一类基于高斯核的动态聚类算法研究

Gaussian kernel-based dynamic clustering algorithm

通过研究核映射机理,提出了用于聚类分析的高斯核聚类算法.采用Mercer核映射将输入空间的样本映射到高维特征空间,在保持样本原有特征的基础上通过核映射使样本的差异性得到一定程度的放大.这样在特征空间中就可以采用传统的k-均值聚类方法,从而弥补k-均值聚类对于各样本的边界是线性不可分以及类分布为非高斯分布或为非椭圆分布时,其聚类效果较差的缺点.这种核聚类方法由于是对样本进行了预处理,增大样本差异的前提下对样本进行聚类,从而提高了聚类精度,获得较好的聚类效果.应用这一算法对几种典型分布的数据进行聚类实验,仿真结果表明高斯核聚类算法性能优于k-均值算法和模糊k-均值算法.

An algorithm of Gaussian kernel clustering is proposed by analyzing kernel mapping theory. Samples in the original space were mapped into a high-dimensional feature space by Mercer kernel mapping , which deference among these samples in sample space were strengthened to some extent at the base of original feature. So k-means clustering could be performed in the feature space, which defect of bad clustering result would be made up when it meet with any samples such as the distribution of which boundary of the sample is non-linear sub, or the distribution of which was non-Gaussian dis- tribution, or the distribution of which was non-elliptical distribution. This algorithm acquired better clustering precision and effect for which original samples were pretreated and the deference among these samples were strengthened. The results of experiment demonstrat that this Gaussian kernel clustering a tering algor gorithm has the advantage over the k-means clustering algorithm and fuzzy k-means clus thm in performance by applying this algorithm to the distribution of some typical experi mental data for clustering.