Abstract: Concerning the problem of wrong partition at fuzzy boundary in Fuzzy C-Means (FCM) clustering algorithm, an improved recalculation technique for fuzzy points was proposed. The new method took into account the data distribution characteristics in different classes. Firstly, it made the hyperspheres central regions by clear data, then defined a new similarity distance based on the clear radius of central region to recalculate the membership of fuzzy point, and finally reassigned the fuzzy points to right category. The experimental results show that the new algorithm can correct some wrong partition and improve the definition of fuzzy point, and also it is a promising algorithm for dataset with significant density differences.

Key words: fuzzy clustering, fuzzy point, similarity distance, central region, second clustering

摘要： 针对模糊C-均值(FCM)聚类算法在模糊边界上容易出现划分错误的问题，提出一种对模糊点进行二次处理的改进算法。该算法以各类中的数据分布密度为依据，首先利用清晰点构成超球体中心区域，然后基于中心区域的清晰半径定义一种新的相似性距离，并利用该距离对模糊点的隶属度进行二次计算，重新确定其类别归属。实验结果显示，改进算法能有效纠正分类错误，提高模糊点的清晰度，在密度差异较大的数据集上具有一定的应用潜力。

关键词: 模糊聚类, 模糊点, 相似性距离, 中心区域, 二次聚类