基于KL散度的信息增益图池化方法  

Information gain graph pooling based on KL divergence

作　　者：张宏宽 朱小草 施浏晟 刘雨星 郭春生[2] ZHANG Hongkuan;ZHU Xiaocao;SHI Liusheng;LIU Yuxing;GUO Chunsheng(SOYEA Technology Co.,Ltd.,Hangzhou Zhejiang 310012,China;School of Communication Engineering,Hangzhou Dianzi University,Hangzhou Zhejiang 310018,China)

机构地区：[1]数源科技股份有限公司,浙江杭州310012 [2]杭州电子科技大学通信工程学院,浙江杭州310018

出　　处：《杭州电子科技大学学报(自然科学版)》2025年第1期53-59,共7页Journal of Hangzhou Dianzi University:Natural Sciences

摘　　要：针对图神经网络池化中不能充分保留图的局部特征的问题,提出一种基于KL散度的信息增益图池化方法。首先,对使用KL散度衡量节点与加权邻域两者之间概率分布的差异得到每个节点的信息增益;然后,信息增益进行直接和局部归一化排序,获得topK个节点对应的索引序列;最后,基于邻域进行边的重构以得到最终的低分辨率图。实验结果表明,与图池化方法AttPool、GXN相比,提出方法的分类正确率分别平均提升了4.69%和0.73%。Aiming at the problem that local features of graph cannot be thoroughly preserved in graph neural network pooling,an information gain graph pooling method based on KL divergence is proposed.Firstly,KL divergence is used to measure the difference of probability distribution between nodes and weighted neighborhood to get the information gain of each node;then,the information gain is sorted by direct and local normalization to obtain the index sequence corresponding to the topK nodes;finally,edge reconstruction is performed based on the neighborhood to get the final low resolution image.Experimental results show that,compared with prevailing graph pooling methods attpool and GXN,the classification accuracy of the proposed method is improved by 4.69%and 0.73%,respectively.

关 键 词：图池化 KL散度 信息增益 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]