摘要
首先将聚类集成问题归结为直观的最佳子空问的求解问题;随后根据线性代数理论将该问题描述为带约束条件的优化问题,通过放松离散约束条件进一步约简为矩阵低秩近似问题;最后通过求解超图的加权邻接矩阵的奇异值分解问题获得最佳子空间的一组标准正交基.据此,设计了一个基于矩阵低秩近似的算法,该算法根据每个对象在低维空间下的坐标使用K均值算法进行聚类,从而得到最终的结果.在多组基准数据集上的实验结果表明:较之于传统的聚类集成算法,本文的算法获得了更好的聚类结果,且效率较高.
As an important extension to conventional clustering algorithms,cluster ensemble techniques became a hotspot in machine learning area.In this paper,cluster ensemble problem was first viewed as a direct problem of seeking the best subspace. And then,we formally described the problem as an optimization problem with constraint according to linear algebra,and further transformed into a matrix low rank approximation problem by relaxing the discrete constraint.Lastly,a set of orthonormal basis of the best subspace was attained by solving the singular value decomposition problem of the hypergraph＇s weighted adjacent matrix. Hereby,a matrix low rank approximation-based algorithm was proposed,which called K-means algorithm to cluster objects according to their coordinates in the low dimensional space and obtained the final clustering result.Experiments on baseline datasets demonstrate the effectiveness of the proposed algorithm,and it outperforms other baseline algorithms.
出处
《电子学报》
EI
CAS
CSCD
北大核心
2013年第6期1219-1224,共6页
Acta Electronica Sinica
基金
国家自然科学基金(No.60970542
No.41006057
No.6110507)
国家863重点项目(No.2008A09701)
国际科技合作聘专重点项目
江苏省高校"青蓝工程"资助项目
盐城工学院人才引进专项基金(No.XKR2011019)
关键词
无监督学习
聚类分析
聚类集成
矩阵低秩近似
unsupervised learning
clustering analysis
cluster ensemble
matrix low rank approximation
