一种基于密度峰值聚类的链路预测算法  

Link Prediction Algorithm Based on Density Peak Clustering

作　　者：邵豪 王伦文 邓健 SHAO Hao;WANG Lun-wen;DENG Jian(College of Electronic Engineering,National University of Defense Technology,Hefei 230031,China;Shijiazhuang Campus,Army Engineering University of PLA,Shijiazhuang 050003,China)

机构地区：[1]国防科技大学电子对抗学院,合肥230031 [2]陆军工程大学石家庄校区二系,石家庄050003

出　　处：《小型微型计算机系统》2020年第5期1007-1012,共6页Journal of Chinese Computer Systems

基　　金：国防科技创新特区项目(17-H863-01-ZT-003-204-03)资助.

摘　　要：传统基于网络结构的链路预测算法只考虑单个节点相似性指标,在结构不同的网络中预测结果差异明显且预测精度低.针对此问题,本文考虑不同指标的互补性,提出一种自适应融合多指标的链路预测算法.首先改进传统路径相似性指标未完全挖掘路径信息的缺点,提出PLD和INR指标,分别考虑路径中间链接及中间节点连通性对预测的贡献以提升预测性能;其次将节点间是否存在链接的预测问题转变为二分类问题,并将上述指标与邻居相似性指标、随机游走指标结合进行链路预测;再次利用密度峰值聚类进行无监督学习,根据学习结果预测链路.仿真实验结果表明该算法在各个网络的预测精度都明显高于传统相似性预测算法.Traditional link prediction algorithms based on network structure only consider the single similarity index The prediction results in different networks have obvious differences and the prediction accuracy is low.To solve this problem,this paper considered the complementarity of different indexes,and proposed a link prediction algorithm based on adaptive fusion of multiple indexes.Firstly,we proposed two improved path similarity indexes PLD and INR,which consider the contribution of path intermediate links and intermediate nodes to prediction respectively;Secondly,we transformed the problem of whether there were links between nodes into a two-class problem and combined the above indexes with neighborhood similarity index,random walk index as multi-dimensional attributes of node pairs;Finally,we classified node pairs by density peak clustering and determined the link properties of each node pair according to the classification results.The simulation results show that the prediction accuracy of proposed algorithm is significantly higher than that of traditional similarity prediction algorithms in various networks.

关 键 词：复杂网络 链路预测 相似性指标 密度峰值聚类 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]