Combining Classification Models

Abstract

Chapter 6 discussed methods for combining several models that are designed to make numeric predictions. For classification models that base their decisions on numeric predictions, the methods of that chapter are often a good choice. However, some models are inherently strict classifiers in that they produce a class decision and nothing more. Also, many number-based classifiers produce numeric predictions that are unstable in some way. In such situations, we must use more specialized techniques. This chapter discusses model-combination algorithms for applications in which the ultimate goal is classification. Component models in which the prediction is and is not numeric are both considered. Also, we see how to take advantage of component classifiers in which class ranks of ordinal scale are produced.