Adapting cost-sensitive learning for reject option

Traditional cost-sensitive learning algorithms always deterministically predict examples as either positive or negative (in binary setting), to minimize the total misclassification cost. However, in more advanced real-world settings, the algorithms can also have another option to reject examples of high uncertainty. In this paper, we assume that cost-sensitive learning algorithms can reject the examples and obtain their true labels by paying reject cost. We therefore analyse three categories of popular cost-sensitive learning approaches, and provide generic methods to adapt them for reject option.