The Case against Accuracy Estimation for Comparing Induction Algorithms

We analyze critically the use of classification accuracy to compare classifiers on natural data sets, providing a thorough investigation using ROC analysis, standard machine learning algorithms, and standard benchmark data sets. The results raise serious concerns about the use of accuracy for comparing classifiers and drawinto question the conclusions that can be drawn from such studies. In the course of the presentation, we describe and demonstrate what we believe to be the proper use of ROC analysis for comparative studies in machine learning research. We argue that this methodology is preferable both for making practical choices and for drawing scientific conclusions.