Sobre o processo de seleção de subconjuntos de atributos - as abordagens filtro e wrapper.

Abstract

Inductive machine learning methods learn the expression of the concept from a training set. Training sets are, generally, composed by instances described by attributevalue pairs and an associated class. The attribute set used for describing the training instances has a strong impact on the induced concepts. In a machine learning environment, attribute subset selection techniques aim at the identification of the attributes which effectively contribute for establishing the class of an instance. These techniques can be characterized as wrappers (if they are associated with a specific machine learning method) or filter and many of them work in conjunction with a search method (there are also embedded feature selection methods, not very representative). This work approaches the attribute subset selection problem by investigating the performance of two families of wrappers the NN (Nearest Neighbor) and DistAl families and three filter families Relief, Focus and LVF. The many members of the NN family (as well as of the DistAl family) differ among themselves with relation to the search method they use. The work presents and discusses the experiments conducted in many knowledge domains and their results allow a comparative evaluation (as far as accuracy and dimensionality are concerned) among the members of the families.