Recomendação automática da estrutura de comitês de classificadores usando meta-aprendizado

Abstract

We are constantly concerned with classifying things, people and making decisions, that when we face problems with a high degree of complexity, we tend to seek opinions from other people, usually from people who have a certain knowledge or even, as far as possible, be specialists in the domain of the problem in question, so that they effectively assist us in our decision-making process. In an analogy to the classification structures, we have a committee of people and or specialists (classifiers) who make decisions and, based on these answers, a final decision is made (aggregator). Thus, we can say that a classifier committee is formed by a set of classifiers (specialists), organized in parallel, that receive input information (standard or instance), and make an individual decision. Based on these decisions, the aggregator chooses the final, single decision of the committee. An important issue in the design of classifier committees is the definition of their structure, more specifically, the number and type of classifiers, and the aggregation method, to obtain the highest possible performance. Generally, an exhaustive test and evaluation process is necessary to define this structure, and trying to assist in this line of research, this work proposes two new approaches for systems of automatic recommendation of the classifier committee structure, using meta-learning to recommend three of these parameters: the classifier, the number of classifiers and the aggregator.