计算机工程与应用 ›› 2015, Vol. 51 ›› Issue (2): 165-170.

• 图形图像处理 • 上一篇    下一篇

基于密度与路径的稳健谱聚类

许洪玮,曹江中,何家峰,戴青云   

  1. 广东工业大学 信息工程学院,广州 510006
  • 出版日期:2015-01-15 发布日期:2015-01-12

Robust density-path-based spectral clustering

XU Hongwei, CAO Jiangzhong, HE Jiafeng, DAI Qingyun   

  1. School of Information Engineering, Guangdong University of Technology, Guangzhou 510006, China
  • Online:2015-01-15 Published:2015-01-12

摘要: 近年来,谱聚类在分类领域得到了广泛的研究,其中基于路径和基于密度的算法是两个重要的研究方向。虽然这两种算法在一些数据集上能取得较好的分类效果,但不能对一些特殊的数据集进行准确分类。融合了这两种方法的优点,通过多级密度约束来寻找路径,根据得到的路径建立新的相似性矩阵。为了加强对噪声的鲁棒性,根据数据集的局部信息加入鲁棒性系数,提出了基于路径与密度的稳健谱聚类算法。实验结果表明该方法在人工数据集和手写体数据集上能取得较理想的分类结果。

关键词: 谱聚类, 基于路径的谱聚类, 基于密度的谱聚类

Abstract: Spectral clustering is wildly studied in the field of class identification in recent years, of which the path-based and density-based algorithms are the two main research topics. These two algorithms have delivered impressive results in some data sets, but are ineffective for some special cases. This paper unites the advantages of these two algorithms and finds the paths under the multi-levels restriction of density, after which a new similarity measure is built. In order to enhance the robustness against noises, robust coefficients are added based on the local information of data set, thus a robust density-
path-based spectral clustering method is proposed. Experimental results on synthetic data sets as well as real world data sets demonstrate that the proposed method can obtain more than acceptable results.

Key words: spectral clustering, path-based spectral clustering, density-based spectral clustering