Classification by fiducial predictive density functions
Yuqi Long, Xingzhong Xu · Communication in Statistics- Theory and Methods · 2020
For a classification problem with given loss function, Bayesian methods lead to the minimal risk among all possible classifiers. However, the prior distribution is often selected with some subjective thoughts. Motivated by this, we consider to replace the posterior distribution by a fiducial distribution, and use the fiducial predictive density to substitute the true but unknown underlying density functions to construct a new classification rule. The newly obtained classifier was proved to have oracle property. Simulations also show that the new classification rule performs better than traditional classifiers.