多标签AdaBoost算法的改进算法  

Improvement on AdaBoost for Multi-label Classification

作　　者：付忠良[1] 张丹普[1,2] 王莉莉[1,2] 

机构地区：[1]中国科学院成都计算机应用研究所,四川成都610041 [2]中国科学院大学,北京100049

出　　处：《四川大学学报（工程科学版）》2015年第5期103-109,共7页Journal of Sichuan University (Engineering Science Edition)

基　　金：四川省科技支撑计划基金资助项目(2011GZ0171;2012GZ0106)

摘　　要：针对多标签AdaBoost系列算法,以尽量减小算法的学习错误率为目的,提出了对其进行改进的2种思路。基于改进思路构造出了改进的多标签AdaBoost算法。一种思路是修改算法的样本分布调整策略,破坏现有AdaBoost算法中样本分布的均匀性,以确保增加每一个弱分类器都能降低学习错误的上界估计,从而实现对多标签AdaBoost算法的改进;另一种思路是训练弱分类器时兼顾后续待学习的弱分类器对学习错误的影响,克服现有算法在训练弱分类器时只考虑当前弱分类器对学习错误的影响,而完全忽略后续待学习的弱分类器对学习错误的影响这一现象,从而改进多标签AdaBoost算法。理论上,对于改进多标签AdaBoost算法,增加每一个弱分类器都能进一步降低学习错误。理论分析和实验结果均表明了提出的改进算法有改进效果。Aiming to decrease the learning error of the series of AdaBoost algorithm for multi-label classification, the AdaBoost algorithm was improved for multi-label classification by two strategies. One idea is to modify the adjustment Strategy of sample distribution, and destroy the sample uniform distribution in the existing AdaBoost algorithm, in order to ensure that the increase of every weak classifier can reduce the learning error bound estimation. Another idea is to consider the effect of subsequent weak classifiers to decrease the learning error w^n ~training current weak classifier, which is different from the existing AdaBoost algorithm. Theoretically, the improved AdaBoost algorithms for multi-label classification increase every weak classifier to reduce more learning error. Theoretical analysis and experimental results showed that all the improved algorithms are effective.

关 键 词：集成学习 多标签分类 REAL ADABOOST算法 多分类问题 分类器组合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]