• 中国计算机学会会刊
  • 中国科技核心期刊
  • 中文核心期刊

J4 ›› 2015, Vol. 37 ›› Issue (08): 1584-1590.

• 论文 • 上一篇    下一篇

基于低秩稀疏图的结构保持投影算法

杨国亮, 罗璐,丰义琴, 梁礼明   

  1. (江西理工大学电气工程与自动化学院,江西 赣州 341000)
  • 收稿日期:2014-08-11 修回日期:2014-11-11 出版日期:2015-08-25 发布日期:2015-08-25
  • 基金资助:

    国家自然科学基金资助项目(51365017,61305019);江西省科技厅青年科学基金资助项目(20132bab211032)

Structure preserving projection algorithm
based on low rank and sparse graph  

YANG Guoliang,LUO Lu,FENG Yiqin,LIANG Liming   

  1. (School of Electrical Engineering and Automation,Jiangxi University of Science and Technology,Ganzhou 341000,China)
  • Received:2014-08-11 Revised:2014-11-11 Online:2015-08-25 Published:2015-08-25

摘要:

在图嵌入理论框架下,能够较好地揭示数据本质特性的图在一些维数约简方法中起到关键性的作用。基于稀疏表示和低秩表示方法,构建了一种低秩稀疏图,能够同时揭示数据的局部结构信息和全局结构信息。然后,利用图嵌入理论方法使这些特性在线性投影的过程中得以保持不变,从而学习出高维数据有效的低维嵌入。在标准的人脸和手写数字数据集(ORL,Yale,PIE,MNIST)上进行实验,同传统的图嵌入方法比较,结果表明了算法的有效性。

关键词: 图嵌入, 稀疏表示, 低秩表示, 低秩稀疏图, 线性投影

Abstract:

In the unifying frameworks like graph embedding,constructing a good graph to represent data properties is critical for dimensionality reduction technology.In this paper,we construct a low rank and sparse graph to reveal local and global structure information of the data based on sparse representation and low rank representation.We first use graph embedding technology to preserve such properties during the linear projections, and then obtain the lowdimensional embedding of the original highdimensional data. The effectiveness of the proposed method is compared with the stateoftheart algorithms and is verified on face and handwritten digit databases (ORL,Yale,PIE,MNIST). Ke

Key words: graph embedding;sparse representation;low rank representation;low rank and sparse graph;linear projections