Modelos híbridos de aprendizaje basados en instancias y reglas para clasificación monotónica

Abstract

In supervised prediction problems, the response attribute depends on certain explanatory attributes. Some real problems require the response attribute to represent ordinal values that should increase with some of the explaining attributes. They are called classification problems with monotonicity constraints. In this thesis, we have reviewed the monotonic classifiers proposed in the literature and we have formalized the nested generalized exemplar learning theory to tackle monotonic classification. Two algorithms were proposed, a first greedy one, which require monotonic data and an evolutionary based algorithm, which is able to address imperfect data with monotonic violations present among the instances. Both improve the accuracy, the non-monotinic index of predictions and the simplicity of models over the state-of-the-art.