Rejection Strategies for Learning Vector Quantization - A Comparison of Probabilistic and Deterministic Approaches

Abstract

In this contribution, we focus on reject options for prototype-based classifiers, and we present a comparison of reject options based on statistical models for prototype-based classification as compared to alternatives which are motivated by simple geometric principles. We compare the behavior of generative models such as Gaussian mixture models and discriminative ones to results from robust soft learning vector quantization. It turns out that (i) reject options based on simple geometric show a comparable quality as compared to reject options based on statistical approaches. This behavior of the simple options offers a nice alternative towards making a probabilistic modeling and allowing a more fine-grained control of the size of the remaining data in many settings. It is shown that (ii) discriminative models provide a better classification accuracy also when combined with reject strategies based on probabilistic models as compared to generative ones. © Springer International Publishing Switzerland 2014.