ABSTRACT

Pattern classification has been a very productive research area in past years. Innumerable applications can be visualized in forecasting, speech recognition, image processing, and bioinformatics, among others. In most cases, a single classification model is trained and evaluated fixing its parameters to maximize the classification accuracy. Still, selecting the best classification algorithm for a given dataset is not always an easy task. Even though cross validation and statistical hypothesis testing (e.g., 5 × 2-fold cross validation with a paired t-test) are often used to perform such selection, there are cases where no significant evidence can be found to assure that one classifier is better than another. Then, considering predictions not from one, but from a set of classifiers, turns out to be a good alternative to enrich a pattern recognition system. This is the main idea behind a multiple classifier system (MCS), which relies on the hypothesis that a set of base classifiers (i.e., the individual expert’s part of the MCS) may provide more accurate and diverse predictions [74]. Of course, MCSs entail additional complexity, not only because a number of classifiers should be trained and evaluated, but also because they require defining criteria to generate a prediction from their outputs.