基于图滤波器的符号属性图链路关系预测算法

来源：计算机技术与发展杂志2023年第09期北京时间：

作者：马华;姜伟;陈明;钟世杰

单位：湖南师范大学 信息科学与工程学院,湖南 长沙 410081 Author(s): MA Hua;JIANG Wei;CHEN Ming;ZHONG Shi-jie School of Information Science and Engineering,Hunan Normal University,Changsha 410081,China 关键词: 图滤波器;符号属性图;图神经网络;节点嵌入;链路关系预测 Keywords: graph filter;attributed and signed graphs;graph neural network;node embedding;link relationship prediction 分类号: TP311 DOI: 10. 3969 / j. issn. 1673-629X. 2023. 09. 019 摘要: 带节点属性的符号网络在信息学、生物学等多个领域存应用广泛,链路符号预测是该类数据分析中的一个热点问题。 基于符号图神经网络的模型是该问题的最新有效解决方案,但现有方法几乎均基于社会平衡理论,且未充分利用节点属性。 针对以上问题,从图信号处理角度设计了一个符号图神经网络,提出了一种端到端的符号属性图链路预测算法。首先,给出了基于低频和高频信号的带通滤波器的符号图神经网络,用于获得基于符号拓扑图的节点嵌入;其次,构造属性相似性图,利用图卷积网络得到属性相似性图节点嵌入;最后,引入注意力机制,融合符号拓扑图与属性相似性图两种节点表达,并将其输入符号判别器,通过 Adam 优化器训练模型。 在三个药物数据集上进行了对比实验与模型设置的影响分析。 与典型的符号图卷积网络与符号图谱嵌入,以及最近提出的基于图滤波的符号卷积网络的对比结果表明,该模型在 AUC 与 F1 指标上比最好的基线方法提升了 8. 68% 与 10. 04% 。 Abstract: Signed networks with node attributes are widely used in many fields such as informatics and biology. Link sign prediction is ahot topic in this kind of data analysis.?The model based on signed graph neural networks is the latest effective solution to this problem,but the existing methods are almost based on social balance theory,and do not make full use of node attributes. Aiming at the aboveproblems,a signed graph neural network was designed from the view of graph signal processing and an end-to-end link relationship prediction algorithm for attributed and signed graphs was proposed based on graph filter. Firstly,based on low frequency and high frequencysignals,a signed graph neural network with band-pass filter was proposed to obtain node embedding based on signed topology graph.Secondly,an attributed similarity graph was constructed and node representation was obtained by graph convolution networks. Finally,theattention mechanism was introduced to fuse two kinds of node representation of signed topology graph and attributed similarity graph. Itsinput was used to train symbol discriminator,and the model was trained by Adam optimizer. Experiments are conducted on three drugdatasets,in order to analyze the comparison to baselines and the effect of model settings. Compared with typical models such as signedgraph convolution network and signed spectral embedding models,and the newly proposed signed graph filtering-based convolutional network,the AUC and F1 metric values are superior to the best baseline method up to 8. 68% and 10. 04% ,respectively.

摘要：带节点属性的符号网络在信息学、生物学等多个领域存应用广泛,链路符号预测是该类数据分析中的一个热点问题。 基于符号图神经网络的模型是该问题的最新有效解决方案,但现有方法几乎均基于社会平衡理论,且未充分利用节点属性。 针对以上问题,从图信号处理角度设计了一个符号图神经网络,提出了一种端到端的符号属性图链路预测算法。首先,给出了基于低频和高频信号的带通滤波器的符号图神经网络,用于获得基于符号拓扑图的节点嵌入;其次,构造属性相似性图,利用图卷积网络得到属性相似性图节点嵌入;最后,引入注意力机制,融合符号拓扑图与属性相似性图两种节点表达,并将其输入符号判别器,通过 Adam 优化器训练模型。 在三个药物数据集上进行了对比实验与模型设置的影响分析。 与典型的符号图卷积网络与符号图谱嵌入,以及最近提出的基于图滤波的符号卷积网络的对比结果表明,该模型在 AUC 与 F1 指标上比最好的基线方法提升了 8. 68% 与 10. 04% 。

关键词：图滤波器;符号属性图;图神经网络;节点嵌入;链路关系预测