A New Diverse Measure in Ensemble Learning Using Unlabeled Data

Ensemble learning has been successfully used in many areas, due to its powerful ability to solve complex problems. In recent years, some researchers have shown that ensemble of some learners instead of all individual learners could get better performances. However, how to select individual learners as diverse as possible is a very important issue. In this paper, a new diversity measure is proposed to achieve a better selection of individual learners. Different from the commonly used diversity measures, it makes full of the data distribution information provided by the cheap and abundant unlabeled data rather than the expensive and scarce labeled data in order to obtain the higher classification accuracy. The selection method based on the new diversity measure is simple in computation and independent of models. Experimental results demonstrate its good performances.