Máquina de aprendizagem mínima para dados intervalares

Abstract

Solving machine learning problems with interval-valued data is a challenging task that can arise in many real world applications, for example, in heart rate prediction. Motivated by this fact, many researchers have proposed nonlinear regression methods and classifiers to deal with interval-valued data in recent years. In this work, four variants of the Minimal Learning Machine (MLM) are proposed for interval-valued data, two focusing on regression problems and two for classification and regression. The choice of MLM is explained by its outstanding performance in many applications and by the need to define a single hyperparameter. To validate the proposed methods, in regression problems, they were compared with five non-linear regressors: three Extreme Learning Machine (iELM) variants and two extensions of kernel regression. In the classification, three logistic regression models were used. Experiments on synthetic datasets (with different configurations) and real data are presented and they show that the variants of the MLM for interval-valued data obtain similar or better results than the others. From this study, it is noted that the proposed methods presented competitive results and can be considered good options.