FCM型聚类算法的统一框架及其核推广

摘要

FCM clustering model is an essential tool for data pattern recognition. Based on FCM, combined with separation measure and spatial constraint measure, a unified framework of the FCM-type algorithm UFCM is proposed. With the help of kernel trick, kernel UFCM is also proposed, which can cope with nonlinear clustering problem efficiently. Furthermore, minimum path is used as the distance measure of data points to improve the clustering acuracy . The experients result show that, compared with FCM and its variants, KUFCM can improve not only the clustering acuracy but also the robust of the algorithm due to noisy data in the data sets.%模糊C-均值(FCM)聚类模型是数据模式识别的一类重要工具.在FCM的基础上,结合了数据的类间离散度信息和空间约束信息,提出了FCM型算法的统一框架--Unified FCM,简称UFCM.针对UFCM算法难以处理非线性分类的问题,运用核技巧,得到核空间的UFCM算法--KUFCM.提出使用最短路作为数据点间的距离度量,提高了算法的聚类精度.实验表明,相对于FCM及其改进算法,KUFCM不仅提高了聚类算法的分类精度,而且改善了FCM型算法对噪声数据的稳健性.