【摘要】 为了在多维聚类分析中运用有效距离度量方法表征数据对象的邻近度,提出一种协方差测距（covariance distance measure analysis,CDM）算法,首先,采用模糊C均值(fuzzy c-means,FCM)方法对数据对象赋予权值,得到每个样本点相对类别特征的隶属度,再依据隶属度计算每个样本的差异度；其次,为了使类别分离最大化,用样本点同关联类别的协方差距离度量代替模糊聚类中欧式距离度量作为优化问题的第一个标准,使相似数据对象更为接近；最后,用样本点间的协方差距离度量作为第二个优化标准,使相异数据相互隔开,交替固定变量迭代计算最优解,使聚类指标和距离度量学习参数同时得到优化,获得更好的聚类结果。在不同数据集上的实验结果表明,与FCM-Sig和UNCA算法相比,CDM算法在聚类准确性和算法收敛性方面均有更好表现。更多还原

【Abstract】 In order to use effective distance measurement methods to characterize the proximity of data objects in multi-dimensional clustering analysis, a covariance distance measurement(CDM) algorithm is proposed. First,fuzzy C-means(FCM) is used to assign weights to the data objects, so that the membership degree of each sample point relative to the category feature is obtained. Based on the membership degree, the difference degree of each sample is calculated. Then, as the first optimization criterion, the variance distance measure is used to replace the Euclidean distance measure in fuzzy clustering to make similar data objects closer. Finally, the covariance distance measure between the sample points is used as the second optimization criterion to make the different data objects separate from each other. The optimal solution is calculated iteratively with alternate fixed variables, so that the clustering index and distance measurement learning parameters are optimized at the same time, and better clustering results are obtained. Experimental results on different data sets show that compared with FCM-Sig and UNCA algorithms, CDM algorithm has better performance in clustering accuracy and algorithm convergence.更多还原