基于对抗学习邻域注意网络的链路预测  

LINK PREDICTION BASED ON NEIGHBORHOOD ATTENTION NETWORKOF CONFRONTATION LEARNING

作　　者：代秀珍[1] Dai Xiuzhen(Baotou Railway Vocational and Technical College,Baotou 014060,Inner Mongolia,China)

机构地区：[1]包头铁道职业技术学院,内蒙古包头014060

出　　处：《计算机应用与软件》2023年第9期78-87,170,共11页Computer Applications and Software

基　　金：内蒙古自治区高等学校科学研究项目自然科学重点项目(NJZZ17539)。

摘　　要：针对传统链路预测模型无法充分挖掘邻域特征信息,且模型的泛化能力不强等缺点,提出一种基于对抗学习邻域注意网络的链路预测。设计一种新的多空间邻域注意机制,通过捕捉邻域的潜在重要性来提取单个邻域特征。进一步提出自邻域注意网络和扩展跨邻域注意网络,前者通过编码和匹配各自的邻域信息来预测两个节点之间的链接,后者设计一个跨邻域注意来实现两个节点之间的直接捕获结构交互作用。另外提出一个对抗性学习框架,设计了一个负样本生成器在对抗性博弈中持续提供高信息量的负样本。在包含多种网络类型的12个基准数据集上评估所提出方法的性能,实验结果证明了提出方法的优越性。The traditional link prediction model can't fully mine the neighborhood feature information,and the generalization ability of the model is not strong.Aimed at this problem,a link prediction method based on neighborhood attention network of confrontation learning is proposed.A new multi-spatial neighborhood attention mechanism was designed to extract the features of a single neighborhood by capturing the potential importance of the neighborhood.Furthermore,self-neighborhood attention network and extended cross neighborhood attention network were proposed.The link between two nodes by encoding and matching their respective neighborhood information was predicted by the former,and a cross neighborhood attention network to realize the direct capture structure interaction between two nodes was designed by the later.In addition,a framework of antagonistic learning was proposed,and a negative sample generator was designed to continuously provide negative samples with high amount of information in antagonistic games.The performance of the proposed method is evaluated on 12 benchmark datasets with various network types,and the experimental results verify the superiority of the proposed method.

关 键 词：链路预测 邻域特征 对抗学习 注意网络 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]