基于缓解特征过度平滑的图神经网络优化算法  

Optimization Algorithm for Graph Neural Networks Based on Alleviating Feature Over-Smoothing

作　　者：林科奥 翁伟 谢小竹 王华伟 文娟[4] LIN Keao;WENG Wei;XIE Xiaozhu;WANG Huawei;WEN Juan(School of Computer and Information Engineering,Xiamen University of Technology,Xiamen 361024,China;Fujian Key Laboratory of Pattern Recognition and Image Understanding,Xiamen 361024,China;School of Information Science and Technology,Xiamen University Tan Kah Kee College,Zhangzhou 363105,China;School of Economics,Xiamen University,Xiamen 361005,China)

机构地区：[1]厦门理工学院计算机与信息工程学院,福建厦门361024 [2]福建省模式识别与图像理解重点实验室,福建厦门361024 [3]厦门大学嘉庚学院信息科学与技术学院,福建漳州363105 [4]厦门大学经济学院,福建厦门361005

出　　处：《厦门理工学院学报》2024年第3期66-73,共8页Journal of Xiamen University of Technology

基　　金：国家社会科学基金项目“复杂网络视角下投入产出数据的应用研究”(21BTJ011)。

摘　　要：为减少过度平滑对传统图卷积网络(graph convolutional network,GCN)模型性能的影响,提出一种优化算法GCN-optimization。该算法通过增强节点特征并在卷积前将特征重新映射的方法,使节点在经过多层GCN传播过程中仍能保持一定的特征差异。在Cora、Citeseer和Pubmed 3个数据集上进行实验,结果显示:在3个数据集中,相比于原版GCN,GCN-optimization算法Accuracy分别提升2.2%、1.5%和0.5%;Macro-F1分别提升1.8%、1.7%和2.1%。表明,相对于基准模型,GCN-optimization算法在节点分类任务中展现出一定的优势,能够有效缓解传统GCN中的过度平滑问题,保持节点特征的差异性,从而提升模型性能。To mitigate the impact of over-smoothing on the performance of traditional graph convolutional network(GCN)model,an optimization algorithm called GCN-optimization is proposed.This algorithm enhances node features and remaps them before convolution,allowing nodes to maintain certain feature differences even after multi-layer GCN propagation.Experimental results on datasets of Cora,Citeseer and Pubmed show that compared to the original GCN,the accuracy of the GCN-optimization algorithm increased by 2.2%,1.5%and 0.5%respectively,and Macro-F1 increased by 1.8%,1.7%and 2.1%respectively.This indicates that the GCNoptimization algorithm demonstrates certain advantages over the baseline model in node classification tasks.It can effectively alleviate the over-smoothing problem in traditional GCN,maintain the diversity of node features,and thereby improve model performance.

关 键 词：图神经网络 优化算法 图卷积网络 过度平滑 节点分类 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]