一种适合于非线性高维数据的谱聚类算法

A SPECTRAL CLUSTERING ALGORITHM FOR NONLINEAR HIGH DIMENSIONAL DATA

谱聚类能识别非线性数据,且优于传统聚类。谱聚类中度量相似性的高斯核函数尺度参数σ和聚类个数k对聚类效果影响较大,但需要人工判断。用向量之间夹角余弦代替σ并且通过特征值的跳跃性确定聚类个数,对于非线性高维数据,提出一种自适应谱聚类算法,将数据通过显式构造映射到随机特征空间,在随机特征空间中实现聚类。实验结果表明,在UCI数据上该算法与传统算法相比效果更好。

Spectral clustering can identify nonlinear data,and it is better than traditional clustering.The Gaussian kernel function scale parametersσand the number of clusters k that measure the similarity in spectral clustering have a great influence on the clustering effect,and face the problem of human determination.This paper used the angle cosine between the vectors to replace theσand determined the number of clusters k by jumping the eigenvalues.In order to solve the nonlinear high-dimensional data clustering,an adaptive spectral clustering algorithm was proposed to map the data to the stochastic feature space through explicit construction,and the clustering was realized in the stochastic feature space.The experimental results show that,compared with the traditional algorithm on UCI data,this algorithm has better effect.