Boosting算法中基分类器权重的动态赋值  

Dynamic Assignment of Base Classifier's Weight in Boosting Method

作　　者：付彬[1] 王志海[1] 王中锋[1] 

机构地区：[1]北京交通大学计算机与信息技术学院,北京100044

出　　处：《广西师范大学学报（自然科学版）》2009年第3期85-88,共4页Journal of Guangxi Normal University:Natural Science Edition

基　　金：国家自然科学基金资助项目(60673089)

摘　　要：Boosting是一种有效的分类器组合方法,其通过加权投票来组合多个基分类器进行分类。在对基分类器进行权重赋值时,该算法采用了以基分类器在当前训练集上的错误率的某种变形来对基分类器进行权重赋值,这是一种静态的赋值方法。介绍一种动态地对基分类器进行赋权重的方法,这种方法利用当前测试实例属于某个被错误分类数据子集的程度,并按照程度的大小给相应的基分类器赋适当的权重。跟静态赋权重相比,这种方法考虑了测试实例属性取值的不同,进而能动态地调整基分类器的权重,从而达到进一步优化分类性能的目的。实验表明,动态权重赋值的方法在大多数情况下跟静态赋值相比具有更好的分类性能。Boosting is an effective classifier combination method by using weighted vote to combine multiple base classifiers. When assigning weight to each classifier,Boosting adopts a form of error rate that the classifier makes on current training set,which this is a typical static method. This paper introduces a kind of dynamic assignment method to decide the base classifier＇s weight. This method uses the current test data to calculate the degrees belong to each data subset that is misclassified and uses those degrees to assign weight to each base classifier. Compared with static method ,this one takes the current test data in consideration and assigns weight accordingly, so it can improve the performance. The experiment shows that it has a better performance in most situations then the stats one.

关 键 词：组合分类器 BOOSTING方法 动态权重赋值 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]