Efficient Minimal Learning Machines with Reject Option

Abstract

Reject option is a widely used technique to improve the reliability of classification algorithms. It consists on withholding the classification of an instance if the classification is not reliable enough. Variants of well known classification algorithms have been proposed on the past years with diverse applications. In this work, we propose two variants of the Nearest Neighbor Minimal Learning Machine (NN-MLM) with reject option. The NN-MLM is an computationally efficient version of the recently proposed supervised learning algorithm called Minimal Learning Machine (MLM). The two variants (rejoNN-MLM and rejoNNwMLM) are evaluated on real world datasets and compared to state-of-the-art classifiers with reject option. Result show that both methods are a valid alternative for problems that require a reject option.