Imbalanced binary classification under distribution uncertainty

Abstract

Imbalanced binary classification plays an important role in many applications. Some popular classifiers, such as logistic regression (LR), usually underestimate the probability of the minority class. Therefore, in this paper, we introduce two novel methods under distribution uncertainty, the idea of which is to modify the predicted probability with an additional uncertainty estimation. We develop the mean-uncertain method and the volatility-uncertain method, respectively, by assuming that the disturbance term follows the maximal and the G-normal distributions, which are the most important distributions within a sublinear expectation framework. Experiments on the simulated dataset and 10real-life datasets are conducted to compare the newly proposed approaches to several existing ones, including two resampling methods and two regression-based methods. The results of experiments show that our methods outperform most of the others in common evaluation metrics, especially the accuracy of the minority class.