Estudio comparativo de los algoritmos de clasificación supervisada empleando datos artificiales.

Abstract

This work reviews the theoretical part and performs its implementation in Matlab of some of the most used supervised classification algorithms: MLP (MultiLayer Perceptron), PNN (Probabilistic Neural Network), K-NN (K -Near Neighbor), NB (Naive Bayes), SVM (Support Vector Machines) and DT (Decision Tree). The main strategy introduced in this work is the use of artificial data, located in a two-dimensional space, which are classified under a nonlinear decision boundary also artificially generated. This allows to have under control the following parameters: the amount of training data assigned to the different classes, the amount of test data assigned to the different classes, the degree of asymmetry between these (there can be more training and/or test data for one of the classes) and their degree of remoteness to the decision boundary (it can be controlled whether the data are close, moderately remote, and far away). This point of view on the creation of artificial data is totally novel and constitutes a contribution to the state of the art that really allows comparing the different supervised classification methods so that firm conclusions can now be drawn regarding the accuracy obtained by them.