Supervised Linear Classification Performance Based on Marginal Probability for Two Groups

Abstract

The conventional technique to determine classification performance for the linear classification techniques strictly depends on the mean probabilities of correct classification or misclassification. Based on the mean probabilities of correct classification, robustness can be determined. In this paper, a new analytic procedure based on the joint and marginal probabilities is applied to determine robustness and the number of sample observations correctly classified. The classification results computed using this approach is unbiased. This technique is applied to investigate the classification performance of the Fisher linear classification analysis and the robust Fisher’s technique based on the minimum covariance determinant. The performance analysis when compared to the conventional procedure revealed that this technique is very informative. Relying on the analysis and the data set used, the recognition rate of the conventional approach is more accurate than the robust Fisher’s technique.