CCE: An Approach to Improve the Accuracy in Ensembles by Using Diverse Base Learners

Building ensembles with good performance depends highly on the precision and on the diversity of the base learners that compose them. However, achieving base learners that are both precise and diverse is a complex issue. In this paper we explore the idea of resolving multiclass classification problems using base learners composed of coupled classifiers that are trained with disjoint datasets. The goal is to achieve an accurate ensemble by using base learners that are relatively accurate but highly diverse. The system resulting from this proposal has been validated on the MNIST dataset, which is a good example for multiclass problem.