Robustness of the Bayes optimal discriminant procedure with 0-1 loss

摘要

This paper studies the robustness of the Bayes discriminant procedure with respect to data which do not follow a Gaussian (normal) probability distribution. The authors present descriminant criteria based on the beta, gamma, and lognormal distribution functions as alternatives to the usual Gaussian rule. These procedure are compared on both actual and simulated data. We find that Bayesian discrimination is sensitive to the incorrect use of the normal distribution.