本文提出了一种度量特征区分度的定义,进而提出一种基于聚类的特征选择方法CBFS.该方法时间复杂度与数据集的大小和特征个数成近似线性关系,适合于大规模数据集中的特征选择;该方法对数据类型没有限制,适用于混合类型数据.在UCI数据集上的实验结果表明,与文献中的方法相比,本文方法具有较好的性能,说明提出的特征选择方法是有效和实用的.
The authors come up with a definition of measuring differentiations between features,and then put forward a method of clustering-based feature selection(Below referred to as CBFS).The time complexity of the method is nearly linear with both the size of dataset and the number of features.Besides,the method is applicable to the selection of features in large dataset.It can particularly handle data with both Nominal and Continuous Features.The results of the experiment on UCI datasets show that the method is effective and practicable.