Neural networks applied to classification of data based on Mahalanobis metrics

This work presents a new algorithm for automatic classification of data that make use of a competitive neural network to aid the classification process. The algorithm basically answer two questions: Given a table where each row is a point of dimension D, in how many classes or clusters these data are disposed in? and given a point out of this set, to witch of this classes or clusters the point belongs to? The number of classes is automatically founded by the algorithm, that cluster according with a similarity measure among points that belong to the classes. The similarity measure used was the Mahalanobis distance, instead of the common Euclidian distance. That measure makes possible the incorporation of the spatial statistics of the data.