基于PageRank传播机制的超图神经网络  

Hypergraph neural network based on PageRank propagation mechanism

作　　者：刘彦北 周敬涛 LIU Yan-bei;ZHOU Jing-tao(School of Electronics and Information Engineering,Tiangong University,Tianjin 300387,China)

机构地区：[1]天津工业大学电子与信息工程学院,天津300387

出　　处：《天津工业大学学报》2023年第2期67-73,共7页Journal of Tiangong University

基　　金：天津市自然基金资助项目(18JCQNJC70600);天津市教委科研计划资助项目(2017KJ087)。

摘　　要：针对传统超图神经网络卷积过程中层数过深过拟合以及传播范围小的问题,将超图神经网络(HGNN)与网页排名机制PageRank相结合,并利用个性化的改进传播方案,构建了基于PageRank传播机制的超图神经网络(HGNNP),在扩大学习领域的同时保持对根节点信息的有效关注,邻域范围扩大且可调节。在ModelNet40数据集和NTU数据集上对HGNNP的分类效果进行了验证。结果表明:HGNNP在ModelNet40数据集和NTU数据集上的最高分类准确率分别达到了93.07%和85.79%,相比HGNN分别提高了0.47%和8.59%,分类效果更好,且克服了过平滑问题;而且,随网络层数加深HGNNP的分类效果趋于稳定。Aiming at the problems that the number of layers is too deep,the propagation range is too small and overfitting of traditional hypergraph neural network(HGNN)in the convolution process,HGNN is combined with PageRank to construct hypergraph neural network based on PageRank propagation mechanism(HGNNP)using the personalized improved communication programs.HGNNP keeps effective attention to the root node information while expanding the learning field,expands and adjusts the neighborhood range.The classification effect of HGNNP is validated on ModelNet40 dataset and NTU dataset.The experimental results show that the highest classification accuracy of HGNNP on ModelNet40 dataset and NTU dataset is 93.07%and 85.79%,which is 0.47%and 8.59%higher than that of HGNN,respectively.Compared with HGNN,HGNNP has better classification effect and overcomes the over-smoothing problem,and tends to be stable with the increase of network layers.

关 键 词：超图 超图神经网络 网页排名机制 半监督分类 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]