Consistency driven feature subspace aggregating for ordinal classification

Abstract

We present a new method for constructing an ensemble classifier for ordinal classification with monotonicity constraints. Ordinal consistency driven feature subspace aggregating (coFeating) constructs local component classification models instead of global ones, which are more common in ensemble methods. The training classification data are first structured using Variable Consistency Dominance-based Rough Set Approach (VC-DRSA). Then, coFeating constructs local classification models in subregions of the attribute space, which is divided with respect to consistency of objects. Our empirical evaluation shows that coFeating performs significantly better than previously proposed ensemble methods on data characterized by a high number of objects and/or attributes.