摘要
为了研究节点特征表示的不确定性对节点分类的影响,提出一种可信的图神经网络节点分类方法。算法使用径向基函数计算节点间距离,得到各类节点质心后,根据距离分配与未标记节点最近质心的类别标签提高节点分类性能,同时定义未标记节点和质心之间的距离为模型输出的不确定性,并使用梯度惩罚损失加强输入变化的可检测性,可以有效地检测分布外节点样本。在Cora、Citeseer和Pubmed这3个公开网络数据集上的结果表明:模型在分类任务的AUROC指标分别达到81.5%、76.2%和74.6%,在分布外样本检测任务中AUROC指标分别达到83.6%、72.8%和70.6%,证明了所提算法在提高节点分类性能的同时,可以有效检测分布外的节点样本,提高了节点分类的可信性。
In order to study the influence of uncertainty of node feature representation on node classification,a node classification method based on trusted graph neural network was proposed.The algorithm used the radial basis function to calculate the distance between nodes,and after obtaining the centroid of various nodes,the classification label of the nearest centroid was allocated according to the distance to improve the classification performance.Additionally,the distance between unlabeled nodes and centroids is defined as the uncertainty of the model′s output.A gradient penalty loss is employed to strengthen the detectability of input variations,loss to strengthen the detectability of input changes,which can effectively detect the distributed outer node samples.The results in classification task are 81.5%,76.2% and 74.6% in terms of AUROC on three public network datasets of Cora,Citeseer and Pubmed,respectively.And the results in the out-of-distribution sample detection task are 83.6%,72.8% and 70.6% in terms of AUROC on three public network datasets of Cora,Citeseer and Pubmed,respectively.It proves that the proposed algorithm can effectively detect the node samples outside the distribution and improve the credibility of node classification,while improving the node classification performance.
作者
刘彦北
马夕然
王雯
LIU Yanbei;MA Xiran;WANG Wen(School of Life Sciences,Tiangong University,Tianjin 300387,China;Tianjin Key Laboratory of Optoelectronic Detection Technology and System,Tiangong University,Tianjin 300387,China;School of Electronics and Information Engineering,Tiangong University,Tianjin 300387,China)
出处
《天津工业大学学报》
CAS
北大核心
2024年第1期82-88,共7页
Journal of Tiangong University
基金
京津冀基础研究合作专项项目(H2021202008)
中国博士后科学基金面上项目(2022M712370)。
关键词
图神经网络
节点分类
分布外检测
不确定性估计
梯度惩罚
graph neural network
node classification
extra-distribution detection
uncertainty estimation
gradient penalty
