基于谱聚类的聚类集成算法

谱聚类是近年来出现的一类性能优越的聚类算法,能对任意形状的数据进行聚类, 但算法对尺度参数比较敏感,利用聚类集成良好的鲁棒性和泛化能力,本文提出了基于谱聚类的聚类集成算法.该算法首先利用谱聚类算法的内在特性构造多样性的聚类成员; 然后,采用连接三元组算法计算相似度矩阵,扩充了数据点之间的相似性信息;最后,对相似度矩阵使用谱聚类算法得到最终的集成结果. 为了使算法能扩展到大规模应用,利用Nystrm采样算法只计算随机采样数据点之间以及随机采样数据点与剩余数据点之间的相似度矩阵,从而有效降低了算法的计算复杂度. 本文算法既利用了谱聚类算法的优越性能,同时又避免了精确选择尺度参数的问题.实验结果表明:较之其他常见的聚类集成算法,本文算法更优越、更有效,能较好地解决数据聚类、图像分割等问题.

Cluster Ensemble Based on Spectral Clustering

Spectral clustering has become increasingly popular in recent years. It can deal with arbitrary distribution dataset, however, it is sensitive to the scaling parameter. Cluster ensemble based on spectral clustering is proposed which utilizes the good robustness and generalization ability of cluster ensemble. Multiform clustering components are generated by exploiting the property of spectral clustering, and the connected triple algorithm which can expand the similarity information among data is used to compute the affinity matrix, then the affinity matrix is used by spectral clustering algorithm to produce ensemble results. In order to make the algorithm extensible to large scale applications, only the similarity among the rand sampling data and the similarity between the random sampling data and the rest data are computed by adopting the Nyström sampling method. The proposed algorithm makes full use of the excellent performance of spectral clustering as well as avoids the selection of the accurate parameter in spectral clustering. Experiments show that compared with other common cluster ensemble techniques, the proposed algorithm is more excellent and efficient, and that it can provide a good way to solve data clustering and image segmentation problem.