基于核协方差矩阵的无监督数据聚类  

UNSUPERVISED DATA CLUSTERING BASED ON KERNEL COVARIANCE MATRIX

作　　者：程宁[1] 戴远泉[1] Cheng Ning;Dai Yuanquan(School of Computer Science,Hubei Institute of Light Industry,Wuhan 430070,Hubei,China)

机构地区：[1]湖北轻工职业技术学院计算机学院,湖北武汉430070

出　　处：《计算机应用与软件》2023年第5期288-296,共9页Computer Applications and Software

基　　金：湖北省教育科学规划2018年度立项课题项目(175)。

摘　　要：为解决数据向量聚类模型过于依赖先验知识以及有监督训练问题,提出一种基于核协方差矩阵的无监督数据聚类方法。将核学习和基于矩阵因子化的聚类问题耦合到一个联合公式中,使得核协方差矩阵的秩等于聚类的数目,从而确保来自不同类的元素弱相关,类内数据强相关;根据数据的信息含量,利用核协方差矩阵的稀疏相关矩阵分解对数据进行聚类,在生成稀疏矩阵因子时引入基于l_(1)-l_(2)的稀疏性度量,从而有效地降低计算复杂性;利用基于凸函数差分算法的优化公式,解决全局的、高度非凸的问题。通过高光谱图像、人类活动和文档分类三个数据集的验证,结果表明该方法能够保证良好的无监督聚类效果,并且不需要参数的调整和选择。To solve the problem that data vector clustering model relies too much on prior knowledge and supervised training,an unsupervised data clustering method based on kernel covariance matrix is proposed.The kernel learning and clustering problem based on matrix factorization were coupled into a joint formula so that the rank of the kernel covariance matrix was equal to the number of clusters,so as to ensure that the elements from different classes were weakly correlated and the data within the class were strongly correlated.According to the information content of the data,the sparse correlation matrix decomposition of the kernel covariance matrix was used to cluster the data,and l_(1)-l_(2)based sparsity measure was introduced to generate the sparse matrix factors,which effectively reduced the computational complexity.The global and highly nonconvex problem was solved by using the optimization formula based on convex function difference algorithm.The experimental results on hyperspectral images,human activities and document classification data sets show that the proposed method can guarantee good unsupervised clustering effect,and does not need parameter adjustment and selection.

关 键 词：聚类 无监督 协方差矩阵 稀疏性度量 

分 类 号：TP311.13[自动化与计算机技术—计算机软件与理论]