Classifier variability: Accounting for training and testing

摘要

We categorize the statistical assessment of classifiers into three levels: assessing the classification performance and its testing variability conditional on a fixed training set, assessing the performance and its variability that accounts for both training and testing, and assessing the performance averaging over training sets and its variability that accounts for both training and testing. We derived analytical expressions for the variance of the estimated AUC and provide freely available software implemented with an efficient computation algorithm. Our approach can be applied to assess any classifier that has ordinal (continuous or discrete) outputs. Applications to simulated and real datasets are presented to illustrate our methods.