基于图神经张量网络的图相似度计算研究  

Research on Graph Similarity Calculation Based on Graph Neural Tensor Networks

作　　者：刘佳俊 庞慧[1,2] 严鑫瑜 LIU Jiajun;PANG Hui;YAN Xinyu(Hebei University of Architecture,Zhangjiakou,Hebei 075000;Big Data Technology Innovation Center of Zhangjiakou,Zhangjiakou,Hebei 075000)

机构地区：[1]河北建筑工程学院,河北张家口075000 [2]张家口市大数据技术创新中心,河北张家口075000

出　　处：《河北建筑工程学院学报》2024年第4期250-254,共5页Journal of Hebei Institute of Architecture and Civil Engineering

摘　　要：以图卷积神经网络为基础结合神经张量网络(NTN)来解决图相似性判断问题,该方法在减轻计算负担的同时也实现了较高的准确率,改进后的方法称为NTNGNN。首先,设计一个可学习的嵌入函数,它将每个图映射到一个嵌入向量,从而提供了一个图的全局特征;其次,应用了一种改进的全局上下文注意力机制,它强调不同节点对图的全局特征影响不同,从而更好地表示出整个图的全局特征。最后,以计算图编辑距离(GED)为例,在AIDS和LINUX图数据集上进行实验,实验结果体现了该方法的有效性和高效性。通过与一系列的基线模型和近似算法的GED计算结果进行比较,此模型均实现了更高的准确率。This paper proposes a method called NTNGNN,which combines Graph Convolutional Neural Networks(GCNNs)with Neural Tensor Networks(NTNs)to address the problem of graph similarity judgment.This approach achieves higher accuracy while reducing computational burden.Firstly,a learnable embedding function is designed to map each graph to an embedding vector,providing global features of the graph.Secondly,an improved global context attention mechanism is applied,emphasizing that different nodes have different impacts on the global features of the graph,thus better representing the overall graph features.Finally,taking Graph Edit Distance(GED)calculation as an example,experiments are conducted on the AIDS and LINUX graph datasets,demonstrating the effectiveness and efficiency of this method.Comparison with a series of baseline models and approximate algorithms in GED calculation shows that this model achieves higher accuracy.

关 键 词：图相似度计算 图神经网络 神经张量网络 全局上下文注意力机制 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]