摘要
针对AdaBoost集成时难分样本权重扩张导致训练样本在更新时分布失衡的问题,提出一种基于正负类样本损失函数(LF)的权重更新策略。权重的调整不仅与训练误差有关,还考虑到基分类器对不同类别样本的正确分类能力,从而避免训练样本过度集中于某一类的异常现象。实验结果表明,基于LF的AdaBoost能在提高收敛性能的情况下,提高算法精度,克服样本分布失衡问题。偏差方差分析的结果显示,该算法在改善偏差的情况下,能有效地减小错误率中的方差成分,提高集成的泛化能力。
As to the issue that the weight expansion for hardest samples can cause imbalance when updating the training sample in AdaBoost algorithm, an improved approach based on the Loss Function （LF） of the different patterns, namely, LF- AdaBoost, was proposed. The weight tuning was affected not only by the training error, but the performance of base classifiers for different classes, thus avoiding the excessive concentration phenomenon. The results based on UCI data sets and different base classifiers have shown that the approach can improve the speed of convergence and overcome the imbalance, as well as promote the generalization ability of ensemble classifier.
出处
《计算机应用》
CSCD
北大核心
2012年第10期2916-2919,共4页
journal of Computer Applications
基金
国家自然科学基金资助项目(60975026)
