核自组织映射竞争聚类

KERNEL SELF-ORGANIZING MAP COMPETITIVE CLUSTERING

基于核方法可在高维特征空间中完成数据聚类,但缺乏对原输入空间聚类中心及结果的直观刻画。提出一种核自组织映射竞争聚类算法。该算法是利用核的特征,导出SOM算法的获胜神经元及权重更新规则,而竞争学习机制依然保持在原输入空间中,这样既解决了当输入样本分布结构呈高度非线性时,其分类能力下降的问题,而且解决了Donald[1]算法导致的特征空间中的获胜神经元在原始输入空间中的原像不存在,而无法对聚类结果利用可视化技术进行解释的问题。实验结果表明,提出的核自组织映射竞争聚类算法在某些数据集中可以获得比SOM算法更好的结果。

In high dimension feature space the data clustering can be completed based on Kernel trick, but the disadvantage of this lies in lack of direct descriptions on clustering＇ s center in primary input space and the result. In this paper, a novel kernel SOM competitive cluste- ring algorithm is proposed. The algorithm uses the property of kernel to educe the winner neuron and weights updating rule of Self-Organizing Map algorithm while still keep competitive learning mechanism in primary input space. In this way the problem of classifying ability decline when the inputted sample has highly non-linear distributed structure can be resolved, moreover, the problem caused by Donald＇ s algorithm is resolved as well in which the winner neuron in feature space does not have its original image existed in primary input space so that the cluste- ring results can not be elucidated with visualised technology Experimental result demonstrated that the kernel SOM competitive clustering algo- rithm presented in the article can perform better in certain data sets than the SOM algorithm.