面向目标的带先验概率的AdaBoost算法  

A Goal-oriented AdaBoost Algorithm with Prior Probabilities

作　　者：赵向辉[1,2] 姚宇[1,2] 付忠良[1,2] 苗青[1,2] 谢会云[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]中国科学院研究生院,北京100049

出　　处：《四川大学学报（工程科学版）》2010年第2期139-144,共6页Journal of Sichuan University (Engineering Science Edition)

基　　金：国家重点基础研究发展规划(973计划)项目(2004CB318003);四川省科技支撑计划资助项目(2008SZ0100;2009SZ0214);中国科学院西部之光人才培养项目资助

摘　　要：针对集成学习算法研究中多个分类器的最佳组合问题,改进了传统的AdaBoost集成学习算法。用于组合的各个分类器通常是基于样本集通过一定的训练得到,样本集中不同类目标的比率可以反映分类目标的先验概率。使用该参数给出了新的组合参数和投票表决阈值计算公式,巧妙的利用样本权值并将其加入到样本属性上进行训练学习,采用新的策略来选择基分类器,给出了面向目标的带先验概率的AdaBoost算法(GWPP AdaBoost算法)和分类器的最佳组合。依据UCI实验数据对传统的AdaBoost算法、Bagging算法、GWPP AdaBoost算法的错误率和性能进行了比较分析,验证了GWPP AdaBoost的有效性。Aiming at the best combination of multiple classifiers in ensemble learning algorithm,the traditional AdaBoost algorithm was improved.The combined classifiers were obtained by training the sample set.Using the sample to centralize the ratio of different kinds of targets can reflect the prior probability of various classifiers.Through utilizing this parameter,a new combination parameter and the computation formula of vote threshold were given.It uses the sample weight skillfully and adds it to sample attributes for training and learning.It adopts new strategies to select base classifiers and gives prior probabilities AdaBoost algorithm of goal-oriented （GWPP AdaBoost algorithm） and the best combination of multiple classifiers.The error rate and performance of ordinary AdaBoost algorithm,Bagging algorithm,and GWPP AdaBoost algorithm were comparatively analyzed based on the UCI datasets.And the results showed the validity of the GWPP AdaBoost algorithms.

关 键 词：集成学习 ADABOOST算法 分类器组合 先验概率 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]