融合DECORATE的异构分类器集成算法

Algorithm of heterogeneous classifiers ensembles based on DECORATE

在基于Stacking框架下异构分类器集成方式分析的基础上,引入同构分类器集成中改变训练样本以增强成员分类器间差异性的思想,提出融合DECORATE的异构分类器集成算法SDE;在1-层泛化利用DECORATE算法,向1-层训练集增加一定比例的人工数据,使得生成的多个1-层成员分类器间具有差异性。实验表明,该方法在分类精度上要优于传统Stacking方法。

Based on the Stacking framework to construct heterogeneous ensembles,this paper introduced manipulating training samples in the context of creating homogeneous ensembles as the mechanism to encourage diversity.It proposed a new algorithm SDE,which used DECORATE to generate level-1 ensembles by adding proportion of artificial data to level-1 training set so as to inject diversity for member classifiers in level-1.Experiment results indicate that the proposed method achieves better performance than classic Stacking.