About the attribute relevance's nature

Abstract

The notion of relevance of an attribute in machine learning is of common use in the construction of classfication rules in inductive learning processes. In this work a formal definition of the relevance concept for a given set of attributes is proposed, which includes the special case of non-relevant attributes or nought attributes. We establish the theoretical conditions that must satisfy the heuristics used to select the potentially more useful attribute to a classification, showing that some of the problems some classical heuristics based upon the information theory present, are actually due to the fact that they do not fulfill those conditions. We propose an heuristic that does satisfy them, and not enhance attributes with a lot of values more accurate than the rest.