Recent advances in error rate estimation
Pattern Recognition Letters
Hi-index | 0.00 |
In preceding studies, error rate estimators have been compared under various conditions and in most cases the population distribution was assumed to be normal. Effects of non-normality of the population have therefore not been studied sufficiently. In this study, we focused on kurtosis as a measure of non-normality and examined the effects of kurtosis for error rate estimators, especially resampling-based estimators. Our simulation results in two-class discrimination using a linear discriminant function suggest that it is necessary to consider non-normality of the population in comparison of estimators.