Analysis of parameter selections for fuzzy c-means
Analysis of parameter selections for fuzzy c-means
复制标题
DOI：
10.1016/j.patcog.2011.07.012
复制
发表时间：
2012
期刊：
影响因子：
--
通讯作者：
Kuo-Lung Wu
中科院分区：
文献类型：
--
作者：
Kuo-Lung Wu
加权指数m称为模糊器，它可以影响模糊c-均值(FCM)的性能。一般认为m∈[1.5，2.5]。在对FCM进行稳健性分析的基础上，提出了一种新的参数m的选择准则。我们将证明，较大的m值将使FCM对噪声和异常值更具鲁棒性。但是，大于理论上限的相当大的m值将使该样本成为唯一的优化器。在模糊聚类中避免这种意外情况的一种简单有效的方法是为每个聚类分配一个聚类核。我们还将讨论一些将FCM扩展为包含模糊聚类中的簇核的聚类算法。对于理论上界较大的情况，我们建议用一个合适的大m值来实现FCM。否则，我们建议使用集群核心来实现集群方法。当数据集包含噪声和离群值时，对于FCM和基于聚类核的方法，在较大的理论上限情况下，推荐使用模糊器m=4。
The weighting exponent m is called the fuzzifier that can influence the performance of fuzzy c-means (FCM). It is generally suggested that m∈[1.5,2.5]. On the basis of a robust analysis of FCM, a new guideline for selecting the parameter m is proposed. We will show that a large m value will make FCM more robust to noise and outliers. However, considerably large m values that are greater than the theoretical upper bound will make the sample mean a unique optimizer. A simple and efficient method to avoid this unexpected case in fuzzy clustering is to assign a cluster core to each cluster. We will also discuss some clustering algorithms that extend FCM to contain the cluster cores in fuzzy clusters. For a large theoretical upper bound case, we suggest the implementation of the FCM with a suitable large m value. Otherwise, we suggest implementing the clustering methods with cluster cores. When the data set contains noise and outliers, the fuzzifier m=4 is recommended for both FCM and cluster-core-based methods in a large theoretical upper bound case.