Exploring diversity and similarity as criteria in ensemble systems based on dynamic selection
Ensemble systems, Dynamic Selection, Diversity, Similarity.
Pattern classification techniques can be considered the most important activitie in pattern
recognition area where aims assing a unknown sample test to a class. Generally individual
classifiers haven’t good recognition rates compared to multiple classifiers. Thus ensemble
of classifiers can be used to increase the accuracy of classification systems. Ensemble
systems provide good recognition rates when the classifiers members of the ensemble
system have uncorrelated errors in different sub-spaces of the problem; This characteristic
is measured by diversity measures. In this context, the present thesis explores ensemble
systems using dynamic selection. Ao contrário de comitês que utilizam seleção estática,
em comitês de classificadores utilizando seleção dinâmica, para cada padrão de teste
estima-se o nível de competência de cada classificador de um conjunto inicial. Apenas
os classificadores mais competentes são selecionados para classificar o padrão de teste.
O presente trabalho objetiva explorar, avaliar e propor métodos para seleção dinâmica
de classificadores baseando-se em medidas de diversidade. Unlike emseble sysetm using
static selection, in ensembles using dynamic selection for each test pattern is estimated
the competence level for the initial set of classifiers. Only the most relevant classifiers are
selected to classify the test pattern. This paper aims to explorer, evaluate and propose
methods for ensemble systems based on diversity measures. To achieve this goal, several
ensemble systems in the literature using dynamic selection are exploited, as well as hybrid
versions of them are proposed in order to quantify, by experiments, the influence of diversity
measure among classifiers members in ensemble systems. Therefore the contribution of this
work is empirically elucidate the advantages and disadvantages of using diversity measures
in dynamic selection of classifiers.