PUB - Publikationen an der Universität Bielefeld

Due to Intuitive training algorithms and model representation, prototype-based models are popular in settings where on-line learning and model interpretability play a major role. In such cases, a crucial property of a classifier is not only which class to predict, but also if a reliable decision is possible in the first place, or whether it is better to reject a decision. While strong theoretical results for optimum reject options in the case of known probability distributions or estimations thereof are available, there do not exist well-accepted reject strategies for deterministic prototype-based classifiers. In this contribution, we present simple and efficient reject options for prototype-based classification, and we evaluate their performance on artificial and benchmark data sets using the example of learning vector quantization. We demonstrate that the proposed reject options improve the accuracy in most cases, and their performance is comparable to an optimal reject option of the Bayes classifier in cases where the latter is available. Further, we show that the results are comparable to a well established reject option for support vector machines in cases where learning vector quantization classifiers are suitable for the given classification task, even providing better results in some cases. (C) 2015 Elsevier B.V. All rights reserved.