适用于复杂结构的多路谱聚类算法的改进

为使多路谱聚类方法对复杂结构数据集有效地聚类,根据矩阵扰动理论,利用局部近邻关系更新谱聚类算法(NJW)中的初始相似度矩阵,得到最终的亲和矩阵.理论分析表明,数据集可划分时,该矩阵是理想块矩阵或接近理想块矩阵,保证了本文算法聚类划分的正确性.将本文算法和基于路径的谱聚类、密度敏感的谱聚类以及基于流平面排序的谱聚类进行了比较,结果表明,本文算法在数据集具有复杂分布结构时可以确定聚类个数,得到正确的聚类结果.进一步将本文算法用于真实数据集上的聚类分析,表明本文算法是有效的.

Improvement on Muti-way Spectral Clustering Algorithm for Complex Distributed Data

To cluster the complex structure dataset effectively using multi-way spectral clustering,based on the matrix perturbation theory,the initial similarity matrix in Ng-Jordan-Weiss(NJW) algorithm was updated by using local neighbor relation and then the last affinity matrix was gained.Theoretical analysis showed that this last affinity matrix was ideal block matrix or near ideal block matrix so that it could make the clustering correct.The method was compared with path-based spectral clustering,density-sensitive spectral clustering and spectral clustering through ranking on manifolds together.Result illustrates that the affinity matrix can decide the clustering number so as to get the correct clustering result.Further,the real dataset is used to check our method,and the result shows that the method is effective.