The interest of reject for classifier optimization has been shown many times. The diversity of the applications requiring this concept makes us to distinguish two main natures of reject with distinct goals: the confusion reject and the distance reject. After the description of this two kinds of reject, we present a unified formalism to define them using reliability functions and reject thresholds. Then we present a generic algorithm dedicated to the automatic learning of these thresholds. Finally, we compare various possibilities of reject to achieve application goals.
展开▼
