Equation for generalization error of Multinomial classifier is derived and tested. Particular attention is paid to imbalanced
training sets. It is shown that artificial growth of training vectors of less probable class could be harmful. Use of predictive
Bayes approach to estimate cell probabilities of the classifier reduces both the generalization error and effect of unequal
training sample sizes.
Keywords BKS rule - Complexity - Generalization error - Learning - Imbalan-ced training sets - Multinomial classifier - Sample size