三种分类算法偏差-方差结构的比较：MCLP,LDA和C5.0  

Comparison of bias-variance structure of three classification algorithms:MCLP,LDA and C5.0

作　　者：朱梅红[1,2,3] 石勇[1,2] 李爱华[4] 张东玲[1,2] 

机构地区：[1]中国科学院研究生院,北京100080 [2]中国科学院虚拟经济与数据科学研究中心,北京100080 [3]首都经济贸易大学统计学院,北京100070 [4]中央财经大学管理科学与工程学院,北京100081

出　　处：《中国科学院研究生院学报》2009年第4期443-450,共8页Journal of the Graduate School of the Chinese Academy of Sciences

基　　金：国家自然科学基金(70621001,70531040,70501030,10601064,70472074,90718042);北京市自然科学基金(9073020);973项目(2004CB720103)资助

摘　　要：基于Domingos的期望预测误差分解框架,在3个数据集上,对MCLP、LDA和C5.0这3种算法的偏差-方差结构特点进行了比较分析.实验结果表明,一般来说,C5.0呈现低偏差-高方差的特点,LDA与之相反,而MCLP则介于两者之间,比较接近LDA.当训练集样本量较小时,MCLP的偏差和方差都相对较高,而随着训练集的增大,MCLP的偏差和方差明显减小,甚至低于其他两者.Based on Domingos＇ bias-variance decomposition framework, on three different data sets, we compared the bias-variance structure of the three classification methods： MCLP, LDA and C5.0. The experimental results showed that, generally speaking, C5.0 has low bias and high variance, LDA has high bias and low variance, and MCLP is in between them but near LDA. When the training set is small, bias and variance of MCLP is comparatively high. However, with the increasing of training set, bias and variance of MCLP obviously decrease and even are lower than those of C5.0 and LDA. This study established the basis for constructing the ensemble suited to MCLP.

关 键 词：多目标线性规划 线性判别分析 C5.0 偏差 方差 

分 类 号：O29[理学—应用数学] N945[理学—数学]