计算机应用研究2024，Vol.41Issue(11)：3343-3349,7.DOI:10.19734/j.issn.1001-3695.2024.03.0072
面向无监督特征提取的结构化图嵌入
Structured graph embedding for unsupervised feature extraction
摘要
Abstract
Feature extraction is one of the most effective tools for processing high-dimensional data.However,existing feature extraction methods suffer from two problems:they do not capture both the local and global structures of the data simultaneous-ly,the constructed graph is disconnected from the number of data cluster and does not have an exact connected component.To address these issues,this paper proposed a SGE for unsupervised feature extraction.By constructing K-nearest neighbor graph for data representation and using least squares regression,SGE can simultaneously respect the local and global correlation structures of the data.Moreover,by enforcing rank constraints on the Laplacian matrix of the representation,SGE constructed the optimal graph of c connected components with c clusters,and thus revealed the clustering structure of the data.Therefore,the proposed SGE can find more discriminative projections.Experiments on real-world datasets show that SGE outperforms other mainstream dimensionality reduction methods.Especially on the PIE dataset,the clustering accuracy of SGE is 18.7％higher than that of LRPP_GRR.These results indicate that the proposed SGE algorithm can effectively reduce the dimensionality of the data.关键词
特征提取/局部结构/秩约束/最小二乘回归Key words
feature extraction/local structure/rank constraint/least squares regression分类
信息技术与安全科学引用本文复制引用
袁凤燕,尹学松,王毅刚..面向无监督特征提取的结构化图嵌入[J].计算机应用研究,2024,41(11):3343-3349,7.基金项目
浙江省高等学校国内访问学者资助项目(FX2023191) （FX2023191）
浙江开放大学312人才培养工程资助项目 （）
浙江省公益技术应用研究项目(LGG22F020032) （LGG22F020032）
温州市基础性公益科研项目(G2023093) （G2023093）
浙江省重点研发计划重点专项资助项目(2021C03137) （2021C03137）
