Discriminative learning for minimum error and minimum reject classification
Haruo Mizutani · 2002
We present a practical learning algorithm for a combined classifier with a reject decision, by extending the original generalised probabilistic descent (GPD) algorithm to use a reject option. Our algorithm simultaneously minimizes the classification error and the reject risk under any reject conditions, in order to design an optimum combined classifier with a reject decision, rather than an optimum reject decision function. We demonstrate that our combined classifier is superior to a conventional classifier with a reject option by using error-reject characteristic.