摘要
针对传统的AdaBoost算法只关注分类错误率最小的问题,在分析传统的AdaBoost算法实质基础上,提出一种基于代价敏感的改进AdaBoost算法。首先在训练基分类器阶段,对于数据集上的不同类别样本根据其错分后造成的损失大小不同来更新样本权值,使算法由关注分类错误率最小转而关注分类代价最小。然后,在组合分类器输出时采用预测概率加权方法来取代传统AdaBoost算法采用的预测类别加权的方法。最后通过实验验证了改进算法的有效性。
In view of the problem that traditional AdaBoost algorithm only concerns the issue of classification error rate minimum, based on analysing the essence of traditional AdaBoost algorithm, an improved AdaBoost algorithm based on cost-sensitive is proposed in this paper. First, in the phase of training the base classifier, for samples of different categories in date set, the value of the sample is updated according to the loss degree of the sample caused by being erroneously classified to other category, this makes the algorithm turn to concern the classification cost minimum but not the classification error rate minimum. Secondly, when the composite classifiers are outputing, the predicted probability weighting method is adopted instead of the predicted category weighting method used by traditional AdaBoost algorithm. Final experiment proves the effectiveness of the improved algorithm.
出处
《计算机应用与软件》
CSCD
北大核心
2013年第10期123-125,138,共4页
Computer Applications and Software
基金
山东省自然科学基金项目(ZR2009GL001)
