다양성 기반 AdaBoost 분류기

초록

The generalization performance of an ensemble classifier depends on diversity between member classifiers as well as performance of each member. However, the existing methods are focused on error minimization of each step through repetition process and they can not suggest a method to add diversity. AdaBoost algorithm is the representative regression ensemble method to make a strong classifier using weak classifiers. In this paper, we suggest the new algorithm, Diverse AdaBoost algorithm, that it can improve performance of the final mixed classifier with adding diversity into the AdaBoost algorithm. In the existing algorithm, it uses a method that it selects a weak classifier to minimize error in repetition training process and includes it as a member of a mixed classifier, while it generates a mixed classifier with totally high performance through adding diversity as sacrificing accuracy of a weak classifier in suggested algorithm. The proposed algorithm suggests a specific method to consider diversity in a mixed classifier design.