Optimal Bayesian Error Estimation
Lori A. Dalton, Edward R. Dougherty · SPIE eBooks · 2020
Given that a distributional model is needed to achieve useful performance bounds for classifier error estimation when using the training data, a natural course of action is to define a prior distribution over the uncertainty class of feature-label distributions and then find an optimal minimum-mean-squareerror (MMSE) error estimator relative to the uncertainty class (Dalton and Dougherty, 2011b).