基于投票信息熵的AdaBoost改进算法  

Improved AdaBoost algorithm based on vote entropy

作　　者：唐焕玲[1] 鲁明羽[1] 邬俊[1] 

机构地区：[1]大连海事大学信息科学技术学院,辽宁大连116026

出　　处：《控制与决策》2010年第4期487-492,共6页Control and Decision

基　　金：国家自然科学基金项目(60773084;J0724003;60603023);高等学校博士学科点专项科研基金项目(20070151009)

摘　　要：针对AdaBoost算法不能有效提升NB(Naive Bayesian)的分类性能,提出一种改进的样本权重维护策略.权重的调整不仅依据样本是否分错,还需考虑前几轮的多个基分类器对它的投票分歧.基分类器的信任度不但与错误率有关,还与基分类器间的差异性有关.这样可以提高基分类器的正确性,增加基分类器的差异性.实验结果表明,改进的BoostVE-NB算法能有效地提升NB文本分类性能.For the problem that AdaBoost works poorly with Naive Bayesian（NB） categorization,an improved re-weighting rule for training examples is proposed.It is not only considering the classification result of the current iteration,but also considering how much those previous classifiers disagree on their decision-making for each training sample.Moreover,the confidence of every base classifier is determined not only by the error rate,but also by the diversity among base classifiers. Thus,the accuracy and instability of the base NB classifiers are increased.Experimental results show that BoostVE-NB is effectively to improve the performance of NB text categorization.

关 键 词：ADABOOST 朴素贝叶斯 文本分类 样本权重 投票信息熵 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]