基于图粗化的层次图池化方法研究  

Research on Hierarchical Graph Pooling Based on Graph Coarsening

作　　者：陈洁[1,2,3] 薛远远[1,2,3] 曹京晶 赵姝 张燕平 CHEN Jie;XUE Yuan-yuan;CAO Jing-jing;ZHAO Shu;ZHANG Yan-ping(Key Laboratory of Computational Intelligence and Signal Processing,Ministry of Education,Hefei 230601,China;College of Computer Science and Technology,Anhui University,Hefei 230601,China;Key Laboratory of Information Materials and Intelligent Sensing of Anhui Province,Hefei 230601,China;Science and Technology Talent Exchange Development Service Center of Ministry of Science and Technology,Beijing 100045,China)

机构地区：[1]计算智能与信号处理教育部重点实验室,合肥230601 [2]安徽大学计算机科学与技术学院,合肥230601 [3]安徽省信息材料与智能传感重点实验室,合肥230601 [4]科学技术部科技人才交流开发服务中心,北京100045

出　　处：《小型微型计算机系统》2023年第3期483-489,共7页Journal of Chinese Computer Systems

基　　金：国家重点研发计划子课题项目(2017YFB1401903)资助;国家自然科学基金项目(61876001)资助;安徽省高校自然科学基金项目(KJ2021A0039)资助。

摘　　要：图神经网络(Graph Neural Networks, GNNs)已被证明能有效对图结构数据进行建模,池化机制在使用GNN模型提取图层次特征过程中至关重要,近年来已经引起了越来越多研究者们的关注.现有基于聚类的层次图池化方法要么需要增加额外的神经网络层以实现特征图的粗化;要么不能从全局角度捕获节点在图中的重要性大小.针对以上问题,本文提出一种基于图粗化的层次图池化方法(Hierarchical Graph Pooling Based on Graph Coarsening, HGP-GC),用于学习图的层次特征表示.该方法主要包括图结构粗化和图属性粗化两个部分.利用结构粗化实现特征图尺寸的缩减;利用属性粗化突显图中重要节点对图级表示的关键作用.通过将HGP-GC池化策略与现有神经网络相结合,在不同规模公共数据集上的图分类实验结果证明了HGP-GC的有效性.Graph Neural Networks have been proved to be effective in modeling graph-structured data.The pooling mechanism is very important in the process of extracting graph hierarchical features using GNN models, which has attracted more and more researchers’ attention in recent years.The existing hierarchical graph pooling methods based on clustering either need to add an extra neural network layer to achieve the coarsening of the feature graph, Or you can’t capture the importance of the node in the graph from a global perspective.To solve the above problems, a Hierarchical Graph Pooling Based on Graph Coarsening is proposed for learning the hierarchical feature representation of graphs.This method mainly includes two parts: graph structure coarsening and graph attribute coarsening.Structure coarsening is used to reduce the size of the feature map.Attribute coarsening is used to highlight the key role of important nodes in the graph to the graph-level representation.By combining the HGP-GC pooling strategy with the existing neural network, the effectiveness of HGP-GC is proved by the experimental results of graph classification on common datasets of different sizes.

关 键 词：图神经网络 图池化 层次图表示学习 节点重要性 图分类 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]