Reduction Stumps for Multi-class Classification

Abstract

Multi-class classification problems are often solved via reduction, i.e., by breaking the original problem into a set of presumably simpler subproblems (and aggregating the solutions of these problems later on). Typical examples of this approach include decomposition schemes such as one-vs-rest, all-pairs, and nested dichotomies. While all these techniques produce reductions to purely binary subproblems, which is reasonable when only binary classifiers ought to be used, we argue that reductions to other multi-class problems can be interesting, too. In this paper, we examine a new type of (meta-)classifier called reduction stump. A reduction stump creates a binary split among the given classes, thereby creating two subproblems, each of which is solved by a multi-class classifier in turn. On top, the two groups of classes are separated by a binary (or multi-class) classifier. In addition to simple reduction stumps, we consider ensembles of such models. Empirically, we show that this kind of reduction, in spite of its simplicity, can often lead to significant performance gains.