带拒绝域的ECOC多类分类

摘要

Aiming at reducing misclassification costs,this paper studies the design of reject options for ECOC multi-classification based on its properties.The first level of reject option is constructed in the process of bipartitions to recognize an instance whose real labels does not belong to the meta-subclasses.Meanwhile,the second reject rule is presented in the dichotomizers based on posterior probabilities and cost matrix to make the minimum-risk Bayesian decision.Finally,different decoding strategies are analysed according to different reject output.The relationship between the number of rejected positions and the minimum Hamming distance of matrix is discussed.The two-stage reject rule makes the ECOC multi-classification with rejection come true and reduce the misclassification error and costs.%针对纠错输出编码分解框架的自身特点、从降低误判风险出发,研究了带拒绝域的ECOC多类分类方法.首先在二类划分过程中引入拒绝域,对不属于正负子类的待识别样本进行拒识;其次,在基分类器内部引入拒绝域,以最小化风险贝叶斯决策为目标,利用后验概率输出和代价矩阵寻找拒绝域阈值,对样本输出值落入拒绝域中的样本进行拒识;最后,研究了不同拒绝域输出的解码方法,并讨论了拒识码字个数和矩阵最小Hamming距离之间的关系.实验结果表明基于二类划分构造的拒绝域能够提高分类正确率,而基于基分类器构造的拒绝域能够减小分类代价.