基于密度和相关分量分析的局部线性嵌入算法  

Density and relevant component analysis distance metric based locally linear embedding algorithm

作　　者：王楠 贾永利 韩淑运 杨裔[1] 李廉[1] 李彩虹[1] WANG Nan;JIA Yong-li;HAN Shu-yun;YANG Yi;LI Lian;LI Cai-hong(School of Information Science and Engineering,Lanzhou University,Lanzhou 730000,China)

机构地区：[1]兰州大学信息科学与工程学院,甘肃兰州730000

出　　处：《计算机工程与设计》2021年第5期1293-1299,共7页Computer Engineering and Design

基　　金：国家重点研发计划基金项目(2018YFB1003205);国家自然科学基金项目(61300230、61370219);甘肃省自然科学基金项目(1107RJZA188);甘肃省科技支撑计划基金项目(1104GKCA037);甘肃省科技重大专项基金项目(1102FKDA010)。

摘　　要：为解决局部线性嵌入算法(LLE)性能受初始邻域值大小和相似性度量选取的制约,提出一种基于密度和相关分量分析(relevant component analysis,RCA)的局部线性嵌入算法(DRLLE)。对每一个样本点计算一个密度缩放因子,根据密度缩放因子对样本点的初始邻域值进行自适应调整,计算RCA距离作为LLE算法的相似性度量,得到样本点的近邻集,进行降维处理。将DRLLE和其它LLE改进算法在Swiss roll、Swiss roll hole和ORL数据库上进行对比实验,其结果表明,DRLLE算法具有良好的降维效果和识别性能。The performance of local linear embedding algorithm(LLE)is restricted by initial neighborhood size and the selection of similarity measurement.To solve the problem,the density and relevant component analysis(RCA)distance metric based locally linear embedding algorithm(DRLLE)was proposed.Density scaling factor was calculated for each sample.The value of k was adaptively adjusted according to these density scaling factors.The RCA distance metric was calculated for the similarity measurement of LLE algorithm to get neighbor set of each points.The comparison result of DRLLE and other improved LLE algorithms on Swiss roll,Swiss roll hole with cavity and ORL data sets verifies that DRLLE is effective in dimensional reduction and images recognition.

关 键 词：降维 流行学习 局部线性嵌入 DRLLE算法 密度缩放因子 相关分量分析 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]