Abstract: The goal of the paper is to investigate what training sample estimate of misclassification probability would be the best one for the histogram classifier. Certain quality criterion is suggested. The deviation for some estimates, such as resubstitution error (empirical risk), cross validation error (leave-one-out), bootstrap and for the best estimate obtained via some optimization procedure, is calculated and compared for some examples.

Keywords: pattern recognition, classification, statistical robustness, deciding functions, complexity, capacity, overfitting, overtraining problem.

ACM Classification Keywords: G.3 Probability and statistics, G.1.6. Numerical analysis: Optimization; G.2.m. Discrete mathematics: miscellaneous.

Link:

ОПТИМИЗАЦИЯ ОЦЕНКИ ВЕРОЯТНОСТИ ОШИБОЧНОЙ КЛАССИФИКАЦИИ В ДИСКРЕТНОМ СЛУЧАЕ1

Виктор Неделько

http://foibg.com/ibs_isc/ibs-08/ibs-08-p07.pdf