自适应参数调整的近邻传播聚类算法  

Affinity Propagation Clustering Algorithm for Adaptive Parameter Adjustment

作　　者：王卫涛 钱雪忠 曹文彬 WANG Wei-tao;QIAN Xue-zhong;CAO Wen-bin(Institute of Intelligent Systems and Network Computing, School of Interact of Things Engineering, Jiangnan University, Wuxi 214122, China;Engineering Research Center of Interact of Things Technology Applications Ministry of Education,Wuxi 214122 ,China;School of Interact of Things Engineering, Jiangnan University,Wuxi 214122, China)

机构地区：[1]江南大学物联网工程学院智能系统与网络计算研究所,江苏无锡214122 [2]物联网技术应用教育部工程研究中心,江苏无锡214122 [3]江南大学物联网工程学院,江苏无锡214122

出　　处：《小型微型计算机系统》2018年第6期1305-1311,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61673193)资助;中央高校基本科研业务费专项资金项目(JUSRP51635B;JUSRP51510)资助

摘　　要：针对近邻传播算法的偏向参数以及聚类类数对聚类结果准确性的影响.本文提出了自适应参数调整的GKAAP算法.首先,为了选取更合适的偏向参数,在传统AP算法的基础上,利用灰色狼群优化算法(GWO)自适应调节偏向参数;然后,为了使得偏向参数能够在合理的区间内搜寻,利用二分查找算法动态更新偏向参数的上限、下限、中间值;最后,为了使得聚类个数更接近真实类数,同时不影响聚类结果的准确性,在算法迭代完成后,通过数据集的真实簇数k来对聚类结果进行约束调整.本文通过10个UCI数据集和ORL人脸数据库来做对比实验,然后从准确率、算法时间、聚类个数三个维度去分析,最终实验结果证明本文所提出的GKAAP聚类准确性更好,算法时间复杂度更低.In view of the parameter preference and Clusters on the accuracy of the Clustering results about Affinity Propagation clustering algorithm. This paper presents a GKAAP algorithm for adaptive parameters adjustment. First,In order to select more suitable preference,Based on the traditional AP algorithm,The gray wolf optimization algorithm（ GWO） is used to adjust the preference adaptively;Second,In order to make the parameter preference p in a reasonable range of search,The upper,lower and intermediate values of the preference are dynamically updated by the binary search algorithm; Last,In order to make the number of clusters closer to the real number,without affecting the accuracy of clustering results,The clustering results are constrained by the real clusters k of the data set after the algorithm iteration. This paper use ten UCI data sets and ORL face database to do comparative experiments,and then from the accuracy rate,the algorithm time,the number of clusters to analyze the three dimensions. Finally,The experimental results show that the GKAAP clustering effect is better and the algorithm time has less time complxity.

关 键 词：近邻传播 灰色狼群 二分查找 约束调整 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]