预测子空间聚类的聚类集成算法  

Cluster Ensemble Algorithm Based on Predictive Subspace Clustering

作　　者：吕红伟[1] 王士同[1] 

机构地区：[1]江南大学数字媒体学院,江苏无锡214122

出　　处：《小型微型计算机系统》2017年第4期845-851,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61272210)资助

摘　　要：鲁棒的预测子空间聚类(robust predictive subspace clustering,RPSC)算法通过在预测子空间聚类中引入RPCA(robust principal component analysis)模型来提高主元分析的鲁棒性,但是算法在执行变量选择和模型选择时,对稀疏参数以及最大子空间维数比较敏感,利用聚类集成良好的鲁棒性和泛化能力,提出预测子空间聚类的聚类集成算法.该算法首先利用RPSC算法的内在特性以及特征重采样技术一起来构造多样性的聚类成员;然后,采用加权连接三元组算法计算簇关联矩阵,发掘出隐藏在簇之间的关联信息;最后,通过谱图分割技术得到最终的集成结果.其优势在于既利用了RPSC算法的优越性能,同时又避免了精确选择参数的问题.实验结果表明,无论在无噪声或加噪声的环境下,新提出的算法都能够提高RPSC算法的聚类质量.RPSC （ robust predictive subspace clustering ） algorithm improved the robustness of principal component analysis by intro- ducing the RPCA model, however, it is sensitive to the sparse parameter and maximum subspace dimension when it carry out variable selection and model selection. Cluster ensemble based on predictive subspace clustering is proposed which utilizes the good robustness and generalization ability of cluster ensemble. First, multiform clustering components are generated by the intrinsic characteristics of RPSC algorithna and the feature resampling technique. Then, the weighted connected-triple algorithm, which can find out the related in- formation hidden in between clusters, is used to compute the refined cluster-association matrix. Finally, a spectral graph partitioning technique is exploited to obtain the final ensemble result. The proposed algorithm makes full use of the excellent performance of RPSC algorithm as well as avoids the problem of accurate selection of parameters in RPSC algorithm. Experiments show that the proposed al- gorithm improved clustering performance both in the absence of noise and noise environment.

关 键 词：预测子空间聚类 变量选择 模型选择 聚类集成 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]