结合提升自举FLD投影的特征提取算法  

Feature Extraction Algorithm Based on Boosting Bootstrap FLD Projections

作　　者：沈道义[1] 俞能海[1] 庞彦伟[2] 王雷[1] 

机构地区：[1]中国科学技术大学电子工程与信息科学系多媒体计算与通信教育部-微软重点实验室,安徽合肥230027 [2]天津大学电子信息工程学院,天津300072

出　　处：《小型微型计算机系统》2009年第4期690-694,共5页Journal of Chinese Computer Systems

基　　金：模式识别国家重点实验室开放课题项目(2005-2)资助;国家自然基金项目(60605005,60672056)资助

摘　　要：针对两类分类问题,提出一种基于自适应提升(Adaptive Boosting,Adaboost)算法的分类特征提取算法.首先对训练样本集进行自举采样从而生成一定数量的训练样本自举子集,然后通过对每个自举子集的Fisher线性判别分析(FLD),得到相应数量的一维自举投影向量.在Adaboost算法迭代的每一步中,具有最小的加权最近邻分类误差的一维自举投影向量被作为分类特征选出,用来构成最终的强分类器.该算法可以有效地克服传统特征提取算法性能依赖于数据分布以及提取准则不直接与训练误差相关的弱点.实验结果表明,相比传统算法,该算法的分类准确度有较明显的提升,并且特别适用于数据分布复杂的情况.In this paper we propose a feature extraction algorithm for two-class problem based on Adaboost algorithm. In this algorithm, we first sample a large number of bootstrap training subsets from the original training set and use FLD in each subset to construct a large number of corresponding one-dimensional （l-d） bootstrap projections. Then at each step of Adaboost algorithm, the bootstrap projection with the minimum weighted 1-Nearest Neighbor （1NN） classification error was selected to combine the final strong classifier. In this manner, our method is able to overcome the limitations of the traditional learning algorithms, such as the classification performance being depending on the data distribution, and the feature extraction rules being not directly related to the training error. Experiment results show that our method is superior to other traditional techniques, especially for the data of complex structure.

关 键 词：集成学习 提升 自举 ADABOOST FISHER线性判别分析 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]