| Reliable classifiers in ROC space (2006) | |||||||||||||||
Abstract | |||||||||||||||
| The performance of a classifier can be improved by abstaining on uncertain instance classifications. The transformation from the original Receiver Operator Characteristic (ROC) curve to the curve obtained by abstention is provided. We include proofs on dominance of this new ROC curve to aid classifier selection and to show the effectiveness of the approach. For specific cost and class distributions we provide an approach in ROC space to transform a classifier into a new one that has a desired precision per class. 1. | |||||||||||||||
Publication details | |||||||||||||||
| |||||||||||||||