摘要
谱聚类算法中,当样本的簇边缘分布不均匀或不同簇边缘分布密度相近时,会导致错分现象。通过对相似度矩阵的改进,提出基于流形距离核的自适应迁移谱聚类算法。使用流形距离作为构造相似度矩阵的度量方法,共享近邻方法对相似度矩阵进行自适应调整,且使用加权距离自适应调节核参数,提高谱聚类对复杂数据集的处理能力。针对样本匮乏或受到污染时聚类效果不佳问题,引入迁移学习,利用源域知识指导目标域进行聚类。经实验验证,该算法性能优于传统谱聚类算法。
In the spectral clustering algorithm,when the cluster edge distribution of the sample is uneven or the density of the cluster edge distribution is similar,it will lead to the phenomenon of misclassification.This paper proposes an adaptive transfer spectrum clustering algorithm based on manifold distance kernel through the improvement of similarity matrix.It used manifold distance as the metric method of constructing the similarity matrix,and the shared neighbor method adaptively adjusted the similarity matrix.The weighted distance was used to adjust the kernel parameters to improve the processing ability of spectral clustering on complex data sets.The transfer learning was introduced to solve the problem of poor clustering effect when samples were scarce or contaminated.The knowledge of source domain was used to guide the target domain for clustering.The performance of the algorithm is better than that of the traditional spectral clustering algorithm by experiments.
作者
齐晓轩
都丽
洪振麒
Qi Xiaoxuan;Du Li;Hong Zhenqi(College of Applied Technology,Shenyang University,Shenyang 110044,China;College of Information Engineering,Shenyang University,Shenyang 110044,China)
出处
《计算机应用与软件》
北大核心
2020年第8期265-273,共9页
Computer Applications and Software
基金
辽宁省重点研发计划项目(2018104012)
沈阳市中青年科技创新人才资助计划项目(RC180362)。
关键词
相似度矩阵
流形距离核
谱聚类
迁移学习
全局一致性
局部结构
自适应
Similarity matrix
Manifold distance kernel
Spectral clustering
Transfer learning
Global consistency
Local structure
Adaptive
