基于特征选择的集成分类器抗噪性能分析

特征选择有助于增强集成分类器成员间的随机差异性，从而提高泛化精度。研究了随机子空间法（RandomSub-space）和旋转森林法（RotationForest）两种基于特征选择的集成分类器构造算法，分析讨论了两算法特征选择的方式与随机差异程度之间的关系。通过对UCI数据集引入噪声，比较两者在噪声环境下的分类精度。实验结果表明：当噪声增加及特征关联度下降时，基本学习算法及噪声程度对集成效果均有影响，当噪声增强到一定程度后。集成效果和单分类器的性能趋于一致。

Anti-noise Performance Analysis of Classifiers Ensembles Based on Feature Selection

Feature selection encourages random differentiation of the members of the ensembles to improve generation accuracy. In this pa per, random subspace and rotation forest, two algorithms based on feature selection for constructing classifiers ensembles were researched and their relationship between ways of selecting features and its affection on diversity was discussed. By introducing noise into UCI data sets,compared anti-noise performance with different noisy level of two algorithms. Experimental results indicate that both base learning algorithms and noisy level affect the accuracy of an ensemble while noise increases and feature correlation decreases. In situation with higher classification noise, both ensembles and single classifier exhibit quite similar performance.