Coupled Transductive Ensemble Learning of Kernel Models

In this paper we propose the concept of coupling for ensemble learning. In the existing literature, all submodels that are considered within an ensemble are trained independently from each other. Here we study the effect of coupling the individual training processes within an ensemble of regularization networks. The considered coupling set gives the opportunity to work with a transductive set for both regression and classification problems. We discuss links between this coupled learning and multitask learning and explain how it can be interpreted as a form of group regularization. The methods are illustrated with experiments on classification and regression data sets.