Máquina de aprendizagem mínima com opção de rejeição

Abstract

The Minimal Learning Machine (MLM) is an inductive learning method applied to supervised classification and regression problems. It is basically a mapping between points in the geometric configurations of the input and output space. With the known configuration to an entry point in the input space corresponding to the output configuration space after obtaining a simple linear model learning distance between arrays of input and output can be estimated. The estimated result is then passed to locate the exit point and thus provide an estimate for response or indication of the class. The MLM has reached a promising performance in various classification and regression problems compared with other classical methods of learning. However, it has not yet been analyzed performance using classification strategy of rejection option. This technique protects the system for decision support in many human activities, especially in the field of medicine, against excessive errors like making difficult decision consequences. In this way, potential errors are converted into rejection, avoiding further confusion and delegating them well for the evaluation of an expert, or even, for more specialized classifiers. Therefore, the purpose of this dissertation is the development of the Minimum Learning Machine (MLM) and its variants with Rejection option in binary classification problems, more specifically, in the classification of diseases Spinal (PVC-2C), Diabetes (Pima Indians diabetes), survival of breast cancer (Haberman) and prediction software defects (KC2). The evaluation of the performance of these techniques is, in general, the analysis of the accuracy-rejection curve compared to more traditional methods of classification with rejection option that are based in methods Perceptron Multilayer (MLP), K-Neighbors More Next (K -NN) and K-Means.