不同类型数据中特征与类别以及特征与特征之间存在一定的线性和非线性相关性.针对基于不同度量的特征选择方法在不同类型数据集上选取的特征存在明显差别的问题,本文选择线性相关系数、对称不确定性和互信息三种常用的线性或非线性度量,将它们应用于基于相关性的快速特征选择方法中,对它们在基因微阵列和图像数据上的特征选择效果进行实验验证和比较.实验结果表明,基于相关性的快速特征选择方法使用线性相关系数在基因数据集上选取的特征集往往具有较好分类准确率,使用互信息在图像数据集上选取的特征集的分类效果较好,使用对称不确定性在两种类型数据上选取特征的分类效果较为稳定.%It has been known that either linear correlation or nonlinear correlation might exist between feature-to-feature and feature-to-class in datasets. In this paper, we study the differences of selected feature subset when different kinds of measures are applied with same feature selection method in different kinds of datasets. Three representative linear or nonlinear measures, linear correlation coefficient, symmetrical uncertainty, and mutual information are selected. By combining them with the fast correlation-based filter (FCBF) feature selection method, we make the comparison of selected feature subset from 8 gene microarray and image datasets. Experimental results indicate that the feature subsets selected by linear correlation coefficient based FCBF obtain better classification accuracy in gene microarray datasets than in image datasets, while mutual information and symmetrical uncertainty based FCBF tend to obtain better results in image datasets. Moreover, symmetrical uncertainty based FCBF is more robust in all datasets.
展开▼