Learning with Nearest Neighbour Classifiers

摘要

This paper introduces a learning strategy for designing a set of prototypes for a 1-nearest-neighbour (NN) classifier. In learning phase, we transform the 1-NN classifier into a maximum classifier whose discriminant functions use the nearest models of a mixture. Then the computation of the set of prototypes is viewed as a problem of estimating the centres of a mixture model. However, instead of computing these centres using standard procedures like the EM algorithm, we derive to compute a learning algorithm based on minimising the misclassification accuracy of the 1-NN classifier on the training set. One possible implementation of the learning algorithm is presented. It is based on the online gradient descent method and the use of radial gaussian kernels for the models of the mixture. Experimental results using hand-written NIST databases show the superiority of the proposed method over Kohonen's LVQ algorithms.