Constructing ECOC based on confusion matrix for multiclass learning problems

Abstract

In the pattern recognition field, error-correcting output codes (ECOC) are a powerful tool to fuse any number of binary classifiers to model multiclass problems, and the research of encoding based on data is attracting more and more attention. In this paper, we are going to propose a new encoding method for constructing subclass Error-Correcting Output Codes, which was first introduced by Escalera et al. To achieve this goal, we first obtain the correlation between each pair of classes with the help of confusion matrix. Then, we select the most easily separated subclasses for classification by following Fisher’s principle. At last, we were able to obtain binary partitions based on subclasses. After finishing this work, a new data-driven coding matrix-Subclass ECOC will be achieved. Experimental results on University of CaliforniaIrvine data sets and three kinds of high resolution range profile data sets with logistic linear classifier and support vector machine as the binary classifiers show that our approach can provide a better performance and the robustness of classification with a little longer but acceptable code length.

创新点

提出一种新型基于数据集构造纠错输出编码解决多类分类问题策略,该方法首先利用混淆矩阵做为衡量多类样本空间中不同类别之间的相似度大小,进而得到类间离散度。其次,基于类间离散度寻找最优子空间划分并得到二类子空间划分集。最后,优化所得到的多个二类子空间(合并和拆分)并最终形成纠错输出编码,基于此编码矩阵划分样本空间集并训练即可得到最优二类分类器,最终有效提高多类分类的准确性和泛化能力。