Target shift awareness in balanced ensemble learning

Abstract

In the balanced ensemble learning for a two-class classification problem, the target values are shifted between [1 ∶ 0.5) or (0.5 ∶ 0] instead of 1 and 0 in the learned error function. Such shifted error function could let the ensemble avoid from unnecessary further learning on the well-learned data points. Therefore, the learning direction could be shifted away from the well-learned data points, and turned to the other not-yet-learned data points. By shifting away from well-learned data and focusing on not-yet-learned data, a good balanced learning could be achieved in the ensemble. Through examining both individual learners and the combined ensembles, this paper is to explore how the target shift awareness could help to decide a decision boundary that is neither too close nor too further to all training samples.