A Median Variant of Generalized Learning Vector Quantization

David Nebel, Barbara Hammer, Thomas Villmann

Abstract: We introduce a median variant of the Generalized Learning Vector Quantization (GLVQ) algorithm. Thus, GLVQ can be used for classification problem learning, for which only dissimilarity information between the objects to be classified is available. For this purpose, the cost function of GLVQ is reformulated as a probabilistic model such that a generalized expectation maximization scheme can be applied as learning procedure. We give a rigorous mathematical proof for the new approach. Exemplary examples demonstrate the performance and the behavior of the algorithm.