Coping with Class Balance Change in Classification: Class-Prior Estimation with Energy Distance

摘要

Due to sample selection bias or non-stationarity of the environment, the class balance often changes between training and test datasets. Naive classifier training under such a situation yields a biased solution. This bias can be corrected by weighted training according to the test class balance, but this test class balance is often unknown in practice. In this paper, we consider a semi-supervised learning setup where labeled training samples and unlabeled test samples are available, and address the problem of class balance estimation. It was shown that the test class balance can be estimated by fitting a mixture of class-wise training input distributions to the test input distribution, and class balance estimators were developed under, e.g., the Kullback-Leibler divergence and the L_2 distance. In this paper, we propose a simple class balance estimator based on the energy distance and demonstrate, its usefulness through experiments.