Combinig classifiers using knowledge rule measures and genetic algortgms

Author(s)

Flávia Cristina Bernardini

• Document type: Doctoral Thesis
• Press: São Carlos.
• Institution: Universidade de São Paulo (USP). Instituto de Ciências Matemáticas e de Computação (ICMC/SB)
• Defense date: 2006-08-29

Examining board members

Maria Carolina Monard; Alexandre Gonçalves Evsukoff; Ana Cristina Bicharra Garcia; Anna Helena Reali Costa; Solange Oliveira Rezende

• Advisor: Maria Carolina Monard

Abstract

The quality of hypotheses induced by most of the available supervised machine learning algorithms depends on the quantity and quality of the instances in the training set. However, several well known learning algorithms are not able to manipulate many instances making it difficult to induce good classifiers from large databases, as are needed in the Data Mining process. One approach to overcome this problem is to construct ensembles of classifiers. An ensemble is a set of classifiers whose decisions are combined in some way to classify new cases (instances). However, although ensembles improve learning algorithms power prediction, ensembles may use an undesired large set of classifiers. Furthermore, despite classifying new cases better than each individual classifier, ensembles are generally a sort of ?black-box? classifier, not being able to explain their classification decisions. To this end, in this work we propose an approach that uses symbolic learning algorithms to construct ensembles of symbolic classifiers that can explain their classification decisions so that the ensemble is as accurate as or more accurate than the individual classifiers. Furthermore, considering that symbolic learning algorithms use local search methods to induce classifiers while genetic algorithms use global search methods, we propose a second approach to learn symbolic concepts from large databases using genetic algorithms to evolve symbolic classifiers into only one symbolic classifier so that the evolved classifier is more accurate than the initial ones. Both proposals were implemented in two computational systems. Several experiments using different databases were conducted in order to evaluate both proposals. Results show that although both proposals are promising, the approach using genetic algorithms produces better results. (AU)