面向无监督特征提取的结构化图嵌入  

Structured graph embedding for unsupervised feature extraction

作　　者：袁凤燕 尹学松 王毅刚[2] Yuan Fengyan;Yin Xuesong;Wang Yigang(Pinghu School,Zhejiang Open University,Pinghu Zhejiang 314200,China;Dept.of Digital Media Technology,Hangzhou Dianzi University,Hangzhou 310018,China)

机构地区：[1]浙江开放大学平湖学院,浙江平湖314200 [2]杭州电子科技大学数字媒体技术系,杭州310018

出　　处：《计算机应用研究》2024年第11期3343-3349,共7页Application Research of Computers

基　　金：浙江省高等学校国内访问学者资助项目(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);温州市基础性公益科研项目(G2023093);浙江省重点研发计划重点专项资助项目(2021C03137)。

摘　　要：特征提取是处理高维数据的最有效工具之一。然而,当前特征提取方法存在两个问题:一是它们没有同时捕捉数据的局部和全局结构;二是构建的图脱离数据的聚类数,没有与聚类相同的连通分量。为了解决这些问题,提出了面向无监督特征提取的结构化图嵌入方法(structured graph embedding, SGE)。通过构建数据表征的K近邻和使用最小二乘回归,SGE能够同时保持数据的局部与全局相关结构。而且,SGE对表征图的拉普拉斯矩阵施加秩约束,使构建的最优图具有与c个聚类一致的c个连通分量,从而能揭示数据的聚类结构。因此,SGE能够找到更有判别力的投影。在多个真实数据集的实验表明,SGE优于其他主流降维方法。特别是在PIE数据集上,SGE的聚类精度比LRPP_GRR的聚类精度高出18.7%。这些结果表明SGE方法可以有效降低数据维数。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 simultaneously,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 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]