Leveraging Sub-class Partition Information in Binary Classification and Its Application

Abstract

Sub-class partition information within positive and negative classes is often ignored by a binary classifier, even when these detailed background information is available at hand. It is expected that this kind of additional information can help to improve the differentiating capacity of a binary classifier. In this paper, a binary classification strategy via multi-class categorization is proposed to leverage sub-class partition information when they are available. Empirical studies on the 20 newsgroups dataset demonstrate the benefits of this strategy. Furthermore, a preliminary application of this binary classification strategy for multi-label classification problem is given with promising results.