Un nuevo enfoque de los algoritmos de construcción de árboles de decisión: Tratamiento de información adicional

Abstract

This paper present two new algorithms for top-down induction of decision trees (TDIDT), (extensions of some previously existing procedures), based on using background knowledge not present in the learning set. Typically, in the different domain applications of the machine learning techniques there exists some knowledge, more or less accurate, about the relationships between attributes and classes. The algorithms presented make use of this additional information, (usually obtained from experts), allowing a better control of the learning process. As result of this, more understandable trees may be obtained, with better generalization capability. The application of handwritten digits is presented. A review of the main TDIDT algorithms is included.