Abstract:
|
Asymmetric binary classification problems, in which the type I and type II errors have unequal severity, are ubiquitous in real-world applications. To achieve a desired classification performance, researchers have proposed two paradigms to deal with such asymmetry: cost-sensitive learning and Neyman-Pearson classification. Since the two paradigms have strength and weakness in different aspects, it remains unclear which paradigm to use in each specific asymmetric classification problem. Our work aims to address this challenge by bridging cost-sensitive learning and Neyman-Pearson paradigms from the perspective of controlling the population type I error. In this article, we for the first time discuss the methodological connection between the two classification paradigms. and we have identified two special cases when the two paradigms lead to equivalent classifiers with the same population type I error. We further propose a TUBE algorithm to estimate the type I error upper bound of cost-sensitive classifiers, and demonstrate that the TUBE algorithm improves cost-sensitive learning by helping objectively select the misclassification costs.
|