Spectral regularization in generalized matrix learning vector quantization

In this contribution we propose a new regularization method for the Generalized Matrix Learning Vector Quantization classifier. In particular we use a nuclear norm in order to prevent oversimplifying/over-fitting and oscillatory behaviour of the small eigenvalues of the positive semi-definite relevance matrix. The proposed method is compared with two other regularization methods in two artificial data sets and a reallife problem. The results show that the proposed regularization method enhances the generalization ability of GMLVQ. This is reflected in a lower classification error and a better interpretability of the relevance matrix.