基于主成份分析的Bagging集成学习方法

PCA Based Bagging Ensemble Methods

机器学习中数据集的冗余特征会影响学习器的泛化能力,一些流行方法如支持向量机和集成学习也难免于此.研究了利用主成份分析进行特征变换对Bagging集成学习算法的影响,提出一种称为PCA-Bagging的算法,并与其它算法比如单个支持向量机、支持向量机Bagging集成、带有特征变换的单个支持向量机等进行了性能比较.在多个UCI标准数据集上的实验表明PCA-Bagging算法具有更好的性能,这说明即使是泛化能力很强的集成学习方法其学习的数据也需要进行适当的特征变换.

Redundant features in data sets hurt the generalization performance of learning machines. Even the state-of-arts algorithms such as support vector machine （SVM） and ensemble learning are not immune to it. This paper studies the effect of feature transformation with principal component analysis （PCA） on bagging method of ensemble learning. This paper proposes a new method termed PCA-bagging, which is compared with other methods such as single SVM, bagging of SVM and SVM with PCA. Experiments on UCI machine learning benchmark data sets show that PCA-bagging has better generalization performance, indicating that even ensemble learning methods with excellent generalization ability also need proper feature preprocessing.