Classifier performance model for the many-class case

摘要

A model is presented for predicting classification performance for systems having a large population of classes. The cases of large and small training set size for each class are treated separately. A method is proposed for measuring classification performance as the mean ranking statistic ρ{sub}E which is derived from the average information content h{sub}E of the system feature vector, which is in turn derived from the system covariance matrices (Σ{sub}W, Σ{sub}B). This method for predicting ρ{sub}E is applied to the large training set case (case 1), explaining why performance is not compromised but improved by adding noisy features. The method is extended for predicting performance in the more difficult small training set case (case 2), explaining why performance may be compromised by the addition of noisy features in that situation.