基于纠错输出编码的多类概率建模  

Probability modeling of multi-class based on error correcting output codes

作　　者：张玉玺[1,2] 王晓丹[1] 周进登[1] 姚旭[1] 

机构地区：[1]空军工程大学防空反导学院 [2]93424部队

出　　处：《控制与决策》2013年第5期671-676,687,共7页Control and Decision

基　　金：国家自然科学基金项目(60975026;61273275)

摘　　要：目前模式识别领域中缺乏有效的多类概率建模方法,对此提出利用纠错输出编码作为多类概率建模框架,将二元纠错输出编码研究的概率输出问题转化为线性超定方程的求解问题,通过线性最小二乘法来求解并获取多类后验概率的结果;而对于三元纠错输出编码的等价非线性超定方程组,提出一种迭代法则来求解多类概率输出.实验中通过与3种经典方法相比较可以发现,新方法求取的概率输出具有更好的分布形态,并且该方法具有较好的分类性能.For the lack of effective method on multi-class＇ probabilities modeling, a posterior probability estimating approach based on error correcting output codes（ECOC） is presented. In this approach, probability estimation is translated into solving overdetermined linear equations with the binary ECOC as a modeling framework. Moreover, when the modeling framework is changed to ternary ECOC, the equivalence problem is overdetermined nonlinear equations. Therefore, an iterative algorithm is proposed to solve this problem for obtaining probabilities. Compared with three classic approaches, the proposed approach has better classification ability and probability distribution of the posterior probability in experiments.

关 键 词：模式识别 多类分类 概率建模 纠错输出编码 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]