Prototype selection: Combining self-generating prototypes and Gaussian mixtures for pattern classification

This paper presents an investigation into prototype-based classifiers. Different methods have been proposed to deal with this problem. There are two main classes of prototype-selection algorithms. The first is merely selective, in which the resulting set of prototypes is formed by well-chosen samples from the training set. The second is known as the creative class of algorithms. This strategy creates new instances and performs adjustments of the prototypes during training. Two methods of the creative strategy are presented here: a self-generating prototype scheme and a fuzzy variation of Nearest Prototype Classification, which uses a Gaussian Mixture ansatz. The respective advantages and problems are discussed. A hybrid method is proposed to overcome difficulties and improve accuracy. The hybrid strategy obtained better results in the experiments when compared to each of two basic approaches and the classic K-Nearest Neighbor.