Classifying with confidence using Bayes rule and kernel density estimation

Abstract

An example in which near infrared spectroscopic data are used to classify animal feed ingredients is used to make the case for the value of probabilistic approaches to classification problems. The accuracy of probabilities given by linear and quadratic discriminant analysis and by a more flexible kernel density approach are examined, and the effect on these probabilities of the use of different tuning criteria is explored. The example involves the classification of multiple particles in a sample, and detailed probability calculations bearing on the inference for both the sample and its parent population are presented.