ECOC多类分类中基分类器non-competence问题研究

在纠错输出编码(error-correcting output code, ECOC)多类分类中,当待识别样本的真实类别不属于对应二类子类划分时,训练得到的基分类器将不具备对此类样本进行分类的能力,此时的基分类器在解码融合时面临着non-competence问题。如何衡量基分类器是否具备对样本的分类能力,以及如果不具备,如何减少此种情况下对分类效果的影响是基于ECOC多类分类面临的新问题。针对解码框架中non-competent基分类器的分类融合问题,提出一种基于基分类器对样本是否具有分类能力的加权解码方法。该方法利用支持向量数据描述衡量待识别样本与各划分子类之间的距离,同时利用加权解码,通过对基分类器权重的学习,进而增强对类别拥有分类能力的基分类器的影响,减少不具备分类能力的基分类器产生的误差。基于UCI数据集的实验表明所提方法的有效性和实用性。

Non-competence reliability in multi-classification based on error correcting output codes

Error correcting output code (ECOC) has been an established technique for multi-classification due to its simpleness and efficiency. However, the non competent classifiers emerge when they classify an instance whose real class does not belong to one of the subclass sets which are used to learn the classifier. In this regard, in order to analyse the non-competence problem in the ECOC decomposing framework, a new weighted decoding strategy based on classifiers’ competence ability is presented as the solution, which can strengthen the influence of competent classifiers and reduce that of non-competent ones on classification performance through learning weight coefficient of base classifiers. Meanwhile, the support vector data description is applied to compute the distance of instances to each class. The statistic simulations based on UCI datasets corroborate the proposed method.