The True Self-adaptive Spectral Clustering Algorithms

XIE Juan-ying; DING Li-juan

doi:10.3969/j.issn.0372-2112.2019.05.004

您当前的位置：

首页 >

文章列表页 >

The True Self-adaptive Spectral Clustering Algorithms

更新时间：2025-07-16

- The True Self-adaptive Spectral Clustering Algorithms
- Acta Electronica Sinica Vol. 47, Issue 5, Pages: 1000-1008(2019)
- 作者机构：
  
  陕西师范大学计算机科学学院,陕西,西安,710062
- 作者简介：
- 基金信息：
  
  National Natural Science Foundation of China (No.61673251);National Key Research and Development Program of China (No.2016YFC0901900);Fundamental Research Funds for the Central Universities (No.GK2017010006);Postgraduate Cultivation Innovation Fund (No.2015CX028, No.2016CSY009)
- DOI：10.3969/j.issn.0372-2112.2019.05.004
  CLC： TP181
- Published Online：25 May 2019，
  
  Published：2019
- 稿件说明：
移动端阅览
XIE Juan-ying, DING Li-juan. The True Self-adaptive Spectral Clustering Algorithms[J]. Acta Electronica Sinica, 2019, 47(5): 1000-1008.
DOI：

XIE Juan-ying, DING Li-juan. The True Self-adaptive Spectral Clustering Algorithms[J]. Acta Electronica Sinica, 2019, 47(5): 1000-1008. DOI： 10.3969/j.issn.0372-2112.2019.05.004.

摘要

针对谱聚类算法self-tuning的局部尺度参数

会受噪音点影响，进而影响聚类结果，及其所使用的K-means算法的不稳定，对聚类结果的影响，提出两种完全自适应的谱聚类算法SC_SD（Spectral Clustering based on Standard Deviation）和SC_MD（Spectral Clustering based on Mean Distance），分别定义样本

的标准差、样本

到其余样本的距离均值，为样本

的邻域半径，统计邻域内的样本数，以样本

的邻域标准差为其局部尺度参数，避免样本

i的

局部尺度参数受噪音点影响，进而影响聚类结果；以方差优化初始聚类中心的SD_K-medoids算法代替K-means算法，克服K-means算法的不稳定，发现数据的真实分布.UCI数据集和人工数据集实验测试表明，提出的SC_SD和SC_MD算法能得到更优聚类结果，不受噪音点影响，有很好的伸缩性.提出的SC_SD和SC_MD能完全自适应地发现数据集的真实分布信息，尤其SC_MD算法很适合较大规模数据集的聚类分析.

Abstract

To avoid the clustering results with the local scaling parameter

of self-tuning may be influenced by outliers

and the unstable clustering results from K-means in self-tuning

two true self-adaptive spectral clustering algorithms were proposed.The two spectral clustering algorithms are respectively named as SC_SD(Spectral Clustering based on Standard Deviation) and SC_MD(Spectral Clustering based on Mean Distance).They respectively define the standard deviation of point

and the mean distance from point

to others

as its radius of neighborhood

then count the number of points in the neighborhood

and use the standard deviation of point

in the neighborhood as its local scaling parameter

so as to avoid the influence from outliers to the local scaling parameter

of point

and the distortion in clustering results of self-tuning.SD_K-medoids are adopted to instead of K-means in self-tuning to avoid the unstable clustering results of K-means

so as to get the true clustering of a dataset.The experimental results on UCI datasets and on synthetic datasets demonstrate that SC_SD and SC_MD can obtain better clustering results than that of traditional spectral clustering algorithm NJW and spectral clustering algorithm self-tuning

and are robust to noises

and has got good scalability.The proposed SC_SD and SC_MD can detect the clustering of a dataset without any given information

and the SC_MD can be used to detect the clustering of a comparable big data.

关键词

Keywords

references

Views

414

下载量

CSCD

Alert me when the article has been cited

提交

Tools

Publicity Resources

Spectral Clustering Ensemble Applied to SAR Image Segmentation Using Nonnegative Matrix Factorization

Neighborhood Definition for Catadioptric Omnidirectional Image Based on Resolution of Position

Automatic Spectral Clustering Based on Eigengap and Orthogonal Eigenvector

Density-Sensitive Spectral Clustering

A Spectral Clustering Method for Big Trajectory Data Mining with Latent Semantic Correlation

Related Author

LU Shan

JIAO Li-cheng

DENG Xiao-zheng

ZHANG Mao-jun

LIN Gao-peng

TAN Shu-ren

SUN Changsihe

DAI Guo-jun

Related Institution

Key Laboratory of Intelligent Perception and Image Understanding of Ministry of Education of China,Xidian University

College of Information Systems and Management, National University of Defense Technology

College of Computer Science,Hangzhou Dianzi University,,Zhenjiang

Department of Electronic and Computer Engineering,HKUST,Kowloon,Hongkong

College of Computer ScienceHangzhou Dianzi UniversityHangzhouZhenjiang 310018China

⁰