基于对偶图正则化非负矩阵分解的链路预测  

Link Prediction Based on Dual-graph Regularized Nonnegative Matrix Factorization

作　　者：陈广福 阎兵早 CHEN Guangfu;YAN Binzao(College of Mathematics and Computer Science,Wuyi University,Wuyishan 354300,China;The Key Laboratory of Cognitive Computing and Intelligent Information Processing in Fujian Insitutions of High Learning,Wuyishan 354300,China)

机构地区：[1]武夷学院数学与计算机学院,福建武夷山354300 [2]认知计算与智能信息处理福建省高校重点实验室,福建武夷山354300

出　　处：《湖北民族大学学报（自然科学版）》2022年第3期336-343,共8页Journal of Hubei Minzu University:Natural Science Edition

基　　金：福建省自然科学基金项目(2021J011146);武夷学院引进人才科研启动基金项目(YJ202017).

摘　　要：现存大部分链路预测方法仅考虑单类型网络结构信息,无法同时保持一阶、局部和全局结构信息,导致预测精度下降.针对以上不足,提出对偶正则化非负矩阵分解的链路预测模型,同时保持一阶、局部和全局结构信息.首先,将无向无权的邻接矩阵映射到低维潜在空间保持网络一阶结构;其次,利用随机游走方法捕获整个网络节点相似度,再启用图正则化技术保持全局结构;此外,利用杰卡尔德系数获得局部相似度去探索网络局部结构;最后,将一阶、局部和全局结构信息相融合构建统一链路预测模型,并启用迭代更新规则学习模型参数获得局部最优.在6个真实网络上进行实验,运用AUC(areas under curve)和AUPR(areas under precision-recall)度量对所提模型进行评估,实验结果表明AUC和AUPR值分别提升了3.1%和8.9%.Most of the existing link prediction methods only consider the single type network structure information and fail to maintain the first-order,local and global structure information at the same time,resulting in the decline of prediction accuracy.In view of the above shortcomings,a link prediction model based on dual-graph regularized nonnegative matrix factorization is proposed to maintain the first-order,local and global structure information at the same time.Firstly,any undirected and unweighted adjacency matrix is mapped to a low dimensional potential space to maintain the first-order structure of the network.Secondly,the random walk method is used to capture the global similarity of the network,and then the graph regularization technology is used to maintain the global structure information.In addition,the local similarity is obtained by using Jaccard coefficient,and then the local structure information is maintained by using graph regularization technology.Finally,the unified link prediction model is constructed by fusing the above three types of structural information.In addition,iterative updating rules are enabled to learn model parameters to obtain local optimization.Experiments are carried out on six real networks,and the proposed model is evaluated by using AUC and AUPR measures.The experimental results show that the values of AUC and AUPR are increased by 3.1%and 8.9%respectively.

关 键 词：复杂网络 链路预测 非负矩阵分解 对偶图正则化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]