计算机科学与探索 ›› 2017, Vol. 11 ›› Issue (5): 802-813.DOI: 10.3778/j.issn.1673-9418.1601005

• 人工智能与模式识别 • 上一篇    下一篇

隐子空间聚类算法的改进及其增量式算法

董  琪+,王士同   

  1. 江南大学 数字媒体学院,江苏 无锡 214122
  • 出版日期:2017-05-01 发布日期:2017-05-04

Improved Latent Subspace Clustering Algorithm and Its Incremental Version

DONG Qi+, WANG Shitong   

  1. School of Digital Media, Jiangnan University, Wuxi, Jiangsu 214122, China
  • Online:2017-05-01 Published:2017-05-04

摘要: 基于稀疏表示的隐子空间聚类(latent subspace clustering,LSC)算法,相对于传统的子空间聚类算法,具有更快的聚类速度,使其适用于更大的数据集,但是其存在字典训练具有随机性,占用内存过多等缺陷。参照LC-KSVD字典训练算法的思想,通过将一部分信号的标签信息添加进字典训练阶段,以此提高了字典的判别性,进而提出了聚类精度更好的ILSC(improved LSC)算法。但相比于LSC算法,ILSC算法在字典训练阶段的耗时却大幅增加,针对此缺陷,参照增量字典训练的思想,提出了ILSC算法的增量式聚类算法I2LSC(incremental ILSC),在确保聚类精度、NMI(normalized mutual information)、RI(Rand index)值高于LSC且与ILSC相当的同时,较之ILSC具有更快的运行速度。

关键词: 子空间聚类, 隐子空间聚类(LSC), 判别式字典训练, 增量式字典训练

Abstract: Compared with the traditional subspace clustering algorithms, the latent subspace clustering (LSC) algorithm based on sparse representation has faster clustering speed, thereby it can be applied into larger data sets. However it still has two shortcomings. One is the randomness and slowness in dictionary training phase and the other is its occupying too much memory. On the basis of the LC-KSVD algorithm, this paper proposes the ILSC (improved LSC) algorithm to obtain its enhanced clustering accuracy by adding some labels into the dictionary training phase to improve its discrimination. However, compared with the LSC algorithm, ILSC algorithm consumes more time in dictionary training phase. In order to circumvent this drawback, based on the idea of incremental training, this paper develops the I2LSC (incremental ILSC) algorithm to achieve comparable clustering performance to ILSC algorithm in the sense of clustering accuracy, NMI (normalized mutual information) and RI (Rand index), but faster speed than ILSC.

Key words: subspace clustering, latent subspace clustering (LSC), discriminant dictionary training, incremental dictionary training