Seleção dinâmica de classificadores baseada em medidas de complexidade

Abstract

Object classification is a task in the field of pattern recognition. It consists of assigning a class, among several possible ones, to an unknown object. This process is accomplished by classifiers. They are mathematical functions built from the training of the classifiers, which make use of the data of a given problem to obtain learning. Monolithic classifiers find difficulty in more difficult classification scenarios. They may not cover the full range of variability of the problem. Multiple Classifier Systems (MCS) have been an option in order to improve the performance of the classification task before the monolithic approach. For this, the MCS’s employ several specialist classifiers in different regions of the problem. The idea is that, from a set of different classifiers, it is possible to absorb the variability presented by the problems. However, not all of them are always used in the classification step, being necessary the step of selection of classifiers from the set. For this task, the dynamic selection has shown to be promising. The architecture with which classification systems are built influences their performance. The difficulty of the classification problem is also related to the complexity of the data. It impacts the way the classifiers act and information about it is obtained from the complexity measures. Studies show that to solve a difficult classification problem, a better understanding of data complexity can be interesting for classifier selection. Thus, in this work, we evaluate the complexity measures as selection criteria for the dynamic selection of classifiers. Modifying classical literature approaches results in higher accuracy of 39.51% of the cases by adding the complexity measures, while maintaining the original approach corresponds to 28.4%. Ties were present in 35.8%. This work also presents the information gain of 14 complexity measures, which suggest some drawbacks of using some of them.