Technical Report, SFB 475: Komplexitätsreduktion in Multivariaten Datenstrukturen, Universität Dortmund 2000,24
When comparing methods for classification, often the rating relies on their prediction accuracy alone. One reason for this is that this is the aspect that can be most easily measured. Yet, often one wants to learn more about the problem than only how to predict. The interpretation of the relation of predictors and classes is often of high interest, but an unique accepted general formalization of interpretability relevant for many classification problems and measurable at least for a wide range of different classification methods does not exist, and - as we believe - is not really what is needed. Instead of trying to measure interpretability as such, standardizing and formalizing typical ways to interpret classification rules and finding performance criteria for this kind of outcomes leads to ratings of classification methods w.r.t. interpretability that can be tailored for the specific problem at hand and the subjective preferences of addressees of results. In this short paper, three results of this kind stemming from a comparative study of various classification methods applied to the classification of German business cycle phases based on 13 economic variables are exemplarily discussed .
classification method performance measures business cycle analysis