Nearest-neighbour ensembles in lasso feature subspaces

Abstract

The least absolute shrinkage and selection operator (lasso) is a promising feature selection technique. However, it has traditionally not been a focus of research in ensemble classification methods. In this study, the authors propose a robust classification algorithm that makes use of an ensemble of classifiers in lasso feature subspaces. The algorithm consists of two stages: the first is a lasso-based multiple feature subsets selection cycle, which tries to find a number of relevant and diverse feature subspaces; the second is an ensemble-based decision system that intends to preserve the classification performance in case of abrupt changes in the representation space. Experimental results on the two-class textured image segmentation problem prove the effectiveness of the proposed classification method.