从拓扑的角度缓解图卷积网络的过平滑问题  

Relieving the over-smoothing problem for graph convolutional networks from the topological view

作　　者：柏玉 宋敏 刘士虎[1] 唐轶[1] 杨昔阳[2] BAI Yu;SONG Min;LIU Shi-hu;TANG Yi;YANG Xi-yang(School of Mathematics and Computer Science,Yunnan Minzu University,Kunming 650504,China;School of Mathematics and Computer Science,Quanzhou Normal University,Quanzhou 362000,China)

机构地区：[1]云南民族大学数学与计算机科学学院,云南昆明650504 [2]泉州师范学院数学与计算机科学学院,福建泉州362000

出　　处：《云南民族大学学报（自然科学版）》2022年第3期280-287,共8页Journal of Yunnan Minzu University:Natural Sciences Edition

基　　金：国家自然科学基金(61966039,61866040);全国高等院校计算机基础教育研究会计算机基础教育教学研究项目(2020-AFCEC-461);福建省自然科学基金(2021J01001)。

摘　　要：运用平滑度度量指标,定量分析了深层GCNs存在的过平滑问题,验证表明:过平滑是造成深层GCNs在半监督节点分类任务上的性能下降的主要原因.从拓扑学的角度,提出了一种DropEdge技术和混合阶传播相结合的方法,DropEdge技术使节点连接更加稀疏,而混合阶传播使模型包含更多的局部信息,从而一定程度上可避免GCNs加深时存在的过平滑问题.实验表明,方法可缓解深层GCNs存在的过平滑问题,使得半监督节点分类的精度更高.Using the problem of over-smoothness in deep GCNs is quantitatively analyzed by using the smoothness index, and the verification shows that over-smoothness is the main reason for the performance decline of deep GCNs in semi-supervised node classification task. From the perspective of topology, a method combining DropEdge technology and mixed-order propagation is proposed. DropEdge technology makes node connections more sparse, while mixed-order propagation makes the model contain more local information, thus avoiding the over-smoothness problem existing in GCNs deepening to a certain extent. Experimental results show that the method can alleviate the over-smooth problem of deep GCNs and improve the accuracy of semi-supervised node classification.

关 键 词：图卷积网络 过平滑 半监督节点分类 DropEdge 混合阶传播 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]