The ROC-Boost Design Algorithm for Asymmetric Classification

Abstract

In many situations (e.g., cascaded classification), it is desirable to design a classifier with precise constraints on its detection rate or on its false positive rate. We introduce ROC Boost, a modification of the Ada Boost design algorithm that produces asymmetric classifiers with guaranteed detection rate and low false positive rates. Tested in a visual text detection task, ROC-Boost was shown to perform competitively against other popular algorithms.