基于证据理论的纠错输出编码解决多类分类问题

针对多类分类问题,利用纠错输出编码作为分解框架,把多类问题转化为多个二类问题加以解决;同时提出一种基于证据理论的解码策略,把每一个二分器的输出作为证据之一进行融合,并讨论在两种编码类型(二元和三元编码矩阵)下证据融合的不同策略.通过实验分别对UCI数据集和3种一维距离像数据集进行测试,并与几种经典的解码方法进行比较,验证了所提出的方法能有效提高纠错输出编码特别是三元编码矩阵的分类正确率.

Error-correcting output codes based on evidence theory for multi-class classification

To model multi-class classification problems, error correcting output codes(ECOC) are used as decomposing
frame to reduce multi-class to binary. A decoding strategy based on DS evidence theory is proposed, which takes every binary
learner’s output as evidence to fusion and discusses different DS evidence fusion strategies based on two kinds of coding
type(binary ECOC and ternary ECOC). Experimental results on UCI and three kinds of HRRP show that the proposed scheme
provides better performance of error correcting output codes especially ternary ECOC than other state-of-the-art encoding
strategies.