GHPN:面向半监督小样本节点分类的图超球面原型网络  

Graph Hypersphere Prototype Network for Semi-supervised Few-shot Node Classification

作　　者：徐祖豪 陈鑫龙 李进 黄益颂 傅仰耿[1] XU Zuhao;CHEN Xinlong;LI Jin;HUANG Yisong;FU Yanggeng(College of Computer and Data Science,Fuzhou University,Fuzhou 350108,China;Information Hub,HKUST(Guangzhou),Guangzhou 511455,China)

机构地区：[1]福州大学计算机与大数据学院,福州350108 [2]香港科技大学(广州)信息枢纽,广州511455

出　　处：《小型微型计算机系统》2025年第3期542-551,共10页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(12271098)资助;福建省高校产学合作科技计划项目(2023H6008)资助.

摘　　要：图神经网络已经成功应用于各种与图相关的任务中.以有监督的方式训练一个图神经网络需要大量标签,而现实世界中受到成本制约难以获取大量标签,因此在小样本学习或半监督学习场景的标签就更为稀少.为了克服这个问题,许多方法通过标签传播的方法来估计标签,但通常会受到图上连接性和同质性假设的限制,容易生成带有噪声的伪标签.为了解决这些限制,本文提出了一个名为图超球面原型网络的新方法GHPN,专注于半监督小样本节点分类.为了减轻图结构对预测结果的影响,GHPN在超球面表示空间中建模类别表示,通过类级别表示在语义空间中传播标签信息.此外,为了利用未标记节点的监督信息,本文设计了一个基于原型网络预测结果的负学习框架,用于补充监督信号,调整各类别原型之间的距离.在5个真实世界的数据集上进行的实验表明,该方法与10个最先进的方法相比能够有效提高性能,在4个数据集上能取得平均排名最佳结果.Graph Neural Networks(GNNs)have shown great success in various graph-related tasks.However,they heavily rely on supervised labels,which can become a problem in few-shot or semi-supervised learning settings where labeled examples are limited and expensive.To overcome this issue,many methods estimate pseudo labels based on label propagation,but they are often restricted by connectivity and homogeneity hypothesis on graphs.To address this,this method proposes a new framework called Graph Hypersphere Prototype Network(GHPN)that focuses on node classification with limited labels.This method models class representations in the hyperspherical space,which effectively mitigates the influence of the graph structure and propagates label information through class-level semantics.Additionally,this method introduces a negative learning framework based on the model′s prediction to improve supervision and fully leverage numerous unlabeled nodes.Results from experiments conducted on five real-world datasets demonstrate that this method enhances the performance of GNNs in node classification with limited labels compared to ten state-of-the-art counterparts.

关 键 词：半监督学习 图表示学习 小样本学习 原型网络 负学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]