graph embeddings.Finally,the motif-level and node-level graph embeddings are concatenated and used for graph classification.This graph embedding representation not only includes the node feature information in the graph but also reflects the structural characteristics of the graph.Experimental results demonstrate that the constructed network model achieves excellent classification accuracy on five public datasets.

关 键 词：图分类 图神经网络 基序 全局注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]