Local Reject Option for Deterministic Multi-class SVM

Johannes Kummert, Benjamin Paassen, Joris Jensen, Christina Göpfert, Barbara Hammer

Abstract: Classification with reject option allows classifiers to abstain from the classification of unclear cases. While it has been shown that global reject options are optimal for probabilistic classifiers, local reject schemes can enhance the performance of deterministic classifiers which do not provide faithful probability estimates [6, 10]. A first efficient scheme how to optimise local threshold parameters has recently been introduced [8]. In this contribution, we improve and simplify this scheme by restricting to a fewer number of possible candidates, and we demonstrate its performance for a one-versus-rest SVM classifier. Further, we have a glimpse at accompanying generalisation bounds.