基于全信息相关度的动态多分类器融合

Dynamic Multiple Classifiers Combination Based on Full Information Correlation

AdaBoost采用级联方法生成各基分类器,较好地体现了分类器之间的差异性和互补性。其存在的问题是,在迭代的后期,训练分类器越来越集中在某一小区域的样本上,生成的基分类器体现不同区域的分类特征。根据基分类器的全局分类性能得到固定的投票权重,不能体现基分类器在不同区域上的局部性能差别。因此,本文基于Ada-Boost融合方法,利用待测样本与各分类器的全信息相关度描述基分类器的局部分类性能,提出基于全信息相关度的动态多分类器融合方法,根据各分类器对待测样本的局部分类性能动态确定分类器组合和权重。仿真实验结果表明,该算法提高了融合分类性能。

The base classifiers trained by AdaBoost combination learning algorithm are produced orderly, diversity and complementarity of base classifiers are assured. But along with the iterative process of AdaBoost, the classifier which represents different area classification performance mainly focuses on a certain small area of input space. The constant weights are obtained according to overall classification performance which can not demonstrate base classifiers＇ classifi-cation performance in different local areas. Based on AdaBoost, a dynamic multiple classifiers combination algorithm based on full information correlation （FIC） which describes base classifiers＇ local classification performance is proposed, the classifiers＇ selection and their weights are determined according to test samples＇ FIC to base classifiers. The simulated experiments show that the combination classification performance is improved greatly.