Parzen窗确定系数的协同模糊C均值算法

Novel collaboration fuzzy C-means algorithm with Parzen window determined collaboration coefficient

协同模糊C均值(collaboration fuzzy C-means,CFC)算法的协同系数通常根据经验人工设定,且在协同过程中保持不变,不能充分利用数据子集之间的协同关系,算法精度有限。提出Parzen窗确定系数的协同模糊C均值(βp-CFC)算法。用模糊C均值(fuzzy C-means,FCM)算法求出各数据子集的隶属度和聚类中心,再用Parzen窗求出各子集在聚类中心处的密度,根据子集间密度的相关性设定变化的协同系数,利用变化的协同系数进行协同聚类。以Matlab为平台,对βp-CFC算法进行了实验,算法聚类准确率可达到80.34%,比模糊C均值算法、固定系数的CFC算法的准确率分别高出11.80%和3.94%。实验证明,βp-CFC算法较为合理,聚类性能较好。

Collaboration fuzzy C-means algorithm （CFC） can improve the performance of fuzzy C-means algorithm by using the collaborative relationship between the sub data sets. But the collaboration coefficient of CFC, in an inadequate using of the collaborative relationship, is always determined by priori knowledge and remains constant during collaboration stages. In order to circumvent this limitation, a novel collaboration fuzzy C-means algorithm with Parzen window determined collaboration coefficient（tip -CFC） was developed. First, fuzzy partition matrix and cluster prototypes of every sub data sets are computed by fuzzy C-means algorithm （FCM）, for the further computing of collaboration coefficient. Second, density of the cluster prototypes is gained by Parzen window method. Third, collaborative coefficient is dynamically adjusted by the correlation of density. Last, objects are clustered with dynamical collaborative coefficient. The algorithm is tested on the matlab platform, achieving a high accuracy of 80.34%, higher than FCM and CFC with 11.80% and 3.94%, respectively. Examples are provided to demonstrate the rationality of collaboration coefficient and the better performance of CFC.