Error-Reject Tradeoff Analysis on Two-Stage Classifier Design with a Reject Option
Eric Michael Vernon, Naoki Masuyama, Yusuke Nojima · 2022 World Automation Congress (WAC) · 2022
When a human is unsure of the answer to a difficult question, a natural response is "I don’t know." It is often useful to allow classification systems this same discretion; the so-called "reject option" allows classifiers to reject an input instead of attempting classification. In this paper, we introduce a simple ensemble design for a two-stage classifier with a reject option. The choice of classification algorithms used for each stage is unrestricted, but directly affects the error and reject rates. We center our analysis around the error-reject tradeoff produced by the user’s choice of algorithms.