Analyzing Dynamic Ensemble Selection Techniques Using Dissimilarity Analysis

In Dynamic Ensemble Selection (DES), only the most competent classifiers are selected to classify a given query sample. A crucial issue faced in DES is the definition of a criterion for measuring the level of competence of each base classifier. To that end, a criterion commonly used is the estimation of the competence of a base classifier using its local accuracy in small regions of the feature space surrounding the query instance. However, such a criterion cannot achieve results close to the performance of the Oracle, which is the upper limit performance of any DES technique. In this paper, we conduct a dissimilarity analysis between various DES techniques in order to better understand the relationship between them and as well as the behavior of the Oracle. In our experimental study, we evaluate seven DES techniques and the Oracle using eleven public datasets. One of the seven DES techniques was proposed by the authors and uses metalearning to define the competence of base classifiers based on different criteria. In the dissimilarity analysis, this proposed technique appears closer to the Oracle when compared to others, which would seem to indicate that using different bits of information on the behavior of base classifiers is important for improving the precision of DES techniques. Furthermore, DES techniques, such as LCA, OLA, and MLA, which use similar criteria to define the level of competence of base classifiers, are more likely to produce similar results.