基于特征空间变换的纠错输出编码

针对基于纠错输出编码多类分类中如何保证基分类器差异性的问题, 提出一种基于特征空间变换的编码方法. 该方法引入特征空间, 将编码矩阵扩展成三维矩阵; 然后基于二类划分, 利用特征变换得到不同的特征子空间, 从而训练得到差异性大的基分类器. 基于公共数据集的实验结果表明: 该方法能够比原始的编码矩阵获得更优的分类性能, 同时增加了基分类器的差异性; 该方法适用于任何编码矩阵, 为大数据的分类提供了新的思路.

Error-correcting output codes based on feature space transformation

The independency between each dichotomizer trained by coding matrix’s bi-partition is the key to using errorcorrecting output codes(ECOC) to solve multiclass problems. Therefore, an error-correcting output codes method based on feature space transformation(FST) is proposed. Inspired by the ensemble learning theory, a third feature space dimension is introduced into the coding matrix. Then, different subspaces are obtained by feature space transformation based on different positive and negative subclasses, so that the diversity between different binary classifiers are promoted to make the classification performance better. The experiment results based on UCI datasets show that the codes based on FST are better than the original codes. Besides, the proposed method can be applied to any kind of coding matrix, and provides new thought to large dataset for its quick training time and simplicity