Abstract: Abstract:SAMME is a multiclass AdaBoost which always fall into retrogression, therefore SAMME.R based on base classifiers selection was proposed by researchers which can resolve the retrogression effectively. The influence of weighed error rate and pseudo loss on the performance of the algorithm was studied in this paper. In order to improve the classification accuracy, a dynamic weighted SAMME.RD algorithm was proposed which based on base classifiers’ performance in effective neighborhood area. The experiment results in UCI dataset show that using real error rate will improve the performance of SAMME.R, using weighed probability will get better performance when the dataset has less class and data distribution equilibrium, using real probability will get better performance in the opposite situation and SAMME.RD algorithm can improve classification accurate effective.

Key words: Keywords: ensemble learning, multi-classify, adaptive boosting(AdaBoost), SAMME, dynamic weighted fusion

摘要： 摘 要: SAMME算法是一种多分类的AdaBoost算法，由于其容易陷入退化，有学者提出基于基分类器筛选的SAMME.R改进算法，很好地解决了退化问题。本文针对使用加权概率和伪损失对算法性能的影响做了研究，并在此基础上提出一种基于基分类器对样本有效邻域分类的动态加权AdaBoost算法，进一步提高了算法的分类正确率。使用UCI数据集进行验证，实验结果表明：使用真实的错误率计算基分类器加权系数效果更好；在数据类别较少且分布平衡时，使用真实概率进行基分类器筛选效果较好；在数据类别较多且分布不平衡时，使用加权概率进行基分类器筛选效果较好；提出的动态加权算法可以有效改善AdaBoost算法的分类正确率。

关键词: 关键词: 集成学习, 多类分类, AdaBoost算法, SAMME算法, 动态加权融合