Interpretable machine learning with reject option
Abstract Classification by means of machine learning models constitutes one relevant technology in process automation and predictive maintenance. However, common techniques such as deep networks or random forests suffer from their black box characteristics and possible adversarial examples. In this contribution, we give an overview about a popular alternative technology from machine learning, namely modern variants of learning vector quantization, which, due to their combined discriminative and generative nature, incorporate interpretability and the possibility of explicit reject options for irregular samples. We give an explicit bound on minimum changes required for a change of the classification in case of LVQ networks with reject option, and we demonstrate the efficiency of reject options in two examples.