A comparison of AUC estimators in small-sample studies

Reliable estimation of the classification performance of learned predictive models is difficult, when working in the small sample setting. When dealing with biological data it is often the case that separate test data cannot be afforded. Cross-validation is in this case a typical strategy for estimating the performance. Recent results, further supported by experimental evidence presented in this article, show that many standard approaches to cross-validation suffer from extensive bias or variance when the area under ROC curve (AUC) is used as performance measure. We advocate the use of leave-pair-out cross-validation (LPOCV) for performance estimation, as it avoids many of these problems. A method previously proposed by us can be used to efficiently calculate this estimate for regularized least-squares (RLS) based learners.